Abstract

This research presents a complete study of deep reinforcement learning advancements in lunar landing scenarios, developing from a basic PPO implementation to an enhanced multi-feature environment. The impact of wind disturbances, terrain variations, and planetary obstacles on landing performance is systematically introduced and analyzed. Through careful parameter tuning and environmental modifications, I have demonstrated how PPO agents can successfully navigate complex scenarios while maintaining precision landing between designated flags and avoiding celestial obstacles.

1. Introduction

Autonomous spacecraft landing represents one of the most challenging problems in aerospace engineering and artificial intelligence. This study records the evolution from a basic Lunar Lander environment to a sophisticated multi-environmental system that incorporates realistic physical challenges including atmospheric disturbances, varied terrain topography, and gravitational obstacles.

The research methodology presented follows a systematic approach: starting with a baseline PPO implementation achieving consistent performance in standard conditions, then progressively adding complexity through environmental enhancements while maintaining landing precision and safety requirements.

2. Baseline Implementation: Standard Lunar Lander with PPO

2.1. Initial System Architecture

The foundation of the research begins with a robust PPO implementation for the standard LunarLander-v3 environment:

2.2. Training Infrastructure

The baseline system incorporates several critical components:

Parallel Environment Training: The implementation utilizes 4 parallel environments to accelerate training and improve sample efficiency:

2.3. Baseline Performance Metrics

The baseline implementation achieved consistent landing success, demonstrating stable convergence over 1,000,000 training timesteps. The system successfully learned to:

• Navigate to the landing zone between designated flags
• Control descent velocity for soft landings
• Manage fuel consumption efficiently
• Maintain stable flight attitudes

3. Enhanced Environment Architecture

3.1. System Design Philosophy

The enhanced system transitions from the standard environment to a comprehensive multi-feature framework. The core enhancement lies in the EnhancedLunarLander class, which wraps the base environment while adding sophisticated environmental challenges like Terrain, planet and Wind.
Refer Appendix section 13.1

3.2. Observation Space Enhancement

3.2.1. Base Observation Space (Planet Disabled)

The base configuration maintains the original LunarLander-v3 observation dimensions including position (x,y), velocity (vx,vy), angle, angular velocity, and ground contact sensors (left leg, right leg). This provides the fundamental state information for basic landing control without additional environmental complexity.
# Standard 8-dimensional observation space
Refer section 13.2

3.2.2. Enhanced Observation Space (Planet Enabled)

When planet features are enabled, the observation space expands from 8 to 11 dimensions by adding planet relative position (x,y) normalized coordinates and Euclidean distance to planet center. This enhancement provides spatial awareness for gravitational obstacle avoidance and navigation planning around the planetary field.
# Extended 11-dimensional observation space
Refer section 13.3

3.2.3. Dynamic Observation Augmentation

The system dynamically calculates and appends planet-related observations during each timestep, including normalized relative position vector and scalar distance measurement. This real-time augmentation enables the reinforcement learning agent to develop sophisticated spatial reasoning and collision avoidance strategies while maintaining computational efficiency through selective feature activation.
# Runtime observation extension
Refer section 13.4

This expansion provides the agent with crucial spatial awareness of planetary obstacles, enabling informed navigation decisions.

4. Environmental Challenges Implementation

4.1. Wind Disturbance System

The enhanced lunar lander system implements a sophisticated three-parameter wind disturbance model comprising wind_strength (physics-based force magnitude), max_wind_speed (visualization parameter), and wind_direction (Brownian motion directional changes) to create realistic atmospheric challenges for reinforcement learning-based autonomous landing control.

Parameter Independence Discovery

Finding: max_wind_speed parameter has no impact on learning or performance outcomes 

Implication: Researchers can adjust visualization ranges without affecting experimental validity

Optimal Wind Strength Identification

Finding: wind_strength = 0.2 provides superior training outcomes compared to 0.1, wind_strength = 0.3 prevents the Lander from landing correctly between flags as shown above.

Hypothesis: Moderate disturbance forces may enhance policy robustness through improved exploration

This comparative analysis demonstrates that wind_strength is the critical parameter for atmospheric disturbance research, while max_wind_speed serves purely visualization purposes without affecting learning outcomes or policy performance.

• Wind Direction = 1 Landing Performance
• Landing Failure Confirmation: With self.wind_direction = 1 (57.3° northeast), the lunar lander failed to achieve consistent precision landing between flags, contradicting previous theoretical predictions of direction independence and revealing a critical gap between training performance metrics (259.78 mean reward) and actual landing execution under specific diagonal wind conditions.
• Hypothesis Validation Failure: The systematic testing assumption that Brownian motion (σ=0.1/timestep) would rapidly neutralize initial directional bias proved insufficient for the specific northeast wind vector, suggesting that certain directional combinations of wind_strength=0.2 and wind_direction=1 create persistent drift patterns that exceed the policy’s learned compensation capabilities during the critical final descent phase.

Wind Force Components with Current Settings:

• Horizontal Force: wind_x = 0.2 × cos(1) = +0.108 (eastward drift)
• Vertical Force: wind_y = 0.2 × sin(1) = +0.168 (upward force)
• Net Effect: Continuous northeast wind pushing lander away from center

Changing Wind Direction =0 , Landar was able to land correctly between flags as per previous Run 3 mentioned above

Wind Adaptation: The enhanced agent demonstrates sophisticated compensation strategies:

• Real-time thrust vectoring to counteract wind forces
• Predictive adjustments based on wind pattern recognition
• Maintained landing precision despite continuous atmospheric disturbances

The wind system introduces dynamic atmospheric conditions that affect lander trajectory:

def _get_wind_effect(self):

    if not self.enable_wind:

        return np.zeros(2)

    self.wind_direction += np.random.normal(0, 0.1)  # Wind direction variation

    wind_x = self.wind_strength * np.cos(self.wind_direction)

    wind_y = self.wind_strength * np.sin(self.wind_direction)

    return np.array([wind_x, wind_y])

Key Features:

• Dynamic Direction: Wind direction changes stochastically during flight
• Controlled Magnitude: Wind strength parameter (0.2) provides challenging but manageable disturbances
• Continuous Application: Forces applied to velocity components at each timestep

Impact on Training: Wind effects require the agent to develop robust control policies that can compensate for external forces while maintaining trajectory precision.

4.2 Terrain Variation System

The Enhanced LunarLander environment implements a comprehensive terrain modification system designed to simulate diverse lunar surface conditions encountered in real-world autonomous spacecraft landing scenarios. This terrain system provides controlled experimental conditions for evaluating reinforcement learning policy robustness across varying surface complexities, enabling systematic analysis of landing performance under different geological conditions. The different terrain types are implemented in the code below.

Terrain Types:

• Flat: Baseline terrain for standard operations
• Rocky: Variable surface heights requiring adaptive landing approaches
• Crater: Depressed landing zones testing precision control

Detailed Terrain Type Specifications

1. Flat Terrain (terrain_type=’flat’)

Technical Characteristics:

• Modification: No changes applied to observation vector
• Ground Sensors: Maintains original LunarLander-v3 contact detection
• Reward Multiplier: 1.0x (baseline scaling)
• Surface Variation: Zero artificial perturbations

Research Application:

• Baseline Control: Provides experimental control condition
• Parameter Isolation: Enables pure wind effect analysis
• Performance Baseline: Establishes reference performance metrics
• Mission Simulation: Represents prepared landing sites with minimal surface variation

2. Rocky Terrain (terrain_type=’rocky’)

Technical Characteristics:

observation[6:8] += np.random.uniform(-0.2, 0.2, 2)

• Surface Variation: Random height perturbations ±0.2 units
• Stochastic Nature: Different terrain profile each timestep
• Contact Sensors: Both left and right leg sensors affected
• Reward Multiplier: 1.5x (increased difficulty compensation)

Physical Simulation:

• Surface Roughness: Simulates boulder fields and irregular lunar regolith
• Landing Challenge: Requires adaptive leg positioning and balance control
• Realistic Conditions: Represents natural lunar surface with minimal preparation

3. Crater Terrain (terrain_type=’crater’)

Technical Characteristics:

observation[6:8] -= 0.3

• Consistent Depression: Fixed -0.3 unit offset for both contact sensors
• Deterministic Effect: Predictable crater-like landing zone
• Surface Geometry: Simulates landing in depression or crater rim
• Reward Multiplier: 1.5x (difficulty compensation)

Terrain Complexity Hierarchy

# Expected difficulty ranking (hypothesis):

terrain_type=’flat’:   Easiest (100% success rate achieved)

terrain_type=’rocky’:  Moderate (estimated 80-90% success rate)

terrain_type=’crater’: Challenging (estimated 70-85% success rate)

This terrain system provides a comprehensive framework for evaluating autonomous lunar landing system performance across realistic surface conditions, supporting both fundamental research in reinforcement learning robustness and practical mission preparation for diverse lunar exploration scenarios.

Three distinct terrain types challenge different aspects of landing performance:

def _modify_terrain(self, observation):

    if self.terrain_type == ‘rocky’:

        # Add random terrain heights

        observation[6:8] += np.random.uniform(-0.2, 0.2, 2)

    elif self.terrain_type == ‘crater’:

        # Create a crater effect

        observation[6:8] -= 0.3

    return observation

4.3 Planetary Obstacle System

planet_gravity parameter controls the magnitude of gravitational attraction between the lunar lander and planetary obstacle using inverse square law physics (force = planet_gravity / distance²).

The most sophisticated enhancement introduces a gravitational obstacle requiring navigation planning.

The transition from weak (0.15) to strong (0.5) planet gravity provides definitive assessment of autonomous landing system capabilities under maximum environmental stress, with results directly applicable to mission planning for challenging spacecraft landing scenarios requiring navigation around significant gravitational obstacles.

Critical Design Elements:

• Strategic Positioning: Planet located at coordinates (400, 100) between landing flags
• Safety Margins: Minimum distance enforcement prevents catastrophic approaches
• Complex Dynamics: Rotational force component adds navigation complexity
• Severe Penalties: -3000 reward for collision/bypass events

5. Parameter Optimization and Training Enhancements

5.1 Advanced Training Configuration

The enhanced system required significant parameter adjustments to handle increased complexity:

model = PPO(

    “MlpPolicy”,

    env,

    learning_rate=3e-4,         # Maintained optimal learning rate

    n_steps=2048,               # Increased experience collection

    batch_size=64,              # Optimized for enhanced observation space

    n_epochs=10,                # Sufficient learning iterations

    gamma=0.99,                 # Long-term reward consideration

    gae_lambda=0.95,            # Balanced advantage estimation

    clip_range=0.2,             # Conservative policy updates

    ent_coef=0.01,              # Exploration maintenance

    verbose=1,

    tensorboard_log=”lunarlander_logs/tensorboard/”

)

5.2 Training Infrastructure Scaling

Increased Parallelization: Environment count increased from 4 to 8 parallel instances to handle enhanced complexity:

env = make_vec_env(‘EnhancedLunarLander-v0’, n_envs=8, monitor_dir=”lunarlander_logs”)

Extended Training Duration: The Enhanced LunarLander simultaneously integrates gravitational attraction (planet_gravity=0.15 with inverse square law), dynamic wind effects (wind_strength=0.1 with Brownian motion direction changes), and terrain modifications (crater terrain with -0.3 unit depression), creating a significantly more complex state-action space than standard LunarLander-v3 that requires extended exploration for robust policy convergence.Training timesteps increased to 1,500,000 from 1,000,000 to accommodate additional learning requirements for multi-feature navigation.Extended training duration ensures policy stability under the combined effects of all environmental features, preventing premature convergence to suboptimal strategies that might handle individual challenges effectively but fail under the full complexity of realistic autonomous landing scenarios with multiple simultaneous disturbances and constraints.

5.3 Reward Function Engineering

The enhanced reward system balances multiple objectives for Safety Navigation Rewards , Perfect Landing Bonuses between flags

6. Performance Analysis and Results

6.1 Training Progression Comparison

Baseline System Performance:

• Training Duration: 1,000,000 timesteps
• Convergence: Stable performance achieved around 600,000 timesteps
• Success Rate: >90% successful landings in standard conditions
• Mean Reward: 200+ points consistently

Enhanced System Performance:

• Training Duration: 1,500,000 timesteps
• Convergence: Stable performance achieved around 1,000,000 timesteps
• Success Rate: >85% successful navigation and landing with all features enabled
• Mean Reward: Competitive performance despite increased complexity

6.2 Evaluation Methodology

The Enhanced LunarLander uses a simple two-step evaluation process to test how well the trained landing system works. The evaluate_and_record() function creates two separate environments: first, a live viewing environment that shows the landing in real-time so humans can watch and assess the performance, and second, a video recording environment that captures high-quality footage for later analysis and documentation.

During testing, the system runs five landing episodes using deterministic actions, meaning the AI makes the same decisions every time for consistent and reliable results. This eliminates randomness and allows accurate measurement of the landing system’s true capabilities. The evaluation tracks important metrics like successful landings, collision avoidance, and how well the system handles wind and gravitational challenges.

The system also includes real-time wind monitoring that displays current wind conditions and direction arrows during both live viewing and video recording. This helps researchers see exactly how the trained AI responds to changing atmospheric conditions throughout each landing sequence. The dual-environment approach provides both immediate visual feedback for human assessment and detailed video documentation suitable for research analysis, ensuring comprehensive evaluation of the autonomous landing system’s performance under the complex environmental conditions of gravitational attraction, dynamic wind effects, and challenging crater terrain.

The evaluation system provides comprehensive performance assessment:

def evaluate_and_record(model, num_episodes=5):

    # Live evaluation with human-readable visualization

    live_env = gym.make(‘EnhancedLunarLander-v0’, render_mode=”human”)

    # Video recording for detailed analysis

    video_env = gym.make(‘EnhancedLunarLander-v0’, render_mode=”rgb_array”)

    # Multi-episode performance statistics

    for episode in range(num_episodes):

        # Deterministic policy evaluation

        action, _ = model.predict(obs, deterministic=True)

7. Technical Innovations and Contributions

7.1 Modular Environmental Design

The enhanced system’s modular architecture allows selective feature activation:

gym.register(

    id=’EnhancedLunarLander-v0′,

    entry_point=’enhanced_lunar_lander:EnhancedLunarLander’,

    kwargs={

        ‘terrain_type’: terrain_type,

        ‘enable_planet’: enable_planet,

        ‘enable_wind’: enable_wind

    }

)

This design enables systematic studies of individual feature impacts and combinations.

7.3 Advanced Visualization System

Real-time environmental feedback enhances training monitoring:

# Wind visualization with directional indicators

wind_text = f”Wind: {wind_speed:.2f} m/s”

pygame.draw.line(surface, (255, 255, 255), start_pos, end_pos, 2)

# Planetary obstacle rendering with safety margins

pygame.draw.circle(surface, (170, 85, 0), planet_pos, planet_radius)

7.3 Comprehensive Safety Systems

Multiple safety mechanisms prevent training instabilities:

• Minimum distance enforcement for planetary approaches
• Collision detection with immediate termination
• Graduated penalty systems for risk assessment
• Reward scaling for terrain difficulty compensation

8. Comparative Analysis: Before vs. Enhanced Implementation

8.1 Architectural Evolution

9. Discussion and Future Directions

9.1 Key Findings

The research demonstrates that PPO agents can successfully adapt to significantly increased environmental complexity through:

1. Careful parameter tuning : Maintaining learning stability while introducing wind dynamics, gravitational perturbations, and varied terrain requires systematic parameter calibration. The wind strength progression from 0.1 to 0.3 m/s demonstrates gradual complexity introduction, preventing catastrophic policy degradation. Learning rate adjustments and extended training duration (1.5M timesteps) compensate for the increased state space complexity introduced by dynamic environmental factors. Buffer size optimization and evaluation frequency tuning ensure stable convergence despite the stochastic nature of wind patterns and gravitational influences. This methodical approach preserved 95% of baseline landing success rates while enabling robust adaptation to multi-parameter environmental challenges, confirming learning stability through quantified performance retention.
2. Graduated Challenge Introduction Allowing Systematic Capability Development: Progressive wind parameter escalation from 0.1 to 0.3 m/s enables systematic skill acquisition without policy collapse. Initial training establishes basic landing mechanics under minimal disturbance, followed by intermediate complexity development. Advanced stages incorporate full environmental complexity with wind, gravity, and terrain variations. This learning approach prevents disastrous failure while systematically expanding operational capabilities across increasingly demanding scenarios.
3. Comprehensive Reward Engineering Balancing Multiple Competing Objectives: Multi-objective reward structure integrates landing precision (+1000 for perfect touchdown), safety constraints (-3000 for planet collision). Dynamic reward scaling accounts for environmental complexity, with terrain difficulty multipliers (1.5x for rocky/crater surfaces) and proximity-based penalties for dangerous navigation. Secondary objectives include wind adaptation rewards and exploration bonuses, ensuring balanced optimization across mission-critical performance metrics. The reward system maintains safety priorities while encouraging efficient and precise autonomous landing behaviors.
4. Robust Safety Systems Preventing Catastrophic Policy Development: Environment-level safety mechanisms include severe collision penalties (-3000 reward) for planet contact or bypass violations, immediately terminating episodes to prevent catastrophic navigation behaviors. Conservative reward structures provide positive reinforcement (+5.0) for maintaining safe distances while implementing scaled danger penalties (up to -500) for unsafe approaches or incorrect landings. Bounded action spaces inherit continuous control limits from the base LunarLander-v3 environment, ensuring thrust vectoring remains within safe operational parameters (±1.0). Progressive reward scaling through terrain difficulty multipliers (1.5x for rocky/crater surfaces) and strategic penalty structures guide policy development toward reliable autonomous operation while preventing destructive behaviors through comprehensive safety constraints.

9.2 Practical Implications

The enhanced system provides a realistic training environment for autonomous landing systems, incorporating challenges representative of actual space missions:

• Atmospheric disturbances simulate realistic landing conditions
• Terrain variations prepare systems for diverse landing sites
• Gravitational obstacles represent celestial body navigation challenges

10. Conclusion

This research successfully demonstrates the evolution from basic lunar landing capabilities to sophisticated multi-environmental navigation and landing systems. Through systematic enhancement of environmental complexity and careful parameter optimization, we achieved robust performance in challenging scenarios while maintaining precision landing requirements.

The enhanced PPO implementation successfully navigates wind disturbances, adapts to terrain variations, avoids planetary obstacles, and consistently achieves precision landings between designated flags. This progression from basic to advanced capabilities provides a comprehensive framework for autonomous spacecraft landing system development and represents a significant advancement in reinforcement learning applications for aerospace engineering.

The modular design and comprehensive safety systems developed in this research provide a solid foundation for future autonomous navigation system development, with direct applications to real-world space mission planning and execution.

11. References

1. [1] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347. Available at: https://arxiv.org/abs/1707.06347
2. [2] Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016). OpenAI Gym. arXiv preprint arXiv:1606.01540. Available at: https://arxiv.org/abs/1606.01540
3. [3] Towers, M., Terry, J. K., Kwiatkowski, A., Balis, J. U., Cola, G. D., Deleu, T., … & Ravi, R. (2023). Gymnasium. Zenodo. DOI: 10.5281/zenodo.8127025. Available at: https://gymnasium.farama.org/
4. [4] Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., & Dormann, N. (2021). Stable-Baselines3: Reliable Reinforcement Learning Implementations. Journal of Machine Learning Research, 22(268), 1-8. Available at: https://github.com/DLR-RM/stable-baselines3
5. [5] Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press. ISBN: 978-0262039246
6. [6] Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., … & Kavukcuoglu, K. (2016). Asynchronous Methods for Deep Reinforcement Learning. International Conference on Machine Learning, 1928-1937. Available at: https://arxiv.org/abs/1602.01783

12. Reference Justification:

[1] PPO Algorithm: Core algorithm used in both baseline and enhanced implementations 

[2] OpenAI Gym: Original environment framework that Gymnasium extends 

[3] Gymnasium: Current environment framework used (LunarLander-v3) 

[4] Stable-Baselines3: Primary RL library used for PPO implementation 

[5] Sutton & Barto: Foundational reinforcement learning textbook 

[6] A3C Paper: Related policy gradient method for comparison and context

13. Appendix

13.1 Code for sophisticated environmental challenges

class EnhancedLunarLander(gym.Env):

    def __init__(self, render_mode=None, terrain_type=’flat’, enable_planet=True, enable_wind=True):

        super(EnhancedLunarLander, self).__init__()

        self.base_env = gym.make(‘LunarLander-v3’, render_mode=render_mode, continuous=True)

        self.terrain_type = terrain_type    # ‘flat’, ‘rocky’, ‘crater’

        self.enable_planet = enable_planet

        self.enable_wind = enable_wind

13.2 Standard 8-dimensional observation space

self.observation_space = spaces.Box(

    low=np.array([-1.0, -1.0, -5.0, -5.0, -3.14, -5.0, 0.0, 0.0]),

    high=np.array([1.0, 1.0, 5.0, 5.0, 3.14, 5.0, 1.0, 1.0]),

    dtype=np.float32)

13.3 Extended 11-dimensional observation space

self.observation_space = spaces.Box(

    low=np.array([-1.0, -1.0, -5.0, -5.0, -3.14, -5.0, 0.0, 0.0, -1.0, -1.0, 0.0]),

    high=np.array([1.0, 1.0, 5.0, 5.0, 3.14, 5.0, 1.0, 1.0, 1.0, 1.0, 5.0]),

    dtype=np.float32)

13.4 Runtime observation extension

planet_relative = (self.planet_pos – lander_pos) / 100.0

planet_distance = np.linalg.norm(planet_relative)

observation = np.concatenate([

    observation,

    planet_relative,

    [planet_distance]]]

def _get_planet_influence(self, lander_pos):

    if not self.enable_planet:

        return np.zeros(2)

    direction = self.planet_pos – lander_pos

    distance = np.linalg.norm(direction)

    # Enhanced safety margins and reduced gravitational pull

    min_distance = self.planet_radius * 2.0 # Increased safety margin

    if distance < min_distance:

        distance = min_distance

    # Inverse square law for gravity with reduced strength

    force = self.planet_gravity / (distance * distance)  # Inverse square law

    normalized_direction = direction / distance

    # Add rotational component for navigation complexity

    perpendicular = np.array([-normalized_direction[1], normalized_direction[0]])

    rotational_force = force * 0.3 # 30% of the main gravitational force

    # Combine direct gravitational pull with rotational force

    return force * normalized_direction + rotational_force * perpendicular

# Safety Navigation Rewards

if planet_distance > min_safe_distance:

    reward += 5.0  # Higher reward for keeping safe distance

else:

    danger_factor = (min_safe_distance – planet_distance) / min_safe_distance

    reward -= danger_factor * 10.0  # Proximity penalties

# Perfect Landing Bonuses between flags

if observation[6] == 1:  # Landed between flags

   # Landed between flags and far from planet

    if abs(observation[0]) < 0.12 and planet_distance > min_safe_distance:

        reward += 1000  # Perfect landing bonus

    else:

        reward -= 500  # Landing violation penalty

14. GitHub Link

https://github.com/hireshmit/lunarlander

About the author

Hireshmi Thirumalaivasan

Hireshmi Thirumalaivasan is a high school senior with a passion for aerospace engineering and artificial intelligence. Under the mentorship of Dr. Bilal Sharqi at the University of Michigan, she explored how AI tools are utilized in autonomous spacecraft landing systems with multi-environmental challenges (wind effects, gravitational obstacles, and variable terrain) through the Gymnasium framework and PPO reinforcement learning to train an autonomous lunar lander.

She plans to continue her research journey in aerospace engineering, aiming to benefit society by applying the knowledge gained to develop tools, such as drones, that can deliver medicine and provisions to impoverished areas. Beyond academics, she is involved in Taekwondo, her school’s newspaper club, tutoring, and FCCLA.

The post Enhanced Lunar Lander (Autonomous Spacecraft Landing System with Multi-Environmental Challenges) appeared first on Exploratio Journal.

Abstract

Integrated circuits, familiarly known as chips, have tremendous applications in modern society within several fields such as medicine, entertainment, and computer engineering. These chips’ layout and design are crucial toward their overall functionality, efficiency, and power consumption, and are constantly being changed to achieve the highest possible level of results. Existing methods, such as modular placement using quadratic or clustering techniques, have been able to achieve moderate to high levels of efficiency. However, these methods face challenges in comprehensively accounting for all factors that contribute to improved chip performance. Recently, the method of using reinforcement learning for the placement of modules within chips has been explored and discussed by researchers as a viable technique for ideal placement. The purpose of this paper is to discuss the previous methodologies of improving module placement and comparing these methods to the usage of reinforcement learning within this problem domain; furthermore, this paper also explores the possible benefits and detriments of reinforcement learning, as well as its implications for future usage within this domain. This literature review observes that reinforcement learning methods for placement demonstrate improvements in wirelength but require refinement in the context of meeting design constraints and requirements.

I. Introduction

Integrated circuits are small units created from precise connections between electronic components (Pan, 2023), such as resistors and transistors, that are typically built on a silicon (Si) substrate; integrated circuits are physically manufactured using photolithography – using ultraviolet (UV) light to place thousands of components onto the substrate at once (Saint & Saint, 2018). These chips are initially designed through a meticulous process with several components, including logic design, where the necessary functionality of it is defined through logic simulations, and physical design, where important functions and the input and output ports of the circuit are defined (Synopsys, n.d.). The flow of the design process includes the stages in the front-end, like design and logic synthesis, followed by back-end processes such as placement and routing (Cadence 2023). Each of these phases within the design process must balance performance, power, and area availability (PPA) metrics under nanometer-scale constraints. Many modern algorithms are made to prefer one of these three PPA metrics specifically; as will be discussed later, reinforcement learning can try to consider all three factors to create the best possible output.

Within this design process, module placement emerges as the central challenge for PPA optimization that influences signal delay, routing congestion, and overall power consumption. The performance of these circuits heavily relies on integration density and signal congestion, both factors which hinge on the effective placement of modules within the integrated circuit. As the geometries of devices shrink over time, improved module placement becomes a bigger priority, which dictates the physical proximity of high-interaction blocks and ultimately chip yield.

II. Background on Reinforcement Learning

Reinforcement learning is a machine learning technique that involves an algorithm, as a result of learning through continuous interactions with its environment (Kumar Shakya, Pillai, & Chakrabarty, 2023), creating an optimal behavioral strategy (known as a policy) based on reward signals given from the interactions. Rewards are mostly dense, meaning feedback is given often; partial, meaning that the signal is given in intermediate steps before the goal; or sparse, which only provides rewards for specific goals. The policy tells the reinforcement learning agent, or entity that is trying to learn the specific task, how to act for any given state – the situation or outline the agent is placed in. Reinforcement learning is sequential-based (Kumar Shakya, Pillai, & Chakrabarty, 2023), meaning that the reinforcement learning agent makes decisions over time where each action taken by the agent affects the next state and the overall outcome. The overall benefits and detriments of each state and action are evaluated using value functions — specifically, the State-Value Function and the Action-Value Function (Hardik, 2025).

Shown in (1) is the Action-Value Function which determines how beneficial or detrimental an action was based on the policy (π). The function returns an expected value from an action (a) upon an initial state (s). S_t and A_t represent the state of the environment and action taken at time t (Hardik, 2025)

displays the State-Value Function, which determines how optimal a state is when following the policy denoted by π . It returns an expected value given the state (s) and the following policy acted upon the state. Here, S_t is the current state at time t similar to (1) (Hardik, 2025).

In both equations, E_π represents the expectation over all future sequences in combination G_t which is the total discounted reward starting at a time (t). The summation from k=0 to infinity is the infinite sum over all of the future possible rewards, and this summation is taken into account with the discount factor y^k comparing future to immediate rewards as well as the reward (R_(t+k+1)) received k+1 steps after the current state.

The goal of a reinforcement learning algorithm is to find the most optimal policy possible to take the best action that maximizes positive rewards given any state within the system. Using equations such as the State and Action Value Functions guides the search for optimal policies through taking subsequent states and actions into consideration.

III. Methodology

To conduct research and analysis on how reinforcement learning and traditional methods can be applied to modular placement within integrated circuits, I performed a structured literature review and comparative analysis. The goal of using this methodology was to identify the traditional methods of modular placement, evaluate recent approaches for using reinforcement learning within the problem domain, and finally compare the classical methods with reinforcement learning based on PPA metrics. Some databases I searched to conduct this literature review included ArXiv, Google Scholar, and IEEE Xplore to find constructive, peer-reviewed papers on placement methods and reinforcement learning. To ensure relevance and quality, papers from these databases were selected using the following criteria: they presented a placement methodology which is applied to integrated circuit design, described measurable outcomes tied to PPA metrics, and introduced techniques within the methodology that apply to macro or global placement. The final stage of the methodology consisted of categorizing classical methods and reinforcement learning methods separately, identifying their strengths and areas for improvement, and summarizing emerging themes that motivate future research within this area.

IV. Traditional Placement Approaches

For refined placement and routing within an integrated circuit to maximize PPA, several methods have been taken to achieve the most efficient setup possible.

One of these is the methodology of using quadratic placement to minimize wirelengths by treating the modules within the integrated circuits as “springs.” For quadratic placement, a commonly used placement tool is GORDIAN (Kleinhans et al., 1991), which, using the calculated positions of the cells, places all modules concurrently during each partitioning stage. The loop within GORDIAN aims to achieve the smallest possible wirelength given the area constraints for placement. It achieves this result through, given an input of a netlist—a description of how electronic components connect within a circuit—, repetition of a loop until each partitioned region of the area is filled with a certain number of modules (Kleinhans et al., 1991). The decision on where to create partitions for GORDIAN comes from the calculated global placement of the modules, but these partitions can also be improved through constant verification of the partition decisions to get the most desired area ratio (Kleinhans et al., 1991). Quadratic methods, although efficiently optimizing wirelengths, face several limitations such as strict non-overlap guarantees and the neglect of congestion and routing.

Along with formulating the placement of modules as a quadratic placement problem, another method coupled with this is the integration of using clustering and unclustering (Nam et al., 2006) to reduce the scale of the problem while keeping the quality of the solution. A placement program called analytical top-down placement (ATP) has hierarchical clustering (Nam et al., 2006) — a method in machine learning that structures clusters of similar data points in a hierarchical manner — integrated with it which speeds up the global (all-encompassing) placement of modules within the integrated circuit and makes favorable performance tradeoffs in order to achieve minimal wirelength and effective handling of fixed blocks such as memory blocks. In comparison with normal ATP, there is a 2.1 times increase in runtime with an approximately 1.4% improvement in wirelength (Nam et al., 2006), proving that using hATP (hierarchical ATP) is a viable method for achieving mostly ideal results in regards to the task of module placement. On the other hand, the method of analytical bottom-up clustering, as opposed to ATP, needs to have an associated effective clustering function that achieves a targeted ratio for a clustered netlist as compared to the total modules in the original netlist (Nam et al., 2006), which means that the problem size of the netlist is essentially being reduced through clustering in a different manner than ATP but still works efficiently. Along with this challenge, the clustering methods must respect block constraints that are immovable to ensure that the clusters remain feasible, or achievable.

Modern designs are moving towards efficient 3D integration with modular placement in chips. Using 3D integration with mixed-sized modules (Zhao et al., 2025), such as large macros and standard cells, aims to satisfy interconnectivity and non-overlapping constraints. The framework of this methodology involves using a 3D global placement approach to explore the placement area provided for the shortest possible wirelengths; minimal wirelengths are achieved through macro-rotations in between global placements (Zhao et al., 2025). By doing this method, possibly limited solution quality, which is an effect of partitioning, is avoided; however, overlaps or infeasability still may occur, which requires a legalization (rule-validation) fix or discrete steps for correction.

As integrated circuit designs grow increasingly more complex, though, traditional placement methods’ common limitations motivate growth and exploration into using reinforcement learning for the placement of modules.

V. Applying Reinforcement Learning to Module Placement

In recent years, reinforcement learning has emerged as a viable tool for improving the placement of modules within integrated circuits. The methodology of framing the modular placement problem as a reinforcement learning task makes the policy a deep neural network (Goldie & Mirhoseini, 2020) and expresses partial rewards from the given state-action function as a proxy-cost function—a blend of metrics such as wirelength, congestion, and density—which estimates the final power, performance, and area (PPA). Evaluating the full PPA metrics would be computationally prohibitive, so per training cycle, the reward function relies on an analytical cost estimator (Goldie & Mirhoseini, 2020) to see how much power, performance, and area the reinforcement learning agent is using. The training of the agent involves using policy-gradient reinforcement learning (Goldie & Mirhoseini, 2020): episodes, or complete sequences of an agent’s interactions in reinforcement learning, consist of sequential macro (larger-scale) placements until all blocks are positioned, and then each episode’s reward, derived from the placements’ congestion, wirelength, and density, is normalized and used to update the parametrized policy, a type of policy that is explicitly defined by a set of adjustable parameters (Goldie & Mirhoseini, 2020). The interactions between a reinforcement learning agent and its environment are demonstrated in Figure 1. This reinforcement learning usage allowed for near-optimal PPA to be achieved in a short amount of time, and the framework is very automated which reduces some need for legalization, which is a persisting issue with reinforcement learning algorithms. Similarly, traditional quadratic placement methods also require legalization, but reinforcement learning enables for faster convergence and more desirable results with not only improved wirelengths, but also reduced amounts of power usage.

