Abstract

Solar-powered unmanned aerial vehicles (UAVs) carry no fuel, meaning their endurance is governed by a balancing act crafted between the energy harvested from sunlight and the energy consumed throughout flight, rather than conventional fuel capacity constraints. Over multi-day missions, this balance becomes hard to predict, as solar harvesting depends on the Sun’s position relative to the UAV whilst energy consumption depends on the aerodynamic forces required to sustain flight. Both of these are also affected by flight conditions and cruise altitudes, meaning a complex optimisation problem emerges. Many simplified models treat the Sun as stationary, an assumption only valid for short missions over small timeframes and distances. In this paper, solar harvesting will be modelled through a fixed panel and gimballed panel approach, where a gimballed panel traces the Sun’s position relative to the UAV and points the panel at the optimal orientation for maximum irradiance.

Overall, this paper develops a model that investigates the energy balance aforementioned, using the example of a transatlantic crossing from London to New York. There are three overarching MATLAB models that are integrated into this energy balance: great-circle navigation, aerodynamic power consumption, and solar energy harvesting. The flight parameters have been derived from steady-level flight mechanics, standardised NACA tables, photovoltaic conversion data and solar geometry. 

Having completed the simulations, it can be seen that an idealised gimballed solar panel increases total harvested energy substantially, by 62%, but much of the additional energy that it harvests cannot be stored, since the battery charge is already full during daylight hours, where there is peak harvesting. This is shown as the minimum state-of-charge only increases by 5% for a gimballed panel, starkly lower than the 62% rise in energy harvesting aforementioned. As a result, it can be said that endurance in this context cannot be governed solely by aerodynamic efficiency or solar collection gains, but by the capacity of the battery to carry energy across extended periods of insufficient insolation.

Future work on this issue should therefore treat battery capacity as a design variable to be considered alongside altitude and cruise conditions, whilst also relaxing assumptions discussed throughout. This concept has potential real-world applications in wildfire monitoring, surveillance/law enforcement, shipping, and ocean monitoring.

1. Introduction

1.1 Motivation

Solar-powered unmanned aerial vehicles (UAVs) offer the ability of extremely long endurance without the constraints provided by conventional fuel. Their endurance is governed simply by the balance and optimisation between the electrical energy harvested by the sun, and the energy consumption by the UAV in flight. As a result, solar flight is attractive for a multitude of applications that require persistent operation, over large and remote areas. Such examples include maritime search and rescue, wildfire detection and atmospheric surveying in remote areas. This concept has been demonstrated experimentally before, with AtlantikSolar achieving an 81-hour continuous flight.(Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015)(Oettershagen, et al., Design of small hand-launched solar-powered UAVs: From concept study to a multi-day world endurance record flight, 2017) The concept explored in this paper also explores the idea of a moveable (on the pitch and roll axis) solar panel on the UAV, allowing it to maximise the electrical energy it harvests from the sun, a concept yet to be quantified by existing studies. The primary reason this has not been integrated into mainstream UAV technology is due to the aerodynamic and mass penalties of such a gimbal system, which are noted as limitations throughout.

In general, the primary difficulty with a solar-powered UAV tends to be that the energy balance of the UAV is directly proportional to the aircraft’s flight conditions and environmental. Solar collection also depends on the position of the sun relative to the aircraft (with a non-gimballed system), which varies with latitude, time of day, and season. Furthermore, aircraft energy consumption varies on the aerodynamic forces required to maintain steady-level flight in a cruise state. These forces vary significantly with changes in airspeed, air density, and lift coefficient, with air density itself varying with altitude. Consequently, due to the balancing act of these two factors, one cannot simply consider aerodynamic or solar conditions alone, creating a complex problem.

Multiple attempts have been made to simplify or solve this problem, with some choosing to limit mission distance and time as being short relative to the sun’s position, allowing it to be treated as stationary. (Klesh & Kabamba, 2009)Others provide validated airframes and power-collection models but lack investigation into cruise-altitude or even route optimisation for a multi-day mission. (Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015)The trade-off between altitude, flight conditions, and battery margin over a multi-day mission therefore remains unestablished, and this paper attempts to address this gap.

2. Preliminaries

2.1 Steady Level Flight

The aircraft is modelled in steady, level cruise. This is primarily due to time constraints and aerodynamical complexities involved in take-off, climb, descent and landing phases, which could further complicate the solar harvesting model mentioned later on. The diagram below contains four forces acting in the aircraft’s x-z plane: lift perpendicular to the flight path, drag opposing the direction of motion, thrust along the flight path, and weight acting vertically downwards.

Figure 1 – Forces Acting on Aircraft in Steady-Level Flight

Steady level flight involves no acceleration, so these forces balance in pairs:

Together, these two conditions establish the relationship between airspeed, lift coefficient and altitude: once aircraft parameters are established, the air density at a given altitude details the airspeed required at a chosen lift coefficient. (Fidowski, Atkins, & Powell, 2019)(Anderson, 1999)

The equations for lift and drag forces are given by Equation [2] below:

Here,  V is true airspeed, ρ air density,  S is the wing reference area over which airflow occurs, and C_D   and C_L  are drag and lift coefficients. The quantity 1/2 ρV²  therefore describes the dynamic pressure. (Fidowski, Atkins, & Powell, 2019)

If we combine Equation [2] with the lift-weight balance condition in Equation [1], we get the following, Equation [3], the airspeed required to sustain a specified lift coefficient at a particular altitude:

The aerodynamic coefficients are derived from two-dimensional NACA 4415 section data. (Abbott & von Doenhoff, 1959)(Airfoil Tools, n.d.) NACA 4415 was selected due to its combination of lift-to-drag performance and sufficient thickness to accommodate both solar-cell installation and endurance structural requirements. We can calculate the Reynold’s Number of the aerofoil using Equation [4] below:

At this Reynolds number the section lift curve is markedly non-linear: the local slope falls from approximately 0.21 per degree near zero incidence to 0.06 per degree at eight degrees, a consequence of a laminar separation bubble whose extent varies with incidence. A single linear fit is therefore not representative, and the tabulated section data are used directly rather than through a lift-curve slope approximation. The drag polar is obtained by fitting the drag coefficient against the square of the lift coefficient over the same data, which does not require the lift curve to be linear. (Airfoil Tools, n.d.)(Drela, 1989)                                     

The two-dimensional data are then corrected for a finite wing. A real wing sheds tip vortices, which induce a downwash and generate induced drag proportional to the square of the lift coefficient. The downwash also reduces the effective angle of attack seen by the section, so the geometric angle of attack required to achieve a given lift coefficient exceeds the two-dimensional value by the induced angle. (Anderson, 1999)

InFor this model, there are two relevant aerodynamic operating points. The first is maximum lift-to-drag ratio, which can be calculated by differentiating   with respect to   and setting the derivative equal to zero. (Fidowski, Atkins, & Powell, 2019) This can be seen below in Eq. [7]. 

This condition occurs when parasitic drag and induced drag are equal. This is at  C_L = 0.593,  kC²_L = 0.00861, matching C_D0  exactly. The condition also occurs at minimum power, given by Eq. [8]:

Here, at the minimum power condition we can see that induced drag is three times the parasitic drag. This is because power is the product of the drag and airspeed, so the additional velocity dependence causes the optimum condition to be shifted to a higher lift coefficient.

The difference between these operating conditions matters for a long-duration solar flight, as maximum L/D ratio would minimise drag and therefore the energy required over a fixed distance. Since that energy is the product of the drag and the distance flown, it corresponds to a maximum range condition. On the other hand, minimum power minimises the electrical demand at each instant and therefore corresponds to a maximum endurance condition. 

 Figure 2 demonstrates the comparison between maximum L/D ratio and minimum power utilising the flight parameters of our aircraft that can be found in the Appendix.

2.2 Standard Atmospheric Conditions

Air density will vary significantly across the altitude range considered. A function of air density as altitude varies can be obtained, combining the hydrostatic relationships and ideal gas laws with a linear temperature profile in the troposphere. This is given in Eq. [9] and [10] below:

The constants in the equations above are:

(International Organization for Standardization, 1975)

This model is valid up to 11km altitude. This function is utilised to calculate the airspeed required to generate lift at different altitudes, as when density falls, the aircraft must fly faster to generate the same lift at a given lift coefficient, changing the power required.

2.3 Propulsion

For a propellor-driven aircraft, electrical power is the most useful measure to determine energy consumption. This is because the battery energy balance depends on the electrical consumption of the propulsion system. Eq. [11] gives the propulsion model, combining aerodynamic power requirements with the propulsion efficiency n_prop  and the constant avionics load P_av . (Fidowski, Atkins, & Powell, 2019)

2.4 Solar Position

The solar power available to harvest by the aircraft is set by the irradiance incidence on the illuminated panel area, and the conversion efficiency of the photovoltaic system. Irradiance will vary continuously with geographic position, time of day, season, and altitude, so solar harvesting must be evaluated along the trajectory of its transatlantic route, rather than being treated as a constant power source.

The Sun’s position relative to the aircraft follows from the solar declination angle, the aircraft latitude, and the solar hour angle. (Duffie & Beckman, 2013)(Cooper, 1969) For a day of year  , the declination can be approximated through the following equation:

The hour angle follows from local solar time, with H = 0  at solar noon:

The solar elevation above the local horizontal can therefore be given through the spherical law of cosines, applied to the triangle formed by the pole, aircraft zenith, and the Sun:

Here,   is the aircraft latitude. Eq. [14] is the same as computing the great-circle central angle between the aircraft and the subsolar point, since the hour angle is the longitude difference between them.

2.5 Atmospheric Attenuation

The irradiance reaching the aircraft is also reduced by absorption and scattering across the atmospheric path, and as the Sun sits lower throughout the day, the longer that atmospheric path becomes. The relative optical air mass can be approximated using the Kasten-Young formulation, which accounts for atmospheric curvature, and remains finite at the horizon. (Kasten & Young, 1989)

Flying at altitude also reduces the mass of atmosphere above the aircraft, which can be accounted for by scaling the air mass by the local pressure ratio from the standard atmosphere model described in Section 2.2. The direct normal irradiance then can be placed in an empirical clear-sky attenuation model:

Here, I₀ = 1361 Wm^-2  is the solar constant, and the coefficient 0.7 corresponds to the fraction transmitted at unit air mass. (Meinel & Meinel, 1976) The exponent 0.678 is an empirical fit accounting for non-uniformity in the distinct atmospheric layers. This model captures both the reduction in available energy as the Sun approaches the horizon and the increase in irradiance at higher cruise altitudes. Diffused and ground-reflected components of irradiance have been assumed to be negligible as part of this investigation.

  Figure 3 demonstrates the effects of the differing insolation models.

2.6 Photovoltaic Conversion

The electrical power generated from an illuminated panel is:

Where A_PV is the photovoltaic cell area, η_PV  the module conversion efficiency,  η_PMMT the efficiency of the maximum-power-point tracking electronics, and   the angle between the solar direction and the panel normal. For the aircraft considered in this paper, A_PV = 1.7223 m² , as it has been taken as equal to the wing planform area. Furthermore, η_PV = .2 and η_PMMT = .95. The conversion process is entirely independent of flight conditions, so variation in harvested power arises entirely from solar geometry and atmospheric attenuation. (Duffie & Beckman, 2013)(Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015)

3. Methods

The following sections detail the mathematics behind MATLAB models utilised in this study. Access to these models can be found through the GitHub repository linked here:https://github.com/rohaannarula/solar-uav-energy-model

3.1 Panel Orientation & Solar Collection

The incidence angle,  ɸ introduced in Eq. [17] follows from the relative orientation of the Sun and the solar panel. Both directions are represented as unit vectors in a local North-East-ηPV (NED) frame. (Fidowski, Atkins, & Powell, 2019) The solar direction is:

Here, a is the solar azimuth, measured clockwise from North, and θs is the solar elevation. The negative third component is due to the downwards axis being positive, so the Sun above the horizon has a negative vertical component. For a panel with roll angle ɸp and pitch angle θp , with yaw fixed at zero, the outward normal is:

Because both of these vectors have unit length, their dot product gives the cosine of the incidence angle directly. The result is fixed above zero, since a negative value implies the Sun lies behind the panel, and no energy is collected:

Two configurations will be evaluated. The first is a flat panel, representing the roughly horizontal photovoltaic surface of an aircraft in steady-level cruise. At a maximum lift-to-drag condition, the required lift coefficient of 0.593 is established using an average cruise angle of 1.11°. This is close to the horizontal, making a flat panel assumption reasonable. Setting both orientation angles to zero reduces Eq. [20] to exactly the sine of the solar elevation:

The second configuration is an idealised gimballed panel; tracing the Sun’s position. The required orientation is obtained by equating the panel normal to the solar direction, which yields three separate component equations into two unknowns. The East component contains the roll angle and yields the pitch by dividing the North component by the Down component, which eliminates the roll term.

Two angles are enough for the three equations because the solar direction is a unit vector, so only two of its components are independent of each other. The yaw angle does not appear, as rotating the panel about its own axis does not alter the theoretical solar panel surface presented to the Sun.

In both configurations, the harvesting of solar energy is reduced significantly when solar elevation is below 5º , at which point the atmospheric path length is long enough that available irradiance becomes negligible. The gimballed panel can be treated as an idealistic case, rather than a realisable configuration, due to multiple assumptions required, such as the requirement for near vertical pitch near sunrise and sunset.

3.2 Great-Circle Route

The aircraft must follow the shortest possible path between London Heathrow and John F. Kennedy International Aiport on a spherical Earth, which is a great-circle arc. For endpoint co-ordinates with longitude difference Δƛ , the central angle subtended at the Earth’s centre is as follows:

(Fidowski, Atkins, & Powell, 2019)

As a result, the great circle distance and initial heading can be calculated, where   is the radius of the Earth:

A four-quadrant arctan2 function can be used throughout, in place of a simple arctan one, since dividing the sine and cosine components would discard their individual signs and therefore lose the directional element of the heading.

As great-circle routes do not maintain constant heading, the aircraft heading will change continuously throughout crossing. Intermediate vectors can be generated through spherical linear interpolation, where   is the fractional distance along the route, and   and   are unit position vectors:

 The result can then be converted back into latitude and longitude for the solar and atmospheric calculations. The route is decomposed into 2000 equally spaced points and is plotted below in Figure 4.

3.3 Cruise State & Energy Balance

For each cruise altitude and flight condition, the energy balance model is calculated independently. Utilising parameters obtained from Section 2; we can calculate the required electrical power and flight time:

Where P_elec is the required electrical power for a given altitude and flight condition.

We can calculate solar power required using the trapezium rule, as solar power harvested varies with both time and position, and is continuously varying:

The battery can be represented as an energy reservoir with limited capacity, updated according to the net electrical power balance:

Where E_{B max} = 850 Wh = 3.06MJ  and the battery is initially fully charged. (Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015) We can therefore calculate the state of charge, and the aircraft’s minimum state of charge over the flight:

The minimum state of charge is utilised as the primary measure of mission feasibility, as it is more informative than comparing net energy expenditure. This is because a positive net balance will not guarantee that sufficient energy is available throughout the flight.

The simulation of state of charge will be repeated for each altitude, with both cruise conditions and panel configurations. I will consider a minimum SOC requirement of 30% to consider feasibility, as assumptions discussed later diminish true viability of such a large margin.

4. Results

4.1 Solar Collection 

The great-circle route from London Heathrow to JFK New York spans 5540km. If the aircraft departs at 12.00 UTC on day 172 of the year, close to June solstice, the crossing at an altitude of 3000m encompasses 5 whole day-night cycles.

Figure 5 compares solar power generated by the flat and idealised gimballed configurations. Peak values appear similar, approximately 329W when gimballed against 291W, but the gimballed curve is broader, unlike the narrow daily peak of the flat configuration. Integrated across the crossing, at maximum L/D ratio, the two configurations collect 77.37MJ and 48.17MJ respectively, so continuous gimballed tracking yields 62% more energy harvested. The differences arise at low solar elevation, where a flat panel projects only 50% of the direct beam at 30 degrees and 17% at 10 degrees, compounding the atmospheric attenuation already present. This is represented by Figure 5.

4.2 Battery State-of-Charge

  Figure 6 below shows the battery state of charge over the same crossing. The battery reaches full capacity for several hours each day, meaning surplus energy harvested cannot be stored given current battery size constraints. Approximately 26.1 MJ is discarded across the maximum-L/D mission, exceeding the 22.02 MJ required to complete the crossing. Storage rather than collection is therefore the binding constraint, and the additional energy gathered by the gimballed panel does not translate proportionally into endurance. The minimum state of charge is a more useful measure of feasibility than net energy balance, as explained in section 3.3.

4.3 Panel Orientation 

   According to simulation, the ideal orientation requires panel roll to sweep between approximately ±60° each day and pitch to reach nearly −80° near sunrise and sunset, a range no practical mechanism could achieve, reinforcing the treatment of the gimballed case as an upper bound. The photovoltaic panel parameters can be found in the Appendix. 

4.3 Effect of Cruise Altitude and Flight Condition

Table 1 above summarises the design variables. The energy required falls only 4.3% across the entire altitude range, because drag at fixed   is independent of air density. On the other hand, battery margin falls much more sharply, from 0.437 to 0.102 under maximum L/D ratio, since greater altitudes raises power required from 43.3 to 71.4 W while night duration is unchanged. The 30% SOC requirement is only met to approximately 5 km. In terms of flight condition chosen; maximum L/D ratio consumes 15 percent less energy at every altitude, while minimum power leaves a larger reserve because it draws less overnight. This disparity grows with altitude, and at 6000m maximum L/D fails the margin requirement at 0.275 while minimum power satisfies it at 0.374. However, maximum L/D has immense benefits over minimum power in terms of practicality. The flight time of a maximum L/D condition cruise at 3000m is 39hr longer, impractical for real-world applications. Furthermore, it can be seen that a gimballed solar panel provides 62% more energy harvested at 3000m (maximum L/D), yet its minimum SOC is only 5% greater. This proves the constraint in this model is battery capacity, as most of the extra energy harvested has been discarded before it can be used, as there is simply not the storage available.

5. Assumptions & Evaluation

The performance of a multi-day crossing is governed by the interactions taking place between aerodynamics, solar geometry and finite storage rather than a specific singular parameter. The distinction between collected and usable energy is the most indicative finding. As 26.1 MJ of excess is unused against 22.02 MJ consumed, increasing collection alone offers diminishing returns unless storage is increased or surplus energy is otherwise employed. Changes in altitude selection is consequently better treated as a storage problem than an aerodynamic or harvesting one.

Several assumptions limit the quantitative accuracy of the study. The aircraft maintains steady level cruise in still air at constant mass, with launch, climb and descent neglected. The aerodynamic model uses a NACA 4415 wing polar rather than a complete aircraft polar, omitting fuselage and tail contributions; the resulting maximum L/D of 34.4 is approximately 13% more optimistic than the value inferred from published AtlantikSolar flight data. (Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015) (Selig, Guglielmo, Broeren, & Giguere, 1995) Moreover, the solar model assumes clear-sky direct radiation, neglecting diffuse and reflected components, and neglects the equation of time (the ‘mean sun’ versus true sun position throughout elliptical orbit). Furthermore, the use of an idealised gimballed panel should be read as an upper bound, since actuator mass, angular limits and airframe shadowing are not represented.(Cooper, 1969)Finally, it is important to note that wind, cloud and other weather variations are not represented, meaning that the calculated 30% battery state-of-charge margin is likely to diminish in real-world applications.

6. Conclusion

Endurance solar flight is limited by the energy balance detailed throughout the paper, and this study’s purpose was to determine the feasibility of a such a concept of near-perpetual flight, and in what conditions it would be possible. The paper’s approach has been to utilise pre-established components where possible, in order to reduce possible error from assumptions and to provide realistic aerodynamic and photovoltaic parameters. It also means that a further exploration of this issue could lead to design models of a bespoke aircraft design, optimised for a solar UAV.

Maximum Lift/Drag ratio is the cruise condition that allows for the lowest energy requirements whilst still fulfilling the 30% SOC requirement, with 22.02 MJ over 123.4hrs at a 3000m altitude. This is measured as the optimum conditions for such a crossing. Increasing altitude reduces the flight time, and atmospheric attenuation too, yet battery margin also decreases due to an increase in power requirements.

The use of an idealised gimballed solar panel increases solar harvesting by 62%, yet most of this amount is not usable due to the battery being full at these daytime hours. We can therefore label storage capacity as a limiting constraint, yet an increase in battery capacity could come with additional complications such as mass changes etc. For applications that require persistent operation over inaccessible areas, such as maritime rescue, marine life monitoring, or wildlife monitoring, this concept could be deployed to improve real-world applications’ efficiency.

7. Bibliography

Abbott, I. H., & von Doenhoff, A. E. (1959). Theory of wing sections: Including a summary of airfoil data. Dover Publications.

Airfoil Tools. (n.d.). NACA 4415 (naca4415-il) Xfoil prediction polar at RE=200,000, Ncrit=9. Retrieved from http://airfoiltools.com/polar/details?polar=xf-naca4415-il-200000

Anderson, J. D. (1999). Aircraft performance and design. McGraw-Hill.

Cooper, P. I. (1969). The absorption of radiation in solar stills. Solar Energy, 12(3), 333-346. doi:10.1016/0038-092X(69)90047-4

Drela, M. (1989). XFOIL: An analysis and design system for low Reynolds number airfoils. In T. J. Mueller (Ed.), Low Reynolds number aerodynamics (pp. 1-12). Springer. doi:10.1007/978-3-642-84010-4_1

Duffie, J. A., & Beckman, W. A. (2013). Solar engineering of thermal processes (4th ed.). John Wiley and Sons.

Fidowski, K., Atkins, E., & Powell, K. (2019). Introduction to aerospace engineering: Aero 201 course notes, Fall 2019. Ann Arbor, MI: University of Michigan.

International Organization for Standardization. (1975). Standard atmosphere (ISO 2533:1975). Geneva: ISO.

Kasten, F., & Young, A. T. (1989). Revised optical air mass tables and approximation formula. Applied Optics, 28(22), 4735-4738. doi:10.1364/AO.28.004735

Klesh, A. T., & Kabamba, P. T. (2009). Solar-powered aircraft: Energy-optimal path planning and perpetual endurance. Journal of Guidance, Control, and Dynamics, 32(4), 1320-1329. doi:10.2514/1.40139

Meinel, A. B., & Meinel, M. P. (1976). Applied solar energy: An introduction. Addison-Wesley.

Oettershagen, P., Melzer, A., Mantel, T., Rudin, K., Lotz, R., Siebenmann, D., . . . Siegwart, R. (2015). A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight. 2015 IEEE International Conference on Robotics and Automation (ICRA), (pp. 3986-3993). doi:10.1109/ICRA.2015.7139756

Oettershagen, P., Melzer, A., Mantel, T., Rudin, K., Stastny, T., Wawrzacz, B., . . . Siegwart, R. (2017). Design of small hand-launched solar-powered UAVs: From concept study to a multi-day world endurance record flight. Journal of Field Robotics, 34(7), 1352-1377. doi:10.1002/rob.21717

Selig, M. S., Guglielmo, J. J., Broeren, A. P., & Giguere, P. (1995). Summary of low-speed airfoil data, Volume 1. SoarTech Publications.

8. Appendix

8.1 Model Parameters

All model parameters are sourced from the following sources:

(Airfoil Tools, n.d.)(Drela, 1989)(Fidowski, Atkins, & Powell, 2019)(Oettershagen, et al., A solar-powered hand-launchable UAV for low-altitude multi-day continuous flight, 2015)(International Organization for Standardization, 1975)(Selig, Guglielmo, Broeren, & Giguere, 1995)(Oettershagen, et al., Design of small hand-launched solar-powered UAVs: From concept study to a multi-day world endurance record flight, 2017)(Kasten & Young, 1989)

8.1.1 Airframe and Propulsion

8.1.2 Aerodynamic Coefficients, NACA 4415 at Re = 200,000

8.1.3 Solar & Electrical Power

8.1.4 Atmosphere & Mission

About the author

Rohaan Narula

Rohaan is a student at King’s College School, Wimbledon in the United Kingdom, with a strong passion for engineering, specifically its applications to environmental concerns. Interweaving aeronautical, mechanical and environmental engineering throughout his research serves as a method of balancing both ethical and efficiency concerns whilst not losing sight of the endless pursuit of optimisation in these fields. Outside of the academic world, he is a keen cricketer and guitarist.

The post Modelling Energy Balance for a Solar-Powered Transatlantic UAV Route appeared first on Exploratio Journal.

1. Introduction

Heating, Ventilation, and Air Conditioning (HVAC) systems are essential across residential, commercial, healthcare, industrial, laboratory, and data-center environments, where maintaining thermal comfort, indoor air quality, and controlled operating conditions is critical (Lee & Chen, 2022). They regulate temperature, humidity, ventilation, and air quality, improving occupant comfort and well-being. As urbanization and living standards continue to rise, the demand for heating and cooling has increased significantly. Consequently, HVAC systems have become one of the largest consumers of energy in modern buildings (Lu et al., 2025). Buildings account for a major share of global energy consumption and carbon emissions. For this reason, improving HVAC efficiency has become an important research area, especially in regions experiencing rapid growth in cooling demand (Adelekan et al., 2022).

HVAC systems have largely relied on fixed schedules, thermostats, rule-based controls, and proportional-integral-derivative (PID) controllers. Although these methods are reliable, they may not respond effectively to change in conditions like room occupancy, outdoor weather and decline in equipment performance (Gharbi et al., 2025). As a result, these systems often continue operating for longer than required, resulting in excess energy consumption.

Recent advances in artificial intelligence (AI), machine learning (ML), the internet of things (IoT), cloud computing, and big data analytics are transforming HVAC operation. Instead of only relying on fixed settings, AI-based systems continuously analyse sensor data, control cooling and heating demands, and automatically adjust operation in real time  (Lee & Tsai, 2020). Studies show that these systems can reduce energy consumption while maintaining occupant comfort (Kannan et al., 2020). AI applications have expanded from traditional ML models to advanced techniques such as deep learning (DL), reinforcement learning (RL), digital twins (DT), and cloud-based optimization, pushing HVAC management toward autonomous operation(Lu et al., 2025). This change has made AI-enabled HVAC systems an important area of research, especially for smart buildings, sustainable development, and achieving carbon neutrality. In addition, AI-based HVAC systems support many United Nations Sustainable Development Goals (SDGs), namely SDG 7 (Affordable and Clean Energy), SDG 9 (Industry, Innovation and Infrastructure), SDG 11 (Sustainable Cities and Communities), SDG 12 (Responsible Consumption and Production), and SDG 13 (Climate Action) (Hauashdh et al., 2024). Figure 1 shows how AI-enabled HVAC systems contribute to these goals. 

Figure 1: Contribution of AI-enabled HVAC systems to the United Nations Sustainable Development Goals

This review explains the basic working of HVAC systems, the main factors affecting their energy use, and how AI is making them more intelligent, adaptive, and energy-efficient.

2. HVAC Principles and Energy Use

2.1 Fundamentals and Working Principles

The main purpose of an HVAC system is to maintain comfortable indoor conditions by controlling temperature, humidity, airflow, and air quality. It works based on principles of thermodynamics like heat transfer, refrigeration, and automatic control. A typical HVAC system consists of four main parts: heating, cooling, ventilation, and a control system that manages their operation.

The cooling process usually uses the vapor-compression refrigeration cycle, which includes a compressor, condenser, expansion valve, and evaporator. These components work together to remove heat from indoor air and release it outside. Heating can be provided by boilers, furnaces, electric heaters, or heat pumps. Ventilation supplies fresh outdoor air while removing stale air, pollutants, and excess moisture to maintain a healthy indoor environment. 

In commercial buildings, HVAC systems are more sophisticated and include equipment such as chillers, cooling towers, air handling units, variable air volume systems, pumps, duct-network, and a building management system (BMS) (Lu et al., 2025). These parts respond to shifts in thermal loads from occupants, equipment, lighting, and the outdoor environment. Since all these parts interact with each other in a complex way, operating the system efficiently is a challenging task (Nashruddin et al., 2025).

2.2 Fundamentals of HVAC Energy Consumption

HVAC systems are one of the largest energy-consuming systems in buildings, this energy is used for cooling, heating, ventilation, pumping, and other supporting controls (Lu et al., 2025). Among these, cooling consumes the highest amount of energy, especially in tropical and subtropical regions where air conditioners operate for most of the year (Chua et al., 2013). 

The energy consumption of an HVAC system depends on several factors, namely:

• Outdoor temperature and weather patterns
• Building orientation and insulation
• Occupancy and schedules
• Internal heat gains from equipment and lighting
• Ventilation requirements
• Equipment efficiency and maintenance condition

Many buildings still use fixed temperature settings and pre-set schedules for operation, which cannot adjust to change in conditions in real time. As a result, problems such as over cooling and ventilation, and inefficient operation are common. Legacy buildings have an increased effect due to outdated equipment, poor insulation, and uneven load distribution (Nashruddin et al., 2025).

Efficiency of HVAC systems is commonly evaluated using performance indicators like the energy efficiency ratio (EER), seasonal energy efficiency ratio, coefficient of performance, and cooling seasonal performance factor (CSPF) are used across the literature. AI-based controls improve the same metrics through predictive scheduling, adaptive control, and demand forecasting (Raza & Khosravi, 2015).

3. AI Application and Enabling Technologies in HVAC Systems

3.1 Functions of AI-Enabled HVAC Systems

AI-based approaches can provide predictive and adaptive capabilities that are limited in traditional non-AI-based systems. Unlike traditional controllers, which work using fixed rules and preset values, AI-based controllers continuously analyze system data and improve their performance over time. Accordingly, depending on the system design applications may include:

• Real-time optimization
• Predictive energy management
• Occupancy-aware control
• Thermal comfort optimization
• Fault detection and diagnosis (FDD)
• Predictive maintenance
• Demand response management
• Self-learning and adaptation

These features allow HVAC systems to respond to changes in building conditions, instead of following a fixed schedule, helping reduce energy consumption while maintaining occupant comfort (Lu et al., 2025). According to Lee & Chen (2022), AI-based energy-saving methods can be grouped into six main categories: model predictive control (coupled with AI) (MPC), thermal comfort control, model-free predictive control, control optimization, multi-agent control (MAC), and knowledge-based systems (KBS). These techniques have been reported to achieve energy-savings ranging from 8.8% to 25.5%, although the actual savings depend on the building type, operating conditions, and control strategy employed (Lee & Chen, 2022).

3.2 Hardware–Software Integration and Model Development

AI-based HVAC systems rely on both hardware and software to collect, process, and analyze data for intelligent control. These components work together to monitor building conditions and optimize HVAC operations.

Hardware Infrastructure

AI-based HVAC system hardware typically includes temperature, humidity,  airflow and air-quality sensors, occupancy sensors, smart thermostats, power meters, edge controllers, and IoT gateways, which continuously collect environmental and operational data and send it to the control system for analysis (Lu et al., 2025). 

Software Architecture

The software design may include the BMS itself, cloud computing platforms, data acquisition modules, ML frameworks, visualization dashboards, DT environments, and optimization engines. Cloud-based systems allow control algorithms to be updated remotely without replacing physical equipment, making upgrades more efficient and potentially reducing maintenance costs (Lee & Tsai, 2020). However, this  may also introduce subscription costs and requires reliable internet connectivity and cybersecurity measures for secure operation. 

Model Development Workflow

AI models are generally developed through data collection, preprocessing, feature extraction, model training, validation, testing, deployment, and continuous optimization. Performance of prediction models depend largely on data quality, feature normalisation, and the model chosen in the first place (Nashruddin et al., 2025).

3.3 Current AI Approaches and Technologies

Different AI techniques are used in HVAC systems depending on the intended application. Applications such as prediction, optimization, system control, fault detection, and energy management require different AI methods. Traditional ML technologies are still widely used because of their simplicity and interpretability. However, as HVAC systems have become more complex and IoT devices generate larger amounts of data, techniques like DL and RL have gained prominence. The table below summarizes the main AI techniques reported in the literature and their common applications in HVAC systems.

Table 1: Major AI approaches reported in the HVAC literature, by application. (Lu et al., 2025; Manmatharasan et al., 2025; Kannan et al., 2020; Lee & Chen, 2022; Nashruddin et al., 2025; Lee & Tsai, 2020)

3.4 Data Requirements and Method Selection

The performance of an AI-based HVAC system depends on both the algorithm, and the quality, quantity and variety of the data it receives. Data collected from IoT sensors, smart meters, weather services, occupancy sensors, and BMS provide the information needed for intelligent decision-making and system optimization. This data typically covers:

• Indoor temperature
• Indoor humidity
• Outdoor weather conditions
• Occupancy information
• Energy consumption records
• Equipment operating status
• Indoor air quality measurements
• User preference and comfort settings

Some AI models are trained using very large datasets. For example, one cloud-based study developed energy-saving recommendations using operational data collected from 37,748 AC units (Kannan et al., 2020).

The choice of an AI technique depends on the type and amount of available data. Traditional ML methods such as SVM, decision trees (DeT), and random forests (RF) are commonly used for energy prediction, load forecasting, fault detection, and system optimization (Nashruddin et al., 2025; Lu et al., 2025). DL methods, LSTM networks specifically, need large time-series datasets to learn changes in weather, occupancy, and energy demand over time. RL learns a control policy through interactions with an environment. This process may require substantial training data or simulation and can create safety concerns when learning directly in occupied buildings (Lu et al., 2025). MPC combines prediction and optimization to plan future control actions. In current HVAC systems, MPC is often coupled with AI techniques like ML and DL to improve prediction accuracy and system adaptability, this makes it one of the most effective control methods (Lee & Tsai, 2020). Hybrid approaches like combining ML with fuzzy logic, genetic algorithms (GA), or Particle Swarm Optimization can further improve system performance (Lu et al., 2025). 

Overall, the most suitable AI method depends on data availability, computing resources, and the application requirements. Among the available methods, AI-based MPC is currently among the most practical solutions (Lee & Chen, 2022). Hybrid frameworks hold promising results as they improve prediction accuracy and control performance, although their effectiveness depends on the application and available computational resources. While DL, RL, transfer learning (TL), federated learning (FL), edge AI, and occupant-centric control hold the potential, hence drive future research (Lu et al., 2025).

4. Key Findings and Inference

This section compares several studies to understand how AI techniques employed in HVAC systems have developed. It examines the most widely used control strategies, applications and reported results. Comparing these findings of different review papers also helps identify the AI techniques and research trends that appear consistently across studies, providing a better understanding of how the field has evolved. 

4.1 Comprehensive Comparison of AI-Based HVAC Studies

Table 2 presents a comparative analysis of the selected review papers, shedding light on their research focus, AI techniques, results, and reported performance improvements. This comparison provides an overview of current advancements in AI-driven HVAC systems and identifies the strengths and contributions of each study. 

Table 2: Comparative summary of key review studies on AI applications in HVAC systems

4.2 AI Techniques: Energy-Saving Potential vs Legacy Building Implementation Difficulty 

Figure 2 provides a comparative overview of the primary AI techniques discussed in the selected studies by relating their reported maximum energy-saving potential to the inferred estimated difficulty of implementation in legacy buildings. The energy-saving values are based on either experimentally reported results or representative ranges summarized in the respective review papers, while the implementation difficulty is qualitatively assessed from factors such as retrofit requirements, sensor infrastructure, computational complexity, and integration with existing HVAC systems. The figure also highlights that simpler AI methods are easier to implement in existing buildings but usually provide moderate energy savings. In contrast, advanced techniques such as predictive control, DL, and RL can achieve much higher energy savings, although they require more complex implementation and greater computational resources. 

Figure 2: Plot of energy-savings and implementation difficulty of AI techniques used in the reviewed studies.

5. Challenges, Limitations, and Future Directions

The selected literature indicates that AI-based approaches have potential to improve HVAC energy management and operational performance. The reviewed studies consistently show that AI-based control performs better than conventional rule-based and PID controllers in terms of energy efficiency, thermal comfort, fault detection, and system reliability. However, several challenges still limit its large-scale implementation. 

The major challenge is data availability and quality. AI models depend on accurate data collected from sensors, BMS, and IoT devices. Many legacy buildings lack this infrastructure, making it difficult to collect reliable data for training AI models (Nashruddin et al., 2025). Retrofitting such buildings with modern sensors, communication networks, and automation systems is often expensive and technically challenging. In addition, differences between building automation platforms create interoperability issues, while concerns over data privacy and cybersecurity further slow down AI adoption (Lu et al., 2025). 

Generalization is another recurring gap. Most models are trained on one building or one climate and often perform poorly when applied to different buildings or environmental conditions (Lu et al., 2025). For this reason, TL and domain adaptation have become important research areas, as they allow knowledge gained from one building to be adapted to another with less retraining.

Improving energy efficiency should not compromise occupant comfort. Several studies considered comfort using indoor temperature stability, humidity control, Predicted Mean Vote, and Predicted Percentage Dissatisfied(Homod et al., 2012). AI techniques such as AI-based MPC, RL, fuzzy logic, and occupant-aware optimization consistently maintained or improved comfort while reducing energy use. Quantitative data supports this claim: Lee & Tsai (2020) reported a 9.12% gain in EER and 7.37% in CSPF without sacrificing indoor stability; Kannan et al. (2020) reached a 57.38% median energy saving by incorporating occupant preferences; and Lee & Chen (2022) found AI-based strategies typically cutting energy use by 8.8% to 25.5% while comfort held steady or improved.

Future research should focus on improving the practical adoption of AI in HVAC and air-conditioning systems through real-world validation, especially in legacy buildings. Greater emphasis is needed on integrating AI with IoT sensors, DT, Building Information Modeling , and 5G-enabled smart infrastructure for real-time monitoring and adaptive control. Advanced techniques such as TL can adapt knowledge from one building to another with less retraining, while FL can train models across multiple buildings without directly sharing sensitive data. Explainable AI can make AI decisions easier to understand and increase user trust, while novel neural network architectures can improve prediction and control for complex HVAC conditions. Surrogate models can provide faster approximations of complex building or HVAC simulations, supporting quicker optimization. However, these approaches still require further research to address limited historical data, differences between buildings, privacy concerns, model complexity, and real-world validation. Researchers should also develop multi-objective optimization frameworks that simultaneously consider energy efficiency, thermal comfort, indoor air quality, and carbon emissions. Occupant-aware control using richer behavioural data, computer vision, and personalized comfort models can further enhance system performance. Finally, more studies should evaluate AI solutions across diverse building types and climates, while addressing security, uncertainty, fault diagnosis, commercialization, and seamless integration with existing HVAC infrastructure.

6. Conclusion

This review shows that AI-based methods  have the potential to improve the adaptability and energy performance of HVAC systems. Unlike most of the traditional controllers that rely on fixed schedules and predefined rules, AI-based systems learn from real-time data and continuously adjust their operation to meet changing building conditions. The reviewed studies demonstrate the successful use of AI-based methods for energy management, thermal comfort optimization, occupancy prediction, fault detection, and predictive maintenance (Lu et al., 2025).

Several studies reported noticeable improvements in HVAC performance, with higher EER, improved energy, and efficiency metrics, while maintaining or improving occupant comfort (Lee & Tsai, 2020). The integration of cloud computing, IoT, and DT technologies has further increased the practicality and effectiveness of AI-driven HVAC systems(Lu et al., 2025). 

Despite these benefits, several challenges remain. Data quality, limited sensing infrastructure, interoperability between different platforms, cybersecurity concerns, and poor model generalization continue to limit wider adoption. In addition, implementing AI in legacy buildings is difficult because many lack the sensors, communication networks, and BMS required for intelligent control, making retrofitting expensive and technically challenging (Akbulut et al., 2025).

Future research should focus on developing more transparent, secure, and adaptable AI systems through the integration of DT, IoT, Building Information Modeling, renewable energy, and smart grids. Technologies such as Explainable AI, FL, TL, and multi-agent RL can further improve system reliability and scalability. As these technologies continue to develop, AI-driven HVAC systems are expected to play an important role in reducing building energy consumption, lowering carbon emissions, and supporting sustainable buildings(Lu et al., 2025). 

7. References

Adelekan, D. S., Ohunakin, O. S., & Paul, B. S. (2022). Artificial intelligence models for refrigeration, air conditioning and heat pump systems. Energy Reports, 8, 8451–8466. https://doi.org/10.1016/j.egyr.2022.06.062 

Akbulut, L., Taşdelen, K., Atılgan, A., Malinowski, M., Coşgun, A., Şenol, R., Akbulut, A., & Petryk, A. (2025). A Systematic Review of Building Energy Management Systems (BEMSs): Sensors, IoT, and AI Integration. Energies, 18(24), 6522. https://doi.org/10.3390/en18246522 

Chua, K. J., Chou, S. K., Yang, W. M., & Yan, J. (2013). Achieving better energy-efficient air conditioning – A review of technologies and strategies. Applied Energy, 104, 87–104. https://doi.org/10.1016/j.apenergy.2012.10.037 

Gharbi, A., Ayari, M., Albalawi, N., Touati, Y. E., & Klai, Z. (2025). Intelligent HVAC Control: Comparative Simulation of Reinforcement Learning and PID Strategies for Energy Efficiency and Comfort Optimization. Mathematics, 13(14), 2311. https://doi.org/10.3390/math13142311 

Hauashdh, A., Nagapan, S., Jailani, J., & Gamil, Y. (2024). An integrated framework for sustainable and efficient building maintenance operations aligning with climate change, SDGs, and emerging technology. Results in Engineering, 21, 101822. https://doi.org/10.1016/j.rineng.2024.101822 

Homod, R. Z., Mohamed Sahari, K. S., Almurib, H. A. F., & Nagi, F. H. (2012). RLF and TS fuzzy model identification of indoor thermal comfort based on PMV/PPD. Building and Environment, 49, 141–153. https://doi.org/10.1016/j.buildenv.2011.09.012 

Kannan, R., Roy, M. S., & Pathuri, S. H. (2020). Artificial Intelligence Based Air Conditioner Energy Saving Using a Novel Preference Map. IEEE Access, 8, 206622–206637. https://doi.org/10.1109/ACCESS.2020.3037970 

Lee, D., & Chen, L. (2022). Sustainable Air-Conditioning Systems Enabled by Artificial Intelligence: Research Status, Enterprise Patent Analysis, and Future Prospects. Sustainability, 14(12), 7514. https://doi.org/10.3390/su14127514 

Lee, D., & Tsai, F.-P. (2020). Air Conditioning Energy Saving from Cloud-Based Artificial Intelligence: Case Study of a Split-Type Air Conditioner. Energies, 13(8), 2001. https://doi.org/10.3390/en13082001 

Lu, S., Zhou, S., Ding, Y., Kim, M. K., Yang, B., Tian, Z., & Liu, J. (2025). Exploring the comprehensive integration of artificial intelligence in optimizing HVAC system operations: A review and future outlook. Results in Engineering, 25, 103765. https://doi.org/10.1016/j.rineng.2024.103765 

Manmatharasan, P., Bitsuamlak, G., & Grolinger, K. (2025). AI-driven design optimization for sustainable buildings: A systematic review. Energy and Buildings, 332, 115440. https://doi.org/10.1016/j.enbuild.2025.115440 

Nashruddin, S. N. A. M., Salleh, F. H. M., Sulaiman, R., Zaiddy, M. I. B. M., Ramasamy, A., & Yahya, A. A. (2025). AI-driven optimization of air-conditioning systems in legacy Buildings: Evaluating machine learning models for enhanced energy efficiency. Journal of Building Engineering, 108, 112839. https://doi.org/10.1016/j.jobe.2025.112839 

Raza, M. Q., & Khosravi, A. (2015). A review on artificial intelligence based load demand forecasting techniques for smart grid and buildings. Renewable and Sustainable Energy Reviews, 50, 1352–1372. https://doi.org/10.1016/j.rser.2015.04.065 

About the author

Kethan Sai N

Kethan is a Grade 12 student majoring in Physics, Mathematics, and Chemistry, with Physics as his favorite subject. His curiosity for how the world works has led him to explore robotics, machine learning, and artificial intelligence, alongside a growing passion for sustainability. His research focuses on how AI and ML can be leveraged to build smarter, more sustainable systems, an interest reflected in his current paper on AI-driven approaches to energy-efficient HVAC systems.