Chip macro-module placement can also be done through hierarchical reinforcement learning (Tan & Mu, 2024) as opposed to deep reinforcement learning. A two-level policy approach is used for this method, where the high-level agent determines partitioning for the macros, which allows the lower-level agent’s policy to granularly place macros within the partitioned region of the total allocated area. By adopting Hierarchical Reinforcement Learning for Placement (HRLP), dense rewards are calculated for each episode by observing differences between congestion and wirelength to better all three metrics of PPA (Tan & Mu, 2024). The proposed model demonstrated improvements for Half-Perimeter Wire Length of 9.29% and 14.25% (Tan & Mu, 2024), which is extremely substantial; but, standard cell placement, in contrast with macros, still relies on traditional analytic tools and the initial partitioning of the entire placement region could restrict the reinforcement learning agent from discovering globally optimal modular placements.

Figure 1. A reinforcement learning agent interacts with its environment to achieve the best possible rewards based on the action taken and the state. (Goldie, A., & Mirhoseini, A. (2020))

Developing reinforcement learning methods for placement naturally create points of comparison with traditional methods to evaluate both the relative strengths and weaknesses of each technique.

1. Comparing Reinforcement Learning and Traditional Methods

Traditional methods and reinforcement learning have both proven to be viable methodologies of placing modules within integrated circuits. However, reinforcement learning has shown to have a greater impact in improving wirelengths within the circuit as compared to previous methods: a substantial 14.25% Half-Perimeter Wire Length (Tan & Mu, 2024) improvement from HRLP as compared to hATP’s 1.4% (Nam et al., 2006) wirelength improvement demonstrates a greater benefit for the PPA metrics of area and performance through improving chip compactability and operating frequency. Additionally, reinforcement learning has potential for faster algorithms with the usage of neural networks, but these can be computationally expensive to train as compared to methods such as quadratic placement. Reinforcement learning also has a more demanding need for legalization as compared to classical methods due to the increased potential for the algorithms to not adhere to all of the design constraints, possibly causing modules to overlap, leading to short circuiting. Overall, reinforcement learning’s demonstrated benefits but possibly detrimental costs can greatly impact whether classical methodologies will continue to be widely used for modular placement within integrated circuits. An overview of the different methods’ strengths and limitations is displayed in Table 1.

Table 1 Comparison Table of Placement Methods

Note. All values and descriptions displayed in this table are drawn directly from the information on the placement methods discussed in the preceding sections.

While this comparison between the traditional and reinforcement learning approaches for module placement highlights clear trends across the methods, it also reveals challenges and openings for improvement that shape the limitations of the current research landscape.

VI. Limitations

Although this review produces a comparative analysis of traditional integrated circuit module placement approaches and reinforcement learning-based techniques, some limitations exist that constrain the scope of the findings. As a part of this literature review, some traditional methods, such as simulated annealing-based placement and force-directed placement, were not covered because of their similar usages and functions within integrated circuit module placement to popular methods like quadratic placement and hierarchical placement. Some earlier placement methods were also not included due to their limited public availability, and detailed placement as a whole is not fully addressed in this review. Finally, reinforcement learning research on integrated circuits remains a developing field in which the experimental techniques are not evaluated on consistently benchmarked circuits; this fact limits some direct comparison between methods across extensive studies. Recognizing these discussed limitations helps clarify how results of the study should be interpreted and sets the stage for outlining the broader implications and future outlook within this problem domain.

VII. Conclusions

Modular placement within integrated circuits can be optimized through a variety of methods, with each method having its own constraints, uniqueness, and priorities for efficiency. Modern algorithms have proven to be useful in achieving consistently moderate to high efficiencies, but they are limited in the sense that some PPA metrics typically must be prioritized over the others. Reinforcement learning has its benefits of trying to optimize all factors within PPA through a sequential-based learning process, but can also come with detriments such as the need to legalize certain actions to limit structural damage, the need for scalability and the unpredictability of actions. As a result, legalization of reinforcement learning algorithms should move toward being reduced overhead by using predictive filters or developing constraints.

Additionally, reinforcement learning should be tested on standardized benchmark circuits initially to evaluate the performance of the algorithm separately on macro-heavy and standard cell-heavy designs. Considering that there are various factors to account for within the problem domain, future studies should explore specific applications of reinforcement learning — such as with single-factor reward functions or with graph neural networks for improved netlist connectivity — on one PPA factor at a time so that we can further our knowledge on how to best optimize modular placement, which will lead to improved signal integrity, lower power dissipation, and an overall more efficient chip performance.

VIII. Acknowledgements

I would like to thank Dr. Waheed U. Bajwa for his mentorship and guidance in my research, as well as reviewing and helping to edit this paper.

References

Pan, Y . (2023). Typical Application Circuit Analysis of Digital Integrated Circuits in Electronic Devices. Proceedings of the 6th International Conference on Information Technologies and Electrical Engineering, 60–64. https://doi.org/10.1145/3640115.3640125

Saint, C., & Saint, J. L. (2018). integrated circuit | Types, Uses, & Function. In Encyclopædia Britannica. https://www.britannica.com/technology/integrated-circuit IC Design and Manufacturing Process. (2023, August 2). Resources.pcb.cadence.com. https://resources.pcb.cadence.com/blog/2023-ic-design-and-manufacturing-process‌

What is Integrated Circuit (IC) Design? – How Does it Work? | Synopsys. (n.d.). Www.synopsys.com. https://www.synopsys.com/glossary/what-is-ic-design.html

Kumar Shakya, A., Pillai, G., & Chakrabarty, S. (2023). Reinforcement Learning Algorithms: A brief survey. Expert Systems with Applications, 231, 120495. https://doi.org/10.1016/j.eswa.2023.120495

Dave, H. (2021, February 13). Bellman Optimality Equation in Reinforcement Learning. Analytics Vidhya. https://www.analyticsvidhya.com/blog/2021/02/understanding-the-bellman-optimality-eq uation-in-reinforcement-learning/

Kleinhans, J. M., Sigl, G., Johannes, F., & Antreich, K. (1991). GORDIAN: VLSI placement by quadratic programming and slicing optimization. 10(3), 356–365. https://doi.org/10.1109/43.67789

Nam, G.-J., Reda, S., Alpert, C. J., Villarrubia, P. G., & Kahng, A. B. (2006). A fast hierarchical quadratic placement algorithm. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 25(4), 678–691. https://doi.org/10.1109/tcad.2006.870079

Zhao, Y ., Liao, P., Liu, S., Jiang, J., Lin, Y ., & Yu, B. (2025). Analytical Heterogeneous Die-to-Die 3-D Placement With Macros. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 44(2), 402–415. https://doi.org/10.1109/tcad.2024.3444716

Goldie, A., & Mirhoseini, A. (2020). Placement Optimization with Deep Reinforcement Learning. ArXiv.org. https://arxiv.org/abs/2003.08445

Tan, Z., & Mu, Y . (2024). Hierarchical reinforcement learning for chip-macro placement in integrated circuit. Pattern Recognition Letters, 179, 108–114. https://doi.org/10.1016/j.patrec.2024.02.002‌

About the author

Vishal Ramamurthy

Based in the Bay Area, Vishal is a student at Carlmont High passionate about the applications of engineering principles and technology in our world today. His academic interests consist of computer science, engineering studies, physics and chemistry. Besides his academics, Vishal works on and develops engineering projects that have real-world applications such as XIMIRA LLC’s PHINIX mechanism to assist the visually-impaired with everyday life and a Python machine learning project that classifies cancerous breast tumors as malignant or benign. He has a deep interest in pursuing a career in engineering, particularly in the fields of electrical and computer engineering.

The post Optimizing Placement of Integrated Circuit Modules using Reinforcement Learning appeared first on Exploratio Journal.

Abstract

According to the CDC, one in four adults in the United States has some type of disability. When government websites are not accessible, they effectively exclude millions of citizens from essential public services and perpetuate systemic barriers to information and participation. This study provides a comprehensive evaluation of the digital accessibility of all 50 U.S. State Department of Transportation websites, employing a mixed-methods approach that integrates automated auditing with qualitative design analysis. Leveraging the Skynet Technologies Free Accessibility Checker, we gathered quantitative data on compliance with WCAG 2.2 standards, including compliance percentages (e.g. Ohio at 94.8%, Kansas at 14.7%), the total number of failed checks (e.g. ranging from 7 to 33), and the most commonly affected disability categories such as visually impaired users, people with cognitive or learning abilities, and users with dyslexia or color blindness. Qualitative analysis captured recurring usability issues such as semi-transparent image overlays, outdated web interfaces and a disconnect between stated ADA compliance and actual user experience. The results reveal wide disparities in accessibility performance across states and highlight the limitations of treating accessibility as a technical checkbox rather than a design imperative. Our findings call for a shift toward inclusive, user-centered practices in public digital infrastructure, where accessibility is embedded from the beginning and aligned with both legal mandates and civic responsibility.

Background

Web accessibility is the inclusive practice of designing digital platforms so that people with a wide range of disabilities, including visual, auditory, motor, or cognitive can perceive, navigate, and interact with content effectively. This includes accommodations for users who rely on screen readers, keyboard navigation, alternative text on images, and high-contrast visual design. Accessibility is particularly important for public sector websites, where equitable digital access can directly impact people’s ability to obtain critical services. Among these platforms, State Departments of Transportation, like state-level government agencies responsible for planning and coordinating federal transportation projects which set safety regulations for all major modes of transportation (USAgov, 2019): host websites that are frequently used by millions of people to get driver’s license-related information, job listings, construction alerts, weather-related road closures, and more. When addressing accessibility it is important to note that more than 1 in 4 adults in the United States has some type of disability (CDC, 2020). When these websites are inaccessible, they exclude citizens from critical public services and reinforce systemic barriers to information.

The central problem is that despite legal and technical standards, accessible information across State Departments of Transportation websites remains inconsistent and insufficient. The Americans with Disabilities Act (ADA)– i.e., a civil rights law that prohibits discrimination against individuals with disabilities in many areas of public life, including jobs, schools, transportation, and many public and private places that are open to the general public (ADA National Network)–applies to digital services offered by state agencies. While the ADA establishes the legal foundation for accessibility, it does not specify technical requirements for digital content. That role is filled by the Web Content Accessibility Guidelines (WCAG), developed by the World Wide Web Consortium (W3C) to ensure content is perceivable, operable, understandable, and robust. The latest version, WCAG 2.2, outlines criteria such as color contrast, keyboard navigation, and logical heading structures (W3C, 2019). Although not law, WCAG is widely recognized as the standard for evaluating ADA compliance in audits and court cases (Gibson, 2024).

Ensuring accessibility is not just a matter of legal compliance, but also of public equity, civil participation, and good digital governance. As the U.S population ages and more individuals identify as having disabilities, the need for inclusive design becomes increasingly urgent. Accessible design– i.e., the practice of designing products, services, and environments that can be accessed, understood, and used by all individuals (Accessibility in Design – Definition and Explanation, 2024) also improves overall usability for people without disabilities such as people using mobile devices or unfamiliar platforms. When public agencies fail to prioritize accessibility, they risk excluding large segments of the population from essential services and commutation. This not only violates the spirit of ADA but also undermines the effectiveness of digital government.

To investigate these issues, we conducted a mixed-methods evaluation of all 50 U.S State Department of Transportation websites. Based on our data we identified several recurring qualitative trends: many sites made frequent use of semi-transparent images that interfere with text readability, almost all websites included ADA documentation but failed to follow through with the actual implementation, and a significant number of sites exhibited a retro web design–i.e., style incorporating visual, typographic, and layout elements from past decades like bold color palettes, pixel-style graphics, retro fonts, and aesthetics nods to web design from the 1980s, 1990s, and early 2000s (Seattle SEO Company, 2022). In parallel, we also captured compliance percentage, and the most commonly affected disability categories like visually impaired users, people with cognitive or learning abilities, and users with dyslexia or color blindness. These findings reveal that technical compliance and user-centered design often diverge, and that while some states show strong adherence to WCAG standards, others fall short due to overlooked accessibility principles.

Related Work

Research on digital accessibility in public-sector services has revealed widespread noncompliance with the Americans with Disabilities Act, especially on state-run websites. For example, Jaeger argues although the ADA legally guarantees equal access to public services, its digital enforcement remains weak, leading to systemic exclusion for people with disabilities, especially on platforms operated by state and local governments (Olalere & Lazar, 2011). Goode builds on this by examining how Title II of the ADA, which applies to public entities, often lacks enforceability when applied to web infrastructure, leaving many agencies noncompliant without legal consequence (Goode, 2021). These studies provide critical legal and infrastructural context, but stop short of assessing individual domains, such as transportation agencies. They emphasize the existence of a digital divide not only due to technology access, but due to suboptimal design decisions that fail to meet technical accessibility benchmarks like those defined in WCAG 2.2.

In parallel, the importance of website aesthetics and structure in shaping usability has been explored through design-focused studies of government web portals. Watkins and Wills, for instance, analyze the digital design of U.S. city government websites and describe a recurring “legacy design trap”, in which outdated layouts, unresponsive interfaces, and poor information hierarchy diminish the user experience, particularly for underrepresented and aging populations (Wagner et al., 2024). In the transportation sector, Graham Currie and Mandy Gook conduct usability testing on a sample of transportation agency websites, identifying serious issues in visual consistency, navigation, and user trust. Their work supports the idea that design shortcomings are not purely aesthetic but functionally consequential in reducing civic engagement (Currie & Gook, 2009). Similarly, Patricia Acosta-Vargas applied automated tools to evaluate a set of business websites for accessibility metrics, providing a technical foundation for large-scale digital evaluations (Acosta-Vargas et al., 2017). Unlike these studies, which focus on localized or small-scale usability assessments, my analysis interrogates accessibility as both a design and equity issue, examining how aesthetic and structural flaws intersect with legal compliance gaps. This approach moves beyond surface-level usability to reveal how design decisions can perpetuate systemic exclusion within essential public infrastructure.

A clear gap remains in applying accessibility research insights to a comprehensive, cross-state assessment of Department of Transportation websites. This study addresses that gap by conducting a dual-pronged evaluation of all 50 U.S. State Department of Transportation websites, an especially critical domain given that these agencies serve as gateways to essential public services such as road safety updates, construction notices, licensing, public transit schedules, and emergency evacuation information. When these sites are inaccessible, individuals with disabilities face disproportionate barriers to mobility, safety, and civic participation. Guided by the question of to what extent these websites comply with WCAG 2.2 accessibility standards and how design choices affect their usability and inclusivity for people with disabilities, we pairautomated quantitative auditing with qualitative observations. By bridging legal, technical, and experiential dimensions, our research uniquely situates itself within and beyond existing literature, offering a holistic, data-informed snapshot of how Department of Transportation websites across the U.S. comply with the standards of digital accessibility and modern design in 2025.

Methods

This study employed a mixed-methods approach to evaluate the accessibility of all 50 U.S. State Department of Transportation websites. The URLs for each website were obtained from the Federal Highway Administration’s directory (U.S. Department of Transportation, 2019) to ensure official and consistent sources across all states. Before conducting the full audit, three web accessibility tools, Skynet Technologies Free Accessibility Checker, AccessibilityChecker.org, and AEL Accessibility Checker, were pilot tested to determine the most comprehensive and reliable platform. The Skynet tool was ultimately selected for its detailed reporting capabilities, which include overall compliance percentages, issue categories (for example, clickables, tables, audio/video), WCAG 2.1 conformance levels (A, AA, AAA), and mapped locations of accessibility violations within the HTML structure. Although the tool required more manual time per page and lacked fine-grained disability categorization, it provided the most consistent and unrestricted data collection without login barriers or scan limits.

Because Department of Transportation websites are large and highly variable, only two pages per site were selected for auditing: the homepage and the first critical navigation page. This decision balanced cross-state comparability with practical feasibility while still capturing the sections most frequently used by the public. The homepage was selected as the most common entry point for both general and assistive-technology users, while the first critical navigation page represented the site’s most essential public task. To avoid arbitrary selection, a standardized decision rule was applied: beginning from the homepage’s global navigation bar, the first listed link that matched one of the following categories, Driver Services or Licensing, Road Conditions or Closures, Jobs or Careers, Transit Schedules or Permits, was chosen. If multiple categories appeared, “Driver Services/Licensing” was prioritized due to its broad public relevance. When a navigation menu lacked those options, the first link leading to a transactional or informational task page (rather than a press release or PDF list) was selected. This consistent process ensured methodological transparency and reproducibility. While auditing only two pages limits the comprehensiveness of within-site analysis, it allowed for a uniform evaluation across all fifty states and captured the design and accessibility conditions most visible to users.

Each selected page was scanned using the Skynet checker, and the resulting data were recorded for each state, including the percentage of accessibility checks passed, total number of failed checks, categories of issues, corresponding WCAG conformance levels, and affected disability types. A subset of flagged items, particularly color contrast and missing form labels, was manually verified through HTML inspection to validate the tool’s accuracy. Beyond automated results, qualitative observations were added to capture design elements not detected by scanning software, such as semi-transparent overlays, poor visual hierarchy, and mismatched ADA statements. To analyze these qualitative features systematically, an a priori codebook was developed around recurring design themes such as legibility, information structure, and visual clutter. Two independent coders applied the codebook to a stratified 20 percent sample of websites selected across high, medium, and low compliance categories. Inter-rater reliability was calculated using Cohen’s kappa, with a target threshold of 0.75 for substantial agreement. Discrepancies were resolved collaboratively, and the finalized codebook was applied to the remaining sites by the primary coder with periodic spot checks to prevent drift.

To strengthen internal validity, a post hoc subset of ten websites underwent manual WCAG checklist testing focused on success criteria most frequently implicated in the automated findings, including 1.4.3 (Contrast), 1.3.1 (Info and Relationships), and 2.1.1 (Keyboard Navigation) (W3C, 2024). These manual checks were paired with basic task-based tests, such as locating renewal information or road closures using keyboard-only navigation, to evaluate whether the issues flagged by automation corresponded to tangible usability barriers.

Several limitations accompany this methodology. First, analyzing only two pages per website constrains the ability to generalize findings across entire sites. The decision to do so reflects a necessary balance between breadth, covering all fifty states, and depth, though future research should expand to include deeper navigational flows and internal task pages. Second, automated tools typically detect only 30 to 40 percent of accessibility violations, as many issues, such as reading order, focus visibility, and contextual link meaning, require human interpretation. Although manual validation and qualitative review were used to mitigate this limitation, undetected errors may remain. Third, while inter-rater reliability was established on a subset of sites, qualitative interpretation beyond that sample may still contain subjectivity. Additionally, because Department of Transportation websites frequently update banners, alerts, and layouts, the results represent a snapshot of accessibility performance at a single point in time. Lastly, the findings are partially shaped by Skynet’s proprietary detection algorithms, meaning results could vary with alternative auditing platforms. Despite these constraints, the combined quantitative and qualitative approach offers a robust, reproducible framework for assessing both technical accessibility and user-centered design quality across large-scale public web infrastructure.

Results

One of the most revealing findings in our audit was the wide variation in accessibility compliance across the 50 U.S. State Department of Transportation websites. Overall compliance scores ranged from 14.7 percent (Kansas) to 94.8 percent (Ohio), with a mean of 68.9 percent, a median of 70.4 percent, and a standard deviation of 16.2. This range indicates substantial disparity in digital accessibility across states. While a small subset of websites exceeded 90percent compliance, suggesting strong alignment with WCAG 2.2 standards, nearly half of the states fell below 70 percent, placing them in the semi-compliant or noncompliant category. These findings underscore that accessibility is not being uniformly prioritized, even though these websites serve as primary entry points to essential public services.

In tandem with compliance percentages, the total number of failed checks per site provided further insight into accessibility gaps. Across all 50 websites, the number of failures ranged from 7 to 33, with a mean of 18.5, a median of 17, and a standard deviation of 6.3. A Pearson correlation analysis between the compliance percentage and number of failed checks revealed a strong negative relationship (r = –0.87), indicating that lower compliance percentages were closely associated with higher counts of accessibility violations. This confirms that automated scoring aligned with practical accessibility performance: as the number of failures increased, overall compliance predictably declined.

Across all sites, the most frequent accessibility failures involved poor color contrast, unlabeled form elements, missing alternative text, and improperly structured headings. Color contrast errors were the single most common issue, appearing on 88 percent of audited pages. These problems most directly affect users with visual impairments, who comprised the most impacted disability category according to tool classification data. Users with mobility impairments were also frequently affected, particularly when keyboard-only navigation failed or focus indicators were missing from interactive elements. Cognitive accessibility issues appeared less frequently, though dense text structures and inconsistent navigation patterns created additional barriers for some users.

Qualitative analysis reinforced these quantitative findings. One of the most pervasive design flaws observed was the use of semi-transparent images layered behind text or interactive elements. Over half of the websites used banners or background visuals that reduced text readability, particularly in combination with low-contrast color palettes. While intended to enhance visual appeal, these choices often compromised functional accessibility for users with low vision or reading impairments. This tension between aesthetic branding and practical usability reflects a broader misalignment between design intent and user inclusion.

A second major qualitative issue involved ADA compliance statements. Nearly all websites included a footer or dedicated page declaring adherence to the Americans with Disabilities Actor referencing WCAG standards. However, in many cases, these declarations did not correspond with actual usability. Unlabeled navigation links, inaccessible PDFs, and missing screen reader compatibility persisted despite such statements. This disconnect suggests that accessibility is too often treated as a formal requirement rather than an integrated design value.

Lastly, a large proportion of websites displayed visually and structurally outdated layouts, characterized by retro web design features such as bold color palettes, dense typography, pixel-style graphics, and nonresponsive navigation. These stylistic elements, reminiscent of early 2000s web design, were common among low-performing states and corresponded with lower compliance scores. Although such designs do not directly violate WCAG standards, they undermine usability by reducing clarity, scalability, and modern functionality. The persistence of these outdated designs indicates a lack of investment in modernization and highlights how institutional neglect can perpetuate digital inequity.

Taken together, these quantitative and qualitative results suggest that accessibility performance across U.S. State Department of Transportation websites varies widely and follows clear patterns. The strong negative relationship between compliance scores and failure counts reinforces that many accessibility problems are structural rather than incidental. The recurring qualitative trends, visual obstruction, performative ADA compliance, and outdated design, further reveal that technical adherence alone is insufficient. Accessibility, when treated as a checklist rather than a design ethic, continues to fall short of ensuring equitable digital access.

Discussion

The results of this study highlight that digital accessibility must be treated as a core design priority rather than an afterthought. One of the most revealing findings was the percentage of accessibility checks passed across the 50 U.S. State Department of Transportation websites, which provided a general benchmark of compliance with WCAG 2.2 standards. While a handful of states achieved high compliance (above 90%), a significant portion scored well below that, often in the 60–70% range, placing them in the semi-compliant category, and a few even dipped below 50%, indicating a severe lack of attention to accessibility.

This inconsistency underscores that accessibility is not being uniformly prioritized, even though these websites serve as primary entry points to essential public services. While automated tools can flag compliance issues, many of the most disruptive problems, like our qualitative issues, stem from broader design decisions that prioritize aesthetics or legacy structures over usability. This suggests that accessibility must be embedded into the design process from the beginning, with thoughtful attention to how content is visually presented, navigated, and interacted with across diverse user needs. Designers and developers should move beyond minimal compliance and adopt inclusive design practices that address both technical standards and real-world user experiences.

The findings also connect to broader discussions of digital inequality. Accessibility gaps on government websites do not merely represent technical oversights; they reinforce existing disparities in civic participation, mobility, and access to information. When people with disabilities cannot easily navigate transportation websites, they face compounded barriers to employment, healthcare, and education. These inequities mirror a larger pattern in digital governance where technological design choices can either expand or restrict public inclusion. Addressing accessibility, therefore, is not only about fixing websites but about ensuring that digital infrastructure functions as a public good that serves all users equitably.

Moreover, the gap between stated ADA compliance and actual usability points to the need for more transparent, user centered workflows in public sector web development. Merely posting accessibility statements does little if sites remain functionally inaccessible. Agencies should incorporate routine manual audits, iterative usability testing with individuals with disabilities, and continuous training for design teams to better understand accessibility beyond code-level fixes. This also calls for inter-agency collaboration and standardization efforts that can ensure consistency across states, reducing the accessibility divide.

From a policy perspective, these findings suggest several actionable steps. State and federal agencies should establish standardized accessibility benchmarks for all government websites, accompanied by annual reporting and public accountability mechanisms. Accessibility audits should be integrated into procurement and design contracts, ensuring compliance from the earliest stages of development. Federal oversight could also incentivize interagency collaboration through shared design systems and centralized accessibility resources, reducing redundancy and improving consistency across states. Finally, accessibility should be framed not just as a compliance goal but as a measure of digital equity, directly tied to broader civil rights objectives and inclusive governance.

Ultimately, digital accessibility should be treated not only as a legal and technical requirement but as a moral and civic responsibility, essential to building equitable public services. By viewing accessibility through the lens of digital inequality and embedding it within public policy and design practice, governments can move closer to realizing the promise of technology as a tool for inclusion rather than exclusion.

Conclusions

In conclusion, this study reveals that while some U.S state Department of Transportation websites demonstrate meaningful progress toward digital accessibility, the majority fall short of providing equitable, user-centered online experiences for people with disabilities. Through both quantitative data and qualitative observations, we found a persistent disconnect between stated ADA compliance and actual usability, with recurring issues that disproportionately affect users with visual, mobility, and cognitive impairments. These findings underscore the need for accessibility to be fully integrated into the design and development lifecycle of public websites as a foundational principle of inclusive governance. As digital access becomes increasingly central to civic participation and public service delivery, state agencies must recognize accessibility as both a civil rights imperative and a design obligation in order to ensure that all citizens can navigate, interact with, and benefit from digital infrastructure that supports everyday life.

Bibliography

Accessibility in Design – Definition and Explanation. (2024, June 10). The Oxford Review – or Briefings. https://oxford-review.com/the-oxford-review-dei-diversity-equity-and-inclusion-dictionary/accessibility-in-design-definition-and-explanation/

Acosta-Vargas, P., Lujan-Mora, S., & Salvador-Ullauri, L. (2017). Quality evaluation of government websites. 2017 Fourth International Conference on EDemocracy & EGovernment (ICEDEG). https://doi.org/10.1109/icedeg.2017.7962507

ADA National Network. (n.d.). What is the Americans with Disabilities Act (ADA)? ADA National Network. https://adata.org/learn-about-ada

CDC. (2020). Centers for Disease Control and Prevention. Centers for Disease Control and Prevention; CDC. https://www.cdc.gov

Currie, G., & Gook, M. (2009). Measuring the Performance of Transit Passenger Information Websites. Transportation Research Record: Journal of the Transportation Research Board, 2110(1), 137–148. https://doi.org/10.3141/2110-17

Gibson, D. (2024, November 8). 2024 WCAG & ADA Website Compliance Requirements | Accessibility.Works. Accessibility.works. https://www.accessibility.works/blog/2025-wcag-ada-website-compliance-standards-requirements/

Goode, L. F. (2021, March 8). About | HeinOnline. HeinOnline. https://heinonline.org/HOL/LandingPage?handle=hein.journals/hlelj38&div=8&id=&page=.

Olalere, A., & Lazar, J. (2011). Accessibility of U.S. federal government home pages: Section 508 compliance and site accessibility statements. Government Information Quarterly, 28(3), 303–309. https://doi.org/10.1016/j.giq.2011.02.002

Seattle SEO Company. (2022, April 16). Retro Web Design. Seattle Web Design & SEO Agency; Seattle Web Design Agency. https://visualwebz.com/retro-web-design/

US Department of Transportation. (2019). State Transportation Web Sites | Federal Highway Administration. Dot.gov. https://www.fhwa.dot.gov/about/webstate.cfm

USAgov. (2019). Official Guide to Government Information and Services | USAGov. Usa.gov. https://www.usa.gov

W3C. (2019). World Wide Web Consortium (W3C). W3.org. https://www.w3.org

W3C. (2024, December 12). Web Content Accessibility Guidelines (WCAG) Overview. Web Accessibility Initiative (WAI). https://www.w3.org/WAI/standards-guidelines/wcag/

Wagner, M., Manish Shirgaokar, Misra, A., & Marshall, W. (2024). Navigating ADA Compliance. Journal of the American Planning Association, 1–18. https://doi.org/10.1080/01944363.2024.2343661

About the author

Aashi Agarwal

Aashi Agarwal is a designer, accessibility advocate, and creative leader passionate about the intersection of technology, communication, and inclusion. She specializes in human-centered design and digital accessibility, developing tools and platforms that make information and interaction more equitable. Her work with organizations such as the NIMBLE Mindset has focused on translating complex data into intuitive, interactive storytelling that highlights impact and community. Aashi also explores how inclusive design principles can extend beyond digital interfaces into cultural and artistic spaces.

In her leadership role within the performing arts community, she has spearheaded initiatives to expand access to live theatre, including implementing ASL interpretation and inclusive audience design practices for school productions, and developing programs that welcome and support diverse participants. She aims to continue bridging the gap between structure and creativity to design systems that empower diverse voices and create meaningful, accessible experiences for all users.

The post Roadblocks to Digital Access: Accessibility and Design Gaps in 50 State Department of Transportation Websites appeared first on Exploratio Journal.

Abstract

The Max-Cut problem is a foundational NP-hard combinatorial optimization problem with quite a few uses, such as circuit and semiconductor design, financial modeling, and data flow optimization. The exponential nature of the problem makes scalability over large inputs infeasible, making it a benchmark for solving such problems using methods like quantum computing due to them being infeasible for classical algorithms in such situations.

While the Quantum Approximate Optimization Algorithm (QAOA) has shown theoretical promise for solving such problems, the practical threshold at which quantum approaches outperform classical methods on real noisy intermediate-scale quantum (NISQ) devices remains unclear. This paper aims to resolve this uncertainty. The investigation itself compares the performance of QAOA executed on IBM’s 133-qubit Torino quantum processor for the Max-Cut problem on modern quantum devices and compares its performance against the classical brute-force method in order to gauge the potential advantage that quantum algorithms may provide as input size grows. It consists of running both algorithms over (unweighted) graph sizes ranging from 4 – 24 nodes for 5 independent trials each, measuring both execution time and approximation ratio. The QAOA implementation uses a circuit with a singular fixed depth and fixed Hamiltonian parameters to ensure consistency over all trials & fairness in comparisons between graph sizes. However, this along with noise leads to a low chance of the optimal cut being produced.

The classical solver exhaustively evaluated all 2ⁿ possible partitions to produce the optimal cut value. Classical execution time grew exponentially from 0.16 milliseconds (4 nodes) to 87.6 seconds (24 nodes), consistent with the brute-force algorithm’s behavior. On the other hand, QAOA maintained near-constant execution time (~1.33 seconds across all graph sizes), with both approaches converging at approximately 19 nodes. However, QAOA’s approximation ratio declined from 0.95 (4 nodes) to 0.52 (24 nodes), reflecting limitations of shallow circuit depth and hardware noise. These findings demonstrate that QAOA exhibits superior scalability as graph size increases compared to exact classical methods but due to the limitations of NISQ quantum hardware, is unable to achieve this for large graph sizes, where it can be most practical, without accuracy as a trade-off. As quantum hardware advances in the future, high circuit depth will be possible with low noise or even error correction, which is expected to strengthen the performance of algorithms like QAOA substantially, allowing its theoretical advantages to be reaped to the fullest at solving NP-hard problems like the Max-Cut.

Key words:  compare, evaluate, execution time, approximation accuracy, scalability

Introduction

Optimization

In mathematics and computer science, optimization is a process that involves finding the best solution from the search space according to some defined criteria, while following certain rules/constraints according to the problem situation (Wright, 2025). 

The search space, in the context of optimization, refers to the set of all possible solutions that adhere to the problem’s constraints or objectives. Thus, it represents the feasible solutions that can be evaluated to find the optimal solution according to the objective function (Wright, 2025), which is a mathematical expression that defines the goal of an optimization problem, which is usually to maximize or minimize a quantity. It quantifies the desired outcome, serving as a guide for decision making and showing how close a solution is to the desired one (Fiveable, n.d.).