Outside academics, he enjoys coding and losing himself in documentaries and videos exploring scientific mysteries and unknowns, a habit that fuels his constant search for new ideas and questions worth solving. He is equally active on the field and court, playing football, badminton, and basketball, believing that sharp thinking is best paired with an active body.

Combining a scientist’s curiosity with an engineer’s drive to build, he is passionate about using emerging technologies not just to innovate, but to create a more sustainable future for everyone. Looking ahead, he aims to build a career in robotics, AI, and machine learning, driven by a desire to turn these technologies into real, sustainable solutions for the world’s biggest challenges.

The post How AI is Transforming HVAC Systems: A Review of Smart Energy-Saving Techniques appeared first on Exploratio Journal.

Abstract

Invasive brain-computer interfaces (BCIs) have the ability to restore deprived functions in people with severe paralysis such as communication and movement. The longevity of said devices hinge on whether implanted electrodes can integrate into the brain’s system and record neural signals over time. However, the brain often walls off these rigid, high-density architectures, which records the best, causing chronic neuroinflammation, degrading signal recordings over time.

In this review, we categorize different neural interfaces into four distinct stages, each with their benefits and drawbacks: Intruder, Neighbor, Partner, and Symbiotic Organ. The Intruder stage offers the highest quality recordings, trading off disturbing tissues locally in the brain due to the rigidity of electrode arrays. The Neighbor stage utilizes flexible and mesh electronics that are more compatible with the softness of brain tissue. The Partner stage employs new materials: graphene, carbon nanotubes, and liquid metals, which constitutes functions beyond basic electrical recording. In the Symbiotic Organ stage, an interface would integrate itself  into the neural tissue and has the ability of self reparation, allowing it to function as a part of the nervous system instead of a separate implant.

This paradigm is not a strict categorization of different device, it offers a way to label different systems to their strengths and limitations. Some devices are often hybridized combinations of different stages, and newer designs are not automatically superior to predecessors. Issues with longevity are not limited to biological reactions such as inflammation or scarring; it can also derive from abiotic failures, such as connector failure, insulation damage, and material degradation. On the other hand, rigid arrays holds some of the strongest long-term human records in signal quality. Overall, optimal systems will come from a combination of co-designing materials as well as adaptive computation together. The goal is not only to record stronger signals, or more electrode channel counts, but also to build interfaces that can remain useful inside the brain for many years.

Fig. 1. The four-stage framework for the evolution of invasive neural interfaces: Intruder, Neighbor, Partner, and Symbiotic Organ. Image generated by the author using ChatGPT (OpenAI).

1. Introduction

Over the past decade, invasive BCIs have become physical experiments. They have restored basic functions, such as control and communication for individuals inflicted with severe paralysis, including cursor operations through neural control (Hochberg et al., 2006; Simeral et al., 2011), high-performance clinical neural prostheses (Gilja et al., 2015), reach-and-grasp with a robotic arm (Hochberg et al., 2012), brain-controlled functional electrical stimulation for reanimating the limb (Ajiboye et al., 2017), and high-bandwidth brain-to-text communication (Willett et al., 2021).

How much of this bottleneck is physical rather than computational remains contested. One line of work treats signal degradation as a hard ceiling: once units are lost to chronic tissue response or device failure, no decoder can recover information absent from the recording (Barrese et al., 2013; Chestek et al., 2011; Salatino et al., 2017). A second line demonstrates that a great deal of apparent degradation is nonstationarity rather than information loss and that closed-loop adaptation (Orsborn et al., 2014) and variability-robust decoder architectures (Sussillo et al., 2016) absorb substantially more drift than the first framing implies. We take the acquisition layer as the primary determinant of longevity while noting that its boundary with computation is precisely what remains unsettled, a point we return to in section 3.3. In the end, how well the systems architecture integrates into the brain is the most important factor: signal quality, longevity, and clinical usefulness all suffer if the connection degrades.

This constraint limits the modality of designs of Brain Computer Interfaces. Non-invasive systems such as electroencephalography (EEG) are safer and more accessible, but volume conduction and the skull’s spatial filtering cap their spatial resolution and signal-to-noise ratio, limiting the neural detail recoverable at the scalp (Srinivasan et al., 1998). Invasive interfaces instead capture local field potentials and spiking activity with millisecond precision and cellular-scale specificity, which is what makes high-performance applications possible (Hochberg et al., 2006; Wolpaw et al., 2002). Rigid intracortical arrays such as the Utah Array (Maynard et al., 1997) and high-density silicon probes of the Neuropixels class (Jun et al., 2017; Steinmetz et al., 2021) are the pinnacle of highest-fidelity approach.

Increased signal fidelity will often sacrifice clinical usefulness after an extended period of time. Many high-density systems that rely on rigid, silicon-based construction, whose elastic modulus are several times bigger in magnitude in comparison to brain tissue. The mechanical mismatch amplifies micromotion-induced strain and chronic injury (Minev et al., 2015; Salatino et al., 2017), which in turn drive persistent inflammation, glial scarring (gliosis), neuronal loss near the implant, and progressive signal decline (Minev et al., 2015; Salatino et al., 2017). This is the paradox at the core of BCI hardware: the structures best suited to high-precision electrophysiology are often the ones most prone to causing long-term biological disruption (Minev et al., 2015; Salatino et al., 2017). The field has therefore started a shift opted in more compatible materials and designs which are less physically disruptive to the brain. As a result, these systems counteract mechanical mismatch and immune response without sacrificing the signal quality needed for useful recording.

This framework is not strictly in a chronological timeline. Some devices hybridize between different stages: Neuralink combines high channel count with flexible materials such as threads and robotic insertion, while mesh electronics may behave as Neighbor or Partner systems depending on whether the emphasis is mechanical integration or multifunctional sensing. The categories are therefore not fixed rules, rather flexible, useful labels.

To organize this response, we introduce a four-stage framework for the evolution of invasive neural interfaces: Intruder, Neighbor, Partner, and Symbiotic Organ (see Fig. 1). The sequence describes a progressive and linear shift from rigid foreign implants toward biologically integrated, adaptive bioelectronic systems, and it gives us a consistent way to compare how each generation navigates the paradox between signal fidelity and biological integration.

2. Evolution of Invasive Neural Interface Technologies

Invasive neural interfaces have developed through distinct paradigms, each a different attempt to reconcile high-fidelity recording with long-term biological compatibility.

2.1 Rigid High-Density Electrode Arrays: The “Intruder”

The Intruder paradigm is often categorized by a pursuit of high-fidelity recording, utilizing rigid, silicon-based microelectrode arrays to receive signals. Devices such as the Utah Array and, more recently, Neuropixels probes are the pinnacle of  spatial resolution, channel counts, and temporal precision, properties that can easily detect single-neuron activity and dynamics in population(Jun et al., 2017; Maynard et al., 1997; Steinmetz et al., 2021). Other devices such as Neuralink have scaled this approach along, further raising the ceiling of channel counts and automating implantation while maintaining signal quality transfusion. (Musk & Neuralink, 2019). However, it must be said that Chronic instability is not an inherent property of rigid devices. Neuropixels 2.0 was designed explicitly around stable long-term recording, pairing a miniaturized probe with methods for tracking the same units across multiple sessions, for an extended period of time in various conditions (Steinmetz et al., 2021), a rigid device pursuing the stability as their objective is usually credited to compliant ones. The paradigms are connecting shared goals from opposite starting points, which is why they are better labeled as design logic instead of chronology. Mechanical rigidity is a key focus: it gives positional stability during implantation and precise targeting between cortical layers, making these arrays the gold standard for short- to medium-term performance in both experimental and clinical settings. The Utah Array and Neuropixels have high signal fidelity, but they often sacrifice biocompatibility.

Additionally, the same rigidity of the paradigm’s design focus is their defining limitation. Because the elastic modulus of silicon-based probes are magnitudes bigger than the brain cells, the implant sits in permanent state of conflict with the soft, viscoelastic tissue around it. Over time this provokes a chronic foreign-body immune response, persistent inflammation, glial scarring, neuronal loss at the electrode surface, and micromotion-induced strain (Minev et al., 2015; Salatino et al., 2017), which steadily lowers the quality of recordings over a long period of time (Chestek et al., 2011; Simeral et al., 2011).

This causal story is, however, is an incomplete summary of attempts at correcting these flaws. Barrese et al.’s (2013) failure-mode analysis of silicon arrays in vivo attributed a large fraction of failures not to the foreign-body response but to abiotic modes, including connector and percutaneous pedestal failure, insulation cracking and delamination, and progressive material degradation. Their retrospective analysis of seventy-eight silicon arrays implanted in twenty-seven macaques found that most complete recording failures were not attributable to the foreign-body response at all. Acute mechanical failure was the single largest class at 48 percent, and 83 percent of those traced to the percutaneous connector rather than to anything occurring at the electrode-tissue boundary; material failures, including broken tips, insulation leakage, and parylene cracking and delamination, accounted for most of the remainder. Grossly observable biological failures made up 24 percent, the most common single mechanism being progressive meningeal encapsulation (14.5%). Mechanical and material modes together thus account for roughly three-fifths of failures (Barrese et al., 2013). A large share of array loss therefore occurs in packaging, wiring, and connector hardware, components that compliance-based redesign of the electrode does not touch.

One qualification protects this claim from over-reading. Barrese et al. (2013) define failure as the absence of extractable action potentials across all channels of an array, so their categories register catastrophic device death rather than the gradual attrition of usable units. Progressive, gliosis-driven decline in unit yield, the degradation most consequential for BCI performance and precisely what the Neighbor paradigm targets, would largely not count as a failure under this definition. The 24 percent figure therefore bounds the biological contribution to total array loss, not to signal degradation. This is the sharpest available illustration of why sources in this literature so often appear to disagree: Barrese et al. (2013) measure when a device dies, whereas chronic-yield studies (Chestek et al., 2011; Simeral et al., 2011) measure how well it performs while alive, and conclusions drawn from one cannot be transferred to the other without argument.

The Intruder paradigm’s limitation is not simply that it is too stiff: it fails along at least two partly independent axes, and the more frequent one is not biological. The Intruder paradigm does not resolve the paradox; it accepts biological instability as the price of maximal instantaneous fidelity.

Fig. 2. Five high-density planar silicon probes ordered by shank width (a 32-channel NeuroNexus polytrode, 128- and 255-channel NeuroSeeker probes, and two Neuropixels variants), with edge recording sites shown in green and center sites in red. Reproduced from Fiáth et al. (2021), Figure 1 (CC BY 4.0).

Fig. 3. Scanning electron micrograph of the Utah Intracortical Electrode Array, a 10 × 10 grid of 1.2-mm silicon needles on a 4.2-mm-square base. Reproduced from Maynard et al. (1997), Figure 1.

2.2 Flexible and Mesh Electronics: Transitioning to a “Neighbor”

The Neighbor paradigm softens the interface to better integrate to the needs of neighbouring brain tissues. Work on syringe-injectable electronics and macroporous mesh architectures showed that ultraflexible, polymer-based electrode networks integrate more seamlessly into tissue, attenuating the chronic immune response and enabling stable recording over extended timescales in animal models (Liu et al., 2015; Xie et al., 2015). By lowering bending stiffness to approach the mechanical compliance of brain tissue, these systems reduce the micromotion-induced damage that defines the Intruder problem (Jeong et al., 2015; Zhang et al., 2020).

Bio-integration is just part of the upside. The porous architecture of mesh electronics allows tissue interpenetration and vascularization, so the implant will not clash and compete(Xie et al., 2015). Neighbor designs therefore prioritize long-term integration and stability over recording density, and this comes at a considerable cost. Flexible architecture, however, raises the issue of not hitting the necessary threshold of electrode counts, and because they cannot penetrate tissue unaided, delivery depends on syringe injection (Liu et al., 2015) or temporary stiffening shuttles, approaches that will inevitable complicate the surgical procedure and whose procedures to the human cortex remains untested. Even so, the flexible and mesh electronics architecture serves as proof of principle that long-term neural implants are viable  without defects such as tissue disruption. This decouples one side of the paradox entirely. Recent reviews furthers this idea, arguing that longevity and soft, flexible structures go hand in hand, rather than opting for rigid recording elements alone (E. Song et al., 2020).

Fig. 4. Syringe-injectable mesh electronics: the mesh design and single unit cell, injection into the mouse brain, and a comparison of the open mesh with a conventional flexible thin-film probe. Reproduced from Zhou et al. (2017), Figure 1.

2.3 Novel Material-Based Electrodes and Hybrid Designs: The Functional “Partner”

The Partner paradigm does not further increase along the biocompatibility pathway, instead it diverges. Where mesh electronics optimize a single figure in its metric, longevity of passive recording systems(Xie et al., 2015), transparent graphene arrays, an example of a partner paradigm, are built to increase a different metric altogether, combining electrophysiology with simultaneous optical access to the same circuit (Kuzum et al., 2014). The two paradigms, Partner and Neighbour, are not contrasting solutions; instead, they sought to answer differently about what an interface is for, and who they are designed for. Resulting in the direct performance comparison between them as largely meaningless, arbitrary numbers. This stage is defined by materials such as graphene, carbon nanotubes, conductive polymers, and liquid metals that make up the interface. Strictly chosen for a combination of purposes: conductivity, optical transparency, stretchability, and chemical sensitivity (Chapman et al., 2018; Hong & Lieber, 2019; E. Song et al., 2020). Partner interfaces excel through its multimodal and biomimetic interactions rather than passive electrical functions represented by previous two paradigms.

Materials illustrate this shift. Compounds such as transparent graphene electrodes permit electrical recording and optical access to the same tissue at once, permitting an alternative pathway to “view” the brain. Whilst, also enabling simultaneous electrophysiology and two-photon calcium imaging of a single circuit, a pairing that opaque metal electrodes from previous paradigms physically obstruct (Kuzum et al., 2014). Additionally, carbon nanotube coatings enlarge effective surface area whilst lowering impedance, improving electrode-tissue coupling and recording quality (Keefer et al., 2008). Liquid metal conductors add extreme deformability while maintaining electrical continuity, so the interface can track tissue motion (Wen et al., 2019). Together, these materials support researchers in a multimodal approach to record the brain’s activity. By measuring electrical activity and neurochemical signals at once, recent reviews stress that interface performance now hinges on micro and nanoscale engineering, not just electrode composition alone (Chapman et al., 2018; Hong & Lieber, 2019).

This expansion in multimodal approach introduces its own functional costs. Fabrication is often intricate and less feasible, long-term material stability under chronic in vivo conditions remains under investigation, and integration with conventional electronics presents abnormal challenges. However, it must be said that Partner systems expand what the interface can measure, but at the cost of fabrication, packaging, long-term reliability, and clinical translation.

Fig. 5. A fully flexible Parylene C neural probe whose PEDOT/silica-nanoparticle-coated chemical sites electrically release glutamate (Glu) or GABA while neighboring microelectrodes record the resulting increase or decrease in spike rate in rat cortex. Reproduced from Malekoshoaraie et al. (2024) (CC BY 4.0).

2.4 Emerging Injectable and Self-Adaptive Interfaces: Toward a “Symbiotic Organ”

The current frontier reframes the interface as an adaptive extension of the nervous system rather than a static implant. The Symbiotic Organ paradigm proposes a structure of self-healing bioelectronic materials, injectable or self-deploying architectures, and closed-loop systems that respond to changes in neural and tissue state. This vision inherits the Partner stage’s costs and compounds them. The fabrication complexity and uncertain chronic material stability that constrain graphene, nanotube, and liquid-metal interfaces (Keefer et al., 2008; Kuzum et al., 2014; Wen et al., 2019) do not disappear when those materials are asked to additionally self-heal, self-deploy, and operate under closed-loop control; they multiply, and they are joined by power and telemetry demands that passive recording never faced. Self-healing interfaces, for instance, restore electrical and mechanical integrity after mechanical damage, a route to durability that does not depend on avoiding damage in the first place (K.-I. Song et al., 2020).

Dynamic adaptation is what distinguishes this stage. Where earlier paradigms rely on fixed geometries and static material properties, these systems continuously renegotiate the balance between signal quality and biological stability; coupling adaptive decoders with closed-loop control lets the interface respond to neural variability and drift in real time (Orsborn et al., 2014; Sussillo et al., 2016). In principle, an interface that evolves alongside the tissue it inhabits dissolves the paradox rather than trading around it, and bioinspired neuron-like designs push toward that goal by minimizing the structural distinction between device and neural tissue (Yang et al., 2019).

For now, this paradigm is still only a proof of concept, with unresolved obstacles in power delivery, wireless telemetry, computational load, long-term reliability, and safety. They define a long-term direction for the field rather than a deployable technology, and the gap between experimental demonstration and clinical translation remains wide.

Fig. 6. Conceptual vision of a symbiotic neural interface combining adaptive materials, tissue integration, intelligent on-chip electronics, and a wireless ecosystem. Image generated by the author using ChatGPT (OpenAI).

3. Discussion and Future Directions

The move from rigid intracortical arrays toward adaptive, tissue-integrated systems reflects a shift in the field’s underlying design philosophy. Early BCI work was driven by maximal signal fidelity and spatial resolution, predicated on the assumption that the highest-quality neural data would translate into robust decoding and clinical utility. Decades of chronic-implant experience overturned that assumption: performance depends not on instantaneous signal quality but on the long-term stability of the interface as a coupled biological-technical system. That realization reoriented the field’s priorities from absolute electrophysiological performance toward sustained harmony among tissue, materials, and computation, the transition the Intruder-to-Symbiotic-Organ progression describes.

3.1 Reconciling the Engineering Paradox: A Multi-Dimensional Optimization Problem

The paradox, that high-precision electrophysiology favors rigid, dense structures while long-term biocompatibility demands soft, compliant, adaptive ones has never been solved outright. Each stage handles this trade-off differently.

The Intruder paradigm buys fidelity by absorbing the biological cost: Utah arrays and Neuropixels probes set the benchmark for recording density and temporal resolution, and they accept chronic tissue disruption as the consequence of achieving it. But the accounting is less clean than the paradox implies. If abiotic failure modes contribute as heavily to array loss as Barrese et al. (2013) report, part of what the Intruder paradigm “pays” is not a biocompatibility cost at all, and no amount of movement along the compliance axis will refund it. The Neighbor paradigm moves the balance the opposite way, matching interface mechanics to tissue to lower inflammation and extend recording lifetimes, at the expense of electrode density, targeting precision, and spatial stability over time. The Partner paradigm widens the problem rather than picking a side: by adding optical transparency, neurochemical sensing, or stretchable multimodality, it redefines “performance” beyond electrical metrics, and in doing so exchanges a binary rigidity-versus-compliance choice for a harder question of fabrication and integration. The Symbiotic Organ paradigm attempts to remove the trade-off entirely, letting self-healing materials, self-deploying architectures, and closed-loop control adapt the interface to changing demands, an approach that, for now, exchanges the old paradox for unresolved problems of power, reliability, and safety. Read together, the four stages show the paradox migrating from a fixed constraint into a multi-dimensional optimization problem, not disappearing.

3.2 Insights from Cross-Technology Comparison and Application-Driven Design

No single paradigm dominates across every performance dimension. Rigid arrays lead on fidelity and spatial resolution, flexible and mesh electronics on chronic biocompatibility, material-based hybrids on functional diversity, and symbiotic concepts on adaptability and autonomy. The absence of a universally superior interface indicates that the field is bounded by fundamental trade-offs, not merely by incremental engineering inefficiencies.

Two implications follow. First, progress will come from context-specific optimization rather than a single “best” interface: applications demanding fine motor control or high-bandwidth speech decoding will likely continue to require high-density invasive arrays (Musk & Neuralink, 2019; Pandarinath et al., 2017; Willett et al., 2023), while long-term neuromodulation or monitoring may favor softer, lower-density systems built for chronic stability. Second, this makes application-driven design central: interface architecture should be matched to the physiological, temporal, and functional requirements of the specific use case rather than optimized in the abstract.

3.3 Persistent Field-Wide Challenges and the Need for Co-Design

Several challenges cut across all four paradigms. Long-term biostability is the central one, spanning not only immune response and tissue remodeling but also material degradation, delamination, and electrochemical drift over multi-year timescales. Manufacturing scalability and reproducibility form a parallel obstacle, especially for complex material systems that perform well in the laboratory but resist high yield consistent industrial fabrication. Power delivery and data telemetry impose further limits, particularly for wireless or minimally invasive implants, where energy efficiency and thermal safety become hard constraints.

Computation faces its own version of these problems. Representational drift, non-stationary signal statistics, and user-specific variability complicate decoder design and limit cross-session generalization. Where the ceiling on algorithmic mitigation actually sits is unresolved: results on closed-loop decoder adaptation (Orsborn et al., 2014) and variability-robust architectures (Sussillo et al., 2016) show that computation absorbs more instability than hardware-centric accounts assume, while chronic-yield data (Chestek et al., 2011; Simeral et al., 2011) show it cannot manufacture units that are no longer recorded. The disagreement is not merely academic, because the two positions imply different investments, one in materials, the other in algorithms, and neither literature can settle it alone. This is the strongest argument for co-design: the allocation question is unanswerable unless hardware and computation are developed and evaluated jointly. This interdependence is the argument for co-design, integrating hardware, materials science, and computation from the earliest stages of development rather than optimizing each in isolation. Recent reviews of electronic and flexible neural interfaces reach the same conclusion: future progress depends on tighter integration across materials, device engineering, and system-level signal processing (Hong & Lieber, 2019; E. Song et al., 2020; Zhang et al., 2020).

3.4 Global Perspectives and Collaborative Trajectories

Regional programs bring complementary strengths. In the United States, academic and industrial work often prioritizes high-performance invasive systems and advanced decoding, driven by clinical-translation initiatives and defense funding; Neuralink exemplifies the emphasis on channel count and surgical automation, while others pursue hybrids that balance performance against longevity. In China, research increasingly emphasizes materials science, flexible electronics, and minimally invasive platforms, drawing on strengths in large-scale manufacturing and material synthesis. European groups tend to lead in soft bioelectronics and neuromodulation, prioritizing safety, regulatory compatibility, and long-term implantation. The complementarity suggests that major breakthroughs are more likely to come from cross-border collaboration than from any single region’s dominance.

3.5 Near-Term Breakthroughs: Hybrid Systems at the Interface of Paradigms

A more cautious near-term expectation is that clinically relevant progress will come from partial hybrids rather than fully symbiotic systems. These platforms aim to combine signal quality sufficient for meaningful decoding with improved chronic stability and manageable fabrication complexity. Fully symbiotic systems, though the ultimate goal, still face barriers in power management, reliability, and regulatory approval that make near-term deployment unlikely. The next wave of translational impact is therefore more likely to come from incremental but durable gains in interface longevity, paired with adaptive decoding and closed-loop control.

4. Conclusion

High-performance BCIs could change neuroscience, medicine, and human-machine interaction, but they still face a basic problem: the clearest signals often come from devices that are difficult for brain tissue to tolerate over time. This review traced the field’s response through four stages: rigid Intruder systems, flexible Neighbors, multifunctional Partners, and emerging Symbiotic Organ concepts, each addressing the paradox through a different set of trade-offs. Chronic stability has improved, functional modalities have expanded, and adaptive behaviors have entered the design space, yet no existing technology optimizes every critical dimension at once. Reaching genuinely symbiotic interfaces will require sustained, interdisciplinary work across materials science, system integration, computational neuroscience, and ethical considerations. Four limitations in particular remain unresolved, and each cuts across all stages of the framework. First, an abiotic failure ceiling: because connector, insulation, and material degradation account for a substantial share of chronic array loss (Barrese et al., 2013), improvements in biocompatibility cannot by themselves extend device lifetime. Second, a translation gap: the chronic-stability case for compliant interfaces rests largely on rodent work (Liu et al., 2015; Xie et al., 2015), while the longest human recordings remain those of rigid arrays (Simeral et al., 2011). Third, power delivery and telemetry impose hard energy and thermal constraints on precisely the wireless, minimally invasive systems the Symbiotic Organ concept requires. Fourth, the division of labor between interface and algorithm is still contested (Chestek et al., 2011; Orsborn et al., 2014; Sussillo et al., 2016), leaving open how much instability future systems must engineer away rather than decode around. The clear direction of travel, however, is toward tissue-aware, adaptive design, interfaces that stop behaving as foreign bodies and begin functioning as durable, integrated collaborators within the nervous system.

References

Ajiboye, A. B., et al. (2017). Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: A proof-of-concept demonstration. The Lancet, 389(10081), 1821-1830. https://doi.org/10.1016/S0140-6736(17)30601-3

Barrese, J. C., et al. (2013). Failure mode analysis of silicon-based intracortical microelectrode arrays in non-human primates. Journal of Neural Engineering, 10(6), 066014. https://doi.org/10.1088/1741-2560/10/6/066014

Chapman, C. A. R., et al. (2018). Multifunctional neural interfaces for closed-loop control of neural activity. Advanced Functional Materials, 28(12), 1703523. https://doi.org/10.1002/adfm.201703523

Chestek, C. A., et al. (2011). Long-term stability of neural prosthetic control signals from silicon cortical arrays in rhesus macaque motor cortex. Journal of Neural Engineering, 8(4), 045005. https://doi.org/10.1088/1741-2560/8/4/045005

Fiáth, R., et al. (2021). Recording site placement on planar silicon-based probes affects signal quality in acute neuronal recordings. Scientific Reports, 11, 2028. https://doi.org/10.1038/s41598-021-81127-5

Gilja, V., et al. (2015). Clinical translation of a high-performance neural prosthesis. Nature Medicine, 21(10), 1142-1145. https://doi.org/10.1038/nm.3953

Hochberg, L. R., et al. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature, 442(7099), 164-171. https://doi.org/10.1038/nature04970

Hochberg, L. R., et al. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, 485(7398), 372-375. https://doi.org/10.1038/nature11076

Hong, G., & Lieber, C. M. (2019). Novel electrode technologies for neural recordings. Nature Reviews Neuroscience, 20(6), 330-345. https://doi.org/10.1038/s41583-019-0140-6

Jeong, J.-W., et al. (2015). Soft materials in neuroengineering for hard problems in neuroscience. Neuron, 86(1), 175-186. https://doi.org/10.1016/j.neuron.2014.12.035

Jun, J. J., et al. (2017). Fully integrated silicon probes for high-density recording of neural activity. Nature, 551(7679), 232-236. https://doi.org/10.1038/nature24636

Keefer, E. W., et al. (2008). Carbon nanotube coating improves neuronal recordings. Nature Nanotechnology, 3(7), 434-439. https://doi.org/10.1038/nnano.2008.174

Kuzum, D., et al. (2014). Transparent and flexible low-noise graphene electrodes for simultaneous electrophysiology and neuroimaging. Nature Communications, 5, 5259. https://doi.org/10.1038/ncomms6259

Liu, J., et al. (2015). Syringe-injectable electronics. Nature Nanotechnology, 10(7), 629-636. https://doi.org/10.1038/nnano.2015.115

Malekoshoaraie, M. H., et al. (2024). Fully flexible implantable neural probes for electrophysiology recording and controlled neurochemical modulation. Microsystems & Nanoengineering, 10, 91. https://doi.org/10.1038/s41378-024-00685-6

Maynard, E. M., et al. (1997). The Utah intracortical electrode array: A recording structure for potential brain-computer interfaces. Electroencephalography and Clinical Neurophysiology, 102(3), 228-239. https://doi.org/10.1016/S0013-4694(96)95176-0

Minev, I. R., et al. (2015). Electronic dura mater for long-term multimodal neural interfaces. Science, 347(6218), 159-163. https://doi.org/10.1126/science.1260318

Musk, E., & Neuralink. (2019). An integrated brain-machine interface platform with thousands of channels. Journal of Medical Internet Research, 21(10), e16194. https://doi.org/10.2196/16194

Orsborn, A. L., et al. (2014). Closed-loop decoder adaptation shapes neural plasticity for skillful neuroprosthetic control. Neuron, 82(6), 1380-1393. https://doi.org/10.1016/j.neuron.2014.04.048

Pandarinath, C., et al. (2017). High performance communication by people with paralysis using an intracortical brain-computer interface. eLife, 6, e18554. https://doi.org/10.7554/eLife.18554

Salatino, J. W., et al. (2017). Glial responses to implanted electrodes in the brain. Nature Biomedical Engineering, 1, 862-877. https://doi.org/10.1038/s41551-017-0154-1

Simeral, J. D., et al. (2011). Neural control of cursor trajectory and click by a human with tetraplegia 1000 days after implant of an intracortical microelectrode array. Journal of Neural Engineering, 8(2), 025027. https://doi.org/10.1088/1741-2560/8/2/025027

Song, E., et al. (2020). Materials for flexible bioelectronic systems as chronic neural interfaces. Nature Materials, 19(6), 590-603. https://doi.org/10.1038/s41563-020-0679-7

Song, K.-I., et al. (2020). Adaptive self-healing electronic epineurium for chronic bidirectional neural interfaces. Nature Communications, 11, 4195. https://doi.org/10.1038/s41467-020-18025-3

Srinivasan, R., et al. (1998). Estimating the spatial Nyquist of the human EEG. Behavior Research Methods, Instruments, & Computers, 30(1), 8-19.

Steinmetz, N. A., et al. (2021). Neuropixels 2.0: A miniaturized high-density probe for stable, long-term brain recordings. Science, 372(6539), eabf4588. https://doi.org/10.1126/science.abf4588

Sussillo, D., et al. (2016). Making brain-machine interfaces robust to future neural variability. Nature Communications, 7, 13749. https://doi.org/10.1038/ncomms13749

Wen, X., et al. (2019). Flexible, multifunctional neural probe with liquid metal enabled, ultra-large tunable stiffness for deep-brain chemical sensing and agent delivery. Biosensors and Bioelectronics, 131, 37-45. https://doi.org/10.1016/j.bios.2019.01.060

Willett, F. R., et al. (2021). High-performance brain-to-text communication via handwriting. Nature, 593(7858), 249-254. https://doi.org/10.1038/s41586-021-03506-2

Willett, F. R., et al. (2023). A high-performance speech neuroprosthesis. Nature, 620(7976), 1031-1036. https://doi.org/10.1038/s41586-023-06377-x

Wolpaw, J. R., et al. (2002). Brain-computer interfaces for communication and control. Clinical Neurophysiology, 113(6), 767-791. https://doi.org/10.1016/S1388-2457(02)00057-3

Xie, C., et al. (2015). Three-dimensional macroporous nanoelectronic networks as minimally invasive brain probes. Nature Materials, 14(12), 1286-1292. https://doi.org/10.1038/nmat4427

Yang, X., et al. (2019). Bioinspired neuron-like electronics. Nature Materials, 18(5), 510-517. https://doi.org/10.1038/s41563-019-0292-9

Zhang, M., et al. (2020). Electronic neural interfaces. Nature Electronics, 3(4), 191-200. https://doi.org/10.1038/s41928-020-0390-3

Zhou, T., et al. (2017). Syringe-injectable mesh electronics integrate seamlessly with minimal chronic immune response in the brain. Proceedings of the National Academy of Sciences, 114(23), 5894-5899. https://doi.org/10.1073/pnas.1705509114

About the author

Anthony Qian

Anthony is a senior at Stevenson School with a strong interest in neuroscience, artificial intelligence, and brain-computer interfaces. His academic interests lie at the intersection of biology, computer science, and machine learning, particularly in understanding brain function and developing technologies that interpret neural activity and human behavior.

Through independent research and hands-on computational projects, Anthony has explored machine learning, behavioral data analysis, and neural signal processing. He is passionate about interdisciplinary research, scientific problem-solving, and applying emerging technologies to advance our understanding of the human brain.

The post From Intruders to Symbiotic Organs: The Evolution of Neural Interface Technologies for Invasive Brain–Computer Interfaces appeared first on Exploratio Journal.

Abstract

Insulation is a crucial part of the building and design process of a structure. There are countless different ways of insulating a building and many are eco-friendly to mitigate how much infrastructure affects the current climate crisis. The process of insulation can improve the energy efficiency of a building significantly. Using the West Side Tennis Club, an existing historical tennis center, as an example, efficiency of the clubhouse could be majorly improved upon through an insulation update. By simulating a whole new insulation design, the heat flow of one wall section went from about 26,520 btu/h to 4,020 btu/h. Insulation is very relevant today and more education on the topic can only improve our infrastructure and environment as a whole.

Introduction

Carbon emission plays a huge role in the built environment. Right now, the built environment is a major contributor to greenhouse gas emission (Architecture 2030). To be specific, 42% of annual global CO2 emissions come from a combination of building operations, building materials, and infrastructure (Architecture 2030). Additionally, there is major growth in the built environment and it will only continue to grow. Embodied carbon is one of the main sources responsible for global warming and infrastructure releases almost half of that carbon (Architecture 2030). That is why sustainability and environmentalism are so important to this topic and why it has such a big section in this paper. There are many low carbon materials that are also biodegradable, non-toxic, and extremely effective that can take us a few steps closer to bettering our environment.

Throughout this paper the topic of insulation is discussed in depth. The focus is on wall insulation, how it is installed and how it works, as well as its application in existing and historic buildings. General wall insulation and its purpose are covered first. Then, explanations of each insulation material as well as their pros and cons. The next section goes into detail about low-carbon insulation materials and their benefits and downsides. This paper then discusses how insulation applies to existing buildings and more specifically, how it can be used in existing historical buildings. Considering the walls in historic buildings are not always accessible, alternative insulation methods are explained as well. Finally, a case study of an existing historical building and how it modified the different kinds of insulation it required is outlined. Calculations of the building insulation statistics are included.

Insulation and Air Sealing

Insulation is an integral part of a building and its general efficiency. Insulation is what allows structures to remain warm in the winter time and cool during summer. Insulation regulates temperature throughout a structure by limiting the amount of heat loss in the walls, floors, and ceilings. Insulation in walls will go in the space between the interior and exterior walls to interrupt any unwanted air coming in or any necessary air moving out. Good insulation materials must be moisture resistant, fire resistant, strong, and stable. While there are many different materials for and methods of insulation, they all must have the strength to withstand a variety of conditions. Each building is different and will require a specific additional set of requirements for its insulation method which is why there are hundreds of different insulation systems.

Air sealing and vapor management are vital elements of the insulation process. Air sealing involves locating and filling gaps or cracks in a building’s envelope or insulation. Air sealing has countless benefits: it improves air quality, humidity, comfort, and energy efficiency of a building. Gaps or cracks in a building allow air to move in and out of the envelope which worsens airflow. Airflow is important for preventing mold and mildew or other bacteria from growing. Air sealing mends the damaged airflow and keeps moisture out which makes it a very important component of a home’s insulation. Although air sealing, and insulation in general, can be expensive, there are programs in certain states like New York that can help with the cost of insulation (“EmPower+.”).

Vapor control is less of a fixed process and more of a context dependent step in insulating a building. Controlling vapor can mean anything from installing drainage planes, sealing nooks and crannies, to allowing specific areas of the house to self dry. One very common method of vapor control is simply installing a vapor barrier in the building’s walls. While effective when necessary, an analysis of the building and its specific vapor patterns may conclude that more personalized methods like the three listed above be implemented. Vapor moves differently through a building depending on climate zone and season so what prevents vapor damage in one area may not do the same in another. Finding the right way to manage vapor is very important because of the damage excess vapor can cause. Moisture will dampen and rot wood, create a comfortable home for mold, mildew, and unwanted bacteria, and ultimately cause safety issues in the building. A vapor barrier stops moist air from entering through the envelope. It can be installed almost anywhere but the attic, basement, and crawl spaces need this management most. While not as important as base insulation itself, vapor management and air sealing are finishing touches that have an extreme impact on a building’s air quality and energy efficiency. They are absolutely essential steps to achieving a comfortable, well insulated structure.

Insulation Methods

There is a great variety of types of insulation each one having its individual pros and cons. First, there is blanket/batt insulation. Blanket insulation involves placing a large sheet of a flexible insulation material, could be fibers or different kinds of wool, into studs, joints, and rafters (Pacific Northwest National Laboratory and Oak Ridge National Laboratory). One of the main pros of blanket insulation is that it is cut-to-size which means you can get custom insulation for whatever space you’re dealing with (Pacific Northwest National Laboratory and Oak Ridge National Laboratory). However this type of insulation can be moisture prone which may lead to mold or mildew (Admin Account).

Loose fill insulation involves filling studs or spaces with small fibers (Pacific Northwest National Laboratory and Oak Ridge National Laboratory). These fibers can be made of all different kinds of materials but are most commonly made up of fiberglass or cellulose (Pacific Northwest National Laboratory and Oak Ridge National Laboratory). Loose fill insulation is a great choice for oddly shaped spaces or areas with obstructions because the material is blown right in using a special blower (Pacific Northwest National Laboratory and Oak Ridge National Laboratory). However, this material cannot as easily be removed or reused like blanket insulation can. Additionally, if not properly sealed, the material can become something of a sandbox for rodents (Ori Pearl).

Foamed-in-place insulation is usually made of polyurethane polymer (Autodesk Support). You can blow this material into any cavity and it will harden to insulate the area (Autodesk Support). Precision is one of the main pros of this material as it can be sprayed into a lot of little areas. This method also has effective airtight abilities (Autodesk Support). However, in many cases the material is too airtight and negatively affects the ventilation of a home (“5 Spray Foam Insulation Problems and a Look at Alternatives 5 Problems With Spray Foam Insulation and Better Alternatives.”). Foamed-in-place insulation can also lead to mold and can have a terrible odor (“5 Spray Foam Insulation Problems and a Look at Alternatives 5 Problems With Spray Foam Insulation and Better Alternatives.”). Additionally, the material must be covered after spray or else toxic smoke becomes a very real and dangerous possibility (“5 Spray Foam Insulation Problems and a Look at Alternatives 5 Problems With Spray Foam Insulation and Better Alternatives.”). While a very effective method, there are a lot of risks that come with this method of insulation (“5 Spray Foam Insulation Problems and a Look at Alternatives 5 Problems With Spray Foam Insulation and Better Alternatives.”).

Board insulation is a rigid piece of material, usually a foamed or recycled plastic, that can be placed between walls (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”). Board insulation is extremely effective and can achieve great R values (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”). They also control moisture and reduce noise (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”). However, there are plenty of drawbacks. In order to install boards you need to have full access to the area and be able to completely open it up making this method extremely hard to use for existing buildings (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”). Boards are also not fire resistant which creates a safety hazard (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”). Lastly, this is one of the more expensive options for insulation (“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..”).

Radiant barriers are foil sheets that can cover plywood and reflect the sun to make the space cooler (“Radiant Barriers: Pros and Cons.”). This method is very effective for keeping attics cooler in very warm climates. While the temperature will not be fully controlled, it will be significantly less hot up there than before. Radiant barriers can also bring the duct surfaces closer to dew point which is the state at which water vapor becomes water (“Radiant Barriers: Pros and Cons.”). This creates moisture and a viable home for mold and fungi (“Radiant Barriers: Pros and Cons.”). While one solution may be to use a dehumidifier in the attic, this risk is not completely unavoidable and makes this method of insulation less than ideal.

Vacuum insulation is when the air between two walls is sucked out to stop heat transfer through air molecules (Alotaibi). This method is very energy efficient and a space saver. However, one of the major disadvantages is that if the envelope gets damaged, the insulation stops working as well and the efficiency of the building decreases significantly (Alotaibi). This is also another one of the more expensive insulation methods (Alotaibi).

Reflective insulation is the process of lining a home with a material made up of different layers of aluminum foil and spacers to reduce heat flow (Hill). The material reflects heat instead of absorbing it or passing it through into the building (Hill). By adding more layers of foil and spacers, greater R values can be achieved (Hill). Reflective insulation works well in tight spaces and can also function alongside other insulation types (Hill). However, this method is not ideal for more open spaces and requires more complicated installation (Hill). The material can also become less effective if the layers are compressed which poses a threat to longterm function (Hill).

Energy Efficient Insulation Methods

Other than comfort, the whole point of insulation is to improve the efficiency of a building. To make the whole process more sustainable low carbon materials are often used to insulate. One of the most popular low carbon materials is cellulose fibre which is made of recycled paper products (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). Added salts make it fire resistant and its installation makes it ideal for lofts and stone walls (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). The material is recyclable as well as biodegradable and provides uniform coverage (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). However, it can absorb moisture which decreases performance and can cause molding (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”).

Glasswool is the lowest embodied carbon material of all mainstream materials (“Pros and Cons of Glass Wool Insulation.”). It is made of 80% recycled glass, trapped air, plant based wool, and, depending on the brand, a low carbon binder (“Pros and Cons of Glass Wool Insulation.”). While very environmentally friendly, glasswool has major downsides (“Pros and Cons of Glass Wool Insulation.”). If handled incorrectly it can cause skin irritation or damage to the respiratory system (“Pros and Cons of Glass Wool Insulation.”). It can also absorb moisture in damp or humid areas. Lastly, installation can be difficult and requires a lot of attention to detail (“Pros and Cons of Glass Wool Insulation.”).

 Wood fibre is made from natural wood, which is a renewable, non-toxic, chemical free, recyclable, and biodegradable source (“Wood Insulation Guide – Pros, Cons, and Cost.”). It is ideal for roofs, drywall, and partition walls as it can regulate humidity (“Wood Insulation Guide – Pros, Cons, and Cost.”). This is a convenient form of insulation as it can be installed at any angle (“Wood Insulation Guide – Pros, Cons, and Cost.”). This method has minimal downsides with the main ones being cost and availability (“Wood Insulation Guide – Pros, Cons, and Cost.”).