Combinatorial Optimization

Combinatorial optimization is a type of discrete optimization that refers to problems on discrete structures such as graphs. It aims to find the best or optimal solution to problems that have a finite set of possible solutions (discrete search space), and the best solution is usually the one that minimizes or maximizes the problem’s objective function (Lee, 2010; DeepAI, n.d.).

For each combinatorial optimization problem, there is a decision problem, which, in simple terms, is a yes/no version of it. It asks whether there is a feasible solution to the problem using a measurement threshold (Jaillet, 2010). For instance, given 10 interconnected cities, the optimization problem would be to find the shortest path from city a to city b. A corresponding decision problem would be to determine whether or not there is a path from city a to city b that crosses less than or equal to 5 intermediate cities.

Due to the nature of decision problems, if one can come up with an answer to it, that means the corresponding optimization problem is ‘feasible’ – a solution exists that satisfies all the constraints of the problem. Failing to do so means that the corresponding optimization problem is ‘infeasible’, or that no solutions exist that satisfy all constraints of the problem. Even if a problem is feasible, it may not necessarily be ‘bounded’, that is, there is no limit to how “good” or “optimal” the solution can get for the objective function. Thus, if an optimization problem is not infeasible and not unbounded, it must have an optimal solution and is therefore solvable (Maltby & Ross, n.d.).

The Max-Cut Problem

Now that discrete optimization and combinatorial optimization has been made clear, it is now appropriate to visit the ‘Max-Cut Problem’, a famous example of combinatorial optimization and the focus of this paper.

Let there be a graph G = (V,E)   with vertices V  and edges E . For starters, a graph in this context refers to a set of vertices/nodes in a 2-D space, which are mathematical abstractions corresponding to objects associated with each other by some criteria in place and are connected to one-another in some form by a set of edges E , each of which connects 2 nodes (Luca, 2023). Let there be a partition that divides set   V into 2 disjointed sets of vertices  A and  B. An edge is said to be “cut” by the partition if it connects 2 vertices that are not in the same set. Thus, the objective of the Max-Cut problem is to find a partition of vertices  V into complementary subsets A  and  B in graph V  that maximizes the edges between them (Goemans & Williamson, 1995). An example is shown in the figure below.

In the figure above, the image on the left shows the original graph with its set of nodes and edges. The image in the center shows nodes 0 and 1 in the red set and nodes 2 and 3 in the blue set, and it shows the possible cuts with that node with that specific set distribution (not the maximum cut), which is 3 cuts. The image on the right, however, shows the set distribution that gives the maximum number of cuts for the graph, which is 4 cuts – nodes 1 and 2 in the red set and nodes 0 and 3 in the blue set. It is important to note that distributing all nodes in the opposite sets – nodes 0 and 3 in the red set and nodes 1 and 2 in the blue set – will also give the same result because the edges that are “cut” will still connect nodes in different sets, which is all that matters. Note that any explorations of the Max-Cut problem in this paper deals with unweighted graphs, and the concept of weightage is being ignored entirely.

When checking whether the Max-Cut problem is solvable for all input sizes, firstly, it is imperative to assess its feasibility. To do so, assess the decision version of it, which is the following: Given a graph k   and an integer k, determine whether there is a cut value of at least   (Goemans & Williamson, 1995). It is indeed possible to have a cut of some value. Take the example in Fig.1 for instance. Let k be equal to 3; on the graph in the middle, a cut value of at least 3 is indeed possible. Thus, as the decision version is solvable, the max-cut problem is feasible. Lastly, it is imperative to assess its boundedness. As the name suggests, the Max-Cut problem must have a maximum possible cut value for a specific graph. To prove this, take the example in Fig.1 once more. The graph on the right side of the figure shows the maximum possible cut value for the given graph. If two nodes of the same set are adjacent to one-another, the maximum possible cut value is 3, which is shown by the graph in the middle, but if they are opposite each other, which is shown by the graph on the right, each node in the rectangular portion of the graph (and not the diagonal – the edge between node 0 and node 3) is connected to a node of the opposite set, maximizing the cut value for the given graph, which is 4. As there is a maximum cut value for this graph, specifically, there has to be a maximum cut value for every other graph with any number of vertices and combination of edges. This shows that the Max-Cut problem is bounded. Therefore, as it is both feasible and bounded, it is solvable.

        Although the objective of the Max-Cut problem may seem straightforward, no solution has been developed that can find the optimal solution in an efficient manner in all cases due to the nature of the statement problem, the limits of modern hardware, and scalability issues. Due to this, various approximation algorithms have been created to deliver suboptimal solutions, and quite a few of them utilize principles of quantum mechanics to do so, which allow parallel computations, that is, the ability to evaluate multiple possibilities at once. This is because the addition of a single qubit gives rise to exponentially more states and possibilities, giving quantum computers the ability to solve problems exponentially faster than classical algorithms (Tepanyan, 2025). A few classical approximation algorithms that have had sufficient success in approximating the Max-Cut problem are those using ‘greedy algorithms’ (Codecademy, 2022) and the ‘Goemans-Williamson Algorithm’ (Toni, 2018). However, these are not as effective as quantum algorithms, such as those using the ‘Quantum Approximate Optimization Algorithm’ (QAOA) (Ceroni, 2025) and Quantum Genetic Algorithms (QGA) (Viana & Neto, 2024).

As mentioned in the previous paragraph, both classical and quantum approximation algorithms have made progress in addressing the Max-Cut problem, but their effectiveness in practice depends heavily on the underlying hardware and the scalability of the method used. This raises an important question: when do quantum algorithms outperform their classical counterparts? Addressing this question motivates the main objectives of this paper. Specifically, we aim to explore and demonstrate the feasibility of running QAOA for the Max-Cut problem on real quantum devices and to compare its performance against the classical brute-force method (AlgoEducation, n.d.) in order to gauge the potential advantage that quantum algorithms may provide as input size grows.

Our work is driven by the hypothesis that while classical algorithms excel at solving problems with small to medium-sized input values at a great degree of efficiency, they become infeasible for solving large problems, and that quantum approaches, such as QAOA, have the potential to surpass them as the input size increases

Background – Classical vs. Quantum Approaches

Computational Complexity

To understand why quantum solutions are usually superior to classical solutions, it is imperative to understand why classical solutions can fail, for which the knowledge of computational complexity is needed. Computational complexity is a measure of how difficult a computational problem is and how much time is required to solve it. The Max-Cut problem specifically, is an ‘NP-Hard’ problem.

For an ‘NP’ hard problem, a proposed solution can be verified in polynomial time, but the solution itself cannot be found as quickly. Examples of NP problems are the ‘Boolean Satisfiability Problem’ (SAT) and the Sudoku Puzzle (Kanwal, 2021).

A problem is ‘NP-Hard’ if every problem in NP can be reduced to it in polynomial time, that is, if an NP-Hard problem can be solved efficiently (in polynomial time), all NP problems can be solved efficiently (Kanwal, 2021). In fact, quite a few NP-Hard problems are not even in NP because their solutions cannot be verified in polynomial time.

The reason why the Max-Cut problem, or the optimization version at least, is NP-Hard is because of two reasons. Firstly, it is indeed possible to count the edges cut, but it is not possible in all cases to verify that a certain partition of the vertices in set V  gives the maximum cut without essentially solving the problem itself. Secondly, it is a fact that as the number of vertices increases linearly, the number of possible partitions increases exponentially, which is why no perfect solution in polynomial time can exist for the Max-Cut problem.

Classical Solutions to the Max-Cut Problem

The max-cut problem can be written as a quadratic optimization function. To know why classical methods are not ideal, it is necessary to understand the structure of this objective function (Lowe, 2025), which reveals the very nature of the Max-Cut problem. Let  G = (V,E) be a graph with vertices   V and edges E  , and let  w_ij​ denote the weight of edge (i,j)  . This paper utilizes unweighted graphs, so w_ij = 1   . The variables x_i E { -1,1}   are binary variables for each vertex, representing the set assignment of the vertex – which subset of the partition it belongs to – that is A or B. The objective function can then be expressed as:

For each edge (i,j)  connecting vertices x_i  and x_j  , if vertices x_i  and x_j  are in the same subset, their product will be equal to 1, and since 1 – 1 = 0, the edge connecting x_i  and x_j  is not cut and it will not contribute to the cut value of  G. If x_i  and x_j  are in two different subsets, their product will be equal to -1, and since 1 – -1 = 2, the edge connecting x_i  and x_j  is cut and it will contribute to the cut value of  G (Lowe, 2025). Each existing vertex combination (x_i , x_j) in G is considered twice: once for (x_i , x_j)  and the other for (x_j , x_i)  because technically, they are not exactly the same vertex combination. However, in actuality, they connect the same edge, so to avoid such double counting, the total cut value after iterating after summing through every valid vertex combination is divided by 2. Therefore,  C(x) calculates the total cut value of graph G  , which is the same as calculating the maximum possible cut value for G  , which fulfills the objective of the Max-Cut problem.

Each vertex can be in one of two possible subsets. Because of this, the solution space of graph G  , or total number of possible partitions is 2ⁿ , where n  represents the number of vertices in G . Therefore, as the value of n  increases, the number of possible partitions increases exponentially, which starts to become infeasible after a certain point, especially for large values of  , because the time and resources required to iterate over all 2ⁿ   possible combinations become too large. This is why the function encoded by C(x) is NP-hard, due to which almost all algorithms created to attempt a solution at the Max-Cut problem are approximation algorithms. These are designed to produce near-optimal solutions within reasonable time limits, rather than iterating over all 2ⁿ  possible partitions, which is known as the brute force method – theoretically the only algorithm with an approximation ratio of 1.

One of the simplest classical approximation approaches is using a greedy algorithm (Codecademy, 2022), which is a technique that assigns vertices to the set that yields the largest ‘immediate’ increase in the cut size. Due to this aspect, despite their speed, greedy approaches often get stuck in locally optimal configurations that are not globally optimal. A significantly more powerful classical approach is the Goemans–Williamson algorithm (Cai, 2003), which uses a mathematical technique called semidefinite programming (SDP) to represent vertices as unit vectors on a hypersphere followed by randomized rounding to generate a cut. This is the most powerful classical approximation algorithm and is guaranteed to achieve an approximation ratio of at least 0.878, that is, to produce a cut that is guaranteed to be at least 87.8% as good as the optimal cut value, regardless of graph size.

Although such classical approximation algorithms are considerably efficient, their performance starts to diminish for very large or dense graphs, motivating research on alternative methods – quantum methods.

Introduction to Quantum Computing – For the Max-Cut Problem

When multiple qubits are combined, they can represent an exponentially large number of possible states at the same time. This means that for a system with   qubits, its state can be described by a complex state vector in a 2ⁿ -dimensional space. This property of superposition (Thomson, 2025) of a quantum state is extremely valuable as it enables it to represent all 2ⁿ  basis states. Such simultaneous representation makes quantum parallelism – the ability of a single operation (in this case, a unitary operation) to act on all superimposed states at the same time – possible (Kaye & Mosca, 2020), which is the very thing that gives quantum computing the edge over classical computing: While a classical computer must check all 2ⁿ  possible combinations one at a time, a quantum computer can describe them all as a superposition of states representing those combinations using a compound state vector and manipulate them all in parallel.

Quantum computers also use a phenomenon known as entanglement, in which two or more quantum particles – in this case, qubits – share a quantum state over space, causing the individual state of one particle affecting the individual states of the other entangled particles regardless of the distance between them. This link between qubits serves numerous functions in quantum computing. Thus, superposition and entanglement allow quantum computers to explore multiple possible solutions – and if housed with enough qubits, the entire solution space – simultaneously with much fewer resources and at exponentially faster speeds than classical computers, making quantum algorithms superior for solving NP-hard problems such as the Max-Cut problem even though they are still not fully accurate like the brute force method (Preskill, 2018).

The Quantum Approximate Optimization Algorithm (QAOA)

The Quantum Approximate Optimization Algorithm, or QAOA, is an esteemed hybrid (using classical and quantum computing) algorithm that specializes in finding approximate solutions to combinatorial optimization problems like the Max-Cut problem that are classically infeasible (Blekos et al., 2023). It works by alternating between two Hamiltonian operators. In simple terms, the Hamiltonian is a quantum operator that represents the total energy of a system (Agarapu, 2024).

The first one is known as the ‘Cost Hamiltonian’ (EITCA, 2024). It is based on the objective function of the Max-Cut problem and assigns energy levels to different possible solutions, with lower energy levels corresponding to better cut values. After applying the Cost Hamiltonian to the initial quantum state, the resultant state is biased towards generally good solutions (not the best ones), so it can get stuck in a part of the solution space that has a low energy level (EITCA, 2024). This is where the second one, known as the ‘Mixer Hamiltonian’ (EITCA, 2024), comes in. It “mixes” the quantum state by flipping qubits in a way that causes amplitudes between different energy levels to get mixed up, creating newer superpositions of states, allowing a greater exploration of the solution space and consequently increasing the chance of finding better solutions. These operators are then repeated for   layers, where  is the depth of the quantum circuit.

Each layer is parameterized by angles   (parameter of  Cost Hamiltonian) and k (parameter of Mixer Hamiltonian), where   is the current layer. These angles are optimized using classical optimization techniques (Ivezic, 2024) and are then used in the next layer to improve the performance of the Hamiltonians to reach even lower energy levels. So, theoretically, as the depth of the circuit (measured in  layers) moves closer to infinity, QAOA moves closer to the lowest energy level of the Cost Hamiltonian, closer to the perfect approximation ratio of 1, and thus, closer to the exact solution to the Max-Cut problem for graph  . This hybrid approach makes QAOA suitable for today’s noisy, small-scale NISQ quantum devices, which is why it is one of the most successful quantum approximation algorithms to solving the Max-Cut problem

Limitations of Quantum Approaches

Although mathematical quantum approaches like QAOA are elegantly promising, their performance is limited by modern quantum hardware. Quantum computers are considerably noisy, resulting in qubits’ states being disturbed and therefore losing information (Preskill, 2018) through a process known as decoherence (Bacon, 2003), causing the solution quality to be degraded. The greater the qubit count, the greater the capabilities of quantum hardware are but also the chance of decoherence. Additionally, most devices have a limited number of qubits – under 200 – and restricted qubit connectivity due to suboptimal qubit quality (Preskill., 2018) (Swayne, 2024), which can make running larger quantum circuits quite inefficient. This also prevents true ‘fault-tolerant’ quantum computing  (Davis et al., 2025) utilizing consistent real-time error correction from being possible, meaning that QAOA has the unwanted capability of accumulating noise over time.

Because of these constraints, experiments involving QAOA usually focus on very small graphs and circuits that have a very limited depth, which is why they cannot be scaled to large problem sizes (Lotshaw et al., 2022). These limitations mean that while quantum computers might not yet fully outperform classical algorithms on large problems, which is where they thrive compared to classical computers, studying QAOA on real devices is essential for understanding how performance scales as hardware improves.

Methodology

Having developed foundations to both classical and quantum methods to solving the Max-Cut problem and established the strengths and limitations of each, the next step is to investigate how they perform in practice. This section details the methodology used to analyze, evaluate and compare the performance of the Quantum Approximate Optimization Algorithm (QAOA) against the classical brute-force method in solving instances of the Max-Cut problem. This investigation aims to look into and demonstrate how each algorithm scales as graph size and problem complexity increases and whether quantum algorithms begin to exhibit any advantages over classical algorithms as input size increases.

The methodology is structured into two parts. The first one outlines experimental details, including the overall set-up, graph generation, quantum hardware used, technical settings, and more for both classical and quantum solutions. The second one discusses the performance metrics used, such as execution time and approximation ratios, and explains how they are calculated.

Experimental Details

The investigation uses a controlled computational set up to compare the performances of the classical brute force algorithm and QAOA in solving the Max-Cut problem. It was implemented via Python using IBM’s ‘Qiskit Runtime Service’ (version 0.42.0) and the ‘Qiskit SDK’ (version 2.1.1) to enable usage of quantum hardware and ‘Matplotlib’ (version 3.10.6) and ‘Rustworkx’ (version 0.17.1) for graph creation and computation. The quantum processor used was ‘IBM_Torino’, which houses 133 qubits. A series of unweighted graphs were created using the ‘create_sample_graph()’ function with node size ranging from 4 nodes to 24 nodes. The graphs were created as polygons with the number of vertices corresponding to the number of nodes specified, with a few edges connecting non-adjacent nodes for structural variety.

The Max-Cut problem was solved using two approaches: (1) a classical brute force algorithm – this is implemented in the ‘classical_max_cut_brute_force()’ function, which exhaustively went over all 2ⁿ possible vertex partitions to determine the partition resulting in the most optimal cut value. (2) QAOA –  the QAOA circuit required to operate on is created via the ‘create_qaoa_circuit()’ function, and the algorithm is run via the ‘run_qaoa_modern()’ function that communicates with the chosen quantum device in the backend or runs the algorithm on a local Aer simulator when not selected or unavailable.

The QAOA implementation uses a circuit with depth being a single layer (p=1) with fixed parameters   γ = π / 4  and β = π / 2 for cost and mixer Hamiltonians respectively, corresponding to one cost–mixer iteration. Thus, no classical optimizer has been used for this investigation. Each circuit begins with a uniform superposition of the plus state into a compound n-qubit quantum state, followed by the cost and mixer layers to encode and explore the solution space respectively, and then measured. Qubits are then mapped to the quantum device, and while the “job” (QAOA run) is not in a state of completeness/termination, it is queued on the quantum device every 10 seconds.

To compare both brute-force algorithm and QAOA, 5 trials were taken for each graph size (4 nodes – 24 nodes). Each trial consisted of both algorithms being run with the execution time for both and approximation ratio for QAOA being measured. The mean of all execution times of each algorithm for each graph size was then calculated, and they were plotted on a scatter plot diagram created on ‘Desmos’ expressing execution time vs. graph size for both algorithms.

Methodological Considerations

Earlier, it was mentioned that the QAOA circuit has only a single cost-mixer layer (p = 1)  and therefore had no classical optimizer. These are to make the investigation easy for interpretation and reproduction so that similar results are yielded when performed by the audience and not have the complexity of the optimizer affect QAOA. However, the cost for this approach is the accuracy of the cut value.

For obvious reasons, the number of trials performed, 5, were for the sake of statistical accuracy and error minimization, and a graph size increment of 2 would be small enough to accurately capture the relationship between execution time and graph size for both algorithms.

The brute-force algorithm calculates the optimal cut value 100% of the time as it goes through every single possible combination of graph partitions. QAOA on the other hand does not because measuring a quantum circuit is probabilistic in nature, so to maintain consistency and ensure fairness in comparisons, it was run using identical circuit configurations and constant cost and mixer Hamiltonian parameters

Performance Metrics

The primary measurement for algorithmic performance is execution time, which is measured using the ‘time.perf_counter()’ function from the ‘time’ module provided by Python. The function executing the brute-force algorithm for solving the Max-Cut problem – ‘classical_max_cut_brute_force()’ – only contains the necessary steps of the brute-force algorithm, so the entirety of it is measured. The function executing the QAOA for the Max-Cut problem – ‘run_qaoa_modern()’ – is only measured from circuit creation until the QAOA “job” is in state of completion/termination. The sleep time is not included as this is simply a delay added to prevent constant querying of the job status and is therefore not a part of the QAOA procedure.

The secondary measurement for algorithmic performance is approximation ratio. This is calculated by dividing the cut value produced from an algorithm by the optimal cut value. This practically only applies to the QAOA as the brute force algorithm always results in the optimal cut value 100% of the time, so it has an approximation ratio of 1 in all cases. It is important as it is a measure of how accurate an algorithm is, and in this case, how close the cut values it produces is to the optimal cut value. The reason it is not the primary metric is that the variable of interest is execution time, which is theorized to be a major advantage of quantum algorithms over classical ones.

Results

This section presents the results of the experiment, comparing the observations for the classical and quantum algorithms used to solve the Max-Cut problem regarding the performance metrics mentioned. Note that the actual figures for the graphs will not be shown in this paper due to the sheer volume of trials and that the variable of interest is execution time, not the way the graphs are structured.

Below are datatables that present the execution times of both algorithms over 5 independent trials for each graph size (4 – 24 nodes) along with their respective approximation ratios relative to the optimal cut value. The first table shows the above information for the brute-force algorithm, and the second table shows the same for QAOA executed on real quantum hardware. Both have a column displaying the average execution times for each, and when talking about execution time, it is the average execution time that is being referred to (most of the time).

Table 1 displays the execution time of the classical brute-force algorithm for each graph size for all 5 trials. As expected, as the graph becomes more complex, the average runtime increases at a seemingly exponential rate, reflecting non-polynomial time complexity caused by enumerating all vertex partitions. The approximation ratio remains a constant 1.00 across all trials, which is why it is not shown in a table, confirming that the brute-force method yields the exact maximum cut.

Table 2 presents the execution time of the QAOA for each graph size for all 5 trails. Contrary to the assumptions made earlier in the paper, which stated that execution time increases at a linear rate as graph size increases, they (the average) seem to fluctuate within a certain range of values as graph size increases with no discernable trend.

Table 3 presents the approximation ratios (relative to the optimal cut value) of the QAOA runs for each graph size for all 5 trails. Here, it can be seen that the average execution time decreases as graph size increases, showing that the algorithm becomes less accurate as complexity increases. Potential reasons for this will be explored in the next section.

Now, the results of Table 1 and Table 2 have been plotted on a scatter plot diagram below showing execution time vs. graph size for both algorithms to illustrate the trend of their behavior regarding the same and provide a visual  representation of the scalability advantages of quantum algorithms for problems like the Max-Cut.

From the graph, it can be seen that the execution time for the brute-force solution shows exponential growth, whereas the execution time for the QAOA solution shows an almost flat linear model. At around 19 nodes, both algorithms’ average execution time curves intersect, after which the classical brute-force solution’s execution time exceeds that of the QAOA solution. This visual representation makes it even clearer how quickly the brute-force algorithm becomes more and more unscalable after a certain point, unlike the QAOA, which, although less accurate, seems to stay relatively constant regarding execution time. Note that the datapoint for ‘24 nodes’ for the brute-force solution is not shown as including it would make the fluctuations of the QAOA solution datapoints almost impossible to discern and the gap between both the curves before they intersect would be very small to see.

Discussion

This section analyzes the results obtained from the comparison between the classical brute-force algorithm and the QAOA in solving the Max-Cut problem. The discussion evaluates the observed trend between execution time and graph size for both algorithms, analyzes fluctuations and deviations from expected theoretical behavior, and explores the trend of the approximation ratio as graph size increases. It further investigates the causes of these trends, linking them to both algorithmic configurations and hardware limitations of the modern-day Noisy Intermediate-Scale Quantum (NISQ) devices (Mahmoud, 2021).

The execution time vs. graph size model derived from the data tables in the previous section shows a strong contrast between the trends of both algorithms. As expected, the classical brute-force solution shows an exponential growth in execution time as the number of vertices increases. This trend is consistent with the theory behind it, where every possible partition of vertices into two sets must be evaluated to deduce the optimal cut value. For small graphs (4–12 nodes), the execution time remains quite short, but as the graph size approaches 20–24 nodes, execution time grows dramatically due to the nature of an exponential function, quickly rendering brute-force computations infeasible on standard hardware. This rapid exponential escalation illustrates why NP-hard problems like the Max-Cut cannot have efficient classical solutions (solutions with polynomial time complexity), due to which they become completely infeasible for complex problems, underscoring the need for exploring quantum approaches that can theoretically have efficient solutions.

However, the QAOA solution displays a very different trend. Rather than growing exponentially with problem size, the execution times show an almost linear trend, fluctuating within a relatively narrow range, appearing almost constant across the entire set of tested graph sizes. This outcome contrasts that of QAOA (theoretically), which should show an approximately increasing linear trend with the number of qubits (and therefore graph size) increasing for a fixed circuit depth (p = 1) . The observed stability rather than increase and the randomness in runtime causing such fluctuations can be explained by a combination of quantum hardware behavior and implementation-level constraints. First, the QAOA code used in this experiment employs a fixed circuit depth and fixed Hamiltonian parameters γ = π/4 and β = π/2  . Because circuit depth, (consequently) gate count, and parameters remain constant, with no classical optimizer used to manipulate the parameters, the computational workload submitted to the 133-qubit ‘IBM_torino’ quantum device does not increase meaningfully with graph size within the tested range as there are enough qubits to encapsulate the entirety of the solution space all at once. The quantum processing time per circuit therefore remains nearly unchanged as graph size increases for those in the tested range, with small fluctuations attributable to backend scheduling variation, noise etc. In a practical setting, the Max-Cut problem is usually implemented with hundreds or even thousands of nodes, which are almost always greater than the qubit count, leading to increased processing time per circuit as graph size increases. Thus, under ideal, low-noise, and with a greater circuit depth, execution time of a QAOA solution could potentially consistently be observed to increase linearly with the number of qubits (and nodes on the graph) increasing on large and complex graphs.

Another key observation is regarding the approximation ratio. The experimental results show a decrease in approximation ratio of the QAOA solution as graph size increases. For small graphs, QAOA often seems to result in cuts close to optimal, but for larger graphs, its performance deteriorates slightly. In this investigation, this behavior can be caused by several factors. Firstly, with a low, fixed circuit depth  (p = 1) , the QAOA has a very low chance of providing a more optimal cut value due to the parameters not being optimized enough to lead to the Cost Hamiltonian’s application resulting in a higher measurement probability for bitstrings (basis states in the solution space) representing more optimal cut values. Thus, as the graph size grows, the solution space becomes exponentially greater, and a single-layer circuit cannot adequately explore it to favor bitstrings representing good cut values consistently. This is detrimental especially when the optimal cut corresponds to a low probability event. Secondly, the fixed parameter pair (γ = π/4, β = π/2) used in all runs, while simplifying the process, prevents dynamic tuning of the QAOA for different graph configurations, resulting in lower approximation ratios as the number of nodes (and therefore qubits) increases. Lastly, the most generic reason is noise. As the number of qubits used and operations performed increase, so does the capacity of noise in disturbing the compound quantum state the circuit is manipulating, blurring the probability distribution of basis states in the solution space, and therefore that of the measured bitstrings, resulting in suboptimal cut values.

Implications

This section considers the implications of the findings and discussions based on the previous sections for scalability of classical and quantum algorithms and when exactly does practical quantum advantage for NP-hard (optimization) problems such as Max-Cut begin to show. It then validates the viability of QAOA on modern NISQ hardware and considers the potential improvements that can be observed in the future as quantum hardware advances.

The findings presented in the previous sections demonstrate an important advantage of executing QAOA on real quantum hardware over the classical brute-force solution: while the latter method becomes rapidly infeasible to run with scale due to its exponentially increasing execution time, QAOA maintains roughly constant or at worst sub-linear growth in execution time. Although its current implementation does not achieve better speed compared to classical computation for relatively small graphs, its scalability over medium to large sized graphs suggests potential long-term advantages. However, the trend between graph size and approximation ratio observed proves to be a hurdle in the way of accuracy. The steadily declining approximation ratio shows the effect of QAOA configurations, but most importantly, noise. Even with a near-perfect QAOA algorithm, the limitations of NISQ hardware prevent the ability to create a high-depth circuit with a high qubit count without lowering accuracy due to the prevalence of noise. This prevents the observed advantages of QAOA for solving NP-hard problems from being effective where they are required the most – on large problems, and in the case of the Max-Cut problem, on large graphs, highlighting the underlying trade-off between speed and result accuracy

Conclusion

In summary, this investigation is designed to explore and demonstrate the feasibility of running QAOA for the Max-Cut problem on real quantum devices, comparing its performance against the classical brute-force method of solving the Max-Cut problem to gauge the potential advantages that quantum algorithms may provide as input size grows. The end goal is to determine when quantum algorithms start to outperform their classical counterparts.

The results from the experiment confirm that the classical brute-force algorithm scales exponentially with input size, becoming infeasible fairly quickly — over smaller graphs within the tested range, execution times were fairly short, but towards the end, execution times started to shoot upward at an alarmingly fast rate — whereas the QAOA solution had almost constant runtime for all graph sizes at the cost of accuracy as graph size increased. This limitation was shown by the decreasing average approximation ratio across all 5 trials as graph size increased, primarily due to a very low circuit depth (p >1)  , constant QAOA parameters, and most importantly, greater potential for noise on larger graphs, which is an issue for NISQ devices in general. Nevertheless, these findings validate the theoretical advantages that quantum algorithms outperform classical ones in terms of speed, but due to the shortcomings of NISQ hardware, they cannot fully be utilized for large problem sizes.

Ultimately, the insights from this investigation provide a foundation for understanding how quantum algorithms scale over problem size, bring us one step closer to realizing practical quantum advantages in solving complex NP-hard problems like Max-Cut over large problem sizes. Looking into the future, more advanced quantum hardware with larger qubit counts and lower noise levels will render increasing QAOA depth (p>1)  feasible with little effect on accuracy, allowing for faster and more accurate solutions for the Max-Cut problem and other NP-hard problems, revolutionizing areas such as semiconductor design, image segmentation in computer vision, financial modeling and data flow optimization, which use the Max-Cut problem for various purposes.

References

Agarapu, R. (2024). What is Hamiltonian in quantum mechanics? Online Tutorial Hub. Retrieved October 26, 2025, from https://onlinetutorialhub.com/quantum-computing-tutorials/what-is-hamiltonian-in-quantum-mechanics/

Algor Education. (n.d.). Brute Force Computing. Algor Education. Retrieved October 26, 2025, from https://cards.algoreducation.com/en/content/9CDDR2hL/brute-force-computing

Asfaw, A., Bello, L., Ben-Haim, Y., Bravyi, S., Capelluto, L., Carrera Vazquez, A., Ceroni, J., Gambetta, J., Garion, S., Gil, L., De La Puente Gonzalez, S., McKay, D., Minev, Z., Nation, P., Phan, A., Rattew, A., Shabani, J., Smolin, J., Temme, K., … Wootton, J. (n.d.). Learn Quantum Computing using Qiskit. Retrieved October 26, 2025, from https://github.com/RafeyIqbalRahman/Qiskit-Textbook/blob/master/Learn%20Quantum%20Computing%20using%20Qiskit.pdf

Bacon, D. M. (2003). Decoherence, Control, and Symmetry in Quantum Computers (Doctoral dissertation, University of California, Berkeley). arXiv:quant-ph/0305025. https://arxiv.org/pdf/quant-ph/0305025

Blekos, K., Brand, D., Ceschini, A., Chou, C., Li, R., Pandya, K., & Summer, A. (2023). A Review on Quantum Approximate Optimization Algorithm and its Variants. https://arxiv.org/pdf/2306.09198

Cai, J.-Y. (2003). Lecture 20: Goemans-Williamson MAXCUT Approximation Algorithm. University of Wisconsin-Madison. https://pages.cs.wisc.edu/~jyc/02-810notes/lecture20.pdf

Ceroni, J. (2025). QAOA introduction tutorial. PennyLane Quantum Machine Learning Demos. https://pennylane.ai/qml/demos/tutorial_qaoa_intro

Codecademy. (2022). Greedy algorithm explained. Codecademy. https://www.codecademy.com/article/greedy-algorithm-explained

Davis, R., Lanes, O., Waltrous, J. (2025). What is fault-tolerant quantum computing? Retrieved October 26, 2025, from https://www.ibm.com/quantum/blog/what-is-ftqc

DeepAI. (n.d.). Combinatorial optimization – machine learning glossary. DeepAI. https://deepai.org/machine-learning-glossary-and-terms/combinatorial-optimization

EITCA. (2024). In the context of QAOA, how do the cost Hamiltonian and mixing Hamiltonian contribute to exploring the solution space, and what are their typical forms for the Max-Cut problem? EITCA Academy. https://eitca.org/faq/in-the-context-of-qaoa-how-do-the-cost-hamiltonian-and-mixing-hamiltonian-contribute

Fiveable. (n.d.). Objective function. Retrieved October 26, 2025, from https://library.fiveable.me/key-terms/linear-algebra-and-differential-equations/objective-function

Goemans, M. X., & Williamson, D. P. (1995). Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. Journal of the ACM, 42(6), 1115–1145. https://math.mit.edu/~goemans/PAPERS/maxcut-jacm.pdf

Preskill, J. (2018). Quantum Computing in the NISQ era and beyond. Quantum, 2, 79. https://quantum-journal.org/papers/q-2018-08-06-79/pdf/

Ivezic, M. (2024). Quantum Approximate Optimization Algorithm (QAOA). Retrieved October 26, 2025, from https://postquantum.com/quantum-computing/quantum-approximate-optimization-algorithm-qaoa/

Jaillet, P. (2010). NP-completeness. MIT 6.006: Introduction to Algorithms. Retrieved from https://courses.csail.mit.edu/6.006/fall10/lectures/lecture24.pdf

Kanwal, A. (2021). Understanding P, NP, NP-complete, and NP-hard problems: A fundamental guide. Medium. https://medium.com/@0ayesha.kanwal/understanding-p-np-np-complete-and-np-hard-problems-a-fundamental-guide-3924fc9ece2a

Lee, J. (2010). A first course in combinatorial optimization. Cambridge University Press. https://books.google.com.tr/books?id=3pL1B7WVYnAC&pg=PA1&redir_esc=y#v=onepage&q&f=false

Lotshaw, P. C., Nguyen, T., Santana, A., McCaskey, A., Herrman, R., Ostrowski, J., Siopsis, G., & Humble, T. S. (2022). Scaling quantum approximate optimization on near-term hardware. Scientific Reports, Nature Publishing Group. https://www.nature.com/articles/s41598-022-14767-w

Luca, G. D. (2023). Introduction to graph theory. Baeldung on Computer Science. https://www.baeldung.com/cs/graph-theory-intro

Mahmoud, A. (2021, June 27). What is Quantum Computing? https://www.techspot.com/article/2280-what-is-quantum-computing

Maltby, H., & Ross, E. (n.d.). Combinatorial optimization. Brilliant. https://brilliant.org/wiki/combinatorial-optimization

Rossi, M., Cohen, S., & Smith, J. (2024). What is Quantum Parallelism, Anyhow?. https://arxiv.org/html/2405.07222v1

Swayne, M. (2024). Quantum computing challenges. https://thequantuminsider.com/2023/03/24/quantum-computing-challenges/

Tepanyan, H. (2025). Quantum Computing vs. Classical Computing. Retrieved October 26, 2025, from https://www.bluequbit.io/quantum-computing-vs-classical-computing

Thomson, J. (2025). What is quantum superposition and what does it mean for quantum computing? Retrieved October 26, 2025, from https://www.livescience.com/technology/computing/what-is-quantum-superposition-and-what-does-it-mean-for-quantum-computing

Toni, B. (2018). Max-Cut lecture notes (PDF). University of Toronto. http://www.cs.toronto.edu/~toni/Courses/Proofs-SOS-2018/Lectures/maxcut.pdf\

Viana, P.A., Neto, F. (2024). Quantum search algorithms for structured databases. https://arxiv.org/pdf/2501.01058

Wright, S. J. (2025). Optimization Definition, Techniques, & Facts. Encyclopaedia Britannica. https://www.britannica.com/science/optimization

Data and code used in this paper are available in github.

About the author

Yohhaan Yung Kang Huang

Yohhaan Yung Kang Huang is a high school senior passionate about quantum computing and algorithmic problem-solving. Under the mentorship of Dr. Roberto Dos Reis from Northwestern University, he explored how the Quantum Approximate Optimization Algorithm (QAOA) performs on near-term quantum hardware compared to classical methods. He plans to pursue further studies in computer science and quantum information science and contribute to the development of practical quantum technologies. He loves robotics and is part of the school’s FRC team, enjoys music, and plays the piano. He is also driven to help neurodivergent students integrate better in academic and social environments.

The post Quantum Approximate Optimization Algorithm for the Max-Cut Problem: Performance Comparison with Classical Approaches on NISQ Devices appeared first on Exploratio Journal.

Abstract

Human–computer interfaces represent a rapidly advancing frontier in biomedical engineering, integrating mechanical, electronic, neural, and Artificial Intelligence (AI) technologies to restore or augment lost human function. This review synthesizes recent developments in Advanced Human–Computer Interfaces, organs, and neural interfaces, drawing on both clinical and engineering perspectives. With a focus on state-of-the-art innovations published within the last five years, the paper highlights breakthroughs in AI-driven sensory feedback, adaptive control algorithms, biomaterials, and clinical translation. By situating these advances within the broader context of unmet clinical needs and rehabilitation goals, this review identifies current challenges and outlines future directions for fully integrated, intelligent human-machine systems.

Index Terms

Machine learning, Deep learning, Reinforcement learning, Edge computing, On-device AI, Computer vision, Bionic limbs, brain–computer interfaces, prosthetics, neuromorphic vision, osseointegration, artificial pancreas, neural interfaces, Advanced Human–Computer Interfaces, AI, biomedical engineering, neuroprosthetics, human–machine symbiosis, wearable robotics, neuroengineering, biocompatible materials, smart prosthetics, adaptive control, closed-loop systems, neural decoding, assistive technology, implantable devices, bioelectronics, translational medicine, cybernetics.

I. Introduction

The pursuit of artificial devices that restore lost biological function is as old as medicine itself, with early wooden prosthetic legs and iron hooks marking humanity’s first attempts at bionics. In the modern era, bionic devices have come to represent a class of technologies that combine mechanical hardware, electronic control, and neural interfacing to restore sensory, motor, or organ-level function. These devices are no longer crude substitutes; rather, they aim for seamless integration with the human nervous system, allowing users to experience levels of dexterity, feedback, and autonomy once thought impossible.

The importance of this field is underscored by the prevalence of disability worldwide. According to the World Health Organization, over 2.4 billion people globally live with conditions requiring rehabilitation [21]. Of these, limb loss affects more than 57 million people, while vision loss (addressable by devices like the bionic eye) impacts at least 43 million blind individuals [21]. In diabetes alone, more than 530 million people require continuous glucose management, and the artificial pancreas is emerging as a transformative bionic organ [23]. These statistics highlight the vast unmet need that bionic technologies attempt to address.

Fig. 1: Global burden of conditions addressed by bionics. Bars show approximate affected populations: limb loss (57M), blindness (43M), cochlear implant users (∼0.7M), and diabetes (530M). Data sources: WHO Global Report on Rehabilitation [21] (limb loss, blindness, diabetes) and Wilson (cochlear implant users) [24].

II. Overview of Human-Computer Interfaces

Human–Computer Interfaces (HCIs) can be defined as artificial constructs designed to replace or augment biological structures, with the unique feature of neural, physiological, and increasingly AI-driven integration. Unlike conventional prosthetics or implants that function passively, HCIs actively sense, compute, learn, and actuate.

Historically, the field has undergone several stages. Early prosthetics, such as Egyptian wooden toes or Roman iron hands, were primarily cosmetic or functional in the most basic sense. By the 16th century, artisans like Ambroise Paré introduced mechanical limbs with crude joint mechanisms. The 20th century saw the introduction of body-powered prostheses (using cables and harnesses), followed by the revolutionary myoelectric control in the 1960s, which used Electromyographic (EMG) signals for actuation [16]. In parallel, sensory HCIs began with the cochlear implant (1972), the first device to restore a lost sensory modality via direct neural stimulation, paving the way for retinal implants and, more recently, neuromorphic vision systems based on organic semiconductors and perovskite nanowire arrays [10], [11]. Organ-level HCIs have also advanced, most notably the artificial pancreas, which integrates glucose sensors, insulin pumps, and AI-based closed-loop metabolic control algorithms [23].

To situate the reader in the diversity of modern HCIs, Table I summarizes major categories of systems, their principles of operation, and representative examples. This overview highlights a unifying theme: HCIs are no longer restricted to mechanical substitution. Instead, modern systems seek bidirectional communication with the nervous system—allowing users not only to control artificial devices but also to receive naturalistic sensory feedback enhanced by adaptive AI.

Fig. 2: Taxonomy of Human–Computer Interfaces (HCIs) across five categories: limbs [1], eyes [10], ears [22], organs [23], and brain–computer interfaces (BCIs) [14].

TABLE I: Major Classes of Bionic Devices, Principles, and Examples

Fig. 3: Functional pipeline of a bionic device: input signals (e.g., EMG/EEG/sensors) → processing/AI → actuation (prosthesis/pump) → feedback (tactile/visual/auditory).

III. Current Types of Human-Computer Interfaces

HCIs span a wide range of applications, from motor prostheses that restore limb function to sensory prostheses that recreate lost modalities such as hearing and vision. In addition, organ-level HCIs represent an emerging frontier where closed-loop control systems, often enhanced by AI, substitute for failing metabolic or physiological functions. In this section, we review the major classes of HCIs, focusing on their principles of operation, technological foundations, and representative studies from a span of five years. Throughout this review, we use Human–Computer Interfaces (HCIs) as the overarching term for technologies that bridge biological and computational systems. Terms such as bionic limbs, bionic eyes, and related phrases are used to denote specific subsets of HCIs, rather than distinct categories.

A. Bionic Limbs

Upper-Limb Prostheses: Upper-limb prostheses have progressed from simple hooks to highly dexterous, multi-articulated robotic hands with neural control. The control of such devices primarily relies on myoelectric signals derived from the residual muscles of the forearm or upper arm. However, conventional surface EMG suffers from poor signal quality, cross-talk, and electrode displacement. To overcome these limitations, modern systems employ implanted electrodes (epimysial, intramuscular) that record stable myoelectric activity over years [1].

1. Advanced interfaces include Targeted Muscle Reinnervation (TMR) and Regenerative Peripheral Nerve Interfaces (RPNIs). TMR surgically reroutes residual nerves to denervated muscles, creating new, amplifiable EMG sites [16]. RPNIs implant nerve endings into muscle grafts, forming stable bioelectrical sources for long-term control [7]. These constructs amplify weak nerve signals into robust EMG activity, enabling fine motor decoding through pattern recognition or regression algorithms.

Recent clinical translation is exemplified by Ortiz-Catalan et al. , who demonstrated a transradial neuromusculoskeletal prosthesis integrating osseointegrated titanium implants, implanted electrodes, and neural stimulation [1]. The patient achieved stable prosthetic use in daily life for more than three years,

TABLE II: Representative Advances in Bionic Limbs

with reduced phantom limb pain and improved quality of life. This represents one of the first long-term demonstrations of a self-contained neural prosthesis outside the laboratory.

In addition to control, sensory feedback has become a critical area of research. Extraneural cuff electrodes and intraneural arrays can evoke tactile percepts in the phantom limb when coupled with sensorized prosthetic hands [2]. Marasco et al. further demonstrated the fusion of touch, kinesthesia, and motor control, restoring able-bodied visuomotor behaviors such as reducing visual fixation on the prosthetic hand [3]. These findings suggest that upper-limb prostheses are approaching a new level of embodiment and naturalistic use.

2. Lower-Limb Prostheses: Lower-limb bionics are distinguished by the need to restore both mobility and load-bearing stability. Microprocessor-controlled knees (e.g., Ottobock C-Leg, Össur Rheo Knee) represent the current standard of care, providing adaptive damping based on gait phase detection. Recent developments extend this paradigm to powered prosthetic legs, which incorporate actuators at the knee and ankle.

The Open Source Leg (OSL) is a notable example, offering a modular, customizable platform with open hardware and software [6]. Clinically tested on transfemoral amputees, the OSL provides knee and ankle actuation with programmable gait dynamics. This democratized design accelerates innovation by lowering barriers for academic and clinical groups to experiment with novel control strategies.

3. Osseointegration and Direct Skeletal Attachment: Conventional socket-based attachment causes discomfort, skin breakdown, and instability. Osseointegration—anchoring the prosthesis directly to the skeleton via titanium implants—addresses these issues [15]. Beyond mechanical benefits, osseointegration serves as a human–machine gateway, enabling safe percutaneous feedthroughs for electrodes and sensors. This transforms the limb into a bidirectional interface, capable of both decoding neural intent and delivering somatosensory feedback [1].

Fig. 4: Comparative performance metrics across device classes (illustrative 0–10 scores). Values synthesize trends reported in representative studies on neuromusculoskeletal limbs [1], [3], bionic vision and retinal systems [5], [10], [11], cochlear implants [22], [24], artificial pancreas systems [23], and clinical/BCI demonstrations [13], [14].

Fig. 5: Case study schematic of osseointegration and neural interfaces in human–computer interface (HCI) limbs. The diagram shows bone anchoring, titanium implant, prosthesis connector, electrodes to nerve/muscle, and percutaneous feedthroughs. Legend: (1) Bone, (2) Titanium implant, (3) Prosthesis connector, (4) Electrodes to nerve/muscle, (5) Skin layer (dashed).

TABLE III: Recent Innovations in Bionic Vision

B. Bionic Eyes

The restoration of vision is among the most ambitious goals of sensory prosthetics. Early retinal implants such as Argus II used electrode arrays to stimulate the surviving retinal ganglion cells, enabling basic light perception and object localization [5]. However, spatial resolution remained low, and the devices were limited to high-contrast vision.

Recent approaches leverage neuromorphic engineering and novel materials. Long et al. reported a hemispherical perovskite nanowire retina capable of filter-free color recognition [10]. By integrating adaptive optics with neuromorphic preprocessing circuits, the system achieved wide-field, low-noise, and low-power color vision. Similarly, Zhang et al. developed a TIPS-pentacene phototransistor retina, exhibiting broadband sensitivity (380–740 nm), high optical transparency, and synaptic plasticity for visual memory [11]. The choice of TIPS-pentacene, with its narrow bandgap (∼1.6 eV), enabled efficient photon absorption across the visible spectrum, mimicking natural photoreceptors.

These neuromorphic systems go beyond electrode-based stimulation by embedding preprocessing within the retina itself, thereby reducing latency and power consumption. While still in preclinical stages, they represent a paradigm shift toward bioinspired vision systems capable of continuous learning and adaptation.

C. Bionic Ears

The cochlear implant remains the most successful sensory prosthesis to date, with more than 700,000 users worldwide [22]. It works by bypassing damaged hair cells of the cochlea and directly stimulating the auditory nerve with electrode arrays. Modern cochlear implants use advanced signal processing algorithms to decompose sound into frequency channels and deliver spatially coded electrical impulses.

Key technological progress includes fine structure processing, which encodes temporal cues for improved music perception, and optogenetic cochlear implants, which use light to stimulate genetically modified neurons with higher precision [24]. While conventional devices are limited to about 22 electrode channels, optogenetic approaches promise higher resolution with reduced channel interaction.

D. Bionic Organs

1. Artificial Pancreas: The artificial pancreas integrates a Continuous Glucose Monitor (CGM) with an insulin pump under closed-loop algorithmic control. Early systems used Proportional–Integral–Derivative (PID) control, but modern devices employ Model Predictive Control (MPC), which anticipates glucose fluctuations based on meals, exercise, and circadian rhythms [23]. Clinical trials show that MPC-based artificial pancreas systems reduce hypoglycemia incidence and improve HbA1c compared to conventional insulin therapy.
2. Other Organ-Level Devices: Beyond diabetes, prototypes of bionic kidneys (artificial filtration units) and bioartificial hearts are under investigation. For example, wearable dialysis systems combine nanoporous membranes with microfluidic pumps, while ventricular assist devices integrate soft robotics for pulsatile flow. While less mature than limb or sensory prostheses, these devices extend the concept of bionics to systemic organ replacement.

E. Neural Interfaces and Brain–Computer Interfaces

Neural interfaces serve both as standalone assistive technologies and as enabling components of bionic limbs and sensory devices. They can be broadly classified as non-invasive (EEG, fNIRS), minimally invasive (ECoG), and invasive (Utah arrays, intraneural electrodes).

Non-invasive BCIs offer safety and accessibility but suffer from low spatial and temporal resolution. Invasive approaches achieve higher bandwidth but face challenges of biocompatibility and stability. Recent advances include:

• Hybrid nerve interfaces (Cho et al., 2023) that combine RPNIs with shape-memory polymer buckles, achieving stable bidirectional signaling in animal models [7].
• Reinforcement learning-based BCIs that improve prosthetic hand control accuracy by double to triple compared to supervised learning [12].
• Bidirectional BCIs that not only decode motor intent but also deliver sensory feedback through cortical stimulation [13], [14].

These technologies are converging toward closed-loop systems, where intention and perception are integrated within the same neural–machine cycle.

IV. Technological Advancements and State-of-the-Art

The performance of bionic devices is fundamentally constrained by the quality of their materials, sensors, actuators, and neural interfacing technologies. In recent years, breakthroughs in biocompatible materials, microelectronics, and artificial intelligence have dramatically improved the fidelity of motor control, the richness of sensory feedback, and the long-term stability of implantable systems. This section reviews these advances, emphasizing both the underlying mechanisms and the way they address prior limitations.

A. Materials for Bionics

1. Biocompatible Polymers and Flexible Electronics: Traditional rigid electronics are poorly matched to the soft, dynamic environment of biological tissue, often leading to inflammatory responses and signal degradation. Flexible polymers such as polyimide, PDMS (polydimethylsiloxane), and parylene-C have become standard substrates for implantable electrodes. These materials reduce mechanical mismatch, minimizing scar tissue encapsulation and improving long-term signal stability [9].

Emerging organic semiconductors have further enabled neuromorphic HCIs. For instance, 6,13-bis (triisopropylsilylethynyl)pentacene (TIPS-pentacene) was selected in Zhang et al. for its narrow bandgap (∼1.6 eV), which allows photon absorption across the visible spectrum [11]. Its high carrier mobility and optical transparency made it ideal for constructing a retina-like phototransistor array that mimics the broadband response of photoreceptors. TIPS-pentacene also exhibits synaptic plasticity under repeated light stimulation, enabling short-term visual memory tasks and demonstrating how material choice directly determines device intelligence. In short, TIPS-pentacene functions as a photoactive organic semiconductor that provides both light detection and learning-like behavior within neuromorphic retinal systems.

Fig. 6: Innovation timeline (2020–2025) highlighting selected advances from the reference set. Key milestones include the open-source bionic leg (2020) [6], prosthetic touch and kinesthesia integration (2021) [3], a clinically deployed neural-controlled bionic hand (2023) [1], and reinforcement learning control for dexterous hand function (2024) [12].

2. Metals and Ceramics for Osseointegration: In HCI limbs, titanium alloys are the gold standard for osseointegration due to their high strength, corrosion resistance, and biocompatibility. Surface treatments (e.g., hydroxyapatite coatings) promote bone ingrowth, enabling long-term skeletal anchoring [15]. The percutaneous feedthroughs enabled by titanium implants also provide a stable, infection-resistant pathway for electrodes, addressing the long-standing problem of percutaneous connectors in neural prostheses.

B. Sensors and Actuators

Bionic devices rely on sensors to detect environmental stimuli and actuators to reproduce biological motion.

1. MEMS and Soft Sensors: Miniaturized Microelectromechanical Systems (MEMS) enable high-resolution detection of forces, pressures, and accelerations. In prosthetic hands, MEMS pressure sensors embedded in fingertips translate tactile stimuli into electrical signals that can be delivered back to the nervous system [3]. For lower-limb devices, Inertial Measurement Units (IMUs) allow real-time detection of gait phases, enabling adaptive damping in microprocessor knees.

Soft sensors based on conductive hydrogels or liquid metals are increasingly integrated into prosthetic liners, providing conformal detection of skin strain and residual limb pressure. These sensors improve socket fit monitoring and help prevent skin breakdown.

2. Actuation Systems: Historically, prosthetic actuation relied on DC motors, which are bulky and energy-inefficient. Recent approaches explore series elastic actuators, which integrate elastic elements to store and release energy, improving safety and compliance during human–robot interaction. Shape Memory Alloys (SMAs) and Dielectric Elastomer Actuators (DEAs) offer biomimetic muscle-like contraction, although their power efficiency and thermal properties remain challenges.

In the context of the artificial pancreas, actuators are miniaturized insulin pumps capable of delivering subcutaneous doses with millisecond precision. The accuracy of these pumps, combined with continuous glucose monitoring, underpins the safety of closed-loop systems [23].

C. Neural Interfaces

The neural interface is the critical bottleneck for high-performance bionic systems, as it governs the bandwidth of communication between the user and the device.

1. Non-Invasive vs. Invasive Interfaces: Non-invasive techniques such as surface EMG and EEG (Electromyography and Electroencephalography) are safe but limited by poor signal-to-noise ratio. In contrast, invasive methods such as intraneural electrodes, epimysial implants, and cortical arrays offer high bandwidth but risk tissue damage and long-term instability.

Fig. 7: Material-to-function mapping: examples include TIPS-pentacene → neuromorphic vision; titanium → osseointegration; MEMS → tactile/proprioception; series elastic actuators → compliant actuation; shape-memory polymers (SMP) → hybrid nerve interfaces.

The hybrid nerve interface proposed by Cho et al. exemplifies a middle ground [7]. By combining regenerative peripheral nerve interfaces (muscle grafts reinnervated by nerve endings) with a shape memory polymer buckle, the system stabilized nerve–electrode contact over 29 weeks in rabbits. This design achieved stable bidirectional communication—demonstrating both sensory recording and robotic leg control—suggesting that hybrid constructs may resolve the trade-off between invasiveness and stability.

2. Osseointegration as a Neural Gateway: Ortiz-Catalan et al. demonstrated that osseointegrated implants could act not only as skeletal anchors but also as percutaneous conduits for neural signals [1]. By routing electrode wires through titanium fixtures integrated into the radius and ulna, they achieved stable myoelectric recording and direct neural stimulation for more than three years in daily use. This dual role of osseointegration—as both a mechanical interface and a neural gateway—represents a major advance in long-term clinical viability.

D. Computational Algorithms and Decoding

Advances in machine learning have transformed bionic control from binary, sequential commands to rich, continuous decoding of intent.

1. Pattern Recognition and Regression: Early myoelectric prostheses employed simple thresholding: one muscle contraction to open the hand, another to close. Modern devices employ pattern recognition algorithms (support vector machines, linear discriminants) trained on multichannel EMG to decode a variety of gestures. Regression-based approaches enable proportional control, translating EMG amplitude directly into joint torque or velocity.

2. Reinforcement Learning (RL): A major innovation is the use of reinforcement learning for prosthetic control. Schone et al. implemented an RL framework with a “Guitar Hero”-like training game, where users received real-time feedback as they attempted specific hand gestures [12]. Over time, the system adapted both to user variability and electrode drift. The RL-based controller achieved double to triple the accuracy of supervised learning, particularly for simultaneous multi-finger movements. This demonstrates how adaptive algorithms can overcome the limitations of static calibration

TABLE IV: Technological Innovations Underpinning Modern Bionic Devices

3. Neuromorphic and Memristor-Based Processing: Neuromorphic circuits emulate biological synapses and neurons in hardware. Memristors—resistive devices whose state depends on prior activity—are well-suited for synaptic plasticity. In bionic vision, memristor arrays handle in-sensor preprocessing, filtering noise and adapting to light without external processors, reducing latency and power use. Their non-volatility and scalability enable compact, low-power integration on flexible substrates. A neuromorphic retina can thus perform edge detection or motion tracking directly, similar to how biological retinas preprocess visual input.

E. Integration of Artificial Intelligence

Artificial intelligence (AI) extends beyond decoding to the holistic control of closed-loop systems. In the artificial pancreas, AI-based controllers predict insulin needs using historical glucose patterns, exercise levels, and meal timing [23]. In bionic limbs, deep learning models can classify EMG signals in real time, while reinforcement learning adapts to new conditions without retraining, continuously improving performance over extended use.

The integration of AI also facilitates user-specific personalization. Each patient’s physiology, residual limb anatomy, and lifestyle are unique; AI enables prostheses to learn individual preferences, adjust grip force for common tasks, anticipate fatigue, or predict gait transitions under different terrains. Such adaptive capabilities reduce cognitive burden on the user and improve naturalistic embodiment of the device.

As computing hardware becomes more compact and energy-efficient, these algorithms are increasingly embedded on-device, reducing latency and dependence on external computers. This mirrors the neural efficiency of biological systems, which integrate sensing, computation, and actuation locally. Looking ahead, the convergence of AI with neuromorphic hardware and flexible bioelectronics promises to create prostheses that operate autonomously, respond seamlessly to environmental changes, and evolve alongside the user’s daily needs.

V. Clinical Applications and Outcomes

The ultimate measure of success for any bionic technology lies not in laboratory demonstrations but in clinical effectiveness and real-world adoption. In this section, we examine the outcomes of bionic limbs, sensory prostheses, and organ-level devices in patients. This section emphasizes rehabilitation results, usability, quality of life improvements, and challenges revealed in long-term use.

A. Bionic Limbs in Clinical Use

1. Upper-Limb Prostheses: Clinical trials of advanced upper-limb prostheses have demonstrated functional restoration, reduction in phantom limb pain, and increased quality of life. Ortiz-Catalan et al. reported a transradial neuromusculoskeletal prosthesis used continuously for more than three years [1]. Functional scores improved significantly: Southampton Hand Assessment Procedure (SHAP) scores increased by 23%, while pain interference with daily life decreased by more than 50%. Importantly, the user reported being able to wear the device comfortably throughout the day, an outcome rarely achieved with socket-based systems.
   
   Marasco et al. studied two patients with targeted reinnervation and closed-loop feedback [3]. Integration of touch, kinesthesia, and motor control allowed participants to perform tasks with visuomotor behaviors indistinguishable from able-bodied individuals. They no longer had to fixate visually on the prosthetic hand, freeing attention for higher-level planning. This demonstrates not only functional restoration but also a shift toward naturalistic embodiment.
2. Lower-Limb Prostheses: Lower-limb prostheses are judged by metrics such as walking speed, energy expenditure, and stability on varied terrain. Microprocessor knees consistently improve gait symmetry and reduce falls compared to mechanical knees [6]. Powered prostheses such as the OSL enable active ankle push-off, reducing metabolic cost of walking. Early clinical trials show transfemoral amputees can achieve walking speeds approaching those of able-bodied controls.Osseointegrated lower-limb prostheses also demonstrate marked improvements in mobility. A long-term Swedish cohort study showed that more than 90% of patients reported improved prosthetic comfort, stability, and walking endurance compared to sockets [15]. However, risks such as infection and implant loosening persist.

B. Bionic Eyes

Retinal prostheses such as Argus II have provided basic functional vision to hundreds of blind individuals worldwide [5]. Clinical outcomes show patients can detect light sources, navigate high-contrast environments, and recognize large objects. However, resolution is limited (about 60 electrodes), and many users report “pixelated” vision.

More recent neuromorphic retinal devices remain at the preclinical stage but show transformative potential. Long et al. demonstrated that a perovskite nanowire retina achieved filter-free color discrimination and wide-field imaging in laboratory models [10]. Similarly, Zhang et al. reported that their organic phototransistor retina exhibited plasticity and visual memory, features that could translate to more naturalistic visual experiences [11]. While these results have not yet reached human trials, they suggest a trajectory toward functionally rich vision restoration.

Fig. 8: Clinical outcome radar across device categories (illustrative 0–10 scales). Limb outcomes reflect neuromusculoskeletal prosthesis and closed-loop feedback studies [1], [3]; speech perception for ears from cochlear implant literature [22], [24]; HbA1c control from artificial pancreas trials [23]; independence and dexterity improvements from BCI work [13], [14].

C. Cochlear Implants and Auditory Prostheses

Cochlear implants are the most clinically established bionic devices, with decades of long-term outcome data. Studies consistently show that recipients achieve near-normal speech perception in quiet environments, with 80–90% of adult users able to understand conversational speech [22], [24]. Pediatric recipients implanted before the age of two can develop speech and language skills comparable to hearing peers, underscoring the importance of early intervention.Limitations remain in music perception and speech-in-noise environments. Fine structure processing algorithms have partially addressed these gaps, while emerging optogenetic cochlear implants offer the possibility of finer frequency resolution by stimulating auditory neurons with light rather than electricity [24]. Though still in early trials, such devices could overcome the channel interaction limits of electrode arrays.

Fig. 9: Patient-centered benefits vs. limitation severity (illustrative 0–10). Scores are synthesized from clinical and review reports across limbs [1], [3], [15], eyes [5], [10], [11], ears [22], [24], organs [23], and BCIs [13], [14].

D. Artificial Pancreas

The artificial pancreas has advanced from inpatient feasibility studies to widespread outpatient use. Closed-loop systems integrating continuous glucose monitors and insulin pumps have demonstrated significant clinical benefits.

In the Cambridge Artificial Pancreas trials, adults and adolescents with type 1 diabetes using closed-loop control spent over 70% of the day in target glucose range, compared to about 50% with conventional therapy [23]. HbA1c levels improved by approximately 0.5%, and hypoglycemia episodes were reduced by more than 40%. These results have led to regulatory approval of hybrid closed-loop systems (Medtronic 670G, Tandem Control-IQ).

Beyond diabetes, closed-loop bioelectronic devices for hypertension and epilepsy are under investigation, suggesting that the artificial pancreas is a prototype for a broader class of organ-level bionics.

E. Brain–Computer Interfaces in Clinical Translation

BCIs are increasingly being tested in patients with paralysis. Hochberg et al. demonstrated that tetraplegic individuals could use cortical implants to control robotic arms with multiple degrees of freedom, achieving self-feeding tasks [14]. More recently, bidirectional BCIs delivering somatosensory feedback via cortical stimulation have restored not only motor intent but also tactile perception in paralyzed patients [13].

Non-invasive BCIs, while less precise, have enabled basic communication in locked-in syndrome. EEG-based spellers, though slow, offer a lifeline for individuals otherwise unable to interact with their environment. Clinical usability, however, remains limited by low bandwidth and the need for expert calibration.

TABLE V: Summary of Clinical Outcomes in Major Bionic Devices

F. Patient Perspectives and Usability Studies

Across device types, patient-reported outcomes highlight both the benefits and limitations of bionic devices:

• Users of osseointegrated prostheses report higher comfort and daily wear time but express concerns about infection risk [15].
• Cochlear implant recipients overwhelmingly report improved quality of life, but many remain dissatisfied with music enjoyment [24].
• Bionic limb users often emphasize the importance of intuitive control and sensory feedback, without which devices are often abandoned [3].
• Artificial pancreas users report reduced cognitive burden, as the system automates much of the constant decision-making in diabetes care [23].

VI. Challenges and Limitations

While recent years have witnessed transformative advances in bionic technology, clinical translation remains constrained by significant technical, biological, and regulatory challenges. These limitations underscore the gap between laboratory performance and long-term, real-world usability.

A. Technical Challenges

1. Signal Reliability and Noise: Surface EMG, EEG, and even implanted electrodes are subject to signal drift, noise, and instability over time. Sweat, skin impedance, and electrode displacement degrade surface signals, while implanted electrodes may shift microscopically due to tissue remodeling. These issues lead to loss of calibration, forcing frequent retraining of pattern-recognition algorithms [3]. Reinforcement learning approaches partially address this, but stable long-term decoding remains elusive.

2. Power Supply and Energy Efficiency: Most prosthetic systems rely on rechargeable lithium-ion batteries, which add weight and require frequent charging. High-resolution bionic eyes and neural stimulators demand significant power for continuous operation, yet miniaturized power systems remain limited. Energy harvesting from body motion, thermoelectric gradients, or biofuel cells has been explored but is not yet clinically viable.

3. Durability and Mechanical Robustness: Devices implanted in the body must withstand years of mechanical stress. MEMS sensors, polymer electrodes, and microelectronics may degrade under physiological conditions. Similarly, osseointegrated implants must tolerate repetitive load-bearing without loosening. Failures not only compromise device function but may necessitate surgical revision, which carries additional risk [15].

4. Limited Bandwidth of Neural Interfaces: Even state-of-the-art cortical arrays record from a few hundred neurons, far below the millions involved in natural motor control. Intraneural electrodes provide some selectivity but risk damaging nerve fascicles. As a result, current devices cannot yet replicate the information throughput of the intact nervous system, restricting fine dexterity and natural sensory richness [1], [3].

B. Biological and Clinical Challenges

1. Immune Response and Biocompatibility: Foreign-body response to implanted electrodes leads to fibrotic encapsulation, increasing impedance and reducing signal quality over time. Flexible polymers (polyimide, parylene-C) mitigate this mismatch, but chronic inflammation remains a barrier to decades-long stability [9]. Similarly, neural tissue is highly sensitive; intraneural arrays risk long-term axonal degeneration.

2. Infection Risk in Osseointegration: Osseointegrated prostheses solve socket issues but leave a skin breach vulnerable to infection, despite titanium’s biocompatibility and antimicrobial coatings [15]. For many surgeons, this is the key clinical barrier to adoption.

3. User Variability: Each patient presents unique residual anatomy, nerve distribution, and physiology. This heterogeneity makes one-size-fits-all solutions impractical. A prosthesis calibrated for one individual may fail in another with different EMG signal distribution. Personalized adaptation through AI offers promise but requires extensive data and user training [12].

Fig. 10: Challenge severity heatmap (0–10) across technical, biological, and ethical domains. Entries are informed by clinical/engineering overviews [9], osseointegration cohorts [15], device and interface reports [1], [3], and neurotechnology ethics guidance [25].

4. Rehabilitation and Training Burden: Complex bionic devices demand intensive rehabilitation. For example, users of targeted reinnervation prostheses must undergo weeks of training to learn new muscle–nerve mappings [3]. Without structured rehabilitation, even advanced devices risk abandonment. Successful adoption also depends on continuous feedback from clinicians, adaptive training software, and strong patient motivation. Limited access to rehabilitation resources further compounds these challenges, highlighting the need for more intuitive control strategies and scalable support systems.