Sheep’s wool is also a renewable, non toxic, and biodegradable insulation material (Local Surveyors Direct). It is a roll of recycled fibre and sheep wool and has low embodied carbon (Local Surveyors Direct). This material absorbs both moisture and sound well (Local Surveyors Direct). One of the main cons of sheep’s wool is its need for chemical treatment (Local Surveyors Direct). Sheep’s wool can attract moths and other insects creating another step in the process (Local Surveyors Direct). There are specific genetically altered rolls of wool that will not attract insects but with the material already being expensive without this extra benefit, this is not an option for every budget (Local Surveyors Direct).

An aerogel insulation blanket is a space saving low carbon insulation material made of silica aerogel, a material that traps air (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). With the trapped air the material ends up being about 90% air content meaning it is very light weight (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). It is also hydrophobic and has low thermal conductivity (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). The main con to this method of insulation is that aerogel and the chemicals used in the material can be hazardous (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”). Animal studies revealed that breathing in this material may cause harm if installed or handled incorrectly (“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.”).

Another low carbon material that can make insulation fibres is hemp. This is a natural, renewable, compostable, and non toxic resource (Schutz). This is a dense and pest-resistant material making it a great choice in the long run (Schutz). It has great insulation properties. This is also a form of insulation that is super easy to install (Schutz). The only downsides to hemp fibre insulation are the availability and cost (Schutz). There is only one major hemp fibre insulation manufacturer in the US which makes the material hard to acquire and expensive (Schutz).

Plaster insulation is a blend of cork, clay, and diatomaceous earth (Sidler). It is a breathable, fire resistant material that is made from natural and renewable sources (Sidler). It is biodegradable and very easy to apply (Sidler). The material itself is very strong, meaning little dents or dings will not be noticeable, and blocks sound (Sidler). The main disadvantage to this product is the fact that plaster often cracks over time (Sidler). This can eventually lead to significant damage and costly repairs (Sidler). This is also a less convenient option if hangers, fixtures, or art needs to be hung considering the material’s strength (Sidler).

The last mainstream low carbon insulation material is straw panels. This is a renewable and biodegradable product made of natural straw (Owens). Minimal processing is required and the panels end up being suitable for most walls (Owens). One very beneficial aspect of this material is its carbon sequestration (“STRAW-BALE – Carbon Smart Materials Palette.”). The straw captures and stores carbon from the atmosphere making it an extremely environmentally efficient insulation option (“STRAW-BALE – Carbon Smart Materials Palette.”). One disadvantage of this material is that it is not included in all local codes as a building material which means if chosen, some extra paperwork may need to be filled out (Owens). Additionally, moisture can be very detrimental to the straw (Owens). If installed in a very humid place or in such a way that it may come into contact with a significant amount of moisture, straw panels most likely will not be an effective form of insulation (Owens).

Insulation in Existing Buildings

Installing insulation into an existing building can vary in difficulty depending on the building. Whether the building has some existing insulation and there is access into the walls makes all the difference when it comes to the methods used. The first step in any case is to review the building’s envelope. Repairing any cracks or leaks on the outside can make all the difference when it comes to energy efficiency (“Insulating an Existing House | Cooperative Extension Service.”). The next step is often to check the basement and attic. These are areas of homes or buildings that face extreme weather conditions and that often have direct contact with snow. Basements can be the main reasons for moisture so it is important that its insulation is up to par (“Insulating an Existing House | Cooperative Extension Service.”). It is preferable that basements are insulated from the outside so that the envelope of the building is protected (“Insulating an Existing House | Cooperative Extension Service.”). The attic can be insulated from the inside because of the protection that the roof provides (“Insulating an Existing House | Cooperative Extension Service.”); however, there are materials better suited for attics than others because of the extreme temperatures that area of the house faces. Next, the walls must be inspected. Whether there is access inside the walls or not can change the direction of the plan. From whatever information there is from the walls, it must be determined if there is existing insulation. If so, you can either remove it and replace it, which is much easier if you can open up the walls completely, or clean whatever is in there and add on to it. Either way, the inside of the wall must be vapor proofed (“Insulating an Existing House | Cooperative Extension Service.”). All moisture and bacteria should be removed so that the new insulation or insulation additions have the best chance of success.

In terms of materials, most insulation methods are technically able to be installed into existing buildings but many require removal of drywall or damage to the envelope. If that is an option for the project then material options are plentiful. If there is not great access to the wall, for example a stone wall, then the insulation options are very limited. Depending on how big a space you can open up, the main options are cut in place boards, loose fill, and foamed in place materials. However, if there is really no way to get into the wall then interior insulation may be necessary depending on the overall efficiency of the building.

Insulation in Historic Buildings

Installing or upgrading insulation in historic buildings is often very challenging because the value of the exterior and interior can restrict access to the inside of the walls. It is typically not effective to improve insulation in a historic building through the walls. Therefore, insulation through the attic and/or basement, depending on whether they are finished or unfinished, and through pipes is more realistic. Pipes and ducts can be wrapped in insulation material like mineral wool, certain rubbers, and foam insulation to improve heat loss (“Weatherization of Historic Buildings: Install Insulation (U.S. National Park Service).”). However, one of the most effective ways of improving a building’s efficiency without creating damage to the envelope is upgrading the windows. Higher quality windows do a great job of limiting unwanted air flow. Improving the windows in a home by adding another pane or adding a vacuum seal can reduce energy bills from 7% to even 30% (“How Much Energy Can New Windows Save? – KobyCo Door & Window.”). Many historic building updates have involved window improvements.

A prime example of historical window replacement was the 2007 Empire State Building renovation (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). The Empire State Building is a famous historic building that is also still very active. Improving efficiency through wall insulation upgrades would have put a lot of peoples’ lives on hold and simply was not an option. Planners decided to improve the envelope with better windows (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). 6,514 windows needed to be updated and brought up to code (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). The Empire State Building was built in 1931 and therefore had a lot that needed to be brought up to code (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). With more and more laws coming out about CO2 emissions, without renovations, the building would have faced fines of about $2.49 million a year (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). Instead of completely replacing each and every window, which would have cost $16.3 million dollars at least, the team decided to take out, inspect, and update each window (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). They installed new panes and spacers and added insulating gasses (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). This ended up costing about $4.5 million dollars, roughly a third of what it would have been had they gone with complete replacement (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). The window U-value went from 0.50 all the way down to 0.125, a sizeable jump (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). As a result, about 54% of the building’s emissions were cut (“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.”). The efficiency of the project, both in terms of how well the windows work and how productive the team’s plan, makes it one of the greatest retrofits in history.

Case Study

Located in Forest Hills, Queens New York The West Side Tennis club was originally built in 1913 (Keane). Its well known stadium was built ten years later (“Outside the Box: The Revival of Forest Hills Stadium | National Trust for Historic Preservation.”). This tennis club is a very historic spot because of the numerous athletic milestones reached here. The U.S. open and Davis Cup were held there for a long time, the color barrier for the Grand Slam was broken there, and famous players like Billie Jean King played on those courts (Keane). In addition to tennis, important musicians have had historic performances at the stadium such as The Beatles, Bob Dylan, and The Rolling Stones (“Outside the Box: The Revival of Forest Hills Stadium | National Trust for Historic Preservation.”). Both structures have unique insulation due to their differing duties.

The club house was built in a classic tudor style to match the surrounding neighborhood. Therefore, without modern day insulative materials, the insulation rested on a thick envelope. Thick materials like brick, stone, and timber helped to create this barrier between the outside and inside. As time has gone on and adjustments have been made to the club, efficiency has had to be upgraded. However, the walls have never been broken into because of their historic value. Therefore, the club frequently updates its plumbing and heating systems to stay on top of insulation (Keane). Adding insulation around the existing pipes and using better quality windows has helped keep energy efficiency up to code without having to damage the walls. However, in order to make a significant improvement to the building’s efficiency, it would be necessary to add more modern insulation into the walls. This is why later on, it is proposed that the walls are opened up.

In terms of the stadium, temperature insulation was not so much of an issue as noise insulation. Originally, in 1923, the stadium was built from concrete because it was made for tennis matches and did not require an extensive amount of sound insulation (“Outside the Box: The Revival of Forest Hills Stadium | National Trust for Historic Preservation.”). However, as the stadium evolved into a concert venue and locals began complaining about the noise, which eventually led to a lawsuit, the need for renovation increased (Keleshian). The stairwells and structural pillars were lined with sound-dampening insulation mesh (Moses). This material is usually a woven mix of polyester or fiberglass (“Engineered Materials Inc..”). It can withstand weather, which is good because this is an outdoor venue, and it can convert sound waves into heat (“Engineered Materials Inc..”). Overtop of all of this is a coat of blue paint to match the aesthetics. Only certain materials could be used for this project because of its outdoor location; however, considering the demanding and specific conditions, the designers did a great job with this project. They were able to keep the rest of the stadium concrete because the bodies that fill it during concerts act as noise insulation themselves.

In order to improve the West Side Tennis Club’s efficiency as much as possible I would break open the walls and add in wool insulation as well as rigid board lining. Air space as well as the wool’s properties will prevent moisture build up. Rigid panels are extremely effective and the wool fibre filling adds protection as well as insulative, acoustic, and environmental benefits (“Rockwool”). The following values are estimated and those used to calculate the building’s measurements were also estimated using the scale image of a trash can as shown in Figure 1. The current U value of the walls, with just brick, stone, and concrete, is about 0.5, which is quite poor (Pullen). Rigid boards have an R value of about 15 for 3.5 inches of space, an estimation. The wool filling has an R value of about 14 which in combination with the rigid boards equals to about 30. This translates to a U value of 0.03 which is much better than 0.5. This improved U value makes a huge difference in wall heat flow. To find the initial heat flow, as shown in Figure 2, the wall area was multiplied by the 0.5 U value which was multiplied by the temperature difference. To find the proposed heat flow the 0.5 U value is replaced with 0.03. Based on these calculations the wall heat flow would go from about 19,370 btu/h all the way down to 1,160 btu/h.

This would be a significant improvement and along with window renovations, the whole building would be much more efficient. The current windows are traditional multi-plane casement-style windows. These probably have an U value of about 0.5. The simplest way of decreasing this value without losing aesthetics would be to purchase triple glazed windows. This would bring the U value down to about 0.20 (“StackPath”). Finding the window heat flow is a similar process by which, as shown in Figure 2, the window area is multiplied by either the initial or proposed U value which is multiplied by the change in temperature. By these calculations the window heat flow would go from about 7,150 btu/h to about 2,860 btu/h. By simple addition, this means that total heat flow of this section would decrease from about 26,520 btu/h all the way down to about 4,020 btu/h.

396 Warren St. Brooklyn is a prime example of how efficient wool insulation can be in tandem with rigid boards. The building was originally built in the 1800s and was used for horse carriages and ice cream truck parking (“Rockwool.”). The goal of the retrofit was to create a comfortable and sustainable living space using low carbon materials (“Rockwool.”). Designers also wanted to keep as much of the original structure as possible (“Rockwool.”). This along with their refusal to use fossil fuel based foams limited their insulation options (“Rockwool.”). The original building had no insulation and therefore needed it on all 6 sides (“Rockwool.”).

The plan was to create entirely new interior walls and insulate them with stone wool materials (“Rockwool.”). They also installed continuous rigid insulation boards to work alongside the wool (“Rockwool.”). The combination allowed for very great thermal, vapor, and air control (“Rockwool.”). The final combined R-value ended up being 31 which is a great jump from having no insulation whatsoever (“Rockwool.”). The blower door test even gave the structure a 0.3 on airtightness (“Rockwool.”). In terms of starting from scratch, this retrofit proves that wool and rigid boards work well to create an efficient insulation structure.

Triple pane windows have become nearly as popular as double pane windows. Lennar corporation, one of nations biggest home building companies, recently chose to install triple pane windows into over 100 homes in the Ridgefields community (“Case Study: National Builder Prioritizes Occupant Health and Well-Being.”). According to them, with these windows they were able to “not only meet but exceed energy codes in this community” (“Case Study: National Builder Prioritizes Occupant Health and Well-Being.”). The company reported little to no difference in availability or installation for these windows compared to double pane (“Case Study: National Builder Prioritizes Occupant Health and Well-Being.”). Considering the major difference in energy efficiency between these two types of windows, triple pane is the most productive choice.

Glass with a low-e coating means the window will have lower emissivity and therefore will become more energy efficient. A low-e coating can change the R value of a triple pane window from R3 to about R5 (Jennings 3). That is a great jump with minimal change to the design of the windows themselves. The window itself is not obnoxiously thick and with the right frame would still fit in with the existing building.

Conclusion

It is clear that an assessment of insulation can make a great difference in a building’s energy efficiency. There are so many different ways of insulating. So many of these methods are energy efficient which makes it unnecessary for infrastructure to be impacting the environment so negatively. In terms of the case study, there are many areas where more precision and accuracy could have been applied. More exact measurements could have been obtained through communication with the establishment but for estimation purposes the values calculated based on perspective work well. Heating and cooling systems in the building itself could have been considered but would have complicated the simple model of insulative energy improvement. Heating and cooling systems create an equilibrium of temperature through the air space and furniture to provide a stable temperature throughout the rooms. When certain insulation works well enough, it can replace these systems altogether; however, those are rare and expensive cases. Ventilation is also a factor that is important to this process but was not considered for simplicity purposes. Ventilation is the management of airflow throughout a structure (“Energy, Weatherization and Indoor Air Quality.”). With faulty ventilation could come uneven temperatures creating damp areas and/or mold and bacteria (“Energy, Weatherization and Indoor Air Quality.”). Depending on the type of insulation used, this could rot or ruin the insulation in place. Even though the scope of this case study was not large enough to include these aspects of the insulative and design process, the proposal and calculations still represent a beneficial change to the structure and the overall lesson that insulation can majorly improve a buildings energy efficiency.

Appendix

Figure 1: Displays estimated measurements of the case study. Measurements estimated by using the trash can as a scale multiplier.

Figure 2: All calculations for window and wall heat flow. Temperature estimated for cold months: When it is 10 degrees outside the club is estimated to be set at a comfortable temperature of 75 degrees. 0.5 initial U-value was estimated based on the fact that the original wall is a stone and brick cavity with some external insulation (Pullen)

Acknowledgements

I acknowledge Kurt Tiechurt for leading me through the production and editing process of this paper. I also acknowledge Danielle Joy for assistance throughout the publishing process.

References

Admin Account. “Blown-In Vs. Batt Insulation: Pros & Cons | Good’s Roofing.” Good’s Roofing, 31 Mar. 2025, https://goodsroofing.com/blown-in-vs-batt-insulation-pros-cons-and-best-uses-for-your-attic/. Accessed 21 Sept. 2026.

Alotaibi, Sultan Sanat, and Saffa Riffat. “Vacuum Insulated Panels for Sustainable Buildings: A Review of Research and Applications.” International Journal of Energy Research, vol. 38, no. 1, 9 Oct. 2013, pp. 1–19, 10.1002/er.3101.

Architecture 2030. “Why the Built Environment – Architecture 2030.” Architecture2030, 2023, https://www.architecture2030.org/why-the-built-environment/. Accessed 21 Sept. 2026.

ASIRI Designs. “Super Efficient Exterior Insulation Details With Rockwool (for Architects + Builders).” YouTube, 2 Mar. 2024, https://www.youtube.com/watch?v=VwUjkSjllPs.  Accessed 21 Sept. 2026.

Autodesk Support. “Insulation Part-2.” Autodesk, 12 Apr. 2018, https://www.autodesk.com/support/technical/article/caas/tsarticles/ts/76EZ1VUHAQv0wk0JrPVmUG.html. Accessed 21 Sept. 2026.

“Case Study: National Builder Prioritizes Occupant Health and Well-Being.” BetterBuiltNW, 12 Mar. 2024, https://betterbuiltnw.com/case-studies/case-study-triple-pane-windows-increase-the-comfort-efficiency-and-value-of-new-homes. Accessed 21 Sept. 2026.

“Choosing Low-Carbon Materials to Help Reduce the Whole-Life Footprint for Your Buildings.” Knauf New Zealand, 2026, https://knauf.com/en-nz/sustainability/choosing-low-carbon-materials. Accessed 21 Sept. 2026.

“Empire State Building Retrofit: How the World’s Most Famous Skyscraper Proves Every Building Can Decarbonize.” Dextall, 12 Feb. 2026, https://www.dextall.com/blog/empire-state-bUilding-retrofit-decarbonization-local-law-97. Accessed 21 Sept. 2026.

“EmPower+.” Nyserda, www.nyserda.ny.gov/All-Programs/EmPower-New-York-Program. Accessed 10 Oct. 2026.

“Energy, Weatherization and Indoor Air Quality.” US EPA, 15 Sept. 2014, https://www.epa.gov/indoor-air-quality-iaq/energy-weatherization-and-indoor-air-quality. Accessed 21 Sept. 2026.

“Engineered Materials Inc..” Engineered Materials Inc., 8 Jan. 2024, https://www.engineeredmaterialsinc.com/articles/differences-sound-control-acoustical-insulation-vs-sound-dampening-materials.  Accessed 21 Sept. 2026.

Hill, Jonathan. “Radiant Barrier Vs Reflective Insulation – What’s the Difference?” Insulation & More, Insulation & More, 30 Oct. 2023, https://insulation-more.co.uk/blogs/the-pipe-duct-lagging-expert/radiant-barrier-vs-reflective-insulation-whats-the-difference. Accessed 21 Sept. 2026.

“How Much Energy Can New Windows Save? – KobyCo Door & Window.” KobyCo Door & Window, 18 July 2025, https://kobycowindow.com/how-much-energy-can-new-windows-save/. Accessed 21 Sept. 2026.

“Insulating an Existing House | Cooperative Extension Service.” Uaf.Edu, 2021, https://www.uaf.edu/ces/publications/database/energy/insulating-existing-house.php. Accessed 21 Sept. 2026.

Jennings, John, et al. Thin Triple Pane Windows: A Market Transformation Strategy for Affordable R5 Windows Contents. 2020, https://neea.org/wp-content/uploads/2025/03/Thin-Triple-Pane-Windows-A-Market-Transformation-Strategy-for-Affordable-R5-Windows.pdf. Accessed 9 Oct. 2026.

Keane, Katharine. “Competition and Community: 124 Years of History at the West Side Tennis Club | National Trust for Historic Preservation.” SavingPlaces, 9 June 2016, https://savingplaces.org/stories/competition-community-124-years-history-west-side-tennis-club.  Accessed 21 Sept. 2026.

Keleshian, Kristie. “Forest Hills Stadium Owners Sued as Neighbors Complain About Noise.” CBS News, CBS New York, 20 May 2023, https://www.cbsnews.com/newyork/news/forest-hills-stadium-lawsuit-noise/. Accessed 21 Sept. 2026.

Local Surveyors Direct. “The Pros and Cons of Sheep’s Wool Insulation for Your Home | Local Surveyors Direct.” Localsurveyorsdirect.Co.Uk, 2021, https://www.localsurveyorsdirect.co.uk/pros-and-cons-sheep%E2%80%99s-wool-insulation-your-home. Accessed 21 Sept. 2026.

Moses, Amanda. “Forest Hills Stadium Makes Changes to Keep Noise Down Following Concerts.” amNewYork, 6 May 2024, https://www.amny.com/news/forest-hills-stadium-noise-changes-concerts/. Accessed 21 Sept. 2026.

Ori Pearl, Uri. “Rodent Proof Insulation: What Actually Works.” Nealoninsulation.Com, 18 May 2026, https://www.nealoninsulation.com/blog/rodent-proof-insulation. Accessed 21 Sept. 2026.

“Outside the Box: The Revival of Forest Hills Stadium | National Trust for Historic Preservation.” Savingplaces.Org, 2016, https://savingplaces.org/stories/match-point-the-revival-of-forest-hills-stadium. Accessed 21 Sept. 2026.

Owens, Ted. “The Pros and Cons of Straw Bale Wall Construction.” Building With Awareness, 27 Apr. 2009, https://www.buildingwithawareness.com/the-pros-and-cons-of-straw-bale-wall-construction-in-green-building/. Accessed 21 Sept. 2026.

Pacific Northwest National Laboratory and Oak Ridge National Laboratory. “BUILDERS CHALLENGE GUIDE TO 40% Whole-House Energy Savings in the Cold and Very Cold Climates.” Building America Best Practices Series, vol. 12, Feb. 2011, https://www1.eere.energy.gov/buildings/publications/pdfs/building_america/cold_climate_guide_40percent.pdf. U.S. Department of Energy. Accessed 10 Oct. 2026.

“Pros and Cons of Glass Wool Insulation.” Huamei Energy-Saving Technology Group Co., Ltd., 8 Sept. 2023, https://www.huameiworld.com/news/pros-and-cons-of-glass-wool-insulation.html. Accessed 21 Sept. 2026.

“Radiant Barriers: Pros and Cons.” Kalos Services, 27 Apr. 2022, https://www.kalosflorida.com/comfort/radiant-barriers-pros-and-cons/. Accessed 21 Sept. 2026.

“Rockwool.” ROCKWOOLInt, 2025, https://www.rockwool.com/north-america/advice-and-inspiration/case-studies/warren-st-retrofit/ . Accessed 21 Sept. 2026.

Schutz, Terry. “Hemp Insulation – Pros and Cons.” Homedit, 30 Aug. 2023, https://www.homedit.com/insulation/hemp-insulation/. Accessed 21 Sept. 2026.