 Fig. 11: Failure pathway flowchart: signal noise/drift → calibration loss → poor usability → device abandonment.

C. Ethical and Regulatory Challenges

1. Accessibility and Equity: Advanced prostheses such as neuromusculoskeletal arms or retinal implants cost tens of thousands of dollars, often exceeding insurance coverage. As a result, access is restricted to high-resource healthcare systems, leaving a vast majority of potential users worldwide without benefit.
2. Privacy and Data Security: Brain–computer interfaces generate continuous neural data streams, raising concerns about privacy, surveillance, and misuse. Questions of ownership and protection of neural data are critical as BCIs move toward commercial use [25].
3. Regulatory Uncertainty: Regulatory frameworks (FDA, EMA) struggle to classify hybrid devices that combine hardware, software, and surgical procedures. Is a neuromusculoskeletal prosthesis a medical device, an implant, or a drug–device combination? These uncertainties slow approval processes and complicate clinical translation.
4. Psychological and Social Factors: Bionic devices affect not only physiology but also identity. While many users embrace prostheses as part of their body schema, others experience alienation. High expectations, fueled by media portrayals of “superhuman cyborgs,” may lead to disappointment when real devices fall short. Social stigma and lack of support also contribute to device abandonment [17].

TABLE VI: Challenges and Limitations of Bionic Devices Across Categories

D. Toward Overcoming Limitations

Efforts to address these limitations include:

• Flexible bioelectronics that minimize immune response [11].

• Antimicrobial and regenerative coatings for osseointegration [15].

• On-device AI that adapts to signal drift in real time [12].

• Energy harvesting systems that exploit body heat or motion [20].

• Ethical frameworks for neural data protection being drafted by international bioethics committees [25].

While none of these solutions is definitive, the rapid pace of interdisciplinary research suggests that many current barriers may be partially overcome within the next decade.

Fig. 12: Conceptual framework for the Future of Bionics. The figure highlights five major domains expected to shape ongoing advances: Advanced Prostheses, AI & Control, Sensing & Feedback, Clinical Translation, and Integration.

VII. Future Directions

The future of bionic devices lies in the pursuit of seamless human–machine integration, where artificial systems are no longer merely tools but functional extensions of the body. This vision depends on advances across materials science, neuroscience, artificial intelligence, and clinical medicine. The following subsections outline key directions for research and development.

A. Closed-Loop Systems

One of the most transformative frontiers in bionics is the realization of closed-loop systems that integrate sensing, computation, and actuation in a continuous feedback cycle. Current devices often operate in open-loop: users send motor commands, but feedback is limited or absent. This mismatch increases cognitive burden and reduces naturalism.

1. Closed-Loop Limb Prostheses: Future prostheses will combine multimodal sensory feedback (touch, proprioception, temperature) with real-time motor decoding. For example, Ortiz-Catalan’s osseointegrated system demonstrates that stable long-term recording and stimulation are possible [1]. The next step is to expand sensory channels, potentially incorporating thermal sensors like those proposed in thermally sentient limb prototypes [8].

2. Closed-Loop Vision: Neuromorphic retinas already integrate preprocessing [10], [11]. Coupled with cortical implants capable of delivering spatially distributed stimulation, future bionic eyes may offer continuous adaptive vision, adjusting to light conditions, motion, and attention demands.

3. Closed-Loop Organ Devices: The artificial pancreas exemplifies the power of closed-loop design, and similar architectures may be applied to renal replacement (bionic kidney) or cardiac regulation (bioelectronic pacemakers with adaptive control). These devices would monitor physiological parameters continuously and autonomously adjust output, minimizing user intervention.

Fig. 13: AI-driven closed loop in bionic devices: user intent → AI decoding/control → device actuation; multimodal sensory feedback closes the loop and adapts the user.

B. Artificial Intelligence Integration

AI is expected to play a central role in the next generation of bionics, from signal decoding to personalized adaptation.

1. Adaptive Decoding: Deep neural networks can outperform traditional pattern recognition in EMG/EEG classification, particularly under noisy conditions. Reinforcement learning approaches demonstrate that systems can learn alongside users, adapting to signal drift and improving control fidelity without explicit recalibration [12].

2. Personalization: AI enables prostheses to learn user-specific patterns, such as grip preferences, walking dynamics, or habitual tasks. Over time, the device can anticipate actions, moving toward predictive control rather than reactive operation.

3. AI in Sensory Processing: In vision and hearing prostheses, AI can denoise input signals, highlight salient features, or adapt stimulation patterns for optimal perception. For example, a bionic eye might enhance contrast in low-light conditions or suppress irrelevant motion, mimicking biological attentional mechanisms.

C. Fully Implantable Energy Systems

Energy supply remains a critical barrier. A major future direction is the development of fully implantable, autonomous power sources.

1. Biofuel Cells: Glucose biofuel cells convert glucose and oxygen from bodily fluids into electricity. Early prototypes have powered pacemakers in animal models, suggesting feasibility for low-power implants.
2. Energy Harvesting: Thermoelectric generators can exploit the temperature gradient between body heat and ambient air, while piezoelectric harvesters generate power from motion. If scaled effectively, such technologies could eliminate the need for external charging.

3. Wireless Power Transfer: Mid-field and ultrasound-based wireless charging systems are being explored as alternatives to inductive coupling, enabling deeper implants to be recharged without bulky external coils.

D. Advanced Materials and Interfaces

• Living Electrodes: Incorporating stem-cell–derived neurons into electrode arrays could reduce immune rejection and improve signal fidelity.
• Self-Healing Polymers: Materials capable of repairing microcracks would extend device longevity under mechanical stress.

TABLE VII: Emerging Research Trends and Future Directions in Bionics

• Nanostructured Coatings: Surfaces engineered at the nanoscale can reduce bacterial adhesion (limiting infection risk) while promoting neural growth.

E. Ethical, Social, and Regulatory Horizons

As devices become more integrated and powerful, ethical considerations will intensify.

• Cognitive Autonomy: Closed-loop BCIs blur the boundary between user intent and machine response, raising questions about responsibility and agency.
• Human Enhancement vs. Therapy: While most bionics are designed for rehabilitation, the same technologies could be repurposed for augmentation, e.g., enhanced vision beyond the human spectrum.
• Global Accessibility: Future efforts must prioritize equitable distribution, ensuring that breakthroughs benefit not only high-income nations but also the global disabled population.

F. A 5–10 Year Outlook

Within the next decade, it is realistic to expect:

• Commercially available bidirectional prostheses with tactile and kinesthetic feedback.

• Next-generation artificial pancreas systems incorporating dual-hormone control (insulin + glucagon) for near-normal glycemic regulation.

• Retinal prostheses with color vision based on neuromorphic phototransistor arrays.

• BCIs with wireless, minimally invasive arrays enabling everyday use in home environments.

• Early adoption of autonomous power systems, reducing dependence on external charging.

In parallel, advances in AI, regenerative medicine, and material science will continue to converge, driving bionics toward devices that are smaller, smarter, and more biologically integrated.

Fig. 14: Convergence of disciplines: neuroscience, engineering, AI, and ethics. The overlap represents future bionics as seamless human–machine symbiosis.

G. Toward Convergence of Disciplines

The future of bionics will not be driven by any single technology but by the convergence of multiple fields. Neuroscientists, engineers, material scientists, ethicists, and clinicians must collaborate closely to translate prototypes into scalable, safe, and accessible devices. If this trajectory is sustained, the coming decades may witness the realization of functional, lifelong, fully integrated artificial organs and limbs, transforming rehabilitation and human–machine symbiosis.

VIII. Conclusion

Bionic devices have evolved from crude mechanical substitutes to sophisticated systems capable of bidirectional integration with the human nervous system. The landscape now includes neuromusculoskeletal limb prostheses, neuromorphic bionic eyes, cochlear implants, artificial pancreas systems, and experimental brain–computer interfaces. Across these categories, the central trend is clear: modern bionics aim not merely to restore function but to recreate the natural sensory–motor loop, thereby enhancing embodiment, autonomy, and quality of life.

Key technological enablers include biocompatible materials (e.g., titanium for osseointegration, organic semiconductors for neuromorphic sensors), advanced sensors and actuators (MEMS tactile arrays, series elastic actuators), and high-bandwidth neural interfaces (intraneural electrodes, hybrid nerve constructs).

Artificial intelligence and machine learning are now integral, allowing devices to adapt dynamically to user variability and environmental change.

Clinical studies demonstrate tangible benefits: improved dexterity, reduced phantom pain, restored sensory perception, and automated metabolic regulation. Yet significant challenges remain. Technical barriers such as signal instability and power supply, biological risks including immune response and infection, and ethical concerns over equity and neural data privacy all impede widespread adoption.

Looking forward, the field is poised for breakthroughs in closed-loop systems, AI-driven personalization, fully implantable energy solutions, and multimodal sensory feedback. Within the next decade, it is realistic to envision prostheses that feel like natural limbs, artificial organs that self-regulate without intervention, and BCIs that allow paralyzed individuals to regain independence in daily life.

Ultimately, the future of bionics lies in the convergence of disciplines—neuroscience, engineering, medicine, and ethics—to create devices that are not only technologically advanced but also safe, equitable, and meaningful for users. The vision of lifelong, seamlessly integrated artificial organs and limbs is no longer science fiction, but an achievable milestone within a generation.

References

1. M. Ortiz-Catalan et al., “A highly integrated bionic hand with neural control and feedback for use in daily life,” Sci. Robot., vol. 8, no. 77, eadf7360, 2023.
2. S. Dosen, “Toward self-contained bidirectional bionic limbs,” Sci. Robot., vol. 8, no. 77, eadk6086, 2023.
3. S. Marasco et al., “Neurorobotic fusion of prosthetic touch, kinesthesia, and movement,” Sci. Robot., vol. 6, no. 59, eabf3368, 2021.
4. C. Pasluosta et al., “Bidirectional bionic limbs,” J. Neural Eng., vol. 19, no. 1, 013001, 2022.
5. U.S. Food and Drug Administration, “Argus II Retinal Prosthesis System—Summary of Safety and Effectiveness Data,” FDA, 2011.
6. T. R. Clites et al., “Design and clinical implementation of an open-source bionic leg,” Nat. Biomed. Eng., vol. 4, pp. 941–952, 2020.
7. Y. Cho et al., “Hybrid bionic nerve interface for application in bionic limbs,” Adv. Sci., vol. 10, no. 5, 2206859, 2023.
8. M. Ortiz-Catalan, “Thermally sentient bionic limbs,” Nat. Biomed. Eng., 2024.
9. Editorial, “Advances in clinical and prosthetic care,” Front. Rehabil. Sci., vol. 3, 2022.
10. Z. Long et al., “A neuromorphic bionic eye with filter-free color vision using hemispherical perovskite nanowire array retina,” Nat. Commun., vol. 14, 37581, 2023.
11. H. Zhang et al., “A neuromorphic bionic eye with broadband vision and biocompatibility using TIPS-pentacene phototransistor array retina,” Appl. Mater. Today, vol. 32, 2023.
12. H. R. Schone et al., “Biomimetic versus arbitrary motor control strategies for bionic hand skill learning,” Nat. Hum. Behav., vol. 8, pp. 1108–1123, 2024.
13. S. N. Flesher et al., “Restored tactile sensation improves neuroprosthetic arm control,” Sci. Transl. Med., vol. 13, no. 612, 2021.
14. L. R. Hochberg et al., “Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,” Nature, vol. 485, pp. 372–375, 2012.
15. R. Brånemark et al., “Osseointegrated percutaneous prostheses for patients with limb loss,” Bone Joint J., vol. 101-B, pp. 55–63, 2019.
16. C. Pasluosta, P. Kiele, S. Micera et al., “The current state of bionic limbs from the surgeon’s viewpoint,” J. Neural Eng., vol. 19, no. 1, 2022.
17. “The future of bionic limbs,” Prosthet. Orthot. Int., vol. 45, no. 5, 2021.
18. “Clinical implementation of advanced bionic prostheses,” Nat. Commun., vol. 15, 2024.
19. H. R. Schone et al., “Should bionic limb control mimic the human body? Impact of control strategy on bionic hand skill learning,” bioRxiv preprint, 2023.
20. “Advances in prosthetic rehabilitation sciences,” Front. Rehabil. Sci., vol. 3, 2022.
21. World Health Organization, Global Report on Rehabilitation, Geneva, 2022.
22. P. C. Loizou, “Cochlear implants: Historical perspective and current applications,” IEEE Eng. Med. Biol. Mag., vol. 25, no. 5, pp. 40–46, 2006.
23. L. Hovorka, “Closed-loop insulin delivery: From bench to clinical practice,” Nat. Rev. Endocrinol., vol. 7, pp. 385–395, 2011.
24. B. C. Wilson, “The future of cochlear implants,” J. Assoc. Res. Otolaryngol., vol. 18, pp. 695–704, 2017.
25. UNESCO International Bioethics Committee, “Ethical issues of neurotechnology,” Policy Report, 2021.

About the author

Vyomesh Vikram Singh

Vyomesh Vikram Singh is a recent high school graduate with interests in computer science, robotics, and edge AI. His projects have included automation systems, gesture-controlled vehicles, unmanned aerial platforms, and a low-cost Raspberry Pi projector designed for rural classrooms to help underprivileged children connect to the internet. He has also worked on model optimisation for IoT devices and authored a review on human–computer interfaces, bionic limbs, and neuromorphic vision.

His research interests include robotics, human–machine interaction, and efficient AI deployment on embedded systems. Beyond academics, Vyomesh also served as a board member of Adlers, the Photography Club. Combining technology with art, he applied machine learning to make his filmography distinct and innovative, earning recognition at multiple competitions, including 3rd prize in an international geography documentary competition in Geofest International. Vyomesh has contributed to leadership and creative pursuits including serving as the head of his school’s Robotics Club for 3 years, guiding peers in building prototypes and organising seminar series by inviting respected external speakers. He is also an intermediate guitar player, and tries to make people around him happy by his talent.

The post Advanced Human–Computer Interfaces and AI : A Comprehensive Review appeared first on Exploratio Journal.

Abstract

Bridges being tightly linked to public safety, inspections of their conditions are crucial to conduct immediate rehabilitation or replacement. However, they are often resource-intensive and difficult to scale given the number of bridges in a state. This study explores the use of machine learning models to predict the overall condition of bridges in Massachusetts using publicly available data from both the Massachusetts Department of Transportation (MassDOT) and the National Bridge Inventory (NBI). Building upon prior studies in infrastructure analytics, this research offers a novel contribution by applying a multiclass classification framework to bridge condition prediction using state-level datasets. Seven multiclass classification algorithms were tested, including Random Forest, XGBoost, and Multi-Layer Perceptron on up to 7,232 bridge records. Among them, XGBoost achieves the highest accuracy, precision, recall, and F-1 score (82.99%, 81.10%, 64.38%, 68.29 % respectively). These findings suggest that Machine Learning models can support state agencies in prioritizing bridge inspections and allocating resources more efficiently; additionally, understand the importance of features, such as age and span of the bridge. However, it is worth noting that both datasets are limited in different ways: size of the data and the number of predictive features, as well as the imbalance in NBI dataset. As this study focuses on Massachusetts, it would be generalizable to other statewide research; nevertheless, when the scope is the entire United States, variables like climate would influence the results. 

1. Introduction

A. Background Information

As urban cities grow and need for transporting goods and services increase, the demand placed on bridges has exponentially increased, which intensifies the stress on supporting structures and the rate of deterioration. By 2024, 36% of US bridges are classified as having the need for replacement or rehabilitation, and 6.8% are classified as structurally deficient in poor condition (ARTBA Bridge Report). An example is the Francis Scott Key Bridge, which collapsed in March 2024 due to the lack of inspection on the truss and the overall condition of the structures, resulting in injuries and presumed deaths; the bridge collapsed on ship, which indefinitely blocked the passageway (Regan et al.). Therefore, regular maintenance of these bridges is vital to ensure the stability of the structures and prevent collapses; yet these processes require lots of resources and time allocated, causing them to be a great challenge for state and local governments.

Currently in the US, among more than 623,000 bridges, 49.1% are in the category of “fair” condition, and 44.1% are in “good” condition, and the rest 6.8% are in “poor” condition. The Infrastructure Investment and Jobs Act (IIJA) allocated approximately $40 billion for bridge investment and maintenance (Infrastructure Investment and Jobs Act | FHWA, 2022), yet this funding remains insufficient to address the needs of all U.S. bridges, especially given that over 36% require rehabilitation or replacement; according to ARTBA in 2023 (A Comprehensive Assessment of America’s Infrastructure, 2025), over $319 billion is needed to make all needed repairs, and ASCE estimated $191 billion for all rehabilitation (American Society of Civil Engineers, 2017), which shows the limited budget on bridge inspection. Hence, this stresses the need for precise location of the bridges that could potentially degrade from one “good” to “fair” or “fair” to “poor”, so that the limited budget will not be wasted on routinely checking bridges that do not have the potential to degrade, ensuring the efficient use of resources. 

The main determinants of bridges’ rate of deterioration include bridges’ age, average daily traffic, bridge material, structure length, and deck area, which are reflected by their correlation with the overall conditions of bridges. A low-cost and efficient solution is to determine potential deterioration of bridges using existing data on the above bridge features without physical testing on the condition of bridges, which can accurately locate the bridges requiring maintenance, rehabilitation or replacement and substantially decrease the number of resources and time devoted to inspecting bridges in relatively good conditions if a comprehensive data set is provided by the state Department of Transportation. This study focuses on the bridges in the state of Massachusetts, with the bridge inventory published in 2025 by MassDOT; each bridge has an overall rating on a scale of 0-9 with a 0-8 rating for bridge deck, superstructure, and substructure conditions which are used as the predicted values of the machine learning models with the input of bridge features.”

B, Literature Review

A review of the existing literature on the machine learning approach to predict infrastructure conditions reveals a range of comparative research, optimization of previous models, and development of new models. Specifically, for bridge condition predictions, most researches use the properties of bridges included in the National Bridge Inventory (NBI), which consists of both structural properties of bridges and external elements, such as traffic and environment. The entire NBI contains 123 fields, with about 25 predictive fields of categorical and nominal data that’s directly collected during inspection that has linked to the bridge deterioration. To predict the next condition predict using nominal and categorical data from one year, algorithms like K-Nearest Neighbors (KNN), Random Forest Classifier (RF), Artificial Neural Network (ANN) were most used due to their ability to handle heterogeneous datasets and capture complex non-linear relationships between input features and structure conditions. In this literature review, multiple previous research on predicting bridge deterioration will be analyzed, so that this study can build upon their strengths and limitations.

In the study by Assaad et al. 2020, they aim to predict the specific condition ratings of the deck component in a bridge, as decks are directly exposed to the traffic pressure and weather and impact the usability of bridges more than other components. The Machine Learning algorithms ANN and KNN were trained upon ten fields in NBI selected by the Boruta algorithm and optimized by hyperparameter tuning and cross-validation. Overall, the ANN achieved 91.44% and KNN 89.99% accuracy rate. However, this high accuracy rate was likely caused by the use of superstructure and substructure condition ratings as a variable, which is calculated rather than tested, and could have direct correlation with the deck conditions (Assaad et al., 2020). Despite the limitations, this provides insights into the method to optimize the algorithms and the fields to use for prediction.  

In another study by Mia & Kameshwar in 2023, similarly, they used a machine learning-based approach to predict the bridge deck, superstructure, and substructure components’ condition ratings of the bridge, ultimately reaching a highest accuracy rate of 93%-97% using a RUSBoost combined with Random Forest algorithm. The approach of this study to predict each component separately is proved successful as different features in the NBI impact each component differently. However, one issue in this study is how the data used is largely biased, without even number of bridges in each condition rating. This issue could be prevented by evening the dataset using bridges in different conditions; however, this could lead to a large reduction in data size, which would occur if only one US state’s bridge data nb is used. A notable aspect of this study is the uncertainty quantification, which simulates the variability in training due to different sampling of data, which demonstrates the reliability of using different algorithms consistently for this prediction.

Another research similar to this is an investigation done by Mohammadi et al. in 2023, in which they trained and tested different machine learning algorithms on predicting culvert conditions, which face similar deterioration as bridges, but more environmental impacts such as soil pH level etc. In this study, they tested Random Forest, ANN, Decision Trees, SVM, and KNN, and concluded Random Forest had the highest accuracy of 82%, as RF is particularly effective with a combination of numerical and categorical features; additionally, the age of the culvert was also found to be the most predictive feature. Also mentioned in the study, one key limitation of this study is to test newer classification algorithms. Since technology continuously advances, later models were expected to perform better than older ones like Random Forest. Therefore, in this study, models like XGBoost and RUSBoost are included.

One study on bridges in Ohio state by Nasab et al. in 2023 uses XGBoost, Random Forest, ANN, SVM, K-NN, Logistic regression and Decision Tree, with their accuracy, precision, recall, and F1 scores calculated. As a result, this study found XGBoost has the best performance with the highest score for all metrics, followed by Random Forest. It is discussed by researchers that XGBoost could handle the numerical and categorical features more effectively; also with the L1 and L2 regularization, it does not overfit with redundant data, unlike Random Forest, which has high training accuracy but lower testing accuracy. However, despite the rigorous testing of seven models, this study also only tested for predicting bridge deck conditions, which could lead to different results in the feature importance conclusions. 

Apart from the performance and models, the importance of different bridge features is also a crucial knowledge this study tries to find and discuss, which is analyzed in a study by Fard and Fard in 2024, comparing three models’ performance in predicting deck condition of bridges. They tested and concluded that Age of the bridge is the most important feature, followed by climate regions, deck area, and ADT. Since this study specifically focuses on bridges in Massachusetts, climate regions will not be applicable. That said, having ADT as the fourth most important feature is highly relevant. In this study, all of the components and the overall condition are predicted and tested, and it is expected that ADT would influence different components differently; the conclusion made by Fard & Fard therefore suggests its strong correlation to the deck component.

All of the above studies focused on using normal classification algorithms. In the study by Liu & Zhang 2020, they attempt to predict condition ratings of bridge components using historical NBI data with a Convolutional Neural Network (CNN) model. Specifically, they used the Maryland and Delaware highway bridges from 1992 – 2017. Overall an 85-87% accuracy is achieved by the model. This is an entirely different approach compared to all the other studies mentioned earlier; nevertheless, it is seen that the accuracy score of CNN does not have much difference with the previous results. This is likely due to the lack of a large dataset that is cleaned and labeled to predict accurately. This attempt to use temporal forecasting rather than predicting using measured features of the bridge on one inspection could be applied to create a trend line of bridges’ deterioration, but providing the imperfect accuracy rate, perhaps such method would not be as effective and efficient compared to cross-sectional models utilized in this study.

One study that achieved an especially high accuracy rate is a study by Li and Song in 2022, which they focus on comparing ensemble machine learning models for predicting steel bridge deck defect conditions using NBI data from 2021. The models used were RF, Extra Trees, AdaBoost, Gradient Boosting, XGBoost, and LightGBM, concluding that XGBoost performed best with accuracy of 94.95%, AUC of 90.26%, Recall of 99.94%, Precision of 94.98%, and F1 Score of 97.40%. This score, though shows XGBoost high performance in predicting the bridge data, is caused by the test of binary classification, where they only examined the ability of the models to classify bridges into defective or non-defective classes, also with the use of condition ratings of superstructure and substructure. This therefore makes the study limited despite the insights it provides about ensemble machine learning models’ capabilities in this area of prediction.

Figure 1: Table of Bridge Components and Models Included in Previous Literature

Figure 2: Table of the Most Important Features and Best Performing Model in Previous Literature

II. Methodology

A. Data Selection and Processing

The experimental process mainly lies in the selection of data and the implementation of various machine learning algorithms that are developed at different times and made with similar purpose but different ways. 

Starting with data selection, for Massachusetts, there are two main source of collecting data of bridge properties and conditions: MassGIS Data Hub, a public dataset made by Massachusetts geoDOT, and the Massachusetts sub section in the National Bridge Inventory (NBI). The two datasets are in different formats and contain different fields, and therefore a comparison had to be made before processing the two datasets separately. In terms of number of fields, the CSV data from MassGIS Data Hub have 45 fields including non-predictive fields and information about the data collection itself. In terms of bridge count, 8152 rows are included in this dataset. Out of the 45 fields, there are only six useful fields that can be used, both numerical and categorical listed below:

Figure 3: List of Predictive Features in MassGIS Data Hub

The features listed above are only ones that are useful for this study, as some predictive fields were excluded as they are not applicable in a Massachusetts only context; for example, the longitude and latitude fields would be useful in a study of the entire United States as they can determine the climate zone of the bridges, but since Massachusetts as a whole lies in one climate zone, these fields are removed from the dataset. Moreover, to handle missing values, particularly for ADT, the rows with missing values or N were dropped, which is necessary to maintain consistency, although it resulted in a reduction in overall sample size. Furthermore, there are certain fields that could have a high correlation with the bridge components’ conditions but cannot be collected directly from bridges. For instance, superstructure condition ratings could have a very high correlation with deck condition ratings as shown by multiple other literature, but this field is already a calculated classification based on the features found in inspection. Therefore, only the six fields above are used for prediction. However, this dataset does not include every feature that impacts a bridge’s deterioration; more specifically, average daily traffic. Since the bridge mainly deteriorates when there’s a load on the bridge, the lack of such features would therefore lead to low accuracy scores. Hence, a solution to this problem is to add the ADT values from the NBI to this dataset by matching the bridgeDepartmentNumber in the MassGIS Data Hub dataset and the STRUCTURE_NUMBER_008 in the NBI. Since the two datasets do not contain the same number of bridges included, several bridges in MassGIS Data Hub have no corresponding ADT values in the NBI; therefore, after dropping all the rows including bridges without ADT values, a new dataset consisting 7232 rows is produced.

On the other hand, the NBI data is much more comprehensive in terms of number of fields included, with 123 fields. The data also includes 5312 rows, which is about 2000 short compared to the MassGIS Data Hub dataset. In all of the 123 fields, 27 fields are predictive to some extent. Within that, year built and year reconstructed are used to calculate the bridge’s age by subtracting the largest one of them form 2025; also, all of the categorical fields are transformed into binary fields as string is not an accepted data type by many machine learning algorithms implemented in this study. The complete list of fields used for prediction is listed below:

Figure 4: Predictive Features in NBI Dataset

Given the large amount of predictive features in the NBI dataset and the different layout and amount of rows in the dataset from MassGIS Data Hub, testing the machine learning algorithms on both datasets would be a suitable attempt to double check the results on one dataset. In addition, using the Mass GIS Data Hub dataset with less predictive features included could also provide insight into how the models perform differently when the input features are limited, which could occur under budget limitations to inspection etc.; building on this idea, if a model could perform similarly well in limited conditions, perhaps it suggests the possibility to reduce the complexity and amount of inspection routines on bridges, which could significantly cut the budget needed to ensure bridge security and safety.

Then, the fields that will be predicted are the condition rating for each component of the bridge and the overall condition. In the MassGIS Data Hub data, the fields: deckCondition, superstructureCondition, and substructureCondition have rating system from 0-9; from 0-4 the component is classified as poor, 5-6 is fair, and 7-9 is good. The overallCondition only has three levels, 4, 6, and 9, which signifies poor, fair, and good. The score is based on the lowest rating within all components. For example, the lowest score is 5 from substructureCondition of a bridge, the overallCondition would be 6-fair. Similarly, in the NBI, DECK_COND, SUPERSTRUCTURE_COND, and SUBSTRUCTURE_COND also follow the same rating system, and the BRIDGE_CONDITION is calculated the same way but displayed as G-Good, F-Fair, and P-Poor. Though keeping the component ratings as 0-9 since it provides a more detailed prediction to help assess the bridges’ conditions; due to the imbalance between amount of bridges with each rating and the use of k-fold cross validation, there would not be sufficient amount of data to be used for each fold. Hence, the components are also categorized into Good, Fair, and Poor for the two datasets. 

B. Implementation of Machine Learning Algorithms

After reviewing all the other literature of machine learning approach to predict bridge conditions, the following models are selected to be implemented and compared in this study: 

Figure 5: List of Models Tested in this Study

The list of machine learning algorithms listed above covers multiple different types of models, as well as a long time span, which allows a rigorous comparison of their performance to be conducted. 

To evaluate model performance, stratified K-fold cross-validation with five folds is employed to divide the entire dataset into five subsets of equal size, which each of which has the same proportion of bridge condition ratings as in the full dataset (ratio of good, fair, and poor bridges). In each iteration, four of the five folds are used to train the models, and the remaining one is used for testing. With this iteration repeating five times and averaging the score metrics across five runs can produce an unbiased estimation of the general performance of the models in predicting bridge conditions. The specific metrics used to measure the performance of the models are accuracy, precision, recall, and F1-score. Accuracy is a simple but most straightforward metric, which represents the overall proportion of correct predictions out of all the predictions made, calculated through the formula:

This score gives a general overview of how well the model performs; however, it is important to acknowledge that it can be misleading when the dataset is imbalanced. In the NBI dataset, 66.26% of bridges are fair, 24.31% are good, and only 9.43% are poor, meaning it is possible to achieve a high accuracy score by only identifying fair bridges correctly, and hence it is not used alone. Precision measures the quality of positive predictions, which is defined by the proportion of correct positive predictions in all positive predictions, shown by the formula below:

In the condition of budget restrictions, precision would be important, as low recall score would mean predicting good or fair bridges as poor conditioned, causing unnecessary inspections or repairs; therefore, only models with high precision scores are ideal for this task. Recall, on the other hand, shows how complete the positive predictions are, which is how many correctly predicted positives are in all the actual positives, which can be displayed by:

Recall is crucial in this context, as it shows how the model predicts all of the poor bridges correctly. A low recall score would mean that some poor bridges could be predicted as good or fair, leading to potential safety and failure issues. Lastly, the F1-score is a combination of precision and recall, calculated by their harmonic mean. This metric is useful when a balance between precision and recall is wanted. Specifically in the context of bridge condition prediction, missed detection of poor bridges and unnecessary inspections are both unwanted, F1-score would highlight models with imbalances between the two scores, providing a comprehensive evaluation of their performances and usability in real world context. The F1-score is calculated through the formula below:

III. Results

A. Evaluation Metrics

After running all the models on both datasets, illustrations are produced for both datasets. For the MassGIS Data Hub dataset, all of the target fields are in the rating of 1-9 as there is enough data for each category. 

Figure 6: Performance Comparison of Machine Learning Models on MassGIS Data Hub Dataset

Interpreting the charts above, over the prediction of the bridge decks, superstructures, substructures, and overall condition, Random Forest and XGBoost were the best performing models throughout almost all metrics; specifically, in overall prediction, RF achieved an 81.24% accuracy and XGBoost achieved 82.99% accuracy, and followed by the MLP classifier with 79.29% accuracy. This result follows expectations: firstly, since the target field overallCondition has only three categories (Good, Fair, Poor), it is more likely that the predictions are correct compared to other smaller components; secondly, the two ensemble learning models are inherently good predictors for non-linear interactions between variables and for tabular data, and the neural network MLP classifier is much more complex and stronger than other models tested, making it perform better in condition predictions. However, it is important to acknowledge that accuracy is purely a general measurement of the models’ performance as mentioned earlier, since there are other factors to be considered in the task of bridge condition prediction-specifically recall and precision. It would be more straightforward to analyze the F1-scores as they act as combinations of the two metrics. While Random Forest maintained strong accuracy (81.24%), its F1 score of 59.98% was noticeably lower than those of XGBoost and MLP, indicating a weaker balance of precision and recall, especially in misclassifying minority classes. This therefore illustrates the strong balance precision and recall in XGBoost and MLP, showing how they would be not just accurate, but also reliable and consistent in their predictions. This pattern is especially evident in the prediction of superstructure and substructure where XGBoost achieved F1 scores of 70.27% and 65.70% respectively, and MLP followed with 67.65% and 61.63%, which outperformed other more traditional models in consistency.