Sidler, Scott. “The Pros & Cons of Plaster Walls.” The Craftsman Blog, 4 May 2020, https://thecraftsmanblog.com/the-pros-cons-of-plaster-walls/. Accessed 21 Sept. 2026.

“StackPath.” Efficientwindows.Org, https://efficientwindows.org/gtypes-3lowe/. Accessed 21 Sept. 2026.

“STRAW-BALE – Carbon Smart Materials Palette.” Www.Materialspalette.Org, https://www.materialspalette.org/straw-bale/. Accessed 21 Sept. 2026.

“The Pros and Cons of Rigid Foam Board Insulation | Ace Insulation Inc..” Aceinsulation.Biz, https://aceinsulation.biz/2022/08/the-pros-and-cons-of-rigid-foam-board-insulation/.  Accessed 21 Sept. 2026.

Pullen, Tim. “How to Insulate Old Homes.” Homebuilding & Renovating, 3 June 2020, https://www.homebuilding.co.uk/advice/how-to-insulate-old-homes. Accessed 9 Oct. 2026.

“Types of Insulation | Department of Energy.” Pacminerals.Com, 2025, https://www.pacminerals.com/news-types-of-insulation-department-of-energy.html. Accessed 21 Sept. 2026.

“Weatherization of Historic Buildings: Install Insulation (U.S. National Park Service).” Www.Nps.Gov, https://www.nps.gov/articles/000/weatherization-of-historic-buildings-install-insulation.htm. Accessed 21 Sept. 2026.

“Wood Insulation Guide – Pros, Cons, and Cost.” Rise, 29 Jan. 2021, https://www.buildwithrise.com/stories/wood-insulation-pros-cons-and-cost. Accessed 21 Sept. 2026.

“5 Spray Foam Insulation Problems and a Look at Alternatives 5 Problems With Spray Foam Insulation and Better Alternatives.” Hempitecture, 9 Oct. 2022, https://www.hempitecture.com/post/problems-with-spray-foam-insulation-in-attics/.  Accessed 21 Sept. 2026.

“8 Insulation Products With Sustainable Properties for a Greener Future – 2050 Materials.” 2050 Materials – Simplifying Sustainable Material Selection, 12 Sept. 2024, https://2050-materials.com/blog/8-insulation-products-with-sustainable-properties-for-a-greener-future/. Accessed 21 Sept. 2026.

About the author

Harper Harrison

Harper a senior at Grace Church School majoring in math and hoping to pursue a career in STEM. She is currently taking an engineering course at Cooper Union to fuel this interest. She loves to design and build things which is why infrastructure has recently become one of her interests. She likes to spend her free time doing sports. She has been playing volleyball for 4 years and is the captain of her school’s cheerleading team.

The post Insulation and Its Relevance in Existing and Historic Infrastructure appeared first on Exploratio Journal.

Abstract

This project investigates the design of a simplified Earth-to-Jupiter spacecraft mission using the patched-conic approximation. The mission is modelled using the Sun, Earth, and Jupiter, with all planetary orbits assumed to be circular and coplanar, and a Hohmann transfer orbit is used to model the spacecraft’s heliocentric trajectory. MATLAB is used to perform the numerical calculations and visualize the resulting trajectories. The transfer takes 997.9 days (about 2.73 years), with the departure burn occurring 191.0 days after the initial planetary alignment. Departure from a 5×10⁷ m Earth parking orbit requires Δv = 6,834 m/s, while capture into a circular 2×10⁸ m orbit at Jupiter requires 10,870 m/s, giving a total Δv of 17,704 m/s, of which capture accounts for 61%. Jupiter’s sphere of influence has a radius of 4.82×10¹⁰ m (0.322 AU). A comparison with the Galileo mission, whose capture burn was only about 630 m/s, suggests that the high capture cost in this model comes mainly from the choice of a circular capture orbit.

1 Introduction

Unlike motion around a single body, an interplanetary transfer involves changes between different gravitational regimes, making the direct modeling of a spacecraft’s motion considerably more complex. Mathematical approximations are therefore required to represent these systems while retaining the fundamental relationships that govern orbital motion. The patched-conic approximation provides one of these approaches by treating the spacecraft’s trajectory as a sequence of two-body problems. Within each region of the trajectory, the gravitational influence of one body is assumed to dominate, allowing the spacecraft’s motion to be described using established orbital mechanics. For transfers between approximately circular planetary orbits, a Hohmann transfer provides a particularly simple model of the heliocentric portion of the trajectory. This investigation applies these approximations to an Earth-to-Jupiter transfer. The spacecraft is initially placed in a circular Earth orbit and ultimately enters a circular orbit around Jupiter. MATLAB is used to examine the resulting trajectory and the conditions required for the transfer.

2 Mathematical Background

2.1 Gravitational Parameter and Circular Orbits

A spacecraft traveling through space is constantly affected by the gravitational attraction of nearby celestial bodies. Newton’s law of universal gravitation gives the attractive force between two bodies as

where G is the gravitational constant, M is the mass of the central body, m is the mass of the spacecraft, and r is the distance between their centers of mass (Curtis, 2021). It is standard in orbital mechanics to combine G and M into a single gravitational parameter,

where μ has units of m³ s⁻². Using this parameter, the gravitational force can be expressed as:

Since μ depends on the central body, the Sun, Earth, and Jupiter each have their own value; this becomes important later, as the model shifts between Earth-centred, Sun-centred, and Jupiter-centred motion. In the case of a circular orbit, gravity supplies the centripetal force needed to keep the spacecraft on its curved path:

After rearranging, the orbital velocity is the following:

The circular orbital velocity therefore depends only on the distance from the central body and on that body’s gravitational parameter.

2.2 Elliptical Orbits

The Earth-to-Jupiter transfer trajectory is not circular but elliptical. An ellipse is described by its semi-major axis a, which represents half of the ellipse’s longest diameter and therefore determines its overall size and its eccentricity e which describes how stretched the ellipse is. In the case of the circle e = 0, while increasing values of e correspond to increasingly elongated ellipses (Atkins, 2019). Along an ellipse, the distance from the central body varies between a closest point, periapsis, and a furthest point, apoapsis. Their distances from the central body are

and

Since these are the two extreme points of the orbit, they can also be used to determine the eccentricity:

On this paper, Earth’s orbital radius is treated as the periapsis of the heliocentric transfer ellipse, while Jupiter’s is treated as its apoapsis. Consequently, the size and shape of the transfer trajectory are set entirely by the two planets’ orbital radii.

2.3 Vis-Viva Equation

Along an elliptical orbit the spacecraft’s speed is not constant: it moves fastest at periapsis and slowest at apoapsis. This behaviour follows from conservation of specific mechanical energy, ε, the sum of specific kinetic and potential energy, which remains constant everywhere on the orbit:

Rearranging for v gives the vis-viva equation (Curtis, 2021):

Because a is fixed for a given orbit, this equation can be applied at different points along the same ellipse to determine how the spacecraft’s speed changes with distance r. For the Earth-to-Jupiter transfer, it is used to find the spacecraft’s heliocentric speed at both Earth’s and Jupiter’s orbital radii. These speeds differ from the planets’ own orbital speeds, which is why burns are needed at departure and arrival.

2.4 Hohmann Transfer

A Hohmann transfer connects two circular orbits about a common central body with a single elliptical trajectory tangent to both: periapsis at the inner orbit, apoapsis at the outer one (NASA, 2023). Taking the Sun as the central body, the transfer ellipse spans Earth’s and Jupiter’s heliocentric radii, giving

and the eccentricity of this transfer is therefore:

where r_E and r_J are the heliocentric orbital radii of Earth and Jupiter. The vis-viva equation then gives the spacecraft’s heliocentric speeds at departure and arrival:

and

Since the transfer covers exactly half of the ellipse, the transfer time is half the ellipse’s orbital period:

The required velocity changes at departure and arrival, Δv = v_final − v_initial, sum to the mission’s total delta-v, Δv_total = Δv_dep + Δv_arr. Knowing the transfer time alone, however, is not sufficient: Earth and Jupiter continue moving while the spacecraft is in transit, so their motion must also be modeled (Atkins, 2019).

2.5 Planetary Orbital Motion and Launch Geometry

For a circular heliocentric orbit, the period and mean angular velocity or mean motion are:

From the period, the mean angular velocity can be found:

Under the assumption of uniform circular motion, the angular position of a planet can therefore be written as

where θ₀ is the initial angular position. This allows the prediction of where Earth and Jupiter will be at a particular time. This is key because the spacecraft’s arrival must be synchronized with Jupiter’s position.

2.6 Patched-Conic Approximation

Treating the Sun as the only gravitational influence is reasonable while the spacecraft is far from the planets, but not close to either one. In reality the spacecraft is affected simultaneously by the Sun, Earth, and Jupiter. The patched-conic approximation avoids the resulting many-body problem by dividing the trajectory into regions, each dominated by a single body: an Earth-centred departure, a Sun-centred transfer, and a Jupiter-centred arrival, treated as a sequence of two-body problems and joined at their boundaries (Schaub, n.d.).

2.7 Sphere of Influence

The boundary between regions is approximated by the sphere of influence (SOI), the region around a planet within which its own gravity is treated as dominant, where its approximate radius is

Using gravitational parameters as an equivalent mass ratio, Jupiter’s SOI radius is

(Lutze, n.d.), giving the model an approximate radius at which the spacecraft’s motion transitions from Sun-centred to Jupiter-centred.

2.8 Hyperbolic Departure and Arrival

Relative to a planet, the spacecraft’s escape or approach trajectory is hyperbolic. Its hyperbolic excess velocity, v_∞, is the spacecraft’s velocity relative to the planet once the planet’s gravity becomes negligible. At departure this is the difference between the spacecraft’s heliocentric transfer velocity and Earth’s own orbital velocity,

and similarly at arrival,

(Atkins, 2021). Closer to the planet, the spacecraft’s speed at periapsis of the hyperbola is

and comparing this to the circular parking-orbit speed, v_park = √(μ/r_p), gives the required maneuvre:

This is applied separately at Earth’s departure and Jupiter’s arrival.

2.9 Hyperbolic Eccentricity and Turn Angle

Eccentricity also describes the shape of a hyperbola, for which e > 1. For the hyperbolic trajectory considered in this investigation, the eccentricity can be determined from the periapsis radius and hyperbolic excess velocity:

As it passes through the planet’s gravitational field, its velocity changes direction. The amount by which the trajectory is deflected is described by the hyperbolic turn angle:

The MATLAB model calculates these quantities for both the Earth departure hyperbola and the Jupiter arrival hyperbola.

2.10 Time of Flight Along the Transfer Orbit

The Hohmann transfer equation gives the total time required to travel from Earth’s orbital radius to Jupiter’s orbital radius. However, the patched-conic model must also determine the time at which the spacecraft reaches the boundary of Jupiter’s sphere of influence. One way of describing the position of a spacecraft on an ellipse is through the eccentric anomaly, E. Unlike the true geometric angle, it makes the relationship between position and time easier to establish. The radial distance from the central body can be written as

Once the spacecraft’s radial distance is known, this relationship can be used to determine the corresponding eccentric anomaly. The eccentric anomaly is then connected to the mean anomaly, M, through Kepler’s equation:

This equation is important because the mean anomaly changes uniformly with time, unlike the spacecraft’s actual angular position along the ellipse. The relationship between mean anomaly and time is

where the mean motion of the transfer ellipse is:

This allows the model to determine the time at which the spacecraft reaches Jupiter’s sphere of influence rather than assuming that the transition occurs at the end of the total Hohmann transfer.

2.11 Hyperbolic Time of Flight

The same idea of relating position to time is required for the Earth-centred hyperbolic departure. Once the spacecraft leaves its circular parking orbit, it follows a hyperbolic trajectory relative to Earth. The model therefore needs to determine how long this phase lasts before the spacecraft approaches its hyperbolic excess velocity. The semi-major axis of a hyperbolic orbit is negative and is given by:

The semi-latus rectum is:

where h is the specific angular momentum. The hyperbolic equivalent of the eccentric anomaly is the hyperbolic anomaly, F. It is related to the hyperbolic mean anomaly M_h by:

The corresponding hyperbolic mean motion is:

Therefore, the elapsed time can be obtained by:

These equations allow the model to connect the spacecraft’s position on its Earth-centred hyperbola with the time elapsed since the departure maneuvre. In this investigation, this is used to determine the time required for the spacecraft’s velocity to reach a value within 2% of its hyperbolic excess velocity.

3 Methodology

3.1 Model Inputs

All calculations were performed in a single MATLAB live script, structured into sequential parts (a)–(i), each of which uses the outputs of the previous parts. The model inputs are listed in Table 1. Gravitational parameters were used directly in every equation, so planetary masses were not required. Planetary orbits were assumed to be circular and coplanar, and all burns were modeled as instantaneous changes in velocity.

3.2 Order of Calculations

1. Transfer orbit. The semi-major axis a_t and eccentricity e of the Hohmann transfer ellipse were calculated from r_E and r_J. The vis-viva equation then gave the spacecraft’s heliocentric speed at periapsis and apoapsis, and the transfer time was taken as half the period of the ellipse.
2. Planetary periods. The orbital periods T_E and T_J and the mean angular speeds n_E and n_J of Earth and Jupiter were calculated, assuming circular orbits.
3. Launch time. The time of the departure burn, t_escape, was found with the search described in Section 4.3.
4. Jupiter’s position. Jupiter’s position at t_escape was calculated in a heliocentric frame whose x-axis points from the Sun to Earth, and the resulting geometry was plotted in Figure 1.
5. Departure hyperbola. The hyperbolic excess speed at Earth was calculated as v_∞,dep = v_p − v_E. From this, the model found the periapsis speed of the departure hyperbola, the departure burn Δv₀ from the parking orbit, the eccentricity e₁, the turn angle δ₁, and the burn location, placed at an angle of −δ₁/2 from the x-axis.
6. Arrival hyperbola. The same steps were applied at Jupiter using v_∞,arr = v_J − v_a, giving the capture burn Δv₃ into the 2×10⁸ m orbit, the eccentricity e₂, the turn angle δ₂, and the burn location, placed at an angle of π + δ₂/2.
7. Time to approach v_∞. On the departure hyperbola, the model found the radius at which the spacecraft’s speed falls to within 2% of v_∞,dep, converted it to a true anomaly and a hyperbolic anomaly F, and used the hyperbolic form of Kepler’s equation to calculate the time elapsed since the burn.
8. Sphere of influence. Jupiter’s SOI radius was calculated from r_J and the ratio μ_J/μ_⊙.

3.3 Launch Time Search

The launch time depends on the relative positions of Earth and Jupiter. At t = 0, Earth was placed at an angle of 0° and Jupiter at −90°, meaning Jupiter starts 90° behind Earth. This starting position is an assumption of the model; a different starting angle would give a different t_escape.

For a Hohmann transfer, the spacecraft arrives at Jupiter’s orbit exactly 180° from its departure point. The model therefore stepped forward in time from t = 0 in increments of 1 hour. At each step it calculated where Jupiter would be after a further t_transfer, and compared that with the point 180° from Earth’s current position. The search stopped at the first time step where the two angles agreed to within 0.5°, and that time was taken as t_escape.

4 Results

Table 1: Gravitational parameters and circular orbital radii used in the MATLAB model.

Table 1 lists the inputs used in every calculation.

Table 2: Hohmann transfer orbit properties.

The spacecraft leaves Earth’s orbit at a heliocentric speed of 38,578 m/s and reaches Jupiter’s orbit at 7,413 m/s. The transfer is half of one orbit around this ellipse, which takes 8.622×10⁷ s, or 997.9 days (2.73 years).

Table 3: Heliocentric orbital periods of Earth and Jupiter.

Earth completes one orbit in 3.156×10⁷ s (365.3 days), and Jupiter completes one in 3.747×10⁸ s (4,336.7 days, about 11.87 years). Because Jupiter moves much more slowly than Earth, the launch has to wait for the right alignment.

Table 4: Launch window (time from t = 0 to the departure burn).

The launch search gives t_escape = 1.6499×10⁷ s, or 191.0 days, which is when the departure burn takes place.

Table 5: Jupiter’s position at t_escape, in the heliocentric frame with the x-axis pointing from the Sun to Earth.

At t_escape, Jupiter is about 97.6° ahead of Earth (Table 5). During the 997.9-day transfer, Jupiter moves a further 82.8° along its orbit, so it reaches the point opposite Earth’s departure position within the 0.5° tolerance used in the launch search.

Table 6: Departure hyperbola at Earth (Earth-centered coordinates).

To escape Earth, the spacecraft needs a departure burn of Δv₀ = 6,834 m/s from its parking orbit. This puts it on a hyperbolic path that leaves Earth with an excess speed of 8,794 m/s. The hyperbola turns the spacecraft’s path through a total of 10.7° (5.36° half-angle). The burn takes place just below the Sun–Earth line.

Table 7: Arrival hyperbola at Jupiter (Jupiter-centered coordinates).

The spacecraft reaches Jupiter with an excess speed of 5,643 m/s. Capture into the 2×10⁸ m orbit requires Δv₃ = 10,870 m/s. That makes it the largest burn of the mission and 61% of the total Δv of 17,704 m/s. The arrival hyperbola turns the path through 144.4° (72.2° half-angle).

Table 8: Time from the departure burn until spacecraft speed is within 2% of v_∞, and radius of Jupiter’s sphere of influence.

The spacecraft’s speed comes within 2% of the excess speed v_∞ at 2.7625×10⁴ s (about 7.7 hours) after the departure burn. After this point, Earth’s gravity barely slows the spacecraft any further, so treating the departure as a separate hyperbola and then a heliocentric orbit is a reasonable approximation. Jupiter’s sphere of influence has a radius of 4.8220×10¹⁰ m, or 0.322 AU. This is where the model switches from the heliocentric transfer orbit to the Jupiter-centered arrival hyperbola.

5 Conclusion

5.1 Discussion of Results

The results from this study show a feasible Earth-to-Jupiter transfer that is fuel-efficient but slow in time. The spacecraft’s heliocentric speed falls from 38,578 m/s at departure to only 7,413 m/s at Jupiter’s orbital radius, reflecting the substantial deceleration associated with outward Hohmann transfers. This explains why the mission’s Δv budget is dominated by the capture maneuver rather than departure: the departure burn required only 6,834 m/s, while capture required 10,870 m/s, nearly 60% more. Since Jupiter’s orbital speed (13,055 m/s) exceeds the spacecraft’s apoapsis speed (7,413 m/s), the spacecraft arrives with a hyperbolic excess velocity of 5,643 m/s. The transfer itself takes 997.9 days (2.73 yr), a direct consequence of Kepler’s third law applied to the large semi-major axis of the transfer ellipse; this long duration is the primary trade-off of the fuel-efficient Hohmann geometry compared with faster, higher-energy trajectories.

The computed SOI radius of 4.82×10¹⁰ m or 0.322 AU is large compared with the planet itself, about 670 Jupiter radii. This relatively large SOI, combined with the shallow 10.725° turn angle at departure versus the much steeper 144.40° turn at arrival, illustrates that Jupiter’s gravity dominates the terminal phase of the mission far more strongly than Earth’s does at departure, which is a natural consequence of Jupiter’s much larger μ and its greater distance from the Sun.

To assess the model’s accuracy, the results can be compared with the Galileo mission (D’Amario et al., 1992). Galileo was launched on 18 October 1989 and reached Jupiter on 7 December 1995, a flight of about 6.1 years compared with the 2.73 years of the direct Hohmann transfer computed here. This longer flight was due to its Venus–Earth–Earth gravity-assist (VEEGA) trajectory, which was needed because its upper stage could not provide the large departure energy of a direct transfer; the model’s departure requires v_∞ = 8,794 m/s at Earth. At arrival, Galileo’s Jupiter orbit insertion burn was about 630 m/s (D’Amario et al., 1992), compared with 10,870 m/s in this model. Galileo was captured into a highly elliptical orbit with a period of about 200 days and a periapsis of about 4 Jupiter radii, whereas this model assumes a circular orbit at 2×10⁸ m. Capturing into a circular orbit requires removing almost all of the spacecraft’s excess energy, while an elliptical orbit only requires slowing it enough to stay bound to Jupiter. Galileo’s Io flyby just before insertion also reduced the burn it needed. This comparison suggests that the circular capture orbit, rather than the patched-conic approximation itself, is the main reason the model’s total Δv is so high.

Overall, the model shows the expected behaviour of an interplanetary transfer: a fuel-efficient but long-duration Hohmann transfer, with most of the Δv required for capture, and a relatively wide but manageable patched-conic boundary near the target planet.

5.2 Limitations

The patched-conic method treats each phase as an independent two-body problem, ignoring simultaneous gravitational effects from all three bodies. Moreover, Earth and Jupiter are assumed to be on circular, coplanar orbits, though both have small real eccentricities and inclinations. Additionally, departure and capture burns are modeled as instantaneous impulses rather than finite-thrust maneuvers, and perturbations such as solar radiation pressure, other planets, and Jupiter’s oblateness are neglected.

5.3 Future Extensions

In future work, a full n-body numerical propagation would capture simultaneous gravitational effects from the Sun, Earth, and Jupiter, giving a more accurate trajectory than the patched-conic approximation. Incorporating the true eccentricities and orbital inclinations of Earth and Jupiter, rather than assuming circular, coplanar orbits, would refine the launch-window and phasing calculations. Moreover, replacing impulsive burns with finite-thrust maneuver models would better represent real propulsion systems. Furthermore, the analysis could be extended to compare alternative trajectories, such as bi-elliptic transfers, to evaluate trade-offs between transfer time and fuel cost. Finally, adding perturbing forces, such as solar radiation pressure, would push the model closer to a mission-realistic trajectory design.