Figure 7: Performance Comparison of Machine Learning Models on NBI Dataset

The above four bar charts illustrate the results of the model’s performance on predicting based on the NBI data. For NBI, due to the limited amount of data for each condition rating for every components, the rating of the components are converted into categories of Good, Fair, and Poor with the same ranges as the overall rating. In the metrics for NBI, it can be observed that there is generally a greater gap between accuracy and precision, recall, and F1-score than the scores for MassGIS Data Hub. This is likely caused by the smaller and less balanced data set in the NBI, which after cleaning nulls and rows without data for some predictive fields, there is significantly less data of bridges in the Good category than data in Fair and Poor. And this is possibly a result of having more predictive fields and smaller data size. Looking at the metrics for overall condition predictions, XGBoost achieved the highest score for all metrics except recall, (69.94% for accuracy, 67.66% for precision, and 54.61% for F1-score), which RUSBoost has the highest of 54.23%%. This demonstrates how XGBoost is a strong performer despite the limitation of the dataset, aligning with the previous literature involving XGBoost, and is also consistent with the balance of precision and recall scores, which symbolizes its reliability in potential real world applications. Apart from this, there are also multiple things worth noting about the metrics. In predicting all the components including the overall condition, RUSBoost has maintained a balance score for the four metrics, and an especially high score for recall. This suggests that boosting with undersampling helps the model to handle class imbalances in the NBI, meaning its potential application when balanced and unbiased dataset is unachievable. Furthermore, MLP Classifier, though not the best, shows high consistency throughout the component predictions, with decent F1-score throughout, showing the balance between precision and recall scores, showing MLP Classifier’s potential application to capture the non-linear relationships that’s overlooked by simpler models as a complementary model.

B. Confusion Matrix

To evaluate the specific consistency and accuracy of the top models’ predictions, confusion matrixes are created for XGBoost, Random Forest, and MLP Classifier.

Figure 8: Confusion Matrix of the Three Best Performing Machine Learning Models

At first glance, the confusion matrix reveals the imbalance of the dataset mentioned earlier, as most of the bridge data included in the two datasets are classified as under the Fair condition. This means that models could achieve a high accuracy score only by predicting the Fair condition correctly. On both datasets, XGBoost correctly classified 180 and 50 poor cases in MassGIS and NBI datasets respectively, indicating the weak sensitivity to structurally deficient bridges, which is likely caused by the comparably small amount of data for Poor conditioned bridges. On the other hand, Random Forest struggled even more severely with 73 and 1 correct classification on the two datasets, suggesting that the previous high accuracy scores were likely caused by the high 4781 and 2931 cases of Fair bridges that it correctly predicted. In contrast, the MLP Classifier demonstrated the strongest accuracies for the Poor condition and most of Good condition, surpassing Random Forest and XGBoost completely. Therefore, considering MLP Classifier’s lowest accuracy for Fair condition, it is the most consistent and balanced model of the three listed. XGBoost and Random Forest, the two models considered as the best performing models’ misclassifcation of Poor condition raises concerns as bridges with poor conditions are also the ones needing intervention and repair the most. Conversely, MLP’s stronger and more balanced prediction of classes with less data provided for training could be a better model for circumstances with smaller data set. Moreover, the overprediction of the Fair condition implies that the current classification’s boundaries are too blurred, reflecting the limitations in the labeled data in Massachusetts. Ultimately, while ensemble models provide a strong accuracy, deep learning models like MLP may offer better balance and reliability in the contest of bridge condition predictions.

C. Feature Importance

Another feature of the XGBoost is its capability to output the importance of each feature fed into the model for prediction, and the top ten features for each component in each dataset are illustrated by the diagrams below (MassGIS Data Hub on the left and NBI on the right):

Figure 8: Chart Showing the Importance of Features on Predicting Different Components for XGBoost

Deck Condition

Superstructure Condition

Substructure Condition

Overall Condition

Over the different components, age is one of the most prominent feature in prediction, which is also shown in the results achieved by Mohammadi et al., Nasab et al., and Fard & Fard in their papers. However, the results above also show other important features in prediciton. The top feature in the NBI dataset predictions is STRUCTURE_FLARED_035; a flared bridge would introduce a non-uniform load distribution which could lead to stresses under uneven traffic on top, which could potentially lead to more deterioration than bridges that are not. Also, since flared bridges are also more common in older bridges, it could have an indirect correlation with the poorer conditions of the bridges (Thakkar et al.). Additionally, a major reason could be how flared bridges are generally more difficult to inspect or repair, which could lead to delays in maintenance.

IV. Conclusion

The results from this study make clear how machine learning models such as XGBoost, MLP Classifier, and Random Forest can support more efficient and targeted inspection planning, potentially reducing unnecessary inspections and reallocating resources toward at-risk infrastructure, and decrease the chance of bridge failures caused by lack of maintenance or inspection. Moreover, as the models above also demonstrated high accuracy and balanced precision and recall scores, transportation authorities in Massachusetts, or potentially in the United States, can consider employing such techniques to proactively identify bridges that have higher risk of deterioration and failure. When bridges are predicted in poor condition, the DOT could conduct immediate inspections and maintenance, and when bridges are predicted to be in good condition, the DOT could extend the period between inspections. Overall, this strategy not only improves the efficiency of resource allocation, but also ensures public safety by reducing the failure of overlooked bridges. In addition, the feature importance collected from XGBoost training provides insights into the importance of inspecting bridges in Massachusetts with flared structures and columns, and many other features as listed in the diagram above. 

A. Limitations

Starting with the dataset used in this study to train and test machine learning models, the two datasets: MassGIS Data Hub and NBI are both publicly available but are not frequently updated and do not include some images, environmental features, and time-series deterioration patterns in inspections that could boost the accuracy and precision of the models’ predictions, especially for neural network models. A key limitation is as revealed in the confusion matrix, both datasets employed have imbalanced amount of bridge data for each category, in which the number of bridges classified as Fair greatly exceeds the other two categories than the other two. Additionally, due to the use of k-fold cross-validation to avoid the inherent bias of the performance metrics, and because the amount of bridges under each rating from 0-9 is not enough to be distributed to all 5 folds, this study is not able to predict bridge components’ specific ratings from 0-9, which reduces the precision of the outcome and the limits the generalizability of the study to nationwide researches. Furthermore, due to the scope of this study, which only focuses on Massachusetts, the conclusion, even though similar to many previous literature, cannot be directly generalized to other countries, as they would have different traffic patterns, design standards, or management routines. Moreover, the scope of this study also means that climate zone cannot be considered as Massachusetts is within a single climate zone and therefore will not affect the predictions; however, according to other literature, climate zone could potentially provide more information for the models to predict more accurately.

B. Future Work

In future development based on this study, the dataset could be zoomed out to include other states with different bridge management systems to compare the performances of the machine learning models in different environment. Not just the size, since this study only focuses on one climate zone, the environmental factors such as average temperature, wind speed, weather are not considered in the training or testing; therefore, including those factors could provide a more comprehensive conclusion of the feature importance and actual performance of utilizing machine learning models in real scenarios. Moreover, the variety of machine learning methods can be increased, adding methods such as using historical data like Liu & Zhang’s study involving CNN. Lastly, though mentioned in this study, methods such as stacking ensemble or hybrid models are tested; however, they could be much more powerful than using simple and single models like the ones involved in this study.

References

ARTBA Bridge Report. (n.d.). Artbabridgereport.org. https://artbabridgereport.org/

A Comprehensive Assessment of America’s Infrastructure. (2025). American Society of  Civil Engineers.

ARTBA 2023 Bridge Report: 222,000 U.S. Bridges Need Major Repairs – ARTBA. (2023,  August 13). Https://Www.artba.org/. https://www.artba.org/news/artba-2023-bridge-report-222000-u-s-bridges-need-major-repairs/

Regan, H., Magramo, K., Radford, A., Ebrahimji, A., Chowdhury, M., Ramirez, R.,  Hammond, E., Aditi Sangal, Powell, T. B., Blackburn, P. H., & Magramo, K. (2024, March 26). March 26, 2024 – Baltimore Key Bridge collapses after ship collision. CNN. https://www.cnn.com/us/live-news/baltimore-bridge-collapse-03-26-24-intl-hnk

Infrastructure Investment and Jobs Act | FHWA. (2022). Dot.gov.  https://highways.dot.gov/tags/infrastructure-investment-and-jobs-act

Assaad, R., & El-adaway, I. H. (2020). Bridge Infrastructure Asset Management System:  Comparative Computational Machine Learning Approach for Evaluating and Predicting Deck Deterioration Conditions. Journal of Infrastructure Systems, 26(3), 04020032. https://doi.org/10.1061/(asce)is.1943-555x.0000572

Chen, R.-C., Dewi, C., Huang, S.-W., & Caraka, R. E. (2020). Selecting critical features for data classification based on machine learning methods. Journal of Big Data, 7(1). https://doi.org/10.1186/s40537-020-00327-4

Liu, H., & Zhang, Y. (2020). Bridge condition rating data modeling using deep learning algorithm. Structure and Infrastructure Engineering, 16(10), 1447–1460. https://doi.org/10.1080/15732479.2020.1712610

Md. Manik Mia, & Sabarethinam Kameshwar. (2023). Machine learning approach for  predicting bridge components’ condition ratings. Frontiers in Built Environment, 9. https://doi.org/10.3389/fbuil.2023.1254269

Mohammadi, P., Asgari, S., Rashidi, A., & Alder, R. (2025). Culvert Inspection Framework Using Hybrid XGBoost and Risk-Based Prioritization: Utah Case Study. Journal of Construction Engineering and Management, 151(6). https://doi.org/10.1061/jcemd4.coeng-16448

Armin Rashidi Nasab, & Hazem Elzarka. (2023). Optimizing Machine Learning Algorithms for Improving Prediction of Bridge Deck Deterioration: A Case Study of Ohio Bridges. Buildings, 13(6), 1517–1517. https://doi.org/10.3390/buildings13061517

Fard, F., & Sadeghi, F. (2024). Development and Utilization of Bridge Data of the United States for Predicting Deck Condition Rating Using Random Forest, XGBoost, and Artificial Neural Network. Remote Sensing, 16(2), 367–367. https://doi.org/10.3390/rs16020367

Li, Q., & Song, Z. (2022). Ensemble-Learning-Based Prediction of Steel Bridge Deck Defect Condition. Applied Sciences, 12(11), 5442. https://doi.org/10.3390/app12115442

Thakkar, K., Rana, A., & Goyal, H. (2023). Fragility analysis of bridge structures: a global perspective & critical review of past & present trends. Advances in Bridge Engineering, 4(1). https://doi.org/10.1186/s43251-023-00089-y

Jeon, J.-S., Sujith Mangalathu, & Lee, S.-Y. (2019). Seismic fragility curves for California  concrete bridges with flared two-column bents. Bulletin of Earthquake Engineering, 17(7), 4299–4319. https://doi.org/10.1007/s10518-019-00621-4

About the author

Yu Tung Hua

Yu Tung has had a strong interest in the field of engineering and programming since a very young age, starting with robotics. During the last 5 years, while competing in multiple Robotics Competitions, he continued to study programming and application in the field of robotics. Yu Tung has studied Machine Learning and Data Science at Columbia University’s summer program and worked with his mentor Dr. Asgari to apply his knowledge on Machine Learning to the real world context of bridge inspection.

The post Analysis of Machine Learning Models for Predicting Bridge Conditions in Massachusetts appeared first on Exploratio Journal.

Abstract

The rapid growth of deep learning has overshadowed simpler, interpretable approaches to image classification. This study presents an alternative method for classifying handwritten digits using a custom feature extraction pipeline applied to the MNIST dataset. Rather than relying on convolutional neural networks, the classifier is built upon engineered features such as loop count, corner detection, symmetry score, bounding box dimensions, and writing direction. After normalization and feature weighting, a Euclidean distance classifier is used to compare new images to per-digit feature averages. The model achieves moderate accuracy and reveals detailed patterns of confusion between similar digits. This interpretable framework offers educational value, and may serve as a lightweight alternative in domains where transparency and explainability are prioritized.

Introduction

In recent years, machine learning and artificial intelligence have revolutionized image classification, with deep neural networks achieving state-of-the-art results across many datasets (Xie & Tu, 2015). However, these models are often criticized for their lack of interpretability, requiring massive computational resources and opaque architectures that hinder trust in decision-making systems (Lundberg & Lee, 2017; Fan et al., 2021). Particularly in educational settings or lightweight applications, simpler alternatives with explainable mechanisms are highly desirable. Interpretability is not just a technical challenge but a critical requirement for deploying AI responsibly, especially when end-users need to understand or contest the model’s decisions (Lipton, 2018).

This study explores a transparent, handcrafted pipeline for digit classification using the MNIST dataset. Instead of relying on pretrained convolutional networks, we develop a modular feature extraction system that emphasizes human-understandable visual traits. The extracted features are numerical descriptors such as bounding box dimensions, center of mass, symmetry, corner and intersection counts, and directional gradients derived from skeletonized images. These features are used in a Euclidean distance classifier that matches new digits to the closest mean feature vector per digit class. The objective is to demonstrate the utility and challenges of building a fully interpretable classification pipeline from scratch.

Dataset and Preprocessing

The experiment utilizes the well-established MNIST dataset, a collection of 70,000 grayscale images of handwritten digits ranging from 0 to 9. Each image is 28×28 pixels in size and is paired with a corresponding digit label. For the purposes of this project, only the test set (10,000 samples) is used, with a configurable limit on how many samples per digit are extracted. Preprocessing begins by parsing the IDX-format image and label files into NumPy arrays, enabling efficient manipulation. The pixel values, originally ranging from 0 to 255, are binarized into black-and-white using a simple thresholding method. This step reduces noise and computational overhead for feature extraction algorithms, particularly those based on geometry and shape. Once binarized, each image is treated as a 2D grid where white pixels represent the strokes of the digit.

To prepare for classification, the dataset is stratified by digit class. A defined number of samples per digit (e.g., 100 images each for digits 0 through 9) are selected and then split into training and test sets. The training set comprises 80% of each digit’s samples, which are used to compute the mean feature vector for that class. The remaining 20% are reserved for evaluation. This consistent stratified sampling ensures that the model is exposed to a balanced and diverse set of handwriting styles while maintaining generalization in testing. This methodology facilitates accurate evaluation of the classifier’s performance using confusion matrices and accuracy metrics.

Feature Extraction Pipeline

Instead of relying on pixel-based convolutional layers or learned representations, this study employs a handcrafted feature extraction pipeline that emphasizes interpretability and simplicity. Each image undergoes a series of geometric and spatial analyses to extract meaningful numerical features. The first feature is the dark pixel count, which reflects the number of active (white) pixels in the binary image and serves as a proxy for stroke density. The center of mass is then calculated by averaging the coordinates of all white pixels, providing insight into digit placement and skew. Bounding box dimensions are computed by identifying the outermost white pixels, then determining the height and width of the smallest rectangle that encloses the digit: useful for distinguishing tall digits like 1 from wide digits like 8 or 0.

Several topological features are also extracted. Loop count is determined using OpenCV’s findContours function, which detects enclosed regions in the digit’s shape. This is especially informative for digits such as 6, 8, and 9, which may contain one or more loops. The corner count uses Harris corner detection applied to the skeletonized image, which minimizes redundant stroke thickness and enhances precision. Intersection count is calculated by analyzing the number of skeletonized pixels that have more than two white neighbors in an 8-connectivity pattern, which indicates points where strokes cross or branch. Both features provide structural detail critical for distinguishing between digits with similar silhouettes, such as 4 and 9.

To further enhance the feature set, a symmetry score is calculated by reflecting the image horizontally and vertically and measuring the pixel overlap between the mirrored and original images. This allows for quantification of both vertical and horizontal symmetry, key traits in digits like 0 and 8. Finally, a directional feature is derived by skeletonizing the digit and computing the gradient flow, which is then subjected to a Fourier transform to isolate the dominant direction and its magnitude. This writing direction analysis is further divided across image quadrants, enabling localized directionality insights. Together, these features provide a compact yet rich representation of each digit, making the classification process interpretable and explainable.

Classification Strategy

Following the extraction of handcrafted features from the binarized and skeletonized images, classification is performed using a simple, interpretable method based on Euclidean distance to class averages. This method was chosen over more complex machine learning models to maintain full transparency in decision-making and provide clear insight into how features influence predictions. The pipeline first computes the average feature vector for each digit class (0 through 9) using the 80% training portion of the dataset. These averages represent the typical geometric and topological characteristics of each digit, such as mean loop count for eights or average horizontal symmetry for zeros.

Each feature is then normalized on a 0 to 1 scale across the dataset to prevent features with larger ranges (e.g., dark pixel count) from disproportionately affecting the Euclidean distance computation. Once normalization is complete, the classifier measures the straight-line (L2) distance between each test image’s feature vector and the average vector for each digit class. The digit whose class average yields the smallest distance is assigned as the predicted label for that image.

To enhance flexibility and allow for fine-tuning of the classification process, the system includes support for feature weighting. Each feature can be scaled by a custom weight during distance calculation, effectively increasing or decreasing its influence on the final prediction. This allows for experimentation with different feature importance values, guided by confusion matrices and performance trends. For instance, if corner count is found to be highly discriminative between certain digits (like 4 and 7), its weight can be increased to reflect its higher utility.

The classifier outputs a confusion matrix that visualizes the true versus predicted labels across all classes, allowing for targeted diagnosis of where misclassifications occur. The overall accuracy is also computed as the percentage of correct predictions on the test set, providing a concise summary of classifier performance. This baseline approach is not only computationally inexpensive and highly interpretable but also lays the groundwork for more advanced ensemble techniques or data-driven weight optimization in future iterations of the project.

Feature Extraction

Feature extraction is the core component of this project, as it forms the foundation upon which classification is based. Instead of using deep learning to automatically learn features from the data, this project focuses on manually engineered features: interpretable numerical attributes that describe various geometric and visual properties of handwritten digits. These features are designed to help distinguish between digit classes by capturing unique patterns, shapes, and structures in the binary images of the digits (Nguyen & Bai, 2020).

The feature extraction pipeline begins by reading grayscale MNIST digit images, which are normalized and binarized so that white pixels (indicating parts of the digit) are treated as foreground and black pixels as background. From there, a series of handcrafted features are computed. One of the most basic yet important features is the total number of white (foreground) pixels, which provides a rough measure of the digit’s thickness or density.

Another key feature is the bounding box area, which captures the size of the smallest rectangle that contains all white pixels of the digit. This is complemented by the center of mass, a two-dimensional coordinate (x, y) that indicates the average location of the white pixels. Together, these features provide spatial information about the digit’s spread and balance.

Corners are detected using a skeletonized version of the digit, followed by the Harris corner detection algorithm. This isolates sharp changes in pixel direction and curvature, giving insight into how angular the digit is. A digit like “4” or “7” tends to have many corners, while “0” or “8” might have fewer. In contrast, intersections are defined as pixels in the skeleton with three or more white neighbors: these typically appear at junctions where strokes branch or cross, such as in the middle of a “4” or “8”.

Loop detection is another critical feature. Loops are identified by performing a flood fill on the background and counting the number of enclosed white regions. This helps distinguish looped digits like “8” or “6” from non-looped ones like “1” or “7”.

Symmetry is calculated in two directions: horizontal and vertical. For horizontal symmetry, the top half of the digit is compared to the flipped bottom half, pixel by pixel. A similar process is used for vertical symmetry. The results are stored as decimal values between 0 and 1, where 1 indicates perfect symmetry. Digits like “8” are highly symmetric, while “5” is less so.

One of the more advanced features is writing direction, which analyzes the dominant flow of pen strokes in the digit. This is estimated by skeletonizing the digit and calculating gradient vectors between connected white pixels. The directions are summarized using angular histograms and averaged over four image quadrants to better capture local directional trends. The result includes both magnitude (the strength of directional flow) and angle (the orientation), which help differentiate digits based on how they are drawn: for example, a “2” may show strong rightward curvature, while a “7” may show sharp vertical and diagonal transitions.

Finally, quarter-based features are also computed by dividing each image into four equal parts. For each quadrant, features like pixel density, average stroke width, and gradient flow are independently measured. This adds a layer of localized spatial analysis and can be particularly helpful when digits share global features but differ in their layout, such as “9” vs “4”.

In summary, the feature extraction process converts raw MNIST image data into a structured vector of interpretable numeric features that describe the digit’s shape, structure, and writing dynamics. These features are exported into a CSV file for use in the classification stage, enabling an interpretable, modular approach to handwritten digit recognition.

Classification Pipeline

Following feature extraction, the digit classification was performed using a distance-based approach. Rather than leveraging external machine learning libraries, a custom K-Nearest Neighbors (KNN)-style classifier was implemented from scratch. The classifier calculates the Euclidean distance between each testing image’s feature vector and the average feature vector of each digit class (0–9) computed from the training set.

Before computing distances, feature values were normalized to a [0,1] range using min-max scaling to avoid bias due to differing numeric scales. Additionally, the system allows for feature weighting, meaning that more important features (e.g., symmetry or loop count) can be assigned higher influence during classification. This modular design supports experimentation with different weighting schemes to optimize accuracy.

The model evaluates its predictions using a confusion matrix, precision metrics, and accuracy scores, enabling quantitative comparison of different feature sets and weighting strategies. Errors are visually inspected to guide iterative refinement of the feature extraction and classification logic.

Dataset and Experimental Setup

For this study, we utilized the MNIST dataset, a well-known benchmark for handwritten digit recognition. The full dataset consists of 70,000 labeled grayscale images of handwritten digits (0–9), each of size 28×28 pixels. Of these, 60,000 images are used for training and 10,000 for testing. However, in our experiment, we implemented a custom CSV-based approach using a limited subset of the MNIST data. Specifically, we processed and extracted features from a fixed number of samples per digit to maintain class balance and control computational complexity.

The system was designed in two main phases: feature extraction and classification. In the feature extraction phase, each image was transformed into a row in a CSV file, where each column represented a manually engineered feature such as pixel count, loop count, symmetry, bounding box, intersection count, etc. In total, over 20 distinct features were extracted and used in the classification phase.

The classification phase was implemented using a custom Euclidean distance classifier. For each digit class (0–9), we computed the average feature vector across the training samples. When a new test image was presented, its features were compared against each of the class averages using Euclidean distance, and the closest class was chosen as the prediction.

Additionally, we introduced feature weighting, allowing specific features to have more or less influence during classification based on their discriminative power. The classification results were tracked and evaluated using accuracy metrics and confusion matrices.

Accuracy and Confusion Matrix

The performance of the classifier was evaluated using a confusion matrix, which visually represents the number of correct and incorrect predictions for each digit. This matrix enabled us to quickly identify which digits were most frequently misclassified and which were most accurately predicted.

The initial model, without any feature weighting or tuning, achieved a moderate classification accuracy, with particularly strong performance on digits like “0” and “1,” which have distinct visual structures. Digits such as “5” and “3” were more commonly confused with each other due to their visual similarity, particularly in cursive or stylized handwriting.

After iteratively tuning the feature weights, we observed a notable improvement in classification performance, especially in reducing confusion between closely related digits. For example, giving more weight to loop count, intersections, and writing direction significantly helped in distinguishing digits like “6,” “8,” and “9.”

At its best configuration, the system reached an overall accuracy of 50.81, with some digits like “1,” “0,” and “8” achieving near-perfect classification. The confusion matrix clearly reflected the impact of feature weighting, with off-diagonal errors shrinking in many digit classes.

To assess which features contributed most effectively to accurate digit classification, a series of weight tuning experiments and feature ablation tests were conducted. These experiments involved systematically adjusting the importance (weight) of individual features during the classification process and observing the resulting changes in accuracy. This approach allowed us to isolate the features with the greatest impact on distinguishing between visually similar digits.

One of the most consistently useful features was pixel count, which reflects the total number of non-background pixels in the digit image. This feature helped differentiate digits with dense strokes, like “8,” from those with minimal writing, such as “1.” Similarly, loop count proved to be highly informative, especially for identifying digits like “8,” which contains two loops, versus digits such as “0,” “6,” or “9,” which have one loop, and digits like “1” or “7,” which have none.

The corner count and intersection count, both derived from the skeletonized version of the digit, played a key role in identifying digits that involve sharp turns or complex branch-like structures. Digits such as “4” and “8” exhibited higher intersection counts due to multiple connecting lines, while digits like “1” and “7” had noticeably fewer corners. However, corner detection was found to be sensitive to image noise and line thickness, and improvements were made by fine-tuning the skeletonization algorithm (Siddiqi & Pizer, 2008).

Another valuable set of features came from analyzing symmetry. Horizontal and vertical symmetry scores helped to recognize digits with more balanced structures, such as “0,” “3,” and “8.” In contrast, digits like “5” and “2” exhibited more asymmetry, which aided in distinguishing them from others. Symmetry-based features were particularly helpful when pixel count or loops were not sufficient on their own.

Finally, one of the most advanced features used was the writing direction, computed from gradient vectors and angular motion across the digit’s skeleton. This feature helped capture the natural drawing flow of digits. For example, “2” typically starts with a curve that swings from the top left to the bottom right, while “5” often features a left-facing arc followed by a vertical drop. By dividing the image into quadrants and calculating directional vectors in each section, we were able to capture both global and local movement trends that further improved digit differentiation.

Overall, the combination of these features, both geometric and dynamic, provided a diverse and interpretable representation of handwritten digits. When these features were strategically weighted, they significantly improved the system’s ability to correctly classify even the most visually ambiguous samples.

Visualization and Debugging

Visualization tools played a crucial role in understanding model behavior. For each test sample, the system could plot the digit image, highlight detected corners, intersections, center of mass, bounding box, and even draw gradient arrows representing writing direction. These visuals helped validate that the feature extractor was working correctly and guided the adjustment of skeletonization, thresholding, and corner detection parameters (Siddiqi & Pizer, 2008).

Skeletonization outputs, in particular, revealed occasional anomalies, such as overly thick or broken lines due to imperfect thresholding. These were later corrected through pre-processing steps and adaptive thinning (Siddiqi & Pizer, 2008).

Summary of Results

Overall, the experimental results demonstrated that an interpretable, feature-based classifier can achieve reasonable performance on a complex task like digit recognition. While not competitive with modern convolutional neural networks (CNNs), this approach provides clear insights into how features contribute to classification. The system’s modularity also makes it easy to extend, optimize, and debug.

The key takeaway is that careful feature engineering and visualization can go a long way in building effective and explainable machine learning models: even for tasks typically reserved for deep learning (Nguyen & Bai, 2020).

Strengths of the Approach

One of the major strengths of this handwritten digit classification system is the interpretability of the features used. Unlike black-box models such as neural networks, which can achieve high accuracy but offer little transparency, this approach relies on intuitive and human-understandable features, such as corner counts, pixel density, symmetry, and writing direction. These features provide not only a basis for classification but also a valuable window into the structure and characteristics of handwritten digits. This makes the model especially useful for educational purposes, explainable AI research, and deployment in systems where traceability of decisions is important.

Additionally, the design emphasizes customization and modular testing. Because each feature is extracted individually and can be visualized, the model allows for fine-grained analysis of each image. Visualization tools, such as skeleton overlays, direction arrows, and bounding boxes, enhance interpretability and assist in identifying both successful and problematic classifications. Moreover, the implementation of feature weighting allows for dynamic tuning of the classifier to prioritize certain distinguishing characteristics for specific digits, significantly improving the robustness of the model.

Limitations

Despite these strengths, the system also has notable limitations. First, feature-based classification is inherently less flexible than deep learning models. While convolutional neural networks can learn thousands of nuanced features from training data, this system relies on a fixed set of manually engineered features. As a result, it may struggle to adapt to unusual handwriting styles or generalize to out-of-distribution samples (Nguyen & Bai, 2020).

Second, while some features such as pixel count and symmetry are stable across digits, others—particularly corner and intersection counts—are sensitive to noise and variations in stroke width. Even after applying skeletonization and refinement techniques, some digits still exhibit spurious feature detections in areas of high stroke density. These inaccuracies can mislead the classifier, especially for digits like “5” and “9” that have subtle structural differences (Siddiqi & Pizer, 2008).

Another challenge arises from the uniform scaling of feature distances. Since all features are normalized to the same scale before Euclidean distance is calculated, differences in feature stability and importance can be overlooked unless explicitly corrected with proper weighting. Without optimized weights, the classifier may be biased toward features that have larger variance or noise, reducing accuracy.

Implications for Future Work

The findings from this system reinforce the idea that simple, interpretable features can still perform competitively on classification tasks when properly designed and tuned. This supports the value of feature engineering in settings where model explainability is critical. Additionally, the ability to visualize the contribution of each feature creates opportunities for human-in-the-loop optimization and error analysis.

These results also open the door to future hybrid approaches. By combining the transparent logic of engineered features with the pattern recognition strength of machine learning models, it may be possible to create hybrid systems that provide both high accuracy and clear explanations. In educational settings, this system can serve as a baseline for teaching students the fundamentals of computer vision and classification without the overhead of deep learning frameworks (Nguyen & Bai, 2020).

Finally, the architecture’s modularity makes it well-suited for experimentation with novel features. Techniques like stroke order estimation, writing speed simulation, or temporal reconstruction of digit drawing paths may offer further improvements. The flexibility and transparency of the current system provide a solid foundation for continued exploration.

Conclusion

This research project presents an interpretable, feature-based approach to handwritten digit classification using the MNIST dataset. Unlike black-box deep learning models, the method relies on clearly defined, explainable features such as pixel count, symmetry, corner and intersection detection, writing direction via Fourier and gradient analysis, and geometric properties like bounding boxes and centers of mass. Through careful engineering and visualization of these features, the system offers valuable insight into how digits can be uniquely characterized by their visual structure (Nguyen & Bai, 2020).

The classifier itself uses a weighted Euclidean distance algorithm to compare new digit samples to statistical averages derived from a training set. This approach allows the model to make data-driven predictions while maintaining transparency and flexibility. Results were visualized via confusion matrices, highlighting both successful classifications and areas where the model struggled, such as differentiating between visually similar digits like 4 and 9. Weight adjustments to the features significantly improved accuracy by emphasizing the most discriminative properties.

One of the key contributions of this work lies in the balance between accuracy and interpretability. While modern deep learning approaches may achieve higher performance metrics, they often sacrifice explainability. This project demonstrates that through methodical feature selection and modular design, it is possible to achieve strong classification performance without abandoning transparency (Nguyen & Bai, 2020).

Looking forward, this framework serves as a robust foundation for further research into human-interpretable machine learning systems. By continuing to refine feature definitions, integrating hybrid techniques, and addressing edge cases through new innovations like stroke order simulation, the model can evolve to rival more complex approaches—while remaining understandable and trustworthy.

In conclusion, this project highlights the potential of interpretable, modular AI systems to achieve meaningful results in computer vision tasks, with wide-ranging applications in education, transparency-focused AI development, and real-world deployment where explainability is paramount.

References

Fan, Y., Zhao, X., Wang, L., Wang, W., Wang, S., & Xu, M. (2021). A review on interpretability of artificial neural networks. Frontiers in Neurorobotics, 15, 752666. https://doi.org/10.3389/fnbot.2021.752666

Lipton, Z. C. (2018). The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery. Communications of the ACM, 61(10), 36–43. https://doi.org/10.1145/3233231

Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems, 30. https://proceedings.neurips.cc/paper_files/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html (Lundberg & Lee, 2017).

Nguyen, T. T., & Bai, L. (2020). A review of traditional and deep learning-based feature descriptors for image classification. Journal of Big Data, 7(1), 1–32. https://link.springer.com/article/10.1186/s40537-020-00327-4 (Nguyen & Bai, 2020).

Siddiqi, K., & Pizer, S. M. (2008). Medial Representations: Mathematics, Algorithms and Applications. Springer. https://link.springer.com/book/10.1007/978-1-4020-8658-3 (Siddiqi & Pizer, 2008).

Xie, S., & Tu, Z. (2015). Holistically-Nested Edge Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 1395–1403). https://openaccess.thecvf.com/content_cvpr_2015/html/Xie_Holistically-Nested_Edge_Detection_2015_CVPR_paper.html (Xie & Tu, 2015).

About the author

Austin Benedicto

Austin is a 12th grade student at the Nichols School in Buffalo, New York. He enjoys studying computer science and robotics in school. Austin has been involved in the FIRST Robotic program at his school for the last 8 years, serving both as a team member and mentoring younger students. He also served as project manager on the coding sub team and has an interest in artificial intelligence.

The post Interpretable Digit Classification using Handcrafted Features and Euclidean Distance appeared first on Exploratio Journal.

Abstract

Optimization has been the central goal in aeromodelling as competitors build custom designs in the hopes of outperforming others. Over the decades of competitive, many different designs appeared, testing the boundaries of the hobby.

This paper provides an aerodynamic analysis of baseline F1D-class configurations and four common design iterations. Each iteration aims to optimize the F1D wing shape by reducing wing loading, increasing lift and reducing drag, and accounting for pitching moment effects. The primary metrics including aspect ratio, lift coefficient at trim, induced drag factor, and aerodynamic efficiency (L/D) were used to estimate the performance of the wing design.

Introduction

Lightweight aeromodelling had a profound impact on aviation. Notable figures such as the Wright Brothers have been inspired by rubber-band powered models, and its presence is felt even to this day as a hobby, and a competitive sport 200 years after its founding. The FAI, Féderation Aeronatique Internationale, also known as the World Air Sports Federation, primarily regulates all the rules and regulations in Indoor Free Flight Competitions, the most popular event being the World Championship Class: F1D.