References

Danciu, B. (2023). Interplanetary mission design [Computer software]. MATLAB Central File Exchange. https://www.mathworks.com/matlabcentral/fileexchange/66192-interplanetary-mission-design

Danciu, B. (2026). Interplanetary trajectory [Computer software]. GitHub. https://github.com/bogdandanciu/interplanetary_trajectory

Lutze, F. H. (n.d.). Summary of patched conic approximations [PDF]. Virginia Tech, Aerospace and Ocean Engineering. Retrieved September 29, 2026, from https://archive.aoe.vt.edu/lutze/AOE4134/patchedconiceqs.pdf

Schaub, H. (n.d.). ScenarioPatchedConics. Basilisk. Retrieved September 29, 2026, from https://hanspeterschaub.info/basilisk/examples/scenarioPatchedConics.html

NASA. (2023). Chapter 4: Trajectories. NASA Science. https://science.nasa.gov/learn/basics-of-space-flight/chapter4-1/

Solar System Dynamics. (n.d.). Planetary physical parameters. NASA Jet Propulsion Laboratory. Retrieved September 29, 2026, from https://ssd.jpl.nasa.gov/planets/phys_par.html

Curtis, H. D. (2021). Orbital mechanics for engineering students (4th ed.). Butterworth-Heinemann. https://doi.org/10.1016/C2020-0-01873-6

Fidkowski, K., Atkins, E., & Powell, K. (2019). Course notes for Introduction to Aerospace Engineering. University of Michigan.

D’Amario, L. A., Bright, L. E., & Wolf, A. A. (1992). Galileo trajectory design. Space Science Reviews, 60, 23–78. https://doi.org/10.1007/BF00216849

Byrnes, D. V., & D’Amario, L. A. (1983). Interplanetary trajectory design for the Galileo mission. AIAA Paper 83-0099. https://ntrs.nasa.gov/citations/19830036690

Comparison of a simple patched conic trajectory code to commercially available software. AAS 07-160 (2007). NASA Technical Reports Server. https://ntrs.nasa.gov/api/citations/20070010447/downloads/20070010447.pdf

Reppert, T. (2006). Extending the patched-conic approximation to the restricted four-body problem [Undergraduate thesis]. Supervised by H. Schaub. https://hanspeterschaub.info/Papers/UnderGradStudents/Reppert2006.pdf

ScienceDirect. (n.d.). Patched conic [Topic overview]. Retrieved September 29, 2026, from https://www.sciencedirect.com/topics/engineering/patched-conic

About the author

Nour Saffar

Nour is a high school student with a strong interest in mathematics, chemistry, and engineering, particularly chemical engineering. Through her research project, Nour realized that although aerospace engineering had previously been her dream, she is now more interested in pursuing chemical engineering.

Academically, Nour has maintained a strong performance and has been part of her school’s Honor Roll. Outside the classroom, Nour is involved in athletics, STEM initiatives, and community activities. She is currently the captain of his school’s FIRST Robotics team, where she aims to create opportunities for students to explore engineering, robotics, and scientific problem-solving. Nour also enjoys writing, reading, and learning about different cultures and perspectives.

The post Modelling an Interplanetary Transfer from Earth to Jupiter via Patched Conics appeared first on Exploratio Journal.

Abstract

Texas maintains the largest state bridge inventory in the United States, creating a need for transparent tools that can support condition screening across a large network. We evaluate the 2025 Texas National Bridge Inventory (NBI), which contains 56,951 bridges labeled Good, Fair, or Poor. We aim to predict the bridge condition using 26 features including inventory, traffic, geometric, service, material, and deck-design while excluding component ratings and other fields that would leak the target. Cumulative-threshold ordinal logistic regression is evaluated with a five-fold stratified cross-validation procedure. We compare three operating approaches: standard ordinal logistic regression, fully class-weighted ordinal logistic regression, and a directional safety-cost decision rule applied to the standard model’s probabilities. The ordinary model provides the highest overall accuracy (72.45%) but identifies only 14 of 680 Poor bridges (2.05%). The class-weighted model identifies 580 Poor bridges (85.29% recall), and achieves 58.32% accuracy. The directional cost rule reduces the most severe Poor-to-Good error to 3 cases (lowest among all models) but identifies only 291 Poor bridges exactly and mislabels 386 as Fair. We therefore select the class-weighted model for a safety-first screening objective that prioritizes exact recognition of actually Poor bridges. Its 10,338 conservative Poor alerts create additional inspection workload, but they are not treated as the same safety failure as overlooking an actual Poor bridge. The study contributes a Texas-only, statewide 2025 Good/Fair/Poor evaluation with leakage control, reproducible validation, class-sensitive training, and explicit directional-error analysis.

Introduction

Bridge inspection data support decisions about maintenance, rehabilitation, replacement, and inspection scheduling. The National Bridge Inventory provides a consistent federal record of bridge attributes and condition information, making it a major source for data-driven infrastructure research. Texas is especially important because the 2025 NBI contains 56,951 Texas bridges, more than any other state. A network of this scale makes manual review of every record difficult and increases the value of transparent screening models that can organize attention without replacing engineering inspection.

Machine learning can relate observed condition labels to age, traffic, geometry, material, location, and other inventory characteristics. However, the apparent performance of a model depends on the target definition, the prediction horizon, the state or national population, and the treatment of class imbalance. In bridge data, Poor-condition examples are rare. A model can therefore achieve respectable overall accuracy by fitting the large Good and Fair groups while failing to recognize the class of greatest safety interest. This makes the confusion matrix, per-class recall and precision, balanced accuracy, and macro-F1 essential complements to accuracy.

The direction and distance of an error also matter because Good, Fair, and Poor are ordered. Any actual Poor bridge assigned Fair or Good is underpredicted; assigning it Good is the more severe two-level error because the label is especially reassuring. Errors in the opposite direction, such as assigning an actual Good or Fair bridge to Poor, are conservative over-alerts that may increase inspection workload but do not hide deterioration. We therefore make exact Poor recall and total Poor underprediction the primary safety measures, while still reporting accuracy, error direction, and false-alert burden for transparency.

We aim to (1) establish what prior research has and has not done with Texas NBI data; (2) distinguish training loss, class weighting, and post-training decision costs; (3) build a leakage-controlled ordinal classifier with 26 predictors; and (4) identify which transparent operating approach most effectively recognizes actually Poor bridges. The primary selection objective is to maximize exact Poor recall and minimize Poor bridges assigned to Fair or Good; Poor-to-Good remains the most severe type of underprediction.

Background

Machine learning for bridge condition prediction using NBI dataset

Research using state and national NBI subsets has examined current condition classification, next-inspection prediction, and long-term deterioration trajectories. The targets are not interchangeable. Some studies predict deck, superstructure, or substructure ratings; others predict structural evaluation or an overall rating derived from the minimum component condition. Studies also differ in whether the original 0-9 NBI scale is retained, ratings are grouped into binary classes, or conditions are grouped as Good, Fair, and Poor. Reported accuracy values should therefore be interpreted inside each study’s target and validation design rather than ranked as if they measured the same task.

State-specific work demonstrates the usefulness of geographically focused modeling. Assaad and El-Adaway (2020) compared computational approaches for Missouri deck deterioration. Rashidi Nasab and Elzarka (2023) evaluated multiple algorithms and feature selection methods for Ohio concrete bridge decks. Mia and Kameshwar (2023) predicted future component-condition classes from Louisiana histories. Liu and El-Gohary (2020, 2022) used Washington bridge data and deep-learning methods for deterioration prediction. These studies show that state-specific models are established, but their component targets, history requirements, and imbalance methods differ from a single-year statewide overall-condition classifier.

National studies provide larger and more varied training populations. Fard and Fard (2024) used more than one million historical U.S. records to predict deck ratings with random forest, XGBoost, and artificial neural networks. Chand and Choe (2026) evaluated 142,265 U.S. reinforced-concrete bridges and reported state-level performance for Texas and other subgroups. Such work includes Texas data but does not constitute Texas-only training and testing. The distinction is important because a national model can learn interstate patterns that may not match a Texas-exclusive population.

What prior research has done specifically in Texas

The 2019 TxDOT report requires careful classification. Its title and abstract identify its main purpose as a synthesis of bridge service-life prediction methods and practices for Texas. It is not primarily a machine-learning bridge-condition paper. Nevertheless, Chapter 4 contains an embedded case study using 1994-2016 Texas NBI records. The case study used 289,529 bridge-year observations, designated NBI Item 67 structural evaluation as the response, compared Naive Bayes, logistic regression, and decision-tree classifiers, and used a 75/25 split. Table 7 reports 0.95 accuracy, precision, and recall for the decision tree. The source is therefore adjacent at the report level but direct within its embedded ML subsection (Gao et al., 2019, PDF pp. 58-60).

Fang et al. (2023) is a direct bridge-condition prediction article. Its introduction states that feature engineering was performed on Texas bridge inventory data. The authors built four binary models for deck, superstructure, substructure, and structural evaluation, grouping ratings 0-6 and 7-9. The model-development sample counts range from 34,764 to 55,476, and the target-specific held-out accuracies range from 77.62% to 88.57%. The implementation case is different: Section 5 applies the structural-evaluation model to South Carolina bridges. Fang et al. is therefore Texas-specific in model development but not Texas-exclusive from data preparation through practical application.

Zhang et al. (2024) is the closest verified Texas-only overall-condition forecasting precedent. The study uses Texas NBI histories from 1992 to 2022, selects features with ReliefF, trains an Elman neural network, and uses a Markov chain to describe future deterioration. The accessible article reports 5,600 eligible bridge histories and a 70/20/10 training, validation, and testing division. Its overall-condition target and Texas-only history are directly relevant, but its longitudinal prediction problem differs from the present cross-sectional 2025 Good/Fair/Poor classification.

Two 2025 studies further limit broad originality claims while clarifying the present gap. Bayat, Kharel, and Li (2025) is an ASCE Technical Paper in the Journal of Structural Design and Construction Practice. It predicts Texas bridge deck condition from TxDOT historical data using autoencoder representations with random forest, XGBoost, and neural networks. Bayat and Kharel (2025) is an open-access research article focused on 1,443 Texas off-system bridge samples, including only eight observations in Condition 4. It combines focal-loss GAN augmentation, Tomek Links, and random-forest classification to improve minority-condition recognition. Both are Texas-only condition studies, but both focus on component or off-system subsets rather than the complete 2025 statewide overall Good/Fair/Poor population.

Loss functions, imbalance, and ordinal safety

Cross-entropy, or log loss, is the standard objective for probabilistic multiclass classification. Class weighting modifies this objective so that errors from rare classes have greater influence during fitting. Focal loss modifies cross-entropy by reducing the influence of easy examples and emphasizing difficult examples. Liu (2019) compared multiclass focal loss with cross-entropy for bridge deterioration, and later studies by Liu and El-Gohary (2020, 2022) extended focal-loss bridge analytics. Bayat and Kharel (2025) provides a direct Texas precedent for focal-loss-based imbalance treatment. Focal loss is therefore not itself a new Texas bridge contribution.

Sampling procedures must be discussed separately. SMOTE and ADASYN synthesize minority observations, undersampling removes majority observations, Tomek Links cleans overlapping class boundaries, and GANs generate additional data. These methods change the training sample rather than the loss. If used, they must be fitted only inside training folds to avoid contamination of validation or test data. The present first experiment uses class weighting because it is easy to audit and does not generate synthetic combinations of coded inventory attributes.

Class weighting and focal loss change how strongly training emphasizes difficult or rare examples. We use cumulative-threshold logistic regression to respect the order Poor < Fair < Good and compare ordinary training with full inverse-frequency class weighting. We also apply a directional cost matrix to the ordinary model’s probabilities as a sensitivity analysis. The cost matrix is a post-training decision rule, not a separately trained third model. Because this study’s primary safety objective is exact recognition of actually Poor bridges, the final approach is selected by Poor recall and total Poor underprediction rather than average cost alone.

Positioning of the present study

The verified literature does not support a claim that machine learning has never been used to model Texas bridge condition. It does support a narrower contribution. None of the verified sources matches the complete combination of the official 2025 Texas population, an overall ordered Good/Fair/Poor target, 26 leakage-controlled predictors, pooled five-fold stratified validation, and an explicit comparison of severe underprediction with false-alert burden. We therefore describe the contribution as a distinct statewide modeling and evaluation design rather than the first use of machine learning on Texas bridges.

Methodology

Dataset, Preprocessing, and Feature Engineering

We use the Federal Highway Administration’s (FHWA) 2025 National Bridge Inventory dataset for Texas. Each row represents one bridge record, and the structure identifier is unique in the analysis file. The final population contains 56,951 bridges. The target is the FHWA bridge-condition category: Good (G), Fair (F), or Poor (P). The distribution is 29,423 Good bridges (51.66%), 26,848 Fair bridges (47.14%), and 680 Poor bridges (1.19%). The severe imbalance of the Poor class motivates class-sensitive evaluation.

Twenty-Six Features and Leakage Control

We use 26 predictors. Fourteen are numeric:
Age_027,
Age_Reconstructed_106,
TRAFFIC_LANES_ON_028A,
TRAFFIC_LANES_UND_028B,
ADT_029,
DEGREES_SKEW_034,
MAIN_UNIT_SPANS_045,
APPR_SPANS_046,
MAX_SPAN_LEN_MT_048,
STRUCTURE_LEN_MT_049,
ROADWAY_WIDTH_MT_051,
APPR_WIDTH_MT_032,
DECK_WIDTH_MT_052,
PERCENT_ADT_TRUCK_109.
Twelve coded fields are categorical:
SERVICE_LEVEL_005C,
BASE_HWY_NETWORK_012,
FUNCTIONAL_CLASS_026,
DESIGN_LOAD_031,
SERVICE_ON_042A,
SERVICE_UND_042B,
STRUCTURE_KIND_043A,
STRUCTURE_TYPE_043B,
DECK_STRUCTURE_TYPE_107,
SURFACE_TYPE_108A,
MEMBRANE_TYPE_108B,
DECK_PROTECTION_108C.
Age_Reconstructed_106 uses the reconstruction year when it is positive and otherwise falls back to the original build year.

Deck, superstructure, substructure, culvert, structural-evaluation, and lowest-rating fields are excluded. The bridge-condition target is derived from component condition information, so using those fields as predictors would allow the model to recover the label directly and inflate performance. The structure identifier is retained only for uniqueness and audit checks, not as a predictor. All preprocessing is contained within the model pipeline so that imputation, scaling, and category encoding are learned separately from each training fold rather than from held-out records.

Model Development and Validation

We use five-fold stratified cross-validation as the primary validation design. Records are shuffled with random seed 42 and divided into five approximately equal folds. Each bridge is evaluated once by a model that was not trained on that bridge, producing one pooled out-of-fold prediction for every record. Imputation, scaling, category encoding, and any class weights are learned only from each training fold. Pooling the five validation folds produces a confusion matrix over all 56,951 bridges while preserving out-of-sample evaluation for every record. We fit cumulative-threshold ordinal logistic regression. One binary logistic equation estimates whether condition is above Poor, and a second estimates whether condition is above Fair; the two outputs are combined into valid Good/Fair/Poor probabilities. We evaluate three operating approaches on identical folds: (1) ordinary ordinal logistic regression with the highest-probability prediction, (2) the same ordinal structure with full inverse-frequency class weighting during training, and (3) a directional safety-cost sensitivity rule applied to the ordinary model’s probabilities. The third approach changes the final decision, not the fitted coefficients. The class-weighted model is the selected safety-first model because it yields the highest exact Poor recall and the fewest total Poor underpredictions.

How to read Table 2: a cost is a decision weight, not a number of bridges. For an actual Poor bridge, predicting Good receives the largest penalty (20), while predicting Fair receives a smaller penalty (10). This ordering reflects the study’s assumption that a Good label is the most falsely reassuring outcome. For each bridge, the selected prediction is the class with the lowest expected cost after the model probabilities are multiplied by the matrix. A zero on the diagonal represents a correct prediction.

Why Logistic Regression?

We use logistic regression because it is a standard baseline for classification, its coefficients and predicted probabilities are inspectable, and the calculation is simple enough to reproduce in a spreadsheet after preprocessing. These properties make the model easier to audit than a more complex black-box method. The ordinal cumulative-threshold construction also uses the known order Poor < Fair < Good rather than treating the labels as unrelated categories.

Evaluation Protocol and Metrics

We compute every metric from the pooled five-fold out-of-fold confusion matrix. For each class k, TP is the number correctly predicted as k, FP is the number incorrectly predicted as k, and FN is the number of actual k bridges assigned another label. Accuracy measures all exact classifications. Precision asks how often a predicted class is correct; recall asks how much of an actual class is found; and F1 balances precision and recall. Balanced accuracy and macro-F1 give each class equal influence. Because Poor is rare and underprediction is the main safety concern, we also report exact Poor recall, the Poor-underprediction rate, Poor-to-Good and Poor-to-Fair counts, and false Poor alerts. The model-selection hierarchy is: first maximize exact Poor recall, then examine the severity of remaining underpredictions and overall accuracy. Conservative false Poor alerts are reported as workload rather than treated as the same safety failure.

Accuracy = (TP_G + TP_F + TP_P) / N

Precision_k = TP_k / (TP_k + FP_k)

Recall_k = TP_k / (TP_k + FN_k)

F1_k = 2(Precision_k x Recall_k) / (Precision_k + Recall_k)

Balanced accuracy = (Recall_G + Recall_F + Recall_P) / 3

Macro-F1 = (F1_G + F1_F + F1_P) / 3

Poor underprediction rate = (Poor-to-Good + Poor-to-Fair) / Actual Poor

Results & Discussion

Five-fold cross-validated performance

The ordinary ordinal model has the highest accuracy, 72.45%, because it performs well on the dominant Good and Fair classes. It correctly identifies only 14 of 680 Poor bridges, however, for 2.06% Poor recall. Twenty-nine actual Poor bridges are labeled Good and 637 are labeled Fair. Overall accuracy alone therefore does not provide adequate evidence for rare-condition screening.

Full class weighting is the selected safety-first model. It identifies 580 of 680 actually Poor bridges as Poor, giving 85.29% Poor recall, and underpredicts only 100 Poor bridges: 63 as Fair and 37 as Good. This is the smallest total Poor underprediction among the three approaches. Its overall accuracy is 58.32%. It also produces 10,338 predicted-Poor alerts for actual Fair or Good bridges, including 1,850 Good-to-Poor conservative two-level errors. Under the study’s stated objective, these are accepted as additional inspection workload because they do not hide an actual Poor bridge.

The directional cost rule is retained as a sensitivity analysis. It reduces actual-Poor-to-predicted-Good errors to 3 of 680, but 386 Poor bridges are labeled Fair and only 291 are labeled Poor. Exact Poor recall is therefore 42.79%, and the total Poor-underprediction rate is 57.21%. Although the rule has the lowest average directional cost under Table 2, that average also rewards avoiding thousands of conservative false-Poor alerts. It does not match the present safety-first preference, which accepts those alerts in exchange for recognizing more actually Poor bridges exactly.

Implications for loss functions and decision costs

The three approaches answer different operational questions, but the paper makes one primary recommendation. If statewide exact accuracy is the only objective, the ordinary model is strongest. For the study’s safety-first objective – assigning as many actually Poor bridges as possible to Poor and accepting conservative alerts – the class-weighted ordinal model is strongest. The directional cost rule is preferable only under a different objective that places greater weight on reducing false-Poor workload and specifically minimizing Poor-to-Good errors, even when many Poor bridges remain labeled Fair.

Reproducibility

The polished analysis notebook downloads the official FHWA TX25.txt file, records its SHA-256 hash, constructs the 26 predictors, fixes randomized operations at seed 42, fits preprocessing only within training folds, and regenerates every table and figure. It also preserves a bridge-level interactive map. The processed dataset and code are available upon request.

Limitations

This analysis is cross-sectional. It classifies 2025 condition labels from 2025 inventory attributes and does not yet demonstrate prediction of a future inspection rating. A future forecasting study should construct bridge histories, use earlier years as predictors, and evaluate a later year or multi-year horizon without allowing future information into training. Five-fold random validation measures within-year generalization; geographic, district-held-out, and temporal validation may be more demanding.

The directional cost matrix is an explicit research assumption rather than a calibrated agency cost model. Its values should be reviewed with bridge engineers and asset managers. Poor precision in the selected class-weighted model is limited because only 1.19% of the inventory is Poor, so conservative screening creates many additional inspections. That workload is a real operational limitation even though it is not the same safety error as overlooking deterioration. The cumulative-threshold logistic model is transparent and respects class order but may not capture nonlinear relationships. The literature review is focused rather than registered and systematic, so the contribution claim remains limited to the verified sources. Future work should add temporal and district-held-out validation and should calibrate the class weights and decision costs with agency inspection capacity.

Conclusion

We conclude that statewide Texas bridge-condition modeling should be evaluated against the decision it must support, especially the consequence of underpredicting an actually Poor bridge. The ordinary ordinal model reaches 72.45% accuracy but identifies only 14 of 680 Poor bridges. The selected class-weighted ordinal model identifies 580 Poor bridges exactly, achieving 85.29% Poor recall and reducing total Poor underpredictions to 100. The directional cost rule reduces Poor-to-Good errors to 3 but identifies only 291 Poor bridges exactly and leaves 386 in Fair. Therefore, class weighting provides the best match to the study’s safety-first objective.