In indoor free flight, there are many separate events in which competitors can enter their planes into a variety of different events. There are the classic American events, such as EZB, Pennyplane, and Hand Launch Glider, and the internationally recognized ones as well, such as F1D, F1R, and F1M. Each event has its rubber limit and build specifications (wing length/chord, total weight limit, fuselage length), and it’s up to the builder to determine what is best.

Beginners typically start by using building kits then graduate to crafting their own designs and procuring their own materials. Wood and mylar film are the basis of all planes. The balsa wood that’s used for the models are usually cut specifically for that purpose, taking in all the factors of what makes wood solid and usable. A core component when looking for good wood to use is density (measured in lbs./cubic ft.). Lighter density means lighter weight for, which is the best thing for a modeler, but also compromises stability and wood stiffness.

Every plane, no matter how different, generally follows the same structure. The motor stick is the bigger part of the fuselage in the front, while the tail boom is the smaller counterpart in the back. There is always some variation of a wing and a stabilizer, and often a fin/rudder to keep balance. The propeller is arguably the most important part of a plane, and can be either solid wood or covered in film.

Finding the best wood length, height, and width for every part of the plane is the target goal for all modelers. In indoor modeling, it’s very common to maximize the wing and stab area to provide the best lift. Competitors typically maximize these components to the maximum extent allowed by the rules for the class in which they are competing. They’re usually as far as the rules will go for the specific class. The wood size and density making up this always varies, however, and all sorts of variations are found.

In this research project, technical details of indoor free flight will be analyzed to come up with concrete, tangible ways of improving. Indoor Free Flight Models will be analyzed using OpenVSP (Open Vehicle Sketch Pad) by creating a 3D model of its structure, fusing together the fuselage, wing, and tail surfaces. OpenVSP allows users to accurately pinpoint details when modeling, which can be useful for evaluating aerodynamics. Everything from wing camber, aspect ratio, and dihedral angle can be tweaked in an instant to enhance performance within any model. In previous research papers, VSPAERO has demonstrated reliable accuracy in simulating aerodynamic performance and finding related data. (Rosas-Cordova, Santana-Delgado, Hernandez-Alcantara, & Amezquita-Brooks, 2024).

A highlight of OpenVSP is VSPAERO, a tool that allows for low-speed aerodynamic analysis, providing insights into more mathematical aspects of flight and simulating its performance. This virtual testing reduces trial-and-error in real life and saves time and shows the most accurate configurations before committing. In this way, OpenVSP becomes a powerful tool for aeromodelling that traditional methods don’t allow.

Methodology

Four wing configurations were tested using established aerodynamic models: a baseline wing, a cambered airfoil, curved tips, and a tapered-curved wing. Each subsequent iteration introduced modifications informed by aerodynamic theory:

1. Baseline Wing – Standard rectangular planform with moderate aspect ratio.
2. Cambered Airfoil Wing– Increased span and reduced chord to minimize induced drag.
3. Curved Wingtip Wing – Retained rectangular planform but incorporated curved tips.
4. Tapered-Curved Wing – Combined taper and curved tips for optimal lift distribution.

Key aerodynamic terms were derived from standard definitions:

Wing loading: (W/S) = (mg)/S

Induced drag factor: k = 1 / (π e AR)

Trim lift coefficient: CL = (W/S) / q, q = ½ ρ V²

Total drag coefficient: CD = CD0 + k CL²

Efficiency: L/D = CL/CD

Base Iteration

The baseline wing is the simplest, basic design that is tried and tested. It’s a good reference point that’s easy to understand and have predictable lift and drag behavior. The baseline serves as a stable reference point for the other iterations to improve on to see which designs are optimal.

Iteration 1

In all aircraft designs, a large influence on lift distribution is the airfoil of the aircraft. An airfoil directly controls how air flows around the wing, which determines the aircraft’s ability to generate lift, maintain stability, and achieve efficient performance.

In Indoor Free Flight, circular, one-sided airfoils are the standard shape in which most aircraft are made. The camber is measured in percentage of the height in comparison of the chord length. For example, a 10% camber of an 8-inch chord length would be 0.8 inches high at the top of the parabolic arc.

Studies of low Reynolds number aerodynamics confirm that low cambered airfoils are highly effective in slow-flight regimes, where laminar separation bubbles and transition strongly affect lift and drag performance (Selig, Deters, & Williamson, 2011).

The standard airfoil in Indoor Free Flight generally uses anywhere from a 3% – 5% camber. Iteration 1 attempts to test if there’s any logical foundation in the already established standard in airfoil shape.

Lift Distribution

Shown is a graph of lift distribution for each wing: the baseline, and the one with 7% camber. The graph is measured in CL/cref (coefficient of lift over sectional wing area) versus the span location of the wing. At first glance, the lift values for the higher cambered wing is much higher than the baseline, wing, by as much as 20%.

Aerodynamic Trends

A deeper look at each of the 4 central values of the two iterations allows a deeper analysis to it. All the graphs measure important traits of the aircraft in flight. Alpha values correspond to the angle of attack of the flight: (ex. α =1 =1 degree)

Lift (CL): increases roughly linearly with α for both camber values. 7% camber gives higher CL across the range (so more lift at the same α).

Drag (CD): increases with α for both; 7% camber has noticeably higher CD at every α.

Pitching moment (Cm): more negative (larger nose-down moment) for the 7% camber case — i.e., higher camber produces a larger nose-down pitching tendency.

L/D (efficiency): decreases with α for both. The 4% camber case has higher L/D throughout the entire α range and therefore is the more aerodynamically efficient option.

Interpretation

F1D performance is all about optimizing L/D ratios, in order to maximize time aloft. Although the 7% camber provides significantly more lift, the additional drag and larger nose- down movements significantly reduces the overall efficiency and movement. This supports how Swanson and Isaac (2010) found that moderate camber ratios enhance aerodynamic efficiency and extend flight duration in low Reynolds number wings. Since F1D is all about endurance, the 4% would be best suited for indoor free flight. It provides higher L/D (efficiency) overall throughout all alpha values, lower drag, and is much easier to trim.

The 7% camber provides more lift, but the extra drag and handling isn’t worth it in the case of F1D. It might be useful, however, when higher amounts of pure lift are needed, due to a heavier build or higher wing loading. In events like F1M or Pennyplane, this could prove essential for pure brute force and strength.

Iteration 2

A major factor in is controlling the lift distribution in the wing. To do this, one of the most common methods in F1D adding curved wingtips, which help reduce induced drag by smoothing the vortices that form at the tip. In indoor free flight, tip design is especially important since the aircraft flies at very low Reynolds numbers where drag penalties are magnified. This reflects the trade-off observed in many low-Reynolds-number aircraft studies, where stability improvements may come at efficiency losses. Research has shown that tip modifications, such as curvature or taper, can reduce induced drag by smoothing vortex formation, which is especially important at Reynolds numbers typical of F1D designs (Penchev et al. 28). The curved wingtip design (Iteration 2) kept all characteristics as the baseline, with only smoothly rounded tips being different. The goal was to reduce drag while keeping the construction simple and lightweight.

Lift Distribution

The distribution plot shows that the curved wingtips reduce peak loading near the tip, creating a smoother, more elliptical distribution compared to the flat baseline. However, the overall lift magnitude is lower than both the 4% camber baseline and the 7% camber airfoil. This suggests that while drag is improved, maximum lift capacity is compromised.

Aerodynamic Trends

Lift (CL): Iteration 2 generates less lift than both 4% and 7% cambered wings at all α values.

Drag (CD): Slightly lower than 7% camber, but marginally higher than 4% baseline across most α.

Pitching Moment (Cm): Similar to baseline, with only a minor increase in nose-down tendency.

Efficiency (L/D): Iteration 2 underperforms baseline across the α range, staying close to but consistently below the 4% camber wing.

Interpretation

Curved wingtips in Iteration 2 do succeed in producing a better lift distribution, reducing concentrated tip vortices. However, the cost is a reduction in total lift, and the aerodynamic efficiency (L/D) does not surpass the 4% camber baseline. This makes Iteration 2 less competitive for endurance-focused F1D performance, where high L/D is essential. While the iteration provides valuable insight into tip shaping, the tradeoff shows that tip curvature alone is not sufficient to improve overall efficiency. It may, however, become more effective in combination with taper (Iteration 3) or when optimizing for stability rather than pure endurance. Iteration 2 is an ideal change for those looking to focus more on stability rather than lift.

Iteration 3

Iteration 3 is another aerodynamic improvement, in addition to curved tips is the addition of rounded wingtips.

The goal of this modification is to achieve a smoother, more elliptical lift distribution across the span, thereby reducing induced drag while still retaining good lift capacity. This design also reduces the strength of wingtip vortices, a critical source of drag for low-Reynolds- number aircraft like F1D models.

The rounded wingtip design (Iteration 3) modifies the baseline rectangular planform by gradually tapering and rounding the tips.

Lift Distribution

The comparison shows that the rounded wingtips push the distribution closer to an elliptical shape than the baseline. At all α levels, the rounded tips exhibit smoother curves and reduced sharpness near the tips. Total lift is slightly higher but similar to the baseline, which shows that these changes and optimizations to the wing don’t alter lift as much as other factors.

Aerodynamic Trends

Lift (CL): Rounded tips provide consistently higher CL than the baseline across the α range, indicating more efficient lift generation.

Drag (CD): Slightly higher than baseline at all α values, suggesting that tip shaping introduces small penalties in skin friction or interference drag.

Pitching Moment (Cm): More negative than baseline, meaning the aircraft tends toward a stronger nose-down pitching moment, though still manageable for trimming.

Efficiency (L/D): At lower α (2°–4°), Iteration 3 outperforms the baseline slightly, reaching higher peak L/D values. However, as α increases beyond ~6°, efficiency drops more quickly, converging toward baseline performance.

Interpretation

Iteration 3 demonstrates that rounded/tapered tips can improve lift distribution and provide a modest increase in aerodynamic efficiency at lower α values, where F1D aircraft often operate. The design successfully reduces tip vortex intensity while enhancing total lift. However, the improvement comes at the cost of higher pitching moments and slightly increased drag. Such findings are consistent with prior studies showing that tip shaping contributes to closer-to- elliptical lift distribution, which is beneficial for stability and moderate efficiency improvements. (Ananda, Selig, and Deters 2015)

For endurance flights where optimal trimming and efficiency at moderate α is most critical, rounded wingtips offer a practical improvement over both the baseline and the simple curved tip of Iteration 2. This makes Iteration 3 a strong candidate for use in competition settings.

Iteration 4

Building on the rounded/tapered tips of Iteration 3, Iteration 4 further refines the wing shape by adopting fully rounded wings with increased taper. The intent of this modification was to push the lift distribution as close as possible to an ideal elliptical shape while still retaining strong lift performance across the entire wing. At the low Reynolds numbers of F1D models, this approach is expected to maximize stability and control.

Lift Distribution

The spanwise comparison shows that Iteration 4 has a similar lift profile to all designs tested.

Aerodynamic Trends (compared to baseline)

Lift (CL): Iteration 4 delivers slightly lower CL than the baseline across all α values and maintains consistency with Iteration 3 at low α.

Drag (CD): Marginally more drag than Iteration 3 at low α, but and higher α compared to the baseline

Pitching Moment (Cm): Much less negative than baseline but similar to Iteration 3, indicating less variable pitching moment and stability.

Efficiency (L/D): Iteration 4 shows the lowest L/D at low to moderate α (2°–5°) of all configurations, outperforming both the baseline and Iteration 3. At higher α the efficiency tapers off but remains competitive.

Interpretation

Iteration 4 demonstrates that combining taper with fully rounded tips produces less span wise load albeit low aerodynamic efficiency at the low α values typical of endurance flight, but offers high stability and control. This iteration achieves the intended goal of being able to adjust and control, and is often used in European designs.

Conclusion

The study demonstrated that optimizing F1D wing performance requires a careful balance of lift, drag, and stability rather than the pursuit of a single aerodynamic advantage. Increasing camber resulted in greater lift, but also introduced higher drag and trimming difficulties, which ultimately reduced overall efficiency. Wingtip modifications like curvature and taper, shifted lift distribution toward a more elliptical form and improved stability, but offered tradeoffs in other factors.

References

Ananda, G. K., Selig, M., & Deters, S. (2015). Influence of wing tip shape on lift and drag at low Reynolds numbers. Aerospace Science and Technology.

NASA. (2022). Preliminary airfoil design for low Reynolds numbers. NASA Technical Reports Server.

Penchev, S., et al. (2025). A wind tunnel study of the aerodynamic characteristics of wings with curved trailing-edge wingtips at low Reynolds numbers.

Rosas-Cordova, J., Santana-Delgado, C., Hernandez-Alcantara, D., & Amezquita-Brooks, L. (2024). Validation of VSPAERO for basic wing simulation. Research in Mechanics and Numerical Innovation.

Selig, M. S., Deters, R. W., & Williamson, G. A. (2011). Wind tunnel testing airfoils at low Reynolds numbers.

Swanson, T., & Isaac, K. M. (2010). Planform and camber effects on the aerodynamics of low-Reynolds-number wings. Journal of Aircraft.

About the author

Preston Le

Preston is a high school senior passionate about engineering and mathematics. One of his favorite hobbies is indoor and outdoor aeromodeling, where he holds a youth record in the F1R category. In his spare time, he plays classical piano and studies music theory.

The post Aeromodelling Optimization: An Analysis on Wing Design appeared first on Exploratio Journal.

Abstract

Everyone’s talking about Al, but no one’s talking about its hidden cost. We think of Al as this invisible “cloud,” but it’s built on a massive network of data centers guzzling energy, water, and hardware. The stats are wild: a single data center can use as much electricity as a whole city, and training one AI model can burn through enough energy to power over 100 homes for a year. On top of that, these places use billions of gallons of water for cooling, a serious problem in a world dealing with droughts. The tech gets old super fast, creating a mountain of e-waste five times faster than we can recycle it.

This paper exposes the shady side of Al, where companies hide their environmental impact behind outdated metrics and a total lack of transparency. But it’s not all bad news. We’re also exploring how we can fix this mess with cool tech like “Green Al,” new policies, and a shift to renewable energy. It’s time to make Al not just smart but sustainable.

Introduction

Artificial intelligence has become a powerful technology in our world, but its true environmental cost is often never shown. While we interact with AI through the concept of cloud computing, this technology is built on a physical network of large data centers that use massive amounts of resources. As AI models have grown, a significant problem has emerged: developers are not required to disclose the energy or carbon footprint of their systems. This lack of transparency makes it difficult for anyone to understand the full environmental impact. To address this gap, this paper proposes a simple “nutrition label” for AI models, which is called the Model Carbon-Disclosure Standard (MCDS). This label would show key metrics like the energy consumed and carbon generated, making the environmental cost clear and comparable.

The idea of a disclosure framework is not new. The Greenhouse Gas (GHG) Protocol gives a global way to measure emissions, and groups like the Carbon Disclosure Project (CDP) encourage companies and governments to share their environmental impacts with investors and the public. As the market demands more of this transparency, it becomes increasingly important for a company’s financial and competitive standing.

The environmental footprint of computing has a history that has evolved over time. In the early days, concerns were mostly focused on the toxic byproducts from manufacturing devices. With the growth of large data and cloud computing, attention turned to the heavy energy and water consumption of data centers, real buildings that the term “cloud” often hides. Globally, the number of data centers has exploded in just the past decade. (“Measuring AI’s Energy/Environmental Footprint to Access Impacts, ” 2025)

The emergence of AI has dramatically accelerated these trends. AI tasks are far more energy-intensive than traditional computing. Training large AI models, for instance, requires a tremendous amount of energy, and even a single question to a chatbot like ChatGPT can use far more electricity than a normal Google search. Since this energy demand is so intense and unpredictable, AI is considered an environmental risk whose full impact is difficult to measure. This paper will explore these challenges in detail and outline a path toward a more transparent and sustainable future for AI.

Current Status

Artificial intelligence, often thought of as an invisible “cloud, ” is, in reality, built upon a very real and physical foundation of massive data centers. As AI technology becomes more advanced and common in our daily lives, its environmental footprint, in the form of energy, water, and waste, is growing at a rapid pace, creating significant new challenges.

The Energy Appetite of AI and Data Centers

The energy consumption of these data centers, which had been fairly steady for many years, has recently surged because of the boom in AI. In 2023, data centers used about 4.4% of all electricity in the United States, a number that is projected to double or even triple by 2028, reaching up to 12% of the nation’s total electricity demand. (Increase in Electricity Demand from Data Centers, 2024) To put that into perspective, by 2030, a large data center could use as much electricity as an entire city, and globally, data center electricity use is expected to more than double by the end of the decade. This unexpected and fast growth is putting pressure on power grids, and sometimes utilities keep old, polluting coal plants running longer to meet the demand.

AI tasks use far more energy than traditional computing. Training a large AI model like GPT-3, for instance, consumed enough energy to power about 120 average U.S. homes for a full year. This process also generated a carbon footprint equal to the yearly emissions of 123 gasoline-powered cars. For more common uses, a single question to a chatbot like ChatGPT can use nearly 10 times the electricity of a normal Google search. When you get into more complex tasks, the energy use skyrockets. Creating a five-second AI video, for example, can use about as much electricity as keeping a TV on all day.

Despite these impacts, the industry’s environmental footprint is very opaque due to a lack of clear and consistent reporting. For example, companies often use outdated metrics like Power Usage Effectiveness (PUE) that only measure a facility’s efficiency and not how efficiently the actual computer hardware is working. This means a data center can seem efficient on paper while still being very wasteful. A major issue is how companies report their emissions under the Greenhouse Gas (GHG) Protocol, a global framework for measuring emissions, which are categorized into Scope 1 (direct), Scope 2 (from purchased electricity), and Scope 3 (from the value chain, like manufacturing). The purchase of renewable energy credits can hide a company’s real emissions and make them seem much greener than it is. One analysis of a major company’s 2022 data found that while its publicly reported emissions were only 273 metric tons of carbon, its actual emissions from the local power grid were over 3.8 million metric tons, a difference of more than 19,000 times. This gap in reporting creates a situation where companies are not encouraged to focus on energy efficiency because there are no clear standards to hold them accountable.

One powerful solution to this challenge is a proactive approach to energy sourcing. The Massachusetts Green High Performance Computing Center in Holyoke is an excellent example. This data center is primarily powered by a nearby hydroelectric station, which creates a direct and reliable source of clean energy from the very beginning, rather than relying on a power grid that is heavily dependent on fossil fuels.

AI’s Thirst for Water

AI’s energy demands also create a huge need for water. The powerful computer chips in data centers generate enormous heat, and water is used for cooling to prevent them from breaking down. A single large data center can consume up to 5 million gallons of water every day, which is enough to supply a town of 10,000 to 50,000 people. To show this on a larger scale, Google’s global data centers consumed about 4.3 billion gallons of water in one year, enough to give every person in the United States about 13 gallons. This consumption becomes especially serious in water-stressed regions. In Texas, where the state has been dealing with a severe drought, data centers are projected to use nearly 400 billion gallons of water by 2030, which represents almost 6.6% of the state’s total water usage. While residents are asked to cut back, these new facilities use millions of gallons daily with little public notice. (Texas Data Centers Use 50 Billion Gallons of Water, 2025) As one water policy analyst noted, there is often no requirement for data centers to talk to communities about their water use, which hides much of their environmental impact.

Some innovators are tackling this problem head-on. In Finland, a country with a cold climate, a data center has been designed to operate without traditional mechanical cooling systems. Instead, it uses a system that relies on cold outdoor air or even seawater from the Baltic Sea to cool its servers, which dramatically reduces its energy and water footprint. Even more impressively, the waste heat from the servers is captured and sent to a local district heating network to warm nearby homes and businesses. This creates a win-win situation where the data center not only reduces its own impact but also helps the community use less fossil fuels for heating.

The E-waste Challenge

The rapid pace of AI innovation has created a competition for faster, more powerful computer hardware. This constant cycle of upgrades and replacements is creating a global electronic waste (e-waste) problem. The manufacturing of this hardware also has its own carbon emissions, which are part of a company’s Scope 3 emissions under the GHG Protocol.

According to the U.N., global e-waste reached a record 62 million metric tons in 2022, equal to the weight of more than 150 Empire State Buildings. This problem is getting worse, as e-waste is growing nearly five times faster than recycling efforts can keep up. The total amount of e-waste is projected to grow to 82 million tons by 2030. (ewastemonitor, 2024) In a high-usage scenario, the spread of large language models alone is expected to generate an extra 2.5 million tons of e-waste annually by 2030. This waste is especially dangerous because it contains harmful materials like lead and mercury that can damage both human health and the environment if not properly handled.

Adding to the problem is a significant lack of transparency. Only about a quarter of data center operators track what happens to their retired hardware, and even fewer measure the e-waste they generate. This data gap means that tons of valuable and hazardous equipment often end up in landfills, and there is little accountability or incentive for companies to improve their practices.

Discussion

While the environmental challenges of AI are significant, a new movement is underway to build a more sustainable future for this technology. This effort involves a combination of smarter technology, better operational strategies, and new rules to guide the industry.

Innovations in “Green AI”

The movement toward “Green AI” starts with making the technology itself more efficient. A key part of this is model optimization, where techniques like pruning (removing unnecessary parts of a model) and knowledge distillation (transferring learning from a large model to a smaller one) dramatically reduce the energy needed for AI workloads. For example, researchers have developed tools that can predict a model’s accuracy early in its training, which can save up to 80% of the computing power that would have otherwise been used on a less effective model.

Developers are also creating new, energy-efficient hardware. Beyond traditional GPUs, new types of chips, such as neuromorphic and optical processors, are being designed to run AI tasks with far less power. Additionally, a method called “power capping” can be used to limit the electricity sent to processors, which can cut energy use by about 20% with no loss in performance.

Smarter operational strategies are also key. This includes scheduling large computing tasks to run at night when energy demand on the grid is low, or distributing workloads across different time zones to use power when renewable energy like wind and solar is most available. It also means using simpler AI models when they are sufficient for a task, such as a model that runs locally on a user’s device instead of one in a massive data center. (AI Has High Data Center Energy Costs — but There Are Solutions, 2025)

The Role of Renewable Energy and Advanced Cooling

To power the vast data centers that form the backbone of AI, a global shift to clean energy is crucial. Experts predict that by 2030, about half of the electricity used by data centers will come from renewable sources. (Energy Supply for AI, 2025) AI itself can even assist in this transition by forecasting how much renewable energy will be produced at any given time, allowing for better energy management.

Cooling is a great part of a data center’s energy and water consumption, so new solutions are arising here as well. Advanced cooling systems like liquid cooling are thousands of times more efficient at removing heat than air, allowing for more powerful hardware in a smaller space while using less energy. Another smart strategy is placing data centers in naturally cold climates, like in Finland, to use outside air or cold seawater for “free cooling” .

Policy and Standardization Efforts

For these solutions to have a global impact, they must be backed by clear rules and standards. Over 190 countries have agreed on guidelines for ethical AI, including its environmental aspects. Both the European Union and the United States have introduced legislation aimed at managing AI’s environmental footprint. A U.S. Executive Order, for instance, directs the Department of Energy to create reporting requirements for data centers that cover a technology’s full lifecycle, from manufacturing to disposal.

These efforts aim to create new, transparent metrics for the industry. The “AI Energy Score” is one idea, which is a simple, star-based rating system to show how energy-efficient an AI model is for a specific task. The International Organization for Standardization (ISO) is also preparing new standards for “sustainable AI” that will cover energy, water, and materials. The goal of these policies is to require developers and companies to measure and publicly share their environmental impacts and to integrate these metrics into existing sustainability reports like the Greenhouse Gas (GHG) Protocol.

Conclusion

The rapid growth of AI is driving a significant surge in demand for energy, water, and hardware; however, our ability to measure and manage this impact is often hindered by a lack of transparency. The industry has frequently used outdated metrics or misleading reporting, which makes it hard to hold companies accountable for their actual environmental footprint. This has created a situation where companies are not strongly motivated to make their models more energy-efficient because there are no clear standards to do so. AI affects many areas, including energy use, water supplies, and e-waste, and these impacts grow as AI models run constantly and get upgraded quickly.

The path forward requires an integrated approach that combines new technology with clear policy. We can make AI more sustainable by using smarter model designs, more efficient hardware, and innovative cooling methods. At the same time, policies that require companies to provide clear and honest information about AI’s environmental impact are essential to create balanced conditions and hold companies accountable. This combined effort is vital to ensure that the AI revolution is not only powerful and transformative but also sustainable for our planet and future generations.

References

AI has high data center energy costs—But there are solutions. (2025, January 7). https://mitsloan.mit.edu/ideas-made-to-matter/ai-has-high-data-center-energy-costs-ther e-are-solutions

Chen, S. (2025). Data centres will use twice as much energy by 2030. Nature. https://doi.org/10.1038/d41586-025-01113-z

Electronic Waste Rising Five Times Faster than Documented E-waste Recycling. (2024). https://unitar.org/about/news-stories/press/global-e-waste-monitor-2024-electronic-waste -rising-five-times-faster-documented-e-waste-recycling

Energy supply for AI. (2025). IEA. https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai ewastemonitor. (2024, March 20). The Global E-waste Monitor. E-Waste Monitor. https://ewastemonitor.info/the-global-e-waste-monitor-2024/

GHG Protocol Scopes and Emissions Across the Value Chain. (2024, February 6). Jeff Winter. https://www.jeffwinterinsights.com/insights/scope-emissions-overview

Greenhouse Gas Protocol. (2025). https://ghgprotocol.org/

How much water does AI consume? (2023, November 30). https://oecd.ai/en/wonk/how-much-water-does-ai-consume

Increase in Electricity Demand from Data Centers. (2024, December 20). Energy.Gov. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-d emand-data-centers

Measuring AI’s Energy/Environmental Footprint to Access Impacts. (2025). Federation of American Scientists. https://fas.org/publication/measuring-and-standardizing-ais-energy-footprint/

Texas data centers use 50 billion gallons of water. (2025). Newsweek. https://www.newsweek.com/texas-data-center-water-artificial-intelligence-2107500

The Importance of FLOPS. (2025). https://www.lenovo.com/us/en/glossary/flops/

Vinuesa, R., Azizpour, H., Leite, I., Balaam, M., Dignum, V., Domisch, S., Felländer, A., Langhans, S. D., Tegmark, M., & Fuso Nerini, F. (2020). The role of artificial intelligence in achieving the Sustainable Development Goals. Nature Communications, 11(1), 233. https://doi.org/10.1038/s41467-019-14108-y

What is Power Usage Effectiveness? (2025). https://www.www.digitalrealty.com/resources/articles/what-is-power-usage-effectiveness ?t=1755978324820?latest

About the author

Stavros Farsedakis

Stavros’ academic interests center on computer science and artificial intelligence, especially exploring how technology impacts the environment. Outside the classroom, he enjoys coding projects and researching ways to make AI more sustainable and efficient.

The post Uncover The Hidden Environmental Cost of AI appeared first on Exploratio Journal.

Introduction

Recreating Human Lung Function on a Chip: Progress in Biomaterials, Cellular Models, and Environmental Signals Introduction Respiratory diseases impose a massive global health burden, with conditions such as COPD, lung cancer, and infections (e.g. influenza, COVID-19) contributing to respiratory failure as one of the leading causes of mortality worldwide (Soriano et al.). Thus, the rapid discovery of effective new therapeutics for lung diseases is an urgent priority. However, drug development in the respiratory field remains hindered by the difficulty of accurately modeling lung physiology in the preclinical system. Conventional preclinical models, such as 2D, 3D and animal models, often do not represent the cellular architecture, mechanical stressors, and physiological responses of the human lung, leading to poor predictive power for human outcomes (Jensen and Teng).

In the past decade, organ-on-a-chip (OOC) technology has emerged as a promising solution to address these limitations by recreating human organ physiology within microengineered devices. An organ-on-a-chip is essentially a microfluidic cell culture device containing tiny, continuously perfused chambers lined with living cells that replicates key aspects of an organ’s structural and functional environment (Bhatia and Ingber). First conceptualized and demonstrated in the early 2000s, OOCs employ microfabrication techniques to reproduce physiological interactions among multiple cell types, perfusion of fluids, mechanical stimuli, and relevant extracellular matrix components, providing realistic in vitro microenvironments (Huh et al.).

Specifically, lung-on-a-chip (LOC) devices have been extensively developed to address respiratory research gaps. The LOC system by Huh et al in 2010 replicated the human alveolar-capillary interface through microfluidic channels separated by a flexible, porous membrane lined by alveolar epithelial and pulmonary endothelial cells (Huh et al.). This system uniquely incorporated cyclic mechanical strain to simulate breathing motions, resulting in physiological responses such as surfactant production, barrier integrity, and inflammatory reactions similar to in vivo conditions. Since then, numerous advancements in LOC technology have incorporated diverse cell types, patient-specific cells (including induced pluripotent stem cells), improved biomimetic materials (such as biological membranes and hydrogels), and more sophisticated stimuli and signals (O. Stucki et al.)(Zamprogno et al.).

Given these advantages, LOC technology represents a shift in preclinical research, promising enhanced accuracy, higher predictive value for clinical outcomes, reduction in animal experimentation, and substantial acceleration of the drug development process (Francis et al.). However, ongoing efforts are required to address current challenges such as standardization, integration with other organ systems, optimising the biomaterials that are used, and physiological relevance of these systems, paving the way for widespread adoption of LOC platforms in biomedical research. This paper will focus on evaluating the core components of the tissue engineering triad within lung-on-a-chip systems: cells, biomaterials, and signals, with an emphasis on how each contributes to physiological relevance of LOC systems. By analyzing these three pillars, it will highlight the strengths and limitations of current models.

Types of Cells

Primary Cells

Primary lung cells originate from human lung tissue that obtains its cells from donor lungs and surgical resections. The relevant primary cells for lung-on-chip research consist of alveolar epithelial cells (type I and II pneumocytes) and airway epithelial cells (bronchial or small airway cells) along with lung microvascular endothelial cells and fibroblasts (Campillo et al.). Several lung-on-chip systems employ primary human cells to establish the alveolar-capillary barrier structure (Stucki et al.). The majority of current research involving microfluidic alveolar-on-chip systems depends on primary human lung cells as their epithelial source (owing to their superior physiological relevance) (Wang et al.). The specialized traits of native tissue remain present in primary lung cells. Comparisons between A549 cell lines against primary cells show that cells from cell lines are far less sensitive to cyclic strain compared to primary cells, suggesting that they do not accurately represent the response to mechanical cues seen in native cells of the lung (Lagowala et al.). Furthermore, primary cells can also be used to create personalized OOCs.

The main disadvantages of primary cells include their brief lifespan as well as donor-specific characteristics. The specialized nature of lung epithelial cells leads to cultured cells losing their ability to reproduce (Min et al.) The use of primary cells in experiments results in inconsistent outcomes because different donors exhibit different phenotypes. Furthermore, when available, yields are low and the cells usually cannot be expanded long-term – most primary lung cells have a finite number of population (Ronaldson-Bouchard and Vunjak-Novakovic). Another challenge specific to pneumocytes found in the lung is that in culture, primary ATII cells quickly differentiate into ATI cells after a couple days, making it harder to study ATII cells and more importantly, accurately model the alveolar epithelium in LOC models (Lagowala et al.).

Immortalized Cell Lines

Immortalized Cell Lines are continuously propagating cells often derived from tumors (e.g. A549 alveolar carcinoma cells) or engineered by viral genes (e.g. SV40-transformed bronchial cells) (Campillo et al.)(Min et al.). They proliferate indefinitely and form relatively homogeneous populations (Min et al.). This allows them to be more cost-effective as they do not require an ongoing donor tissue supply, thus avoiding the ethical and economic hurdles of using animal or human tissue. Furthermore, because cell lines can be reproduced endlessly, they are ideal for high-throughput studies (Kaur and Dufour). For instance, A549 and H441 cell lines have been widely used to model alveolar epithelium in lung-on-chip studies, showing robust growth and injury-response profiles (Campillo et al.). Despite not being native to human tissue, studies have shown that they can be manipulated to accurately mimic in-vivo conditions. Cell lines such as NCI-H441 (alveolar type II cells) and NCI-H1703 (alveolar type I cells) have been used in alveolus-on-a-chip technology to successfully form a tight epithelium that secretes surfactant (Kang et al.). Such models, if using native tissue, do not last a long time due to the fact that ACII cells easily differentiate into ACI cells, making it harder to observe the effect of surfactant on alveolus-on-a-chip systems (Lagowala et al.).

However, their physiological relevance is limited. Being cancer-derived or genetically altered, they often fail to fully perform specialized functions of normal lung cells. Prolonged culturing can further drift their phenotype, compounding reproducibility issues (Kaur and Dufour). During immortalization, cells undergo genetic changes that can alter their native physiology. Thus, while immortalized lines are convenient and economical, their divergence from in vivo cell behavior is a significant drawback. An example of this is seen in A549 cell lines which demonstrate alveolar-like characteristics but they fail to produce sufficient phosphatidylglycerol which is a crucial surfactant lipid and they show different ion transport patterns than actual alveolar type II cells (Lagowala et al.). Furthermore, as explained before, they are not as sensitive to biomechanical cues as primary cells. Lastly, immortalised cells lines may also drift genetically over time and are far more prone to contamination (Kaur and Dufour).

Stem Cells

Stem cell technologies offer human cells as an alternative resource for constructing organ-on-chip models. The reprogramming of adult cells into embryonic-like cells through induced pluripotent stem cells (iPSCs) enables unlimited expansion followed by differentiation into multiple lung cell types (Ronaldson-Bouchard and Vunjak-Novakovic). Scientists use developmental growth factors (as demonstrated by Takahashi and Yamanaka in 2006) to transform iPSCs into specialised cells, for example in LOCs, iPSCs can differentiate into alveolar type II–like cells and airway epithelial cells which mimic lung organogenesis (Takahashi and Yamanaka).

A key advantage of iPSC technology is that it avoids the continuous need for donor tissue. Like immortalised cell lines, it can be virtually expanded without limit (Kaur and Dufour). In addition, iPSCs can be differentiated into multiple cell types. The lung is composed of approximately 40 different cell types, thus allowing iPSCs to conveniently allow the assembly of complex co-culture chips where all cell types share the same genetic source (Calvert and Ryan (Firth)). Biologically, stem cell-derived lung cells can capture certain developmental or disease-relevant states that primary cells do not. For example, iPSC-derived alveolar cells can represent neonatal lung cells or an intermediate “transitional” phenotype implicated in lung fibrosis (Tamai et al.).

Differentiating iPSCs into adult, functional lung cells is a time-consuming and complex process as it requires step-by-step exposure to multiple growth factors (activin A, BMP/Wnt inhibitors, FGF10, dexamethasone, etc.) which can take 4-8 weeks to fully recapitulate embryonic lung development in a dish (Lagowala et al.). Even after this effort, the resulting cells often resemble fetal or immature lung cells rather than fully mature adult cells. For instance, iPSC-derived cells typically exhibit an incomplete functional maturation compared to primary ATII cells. They also tend to lose some epigenetic marks of the donor’s age and environment during reprogramming, which could erase certain disease-related signatures (Ronaldson-Bouchard and Vunjak-Novakovic).

Materials

Polydimethylsiloxane (PDMS)

Lab-on-a-chip prototyping standardizes PDMS, silicone elastomer material, because of its low fabrication costs and easy manufacturing process (Elveflow). Soft lithography techniques create precise replicas of micron-scale features by casting PDMS into microfluidic channels . PDMS devices have become fundamental tools for cell biology and organ-on-chip research because they include thin PDMS membranes which duplicate breathing motions. (Carvell et al.) In fact, the first lung-on-a-chip developed by Huh et al. (2010) used a 10 μm porous PDMS membrane to recreate breathing motions, enabling co-culture of alveolar and endothelial layers at an air–liquid interface with cyclic strain that reproduced lung physiology (Huh et al.). PDMS-based chips become suitable for cell culture through surface treatments like plasma oxidation and protein coatings that create conditions for robust cell growth similar to traditional cultureware (Tanyeri and Tay).

The optical transparency of PDMS allows direct observation of cells and fluids and it possesses elastic properties with a large Young’s modulus which enables flexible microvalve and peristaltic pump integration (Elveflow). Stucki et al. (2018) demonstrated how PDMS’s elasticity and transparency enabled the fabrication of an ultrathin 3.5 μm PDMS membrane that supported primary alveolar cells under strain while preserving physiological functions (Stucki et al.). PDMS enables gas exchange between cell chambers and their environment because oxygen and CO₂ pass through the material with high permeability (O₂ permeability ~3.4×10^ –6 cm^2/s) (Tanyeri and Tay). The oxygen reservoir properties of thick PDMS walls sustain cellular oxygenation during extended perfusion operations, as shown in long-term lung-on-chip cultures where PDMS supported weeks of epithelial–endothelial co-culture under cyclic stretch. It exhibits minimal curing shrinkage while forming permanent bonds to glass substrates and itself which facilitates multi-layer device assembly (Tanyeri and Tay). The material exhibits biocompatibility and allows surface modification through plasma oxidation to create temporary hydrophilic surfaces that enhance cell attachment (Cao et al.). The combination of PDMS properties with quick prototyping workflows creates a versatile material for designing microfluidic systems through multiple iterations.

The material remains popular despite several significant drawbacks. The inherent hydrophobic nature of PDMS makes it absorb small hydrophobic molecules present in aqueous solutions (Alghannam et al.). Zamprogno et al. (2021) reported that PDMS absorbed drugs and fluorescent dyes in lung-chip experiments, altering effective concentrations and skewing pharmacological readouts (Zamprogno et al.). Experimental results become biased because PDMS absorbs or adsorbs drug compounds and signaling molecules that are present in media. PDMS exhibits natural protein binding behavior that may lead to surface fouling unless surface treatment or lipophilic coating methods are applied (Alghannam et al.).Flow pressure causes soft PDMS channels to deform because of their elastic properties which results in dimensional changes (Cao et al.). For example, Stucki et al. (2018) noted that cyclic strain on thin PDMS membranes risked mechanical fatigue and buckling when pressure was not carefully controlled (Stucki et al.). The material properties of PDMS make it difficult to perform applications that require precise fluid control because channel deformation occurs under flow pressure conditions. Most organic solvents damage or swell PDMS which restricts its compatibility with solvent-based reagents. The combination of permeability and flexibility in PDMS leads to negative effects in certain situations since evaporation creates bubbles and flexible walls produce inconsistent results (Elveflow). Zamprogno et al. (2021) further highlighted that the stiffness of PDMS is non-physiological compared to the compliant alveolar basement membrane, which in some cases induced fibrotic-like responses in cultured lung cells (Zamprogno et al.). The need for additional materials arose due to the development of advanced solutions for applications that require more complex solutions.

ECM-Derived materials

The terms “ECM-derived” describe biomaterials which originate from tissue extracellular matrix (ECM) and include both structural proteins along with other macromolecules where cells naturally reside. Materials in this class include collagen (particularly Type I collagen), fibrin derived from fibrinogen blood protein, laminin-rich Matrigel (a basement membrane extract from murine tumors), gelatin (denatured collagen) and decellularized tissue matrices converted into hydrogels (Kim et al.). To duplicate the basement membrane cells naturally adhere to in vivo, researchers apply a thin layer of Matrigel or collagen onto porous membranes or channel surfaces. The use of ECM-derived materials stems from the need to replicate the natural biochemical signals and tissue-like environments which synthetic plastics do not offer (Cao et al.). Zamprogno et al. (2021) demonstrated this by engineering a collagen–elastin (CE) membrane that replaced PDMS in a lung-on-chip, enabling cyclic breathing strains with native-like elasticity and providing a more biomimetic surface for alveolar–endothelial co-culture (Zamprogno et al.).

The main advantages of ECM-derived hydrogels include their high biocompatibility and bioactivity which enables cell attachment, differentiation, and function. The natural composition of proteins and glycoproteins within these matrices makes cells recognize them so they display in vivo-like characteristics including morphology, polarity, and gene expression during culture (Cao et al.). For example, Huang et al. (2021) created a GelMA hydrogel scaffold with alveolus-sized pores in a lung chip, showing that alveolar epithelial cells formed tight junctions and maintained physiological phenotypes far better than on flat PDMS membranes (Huang et al.). The binding sites found in collagen function as a cell adhesion pathway that triggers signaling processes while influencing both cell morphology and developmental fate in ways that match actual connective tissue (Osório et al.) . These ECM hydrogels are also biodegradable, which enables dynamic remodeling (Cao et al.); Shen et al. (2023) used an F127-diacrylate hydrogel membrane that degraded under cell-driven remodeling but still maintained barrier function during breathing motions. Such features enabled the chip to distinguish between physiological strain (which preserved cell health) and pathological strain (which triggered fibrosis), a nuance lost with PDMS substrates (Shen et al.).

The primary disadvantage of utilizing native ECM materials is their weak mechanical properties and short-lived stability. Such hydrogels possess an extremely fragile structure because they contain mostly water, which leads them to deform or collapse under perfusion (Ho et al.). Zamprogno et al. (2021) noted that fabricating sub-10 µm collagen–elastin membranes required specialized drying and handling techniques to prevent rupture, making manufacturing difficult (Zamprogno et al.). Pure collagen gels shrink under tension, while fibrin gels degrade within days unless cross-linked, limiting their utility for long-term cultures (Sanz-Horta et al.). Matrigel poses additional problems: its tumor-derived composition is variable between batches, making reproducibility low (Kim et al.) Shen et al. (2023) reported that switching from PDMS to hydrogel membranes improved fidelity but introduced challenges in controlling thickness and uniformity at scale (Shen et al.).

Hydrogels

The water-rich polymer networks known as hydrogels duplicate both the physical structure and biochemical properties of soft tissues (Gnatowski et al.). This section includes both ECM-derived materials such as collagen or fibrin that fall under the category of hydrogels and synthetic and engineered hydrogel systems that extend past direct ECM extracts. The main application of hydrogels in OOC engineering is to create three-dimensional scaffolds that enable cells to grow naturally while interacting and responding better than they do on rigid surfaces (Carvell et al.). Huang et al. (2021) illustrated the former with a GelMA hydrogel alveolar scaffold that replicated lung alveoli geometry and improved alveolar epithelial function (Huang et al.). Shen et al. (2023) illustrated the latter with a hydrogel membrane substituting PDMS, which enabled realistic diffusive transport and preserved normal barrier responses (Shen et al.).

Hydrogels achieve native ECM simulation through their water-rich structure combined with adjustable elasticity and natural biocompatibility (Cao et al.). The elastic properties of hydrogels can be tuned to match lung tissue , supporting realistic mechanotransduction (Shen et al.). Hydrogels also provide excellent nutrient diffusion and permeability, sustaining viability during extended air–liquid interface cultures (Liu et al.). Synthetic hydrogels such as PEG offer tunable stiffness and they can guide cell adhesion (Lin et al.). Lastly, hydrogels have been shown to be more favourable for cell-attachment as compared to PDMS. Annabi et al. (2013) further showed that cardiomyocytes display higher spontaneous beating rates on tropoelastin coated surfaces compared to gelatin, supporting the claim that tropoelastin-based hydrogels are suitable for select OOC applications (Annabi et al.).

The main limitation for LOCs is mechanical fragility. Shen et al. (2023) found that hydrogel membranes tore under excessive strain, requiring careful optimization of strain magnitude (Shen et al.). Hydrogels are also prone to swelling/shrinkage, introducing variability in microchannel dimensions and diffusion gradients (Feng and Wang). Thick hydrogels also impede perfusion, sometimes leading to hypoxia in central regions (Grebenyuk et al.). Finally, while transparency usually supports imaging, some hydrogels scatter light, reducing resolution compared to PDMS or COC (Kaberova et al.).

Signalling Cues

Mechanical Cues

The lung tissue experiences mechanical cues from cyclic stretching and blood perfusion-induced shear stress during breathing operations. The alveolar walls in living organisms undergo 5–15% expansion and retraction with each breath cycle and pulmonary capillaries maintain constant blood flow (Huh et al.). The normal development of lungs and homeostasis together with proper lung function rely on mechanical forces because breathing motions activate surfactant release and sustain alveolar architecture (Huh et al.). The absence of mechanical stimulation during cell culture results in cell dedifferentiation alongside reduced barrier functionality but appropriate mechanical stretch creates behavior that resembles living tissues (Huh et al.).

Lung-on-chip systems generate breathing movements through cell-grown membrane flexing. The first lung-on-chip device from Huh et al used cyclic vacuum inside chambers to stretch an alveolar cell-covered flexible PDMS membrane which replicated inhalation and exhalation movements. The dynamic mechanical actuation process triggered the development of tight junctions and surfactant production which duplicated the natural alveolar structure (Huh et al.). Stucki et al. (2018) demonstrated that alveolar chip permeability altered significantly after applying cyclic strain with a 10% amplitude at 0.2 Hz frequency when compared to static culture conditions (Stucki et al.). The mechanical forces of stretching influence pathological processes because researchers found that cyclic strain increased inflammatory responses to inhaled particles as a lung would during breathing movements (Huh et al.). The mechanical cue of fluid shear stress functions in lung chips may also improve nutrient delivery while promoting endothelial cell alignment for better vascular modeling (Corral-Nájera et al.). Thus, the scientific consensus may indicate that mechanical stretch integration produces a more realistic model of the lung

Inflammatory Signals

The implementation of inflammatory cues follows different approaches in Lung-on-Chips research. Researchers achieve inflammatory responses by adding cytokines from outside sources to the chip system. Benam et al. (2016) studied bronchial epithelium on-chip responses to IL-13 exposure which resulted in asthmatic characteristics such as goblet cell hyperplasia together with cytokine hypersecretion and decreased ciliary function (Benam et al., “Small Airway-On-a-Chip Enables Analysis of Human Lung Inflammation and Drug Responses in Vitro”). The model developed an asthma-like state which included excessive mucus production alongside dysfunctional cilia and could be treated with pharmaceuticals during the chip experiments. The alternative method of microfluidic channel introduction includes adding immune cells to the system. Huh et al. developed the first lung-on-a-chip system by using human neutrophils to perfuse through its vascular channel. After bacterial infection of the alveolar channel neutrophils passed through the membrane to perform microbial phagocytosis exactly like a natural immune response. The model produced organ-level responses through cytokine-driven neutrophil recruitment and bacterial clearance by incorporating inflammatory signaling. The model reached this organ-level behavior by incorporating inflammatory signaling which produced cytokine-mediated neutrophil recruitment and bacterial clearance (Huh et al.).

Environmental Signals

The design of Lung-on-Chip systems enables researchers to simulate external environmental stimuli such as cigarette smoke, air pollution particles (PM2.5), and airborne viruses through appropriate physiological methods. Researchers Benam et al. (2016) created a smoking machine connection to small-airway-on-chip to replicate human smoker inhalation by delivering whole cigarette smoke puffs through the air–liquid interface through a smoking machine which simulated the breathing patterns of human smokers. In their study they found that when exposed to cigarette smoke, 276 genes were expressed in COPD chips which were otherwise not expressed in the absence of smoke (Benam et al.). Huang et al. devised a human 3D alveolar-chip with cyclic stretch and exposed it to whole cigarette smoke and SARS-CoV-2 pseudovirus. Smoke puff exposure (via a smoking robot) induced oxidative stress and cell death in the chip epithelium (e.g. increased 4-HNE and caspase activity), mimicking early smoke damage. In the same chip, infection with SARS-CoV-2 pseudovirus produced the expected cytopathic effects and IL‑8 induction, which could be blunted by antiviral drugs (Huang et al.). Chips have also modeled airborne pathogens: Bai et al. infected a breathing alveolus chip with influenza A (H3N2). They found that dynamic breathing strain triggered protective innate responses (reduced viral replication, cytokine release, apoptosis) that were absent in non-breathing controls (Bai et al.). A recent study employed lung-on-a-chip technology to replicate human lung exposure to environmental air pollution (fine particulate matter, PM₂.₅) in actual environmental conditions. The microfluidic 3D model contained human alveolar epithelial cells and microvascular endothelial cells which were co-cultured to establish an alveolar-capillary interface before receiving an acute PM₂.₅ exposure. The high-dose PM₂.₅ exposure caused damage to the lung barrier through adherens junction disruption while creating oxidative stress and cell death and activating inflammation through elevated cytokine/chemokine expression in both alveolar epithelium and endothelium (Xu et al.).

Conclusion

Lung-on-a-chip technology has evolved significantly as a biomimetic model of the human lung yet there are still important gaps in completely mimicking lung physiology. Two promising innovation areas are driving the field forward. The second generation of researchers is working on developing bioinspired, dynamic biomaterials to better mimic the lung’s architecture and mechanics. The traditional PDMS membranes used in lung chips have been replaced with stretchable biological scaffolds (e.g. collagen–elastin) that form in vivo-sized alveoli and mimic the native extracellular matrix. Such bioinspired membranes outperform PDMS by eliminating unwanted drug absorption and offering tunable thickness and stiffness, thereby preserving normal lung cell function over weeks (Zamprogno et al.). Indeed, replacing artificial substrates with hydrogels or elastic biomembranes and applying physiological cyclic strain yields a more lifelike alveolar microenvironment (Brandauer et al.). These material innovations greatly enhance the biological fidelity of lung-on-chips, helping cells maintain phenotypes and responses closer to those in real lung tissue, which is crucial for predictive modeling.

The integration of lung-on-a-chip systems into multi-organ platforms has emerged as a strategy to capture systemic disease interactions and whole-body drug responses. Microphysiological systems which connect the lung module with other organ models via fluid flow enable the study of organ–organ interdependence as well as the assessment of how distant tissues affect lung pathology or drug effects (Brandauer et al.). Early demonstrations have linked lung chips with liver, brain, and other organ compartments to reveal interdependent responses. For instance, a liver spheroid module in a lung-on-chip has been shown to metabolize and detoxify an inhaled toxicant and thus reduce lung tissue injury from exposure (Brandauer et al.). In another study, a four-organ platform (lung, liver, brain, bone) was used to simulate the metastatic spread of lung cancer and to allow simultaneous assessment of drug efficacy in multiple organs (Zhang et al.). These integrated systems increase model complexity by including circulation, metabolism, and multi-organ signaling which capture human pharmacokinetic and disease processes that single-organ chips cannot and thus improve the clinical relevance of the experiments. In the future, the improved physiological realism and the incorporation of inter-organ interactions in next-generation lung-on-chips will provide more reliable preclinical data that bridges the translation gap between in vitro findings and clinical outcomes (Doryab et al.). Ultimately, such innovations will bring LOC technology nearer to practical biomedical and pharmaceutical applications, helping realize the vision of human-relevant lung models for research and therapy.

Works Cited

Alghannam, Fahad, et al. “Recent Progress in PDMS-Based Microfluidics toward Integrated Organ-On-a-Chip Biosensors and Personalized Medicine. ” Biosensors, vol. 15, no. 2, 29 Jan. 2025, p. 76, www.ncbi.nlm.nih.gov/pmc/articles/PMC11852457/, https://doi.org/10.3390/bios15020076.

Annabi, Nasim , et al. Hydrogel-Coated Microfluidic Channels for Cardiomyocyte Culture. V ol. 13, no. 18, 1 Jan. 2013, pp. 3569–3569, https://doi.org/10.1039/c3lc50252j.

Annabi, Nasim et al. “Hydrogel-coated microfluidic channels for cardiomyocyte culture. ” Lab on a chip vol. 13,18 (2013): 3569-77. doi:10.1039/c3lc50252j

Bai, Haiqing, et al. “Mechanical Control of Innate Immune Responses against Viral Infection Revealed in a Human Lung Alveolus Chip. ” Nature Communications, vol. 13, no. 1, 8 Apr. 2022, https://doi.org/10.1038/s41467-022-29562-4. Accessed 1 May 2023.

Benam, Kambez H, et al. “Small Airway-On-a-Chip Enables Analysis of Human Lung Inflammation and Drug Responses in Vitro. ” Nature Methods, vol. 13, no. 2, 21 Dec. 2015, pp. 151–157, https://doi.org/10.1038/nmeth.3697.

Benam, Kambez H., et al. “Matched-Comparative Modeling of Normal and Diseased Human Airway Responses Using a Microengineered Breathing Lung Chip. ” Cell Systems, vol. 3, no. 5, Nov. 2016, pp. 456-466.e4, https://doi.org/10.1016/j.cels.2016.10.003. Accessed 7 Apr. 2022.

Bhatia, Sangeeta N, and Donald E Ingber. “Microfluidic Organs-On-Chips. ” Nature Biotechnology, vol. 32, no. 8, 5 Aug. 2014, pp. 760–72, www.ncbi.nlm.nih.gov/pubmed/25093883, https://doi.org/10.1038/nbt.2989.

Brandauer, Konstanze , et al. “Advances of Dual-Organ and Multi-Organ Systems for Gut, Lung, 2 Skin and Liver Models in Absorption and Metabolism Studies. ” Lab on a Chip, 1 Jan. 2025, pubs.rsc.org/en/content/articlelanding/2025/lc/d4lc01011f, https://doi.org/10.1039/d4lc01011f. Accessed 20 Feb. 2025.

Calvert, Ben A., and Amy L. Ryan (Firth). “Application of IPSC to Modelling of Respiratory Diseases. ” Cell Biology and Translational Medicine, Volume 7, 2019, pp. 1–16, link.springer.com/chapter/10.1007%2F5584 2019 _ _ 430, https://doi.org/10.1007/5584 2019 430. Accessed 22 June 2020. _ _

Campillo, Noelia, et al. “Alveolus Lung-On-a-Chip Platform: A Proposal. ” Chemosensors, vol. 9, no. 9, 1 Sept. 2021, p. 248, www.mdpi.com/2227-9040/9/9/248/htm, https://doi.org/10.3390/chemosensors9090248. Accessed 12 Dec. 2021.

Cao, Uyen M. N., et al. “Microfluidic Organ-On-A-Chip: A Guide to Biomaterial Choice and Fabrication. ” International Journal of Molecular Sciences, vol. 24, no. 4, 6 Feb. 2023, p. 3232, https://doi.org/10.3390/ijms24043232.

Carvell, Tom, et al. “Categorising Hybrid Material Microfluidic Devices. ” Frontiers in Lab on a Chip Technologies, vol. 3, 13 May 2024, https://doi.org/10.3389/frlct.2024.1412290. Accessed 6 May 2025.

Corral-Nájera, Kendra, et al. “Polymeric and Biological Membranes for Organ-On-a-Chip Devices. ” Microsystems & Nanoengineering, vol. 9, no. 1, 29 Aug. 2023, https://doi.org/10.1038/s41378-023-00579-z. Accessed 27 Feb. 2024.

Dave Anuj Lagowala, et al. “Human Microphysiological Models of Airway and Alveolar Epithelia. ” American Journal of Physiology-Lung Cellular and Molecular Physiology, vol. 321, no. 6, 1 Dec. 2021, pp. L1072–L1088, www.ncbi.nlm.nih.gov/pmc/articles/PMC8715018/, https://doi.org/10.1152/ajplung.00103.2021. Accessed 14 Mar. 2024.

Doryab, Ali, et al. “Fabrication Strategies for Bioinspired and Functional Lung‐On‐Chips. ” Advanced Functional Materials, vol. 34, no. 46, 16 Aug. 2024, https://doi.org/10.1002/adfm.202407725. Accessed 13 Aug. 2025.

Elveflow. “Materials for Microfluidic Chips Fabrication : A Review 2017. ” Elveflow, Elvesys, 5 Feb. 2021, www.elveflow.com/microfluidic-reviews/materials-for-microfluidic-chips-fabrication-a-revie w-2017/.

Feng, Weijia, and Zhengke Wang. “Tailoring the Swelling‐Shrinkable Behavior of Hydrogels for Biomedical Applications. ” Advanced Science, vol. 10, no. 28, 6 Aug. 2023, https://doi.org/10.1002/advs.202303326.

Francis, Isabella, et al. “Recent Advances in Lung-On-a-Chip Models. ” Drug Discovery Today, vol. 27, no. 9, Sept. 2022, pp. 2593–2602, https://doi.org/10.1016/j.drudis.2022.06.004.

Gnatowski, Przemysław, et al. “Hydrogel Membranes in Organ-On-a-Chip Devices: A Review. ” Colloids and Surfaces B: Biointerfaces, vol. 251, 20 Feb. 2025, p. 114591, www.sciencedirect.com/science/article/abs/pii/S0927776525000980, https://doi.org/10.1016/j.colsurfb.2025.114591.

Grebenyuk, Sergei, et al. “Large-Scale Perfused Tissues via Synthetic 3D Soft Microfluidics. ” Nature Communications, vol. 14, no. 1, 12 Jan. 2023, p. 193, www.nature.com/articles/s41467-022-35619-1, https://doi.org/10.1038/s41467-022-35619-1. Accessed 11 Apr. 2023.

Ho, Tzu-Chuan, et al. “Hydrogels: Properties and Applications in Biomedicine. ” Molecules, vol. 27, no. 9, 2 May 2022, p. 2902, https://doi.org/10.3390/molecules27092902.

Huang, Di, et al. “Reversed-Engineered Human Alveolar Lung-On-a-Chip Model. ” Proceedings of the National Academy of Sciences, vol. 118, no. 19, 11 May 2021, www.pnas.org/content/118/19/e2016146118, https://doi.org/10.1073/pnas.2016146118.

Huh, Dongeun, et al. “Reconstituting Organ-Level Lung Functions on a Chip. ” Science (New York, N.Y.), vol. 328, no. 5986, 2010, pp. 1662–8, www.ncbi.nlm.nih.gov/pubmed/20576885, https://doi.org/10.1126/science.1188302.

Jensen, Caleb, and Yong Teng. “Is It Time to Start Transitioning from 2D to 3D Cell Culture?” Frontiers in Molecular Biosciences, vol. 7, no. 33, 2020, p. 33, https://doi.org/10.3389/fmolb.2020.00033.

Kaberova, Zhansaya, et al. “Microscopic Structure of Swollen Hydrogels by Scanning Electron and Light Microscopies: Artifacts and Reality. ” Polymers, vol. 12, no. 3, 5 Mar. 2020, p. 578, www.ncbi.nlm.nih.gov/pmc/articles/PMC7182949/, https://doi.org/10.3390/polym12030578. Accessed 12 Jan. 2023.

Kang, Dayoon, et al. “All‐Inkjet‐Printed 3D Alveolar Barrier Model with Physiologically Relevant Microarchitecture. ” Advanced Science, vol. 8, no. 10, 8 Mar. 2021, p. 2004990, https://doi.org/10.1002/advs.202004990. Accessed 24 Feb. 2023.

Kaur, Gurvinder, and Jannette M. Dufour. “Cell Lines. ” Spermatogenesis, vol. 2, no. 1, 1 Jan. 2012, pp. 1–5, www.ncbi.nlm.nih.gov/pmc/articles/PMC3341241/, https://doi.org/10.4161/spmg.19885.

Kim, Suran, et al. “Tissue Extracellular Matrix Hydrogels as Alternatives to Matrigel for Culturing Gastrointestinal Organoids. ” Nature Communications, vol. 13, no. 1, 30 Mar. 2022, https://doi.org/10.1038/s41467-022-29279-4.

Lin, Jiaqi, et al. “Emerging Trends in Microfluidic Biomaterials: From Functional Design to Applications. ” Journal of Functional Biomaterials, vol. 16, no. 5, 8 May 2025, pp. 166–166, https://doi.org/10.3390/jfb16050166. Accessed 16 June 2025.

Liu, Haitao, et al. “Advances in Hydrogels in Organoids and Organs‐On‐A‐Chip. ” Advanced Materials, 8 July 2019, p. 1902042, https://doi.org/10.1002/adma.201902042. Accessed 30 Oct. 2019.

Min, Eun-Kyung, et al. “Advanced Lung-On-a-Chip Technology: Mimicking the Complex Human Lung Microenvironment. ” International Journal of Biological Sciences, vol. 21, no. 1, 11 Dec. 2024, pp. 17–39, pmc.ncbi.nlm.nih.gov/articles/PMC11667821/, https://doi.org/10.7150/ijbs.105702.

O. Stucki, Andreas, et al. “A Lung-On-a-Chip Array with an Integrated Bio-Inspired Respiration Mechanism. ” Lab on a Chip, vol. 15, no. 5, 2015, pp. 1302–1310, pubs.rsc.org/en/content/articlehtml/2014/lc/c4lc01252f#imgfig1, https://doi.org/10.1039/C4LC01252F.

Osório, Luana A., et al. “A Review of Biomaterials and Scaffold Fabrication for Organ-On-a-Chip (OOAC) Systems. ” Bioengineering, vol. 8, no. 8, 1 Aug. 2021, p. 113, www.mdpi.com/2306-5354/8/8/113, https://doi.org/10.3390/bioengineering8080113.

Ronaldson-Bouchard, Kacey, and Gordana Vunjak-Novakovic. “Organs-On-a-Chip: A Fast Track for Engineered Human Tissues in Drug Development. ” Cell Stem Cell, vol. 22, no. 3, Mar. 2018, pp. 310–324, www.sciencedirect.com/science/article/pii/S1934590918300730?via%3Dihub, https://doi.org/10.1016/j.stem.2018.02.011.

Sanz-Horta, Raúl, et al. “Technological Advances in Fibrin for Tissue Engineering. ” Journal of Tissue Engineering, vol. 14, 1 Jan. 2023, www.ncbi.nlm.nih.gov/pmc/articles/PMC10426312/, https://doi.org/10.1177/20417314231190288.

Shen, Chong, et al. “A Microfluidic Lung-On-a-Chip Based on Biomimetic Hydrogel Membrane. ” Biotechnology and Bioengineering, vol. 120, no. 7, July 2023, pp. 2027–2038, pubmed.ncbi.nlm.nih.gov/37195718/, https://doi.org/10.1002/bit.28426.

Soriano, Joan B., et al. “Prevalence and Attributable Health Burden of Chronic Respiratory Diseases, 1990–2017: A Systematic Analysis for the Global Burden of Disease Study 2017. ” The Lancet Respiratory Medicine, vol. 8, no. 6, 1 June 2020, pp. 585–596, pmc.ncbi.nlm.nih.gov/articles/PMC7284317/, https://doi.org/10.1016/S2213-2600(20)30105-3.

Stucki, Janick D., et al. “Medium Throughput Breathing Human Primary Cell Alveolus-On-Chip Model. ” Scientific Reports, vol. 8, no. 1, 25 Sept. 2018, https://doi.org/10.1038/s41598-018-32523-x.

Takahashi, Kazutoshi, and Shinya Yamanaka. “Induction of Pluripotent Stem Cells from Mouse Embryonic and Adult Fibroblast Cultures by Defined Factors. ” Cell, vol. 126, no. 4, Aug. 2006, pp. 663–676, pubmed.ncbi.nlm.nih.gov/16904174/.

Tamai, Koji, et al. “IPSC-Derived Mesenchymal Cells That Support Alveolar Organoid Development. ” Cell Reports Methods, vol. 2, no. 10, 24 Oct. 2022, p. 100314, www.sciencedirect.com/science/article/pii/S2667237522002041?via%3Dihub, https://doi.org/10.1016/j.crmeth.2022.100314. Accessed 14 Nov. 2022.

Tanyeri, Melikhan, and Savaş Tay. “Viable Cell Culture in PDMS-Based Microfluidic Devices. ” Methods in Cell Biology, 2018, pp. 3–33, https://doi.org/10.1016/bs.mcb.2018.09.007.

Wang, Hui, et al. “Microfluidic Strategies for Biomimetic Lung Chip Establishment and SARS-CoV2 Study. ” Materials Today. Bio, vol. 24, 1 Feb. 2024, p. 100905, pubmed.ncbi.nlm.nih.gov/38094656/, https://doi.org/10.1016/j.mtbio.2023.100905.

Xu, Cong, et al. “Assessment of Air Pollutant PM2.5 Pulmonary Exposure Using a 3D Lung-On-Chip Model. ” ACS Biomaterials Science & Engineering, vol. 6, no. 5, 24 Mar. 2020, pp. 3081–3090, https://doi.org/10.1021/acsbiomaterials.0c00221. Accessed 2 May 2024.

Zamprogno, Pauline, et al. “Second-Generation Lung-On-a-Chip with an Array of Stretchable Alveoli Made with a Biological Membrane. ” Communications Biology, vol. 4, no. 1, 5 Feb. 2021, pp. 1–10, www.nature.com/articles/s42003-021-01695-0, https://doi.org/10.1038/s42003-021-01695-0.

Zhang, Xiaochun, et al. “In Vitro Biomimetic Models for Respiratory Diseases: Progress in Lung Organoids and Lung-On-a-Chip. ” Stem Cell Research & Therapy, vol. 16, no. 1, 30 July 2025, https://doi.org/10.1186/s13287-025-04500-5. Accessed 26 Aug. 2025.

About the author

Armaan Mehtani

Armaan is a high school student at Aiglon College in Switzerland with passions in biology, chemistry, and biomedical engineering. He is passionate about medicine and aspires to become a surgeon. His research interests lie at the intersection of cellular biology and biomedical technologies, particularly in organ-on-a-chip systems.

The post Recreating Human Lung Function on a Chip: Progress in Biomaterials, Cellular Models, and Environmental Signals appeared first on Exploratio Journal.