Our recommendation is qualified. The selected model’s 58.32% accuracy is lower than the ordinary baseline, its Poor precision is 5.31%, and it generates 10,338 conservative Poor alerts for actually Fair or Good bridges. It should therefore be used as a screening model whose Poor predictions trigger review, not as an autonomous condition assignment or a replacement for inspection. The ordinary model remains stronger when raw accuracy is the sole objective, while the directional cost rule remains useful when minimizing Poor-to-Good errors and inspection workload is more important than maximizing exact Poor recall.

The literature establishes that Texas bridge-condition machine learning is not entirely new. Texas-only precedents include Zhang et al. (2024) and recent deck-focused studies, while Fang et al. (2023) uses Texas data for model development and a South Carolina implementation case. Our work remains distinct in its combined use of the complete 2025 Texas inventory, an overall ordered Good/Fair/Poor target, 26 leakage-controlled predictors, pooled five-fold out-of-fold evaluation, class-weighted ordinal modeling, and a directional cost sensitivity analysis. Future work should test future-year forecasting, district-held-out validation, nonlinear ordinal models, and engineer-reviewed class-weight and cost calibration.

References

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). https://doi.org/10.1061/(ASCE)IS.1943-555X.0000572

Bayat, M., & Kharel, S. (2025). Leveraging artificial intelligence for predictive maintenance and condition rating of off-system bridges. Applied Sciences, 15(21), 11301. https://doi.org/10.3390/app152111301

Bayat, M., Kharel, S., & Li, J. (2025). An autoencoder-based machine- and deep-learning approach for predicting bridge deck conditions in Texas. Journal of Structural Design and Construction Practice, 30(4), 04025080. https://doi.org/10.1061/JSDCCC.SCENG-1799

Chand, P., & Choe, D.-E. (2026). Bridging the gap: A comprehensive analytical study of traditional, hybrid, and explainable AI for bridge condition prediction. Results in Engineering, 29, 109791. https://doi.org/10.1016/j.rineng.2026.109791

Fang, J., Hu, J., Elzarka, H., Zhao, H., & Gao, C. (2023). An improved inspection process and machine-learning-assisted bridge condition prediction model. Buildings, 13(10), 2459. https://doi.org/10.3390/buildings13102459

Fard, F., & Fard, F. S. N. (2024). Development and utilization of bridge data of the United States for predicting deck condition rating using RF, XGBoost, and ANN. Remote Sensing, 16(2), 367. https://doi.org/10.3390/rs16020367

Federal Highway Administration. (2025). National Bridge Inventory: Texas delimited data file (TX25.txt). https://www.fhwa.dot.gov/bridge/nbi/2025/delimited/TX25.txt

Federal Highway Administration. (n.d.). Specifications for the National Bridge Inventory. https://www.fhwa.dot.gov/bridge/snbi.cfm

Gao, L., Mo, Y.-L., Dhonde, S., Saldarriaga, D., Song, L., & Senouci, A. (2019). Synthesis of service life prediction for bridges in Texas: Final report (FHWA/TX-19/0-6938-1). Texas Department of Transportation. https://rosap.ntl.bts.gov/view/dot/63293/dot_63293_DS1.pdf

Liu, K. (2019). Machine learning-based analytics of structured and unstructured bridge data for data-driven bridge deterioration prediction [Doctoral dissertation, University of Illinois Urbana-Champaign]. https://www.ideals.illinois.edu/items/113975

Liu, K., & El-Gohary, N. (2020). A smart bridge data analytics framework for enhanced bridge deterioration prediction. Construction Research Congress 2020. https://par.nsf.gov/servlets/purl/10295991

Liu, K., & El-Gohary, N. (2022). Deep learning-based analytics of multisource heterogeneous bridge data for enhanced data-driven bridge deterioration prediction. Journal of Computing in Civil Engineering, 36(3). https://doi.org/10.1061/(ASCE)CP.1943-5487.0001018

Mia, M. M., & Kameshwar, S. (2023). Machine learning approach for predicting bridge components’ condition ratings. Frontiers in Built Environment, 9, 1254269. https://doi.org/10.3389/fbuil.2023.1254269

Rashidi Nasab, A., & Elzarka, H. (2023). Optimizing machine learning algorithms for improving prediction of bridge deck deterioration: A case study of Ohio bridges. Buildings, 13(6), 1517. https://doi.org/10.3390/buildings13061517

Scikit-learn developers. (n.d.). Model evaluation: Quantifying the quality of predictions. https://scikit-learn.org/stable/modules/model_evaluation.html

Zhang, T., Chen, H., Cui, X., Li, P., & Zou, Y. (2024). Condition rating prediction for highway bridge based on Elman neural networks and Markov chains. Applied Sciences, 14(4), 1444. https://doi.org/10.3390/app14041444

About the author

Temitope Ogunyomi

Temitope (Tope) O. Ogunyomi is a 12th-grade student at Foster High School in Richmond, Texas. Tope is a 3x Junior Olympian and Junior Olympic bronze medalist in track and field. Beyond athletics, he is interested in engineering, technology, machine learning, and using research to address real-world problems. Tope hopes to continue pursuing academic research and athletics in college while developing solutions that positively affect communities.

The post Safety-Aware Ordinal Logistic Regression for Bridge-Condition Classification Using the 2025 Texas National Bridge Inventory appeared first on Exploratio Journal.

Abstract

The automotive industry is rapidly changing together with increasingly tougher environmental regulations; including, among others, the shift towards use of low-carbon-footprint products, and ban of per- and polyfluoroalkyl substances (PFAS). The remarkable thermal stability, chemical resistance, and mechanical durability of fluorocarbon elastomers (FKMs) are essential in seals, hoses and tubings of internal combustion engines. These are difficult to match by any other materials making them so far irreplaceable.  The new regulations on PFAS, however, stimulated researchers to test the alternative non-fluorinated elastomers. These materials are additionally challenged by developments of new fuels, i.e. higher energetic synthetic e-fuels, which compromise their long-term performance. This manuscript summarizes use of FKMs in passenger cars, regulatory challenges and environmental concerns on PFAS, as well discussed technical challenges associated with substitution of fluorinated materials to be exposed with ethanol-based fuel, such as E85. Material replacement strategies, implications for road and race cars, and future developments in sealant technologies are discussed.

Introduction

Fluoroelastomers in Automotive Sealing Applications

Fluoroelastomers (FKMs) are a class of high-performance synthetic elastomers characterized by a fluorine-rich polymer backbone (60-75 wt.% fluorine), typically comprising vinylidene fluoride and hexafluoropropylene units (Scheme 1). FKMs were specifically developed to address sealing challenges in environments involving high temperatures, aggressive chemicals, and long service lifetimes, particularly in the automotive and aerospace sectors (Drobny, 2007). Places where, FKM-based parts are used in a combustion engine car are presented in Figure 1. Sealing materials used in the automotive engines and fuel systems are subjected to high temperatures, variable pressure, aggressive lubricants, and fuel compounds that are growing in complexity (Nishina, 2008). The FKMs, which are elastomers with fluorine-based polymer backbone, are characterized by an exceptional mix of heat resistance, chemical inertness, low permeability, and long service life (Moon et al., 2021). These properties have made FKMs the material of choice in critical components like O-rings, gaskets, shaft seals, injector seals, and fuel hoses, especially in areas of the engine where a failure would lead to leakage, emissions, or even safety hazards. Their elasticity and resistance to swelling and degradation have seen them become a standard material in the sealing technology of automobiles. However, currently FKM alternative sealants are also considered, such as silicone-based or nitrile-butadiene rubbers (Scheme 1). Their performance will be discussed in the sections below.

Influence of Alternative and Ethanol-Based Fuels on Elastomer Performance

The relevance of FKMs has escalated further with the diversification of automotive fuels. Blended fuels like E5, E10, and E85, biodiesel, and new synthetic e-fuels have been progressively replacing or supplementing conventional gasoline and diesel (Czerwinski et al., 2016; Muelaner, 2023). Fuel containing alcohol, e.g. higher concentration of ethanol, poses new threats to elastomeric materials because they are polar, hygroscopic, and may behave as swelling solvents. Ethanol can enter into elastomer networks, remove additives, and enhance swelling or embrittlement, ultimately degrading seal integrity (U.S. Department of Energy (DOE), n.d.). FMKs with high concentrations of fluorine (usually 69-71% fluorine) are less susceptible to these effects and have longer-lasting dimensional stability and mechanical performance even in long-term contact with alcohol-containing fuels (Stevens, 2006). This is the reason why they find extensive application in the fuel system parts, such as in-tank seals and tubing, where they are likely to be subjected to hostile fuel environments over extended periods.

PFAS Regulatory Pressure and Environmental Concerns

Although fluorinated materials have technical benefits, there is a growing wave of questioning the use of fluorinated materials because of the environmental and regulatory factors related to per- and polyfluoroalkyl substances (PFAS) (Améduri, 2023). PFAS have been referred to as “forever chemicals” due to their environmental persistence and their resistance to degradation and low bioaccumulation possibilities (Brun et al., 2023; Lee et al., 2025). Regulatory agencies, especially the European Union, are working towards general bans on PFAS that have cast doubt on the future use of fluorinated polymers in industries (European Commission, 2025). One of the main controversies is whether high-molecular-weight fluoropolymers like FKMs should be considered together with low-molecular-weight PFAS that are already known to be dangerous to the environment and health. Fluoropolymer can be subdivided into polymers where perfluorination resides either in the main chain or in the side chain (Figure 2). This differentiation is important, as in the advent of hydrolytic degradation, the perfluorinated side chains are released to the environment and their mobility and accumulation may be similar to other small molecular weight PFAS. In contrast, main-chain fluoropolymers, such as FKMs, do not hydrolyse and the risk of release of similar small molecular weight compounds is rather small.In turn, FKMs do not have high mobility and have low bioaccumulation; nevertheless, they are fluorinated in nature, which puts them in an ever-tighter regulatory environment.” FKMs do not have high mobility and have low bioaccumulation; nevertheless, they are fluorinated in nature, which puts them in an ever-tighter regulatory environment.

Sustainability Challenges and the Need for Material Substitution

This regulatory pressure is in line with an overall shift to sustainable mobility, such as the use of biofuels and e-fuels as solutions to lower greenhouse gas emission levels in internal combustion engines (Yılbaşı, 2025). Although these fuels have environmental advantages, they tend to cause compatibility challenges with materials, because they are more corrosive and chemically aggressive. This, in effect, makes it a significant technical challenge to replace fluorinated materials, as their unique performance characteristics remain difficult to replicate with non-fluorinated alternatives (Améduri, 2023). Any alternative substance cannot be inferior in terms of thermal stability, chemical resistance, elasticity, low permeability, and long durability, especially when subjected to fuels that are ethanol-based, like E85 (Sahu et al., 2022). This paper discusses the multifaceted nature of the relationship between material operation, fuel development, and regulatory requirements and explains why the substitution of FKMs remains one of the most pressing and unsolved issues in automotive material development.

Discussion

Currently Used Sealant Materials

Elastomers in Use: FKM and NBR

Fluoroelastomers (FKM) and nitrile butadiene rubber (NBR) (Scheme 1) are the most common elastomers used in automotive fuel and engine systems, chosen through a performance-to-cost ratio (Białecki et al., 2021). FKMs are normally utilized in the critical sealing processes that are subjected to high temperature, aggressive fuel, and extended working time (Asthana  et al., 2025). Their fluorine content is high, and this gives them high thermal stability, which means that they can run continuously above 200 C, and are also highly resistant to oils, fuels, and additives. Such characteristics render FKMs applicable in injector seals, O-rings, shaft seals, and in-tank fuel components. Conversely, NBR has found extensive use as a cost-sensitive component in certain applications because it is less expensive and good enough for non-polar hydrocarbon resistance of petroleum-based fuels and lubricants.

NBR reveals evident limitations despite the benefits in its economy, which is relevant in the contemporary operating conditions (Akhlaghi et al., 2015). It is rather thermally unstable with a maximum service temperature of approximately 120 °C, above which it becomes hardened and loses elasticity and acquires a high compression set. “Compare static thermal stability of NBR with FKM in TGA, Figure 3.” NBR is also highly susceptible to polar solvents and, therefore, is prone to swelling and degrading when subjected to fuels containing ethanol, like E10 or E85, which reduces the life of seals. On the other hand, FKMs are highly mechanically stable, have low fuel permeability, recover elastically at high temperature and alcohol-based fuels, and have a high service life covering mileage of over 100,000 km (Lee et al., 2022). Compare mechanical properties of NBR and FKM in Table 1. In turn, NBR is suitable for low-stress, low-cost applications, whereas FKMs are needed in high-performance, high-reliability automotive sealing system applications.

Fuel Types and Their Impact

The modern internal combustion engines use a wider variety of fuels, such as gasoline-ethanol mixtures (E5 and E10), high-ethanol fuels (E85), biodiesel, and synthetic fuels (Laskowski & Zimakowska-Laskowska, 2025). Chemical and physical characteristics of each fuel also have different values that affect engine performance and material behavior (Table 2). The ethanol fuels have a lower caloric value compared to gasoline, and thus, they consume more fuel to produce the same amount of energy (Yakın et al., 2022). On the other hand, higher octane rating of ethanol enables some engines to operate at higher compression ratios and even better efficiencies. Although E5 and E10 produce moderate variation of fuel system conditions, E85 produces a significant exposure of polar solvents and water absorption that can aggravate the chemical stress on elastomeric fuel-system components.

Synthetic fuels and biodiesel present a new range of problems. Biodiesel is more lubricious than standard diesel, which may be advantageous in fuel pumping and injecting, but it oxidizes more easily and is susceptible to microbial growth where there is water. These may result in the production of acidic by-products, which hasten the breakdown of elastomers, especially those of low chemical resistance. Biofuels are classified according to their physical state, technology maturity, the generation of feedstock, and the generation of products (Awogbemi et al., 2023). Synthetic fuels, such as e-fuels made of renewable electricity and captured carbon dioxide, are developed to attain similar properties to the conventional hydrocarbons without causing lifecycle greenhouse emissions (US Department of Energy, 2019). Despite being relatively cleaner and relatively chemically inert, they react to certain extent with elastomer alternatives according to certain formulations and additive packages

FKMs, PFAS, and Regulatory Debate

Fluoroelastomers are extremely resistant to alcohol-based fuels, have low permeability, and have a long service life, so they are an essential component in the automotive fuel system (Ameduri and Sawada, 2016). However, their fluorinated structure makes them the object of the extended regulations of PFAS. Even though the FKMs do not have the same mobility, bioavailability, and environmental behavior as low-molecular-weight PFAS, the regulatory definitions are generic. This has brought about confusion among manufacturers and suppliers, where alternative materials and fluorine-reduction strategies are being researched, even though there are no similar non-fluorinated substitutes with the same level of performance.

Strategies to Replace FKMs

Several alternatives are under development that will eliminate or limit the use of FKMs in automobiles. These consist of hydrogenated nitrile butadiene rubber (HNBR, Scheme 1), high-performance thermoplastic elastomers, and silicone elastomers that have improved fuel resistance (Joshi, 2025; Gao et al., 2025). Other mitigation strategies include hybrid seal or multilayer seal construction or surface coating, whereby the traditional elastomers are coupled with thin barrier seals to restrict fuel permeation (Figure 4). The compatibility of elastomers with various fuels indicates that most alternative rubbers (including HNBR and other engineered types) change their mechanical properties significantly under exposure to renewable and blended fuels, which makes it challenging to find alternatives with more or less the same performance as FKM ( Müller et al., 2024). Although they may partially reduce the effects of chemical attack and swelling, none of these solutions can be considered as having the same thermal stability, chemical resistance, and durability as FKMs, especially when exposed to ethanol-rich fuels like E85. Other alternative engineered elastomers such as HNBR, as well as most alternative rubbers, vary their mechanical properties greatly when in contact with renewable and blended fuels, especially when in extended contact with fuels high in ethanol (Conen, Haefele and Dahlmann, 2025).

Road Cars versus Race Cars

The operational and regulatory aspects vary significantly, and this difference in priorities leads to a significant difference between the selection of materials used in road vehicles and race cars. In racing use, performance, high temperature performance, and high temperature performance are crucial factors, and the need to match rapidly evolving fuel compounds is often the reason why FKMs or perfluoro elastomers are continued to be employed despite the cost or regulatory pressure (Katon, 2025). In contrast, road vehicles have to weigh between performance and cost-effectiveness, compliance with regulations, and durability during the long run due to the long service life (Adebowale, 2025). Although race cars still almost entirely use FKMs, the continued development of fluorine-free polymers and multilayer seal technologies implies that road vehicles can use other types of elastomer systems in the future (Mandlekar, Joshi & Butola, 2022; Valentini & Lopez-Manchado, 2020).

Summary

Fluor elastomers have been very important in providing reliability and safety of automotive engines, especially those with high temperatures in the fuel system and hostile fuels (Drobny, 2023). The shift to the use of ethanol-based and synthetic fuels, coupled with stricter regulations on PFAS, has put a lot of pressure on the need to substitute or redesign fluorinated substances. Although the FKMs are still unrivaled in terms of total performance, more so when it comes to accommodating E85 compatibility, regulatory and environmental issues are driving the development of alternative elastomers and hybrid solutions. Even in the near future, it is not probable that FKMs can be fully replaced without performance loss (Puga et al., 2025). Rather, the future of automotive sealants will be characterized by incremental changes in materials and an application-specific approach in both road and race applications. 

References

Adebowale, O. J. (2025). Battery module balancing in commercial EVs: strategies for performance and longevity. Int J Eng Technol Res Manag, 9(4), 162.). https://doi.org/10.5281/zenodo.15186621

Akhlaghi, S., Gedde, U. W., Hedenqvist, M. S., Braña, M. T., & Bellander, M. (2015). Deterioration of automotive rubbers in liquid biofuels: A review. Renewable and Sustainable Energy Reviews, 43, 1238-1248. https://doi.org/10.1016/j.rser.2014.11.096

Améduri, B. (2023). Fluoropolymers as unique and irreplaceable materials: challenges and future trends in these specific per or poly-fluoroalkyl substances. Molecules, 28(22), 7564.

Ameduri, B., & Sawada, H. (Eds.). (2016). Fluorinated Polymers: Applications: Volume 2. Royal Society of Chemistry. https://doi.org/10.1039/9781782626718-FP001

Asthana, P., McCormick, J., Salazar, O., & Verma, H. (2025). Application specific rapid gas decompression testing for oilfield elastomers under ultra-high H2S environment. Middle East Oil, Gas and Geosciences Show (MEOS GEO). https://doi.org/10.2118/227003-ms

Awogbemi, O., Kallon, D. V., Onuh, E. I., & Aigbodion, V. S. (2021). An overview of the classification, production and utilization of Biofuels for internal combustion engine applications. Energies, 14(18), 5687. https://doi.org/10.3390/en14185687

Białecki, T., Sitkiewicz, A., Giemza, B., Sarnecki, J., Skolniak, M., & Gawron, B. (2021). Compatibility of different automotive elastomers in paraffinic diesel fuel. Applied Sciences, 11(23), 11312. https://doi.org/10.3390/app112311312

Brunn, H., Arnold, G., Körner, W., Rippen, G., Steinhäuser, K. G., & Valentin, I. (2023). PFAS: Forever chemicals—persistent, bioaccumulative and mobile. Reviewing the status and the need for their phase out and remediation of contaminated sites. Environmental Sciences Europe, 35(1). https://doi.org/10.1186/s12302-023-00721-8

Conen, T., Häfele, B., & Dahlmann, R. (2025). Review on compatibility considerations of fuel system elastomers with paraffinic fuels. SAE International Journal of Fuels and Lubricants, 18(3). https://doi.org/10.4271/04-18-03-0017

Dhiman, S., & Ansari, N. G. (2024). A review on extraction, analytical and rapid detection techniques of per/poly fluoro alkyl substances in different matrices. Microchemical Journal, 196, 109667. https://doi.org/10.1016/j.microc.2023.109667

Drobny, J. G. (2007). Fluoropolymers in automotive applications. Polymers for Advanced Technologies, 18(2), 117-121. https://doi.org/10.1002/pat.807

Drobny, J. G. (2023). Fluorocarbon elastomers. Technology of Fluoropolymers, 133-186. https://doi.org/10.1201/9781003204275-12

European Commission. (2025). Biofuels. Energy. https://energy.ec.europa.eu/topics/renewable-energy/bioenergy/biofuels_en

Gao, X., Wang, X., Pan, W., Wang, M., Xu, X., Li, Z., & Zhao, S. (2025). Highly hydrogenated, solvent-resistant, low-temperature conductive nitrile rubber for multifunctional sensors. Chemical Engineering Journal, 504, 158735. https://doi.org/10.1016/j.cej.2024.158735

Joshi, A. M. (2024). Rubbers and elastomers in specialty applications. Specialty Polymers and Materials, 293-314. https://doi.org/10.1201/9781003561361-15

Katon. (2025). FFKM vs FKM: Key differences, properties, and applications. FFKM O-Ring Supplier | Advanced FFKM O-Rings at Smarter Prices – Katon®. https://katonseal.com/en/news/3/20

Kitajima, Y. (2023). Applying seal technology to meet needs: Supporting automobiles with NOK. NOK × Nikkan Kogyo Shimbun Journagram. https://www.nokgrp.com/en/technology/tec_column/000005/

Laskowski, P., & Zimakowska-Laskowska, M. (2025). Simulation study of the effect of ethanol content in fuel on petrol engine performance and exhaust emissions. Combustion Engines, 201(2), 129-135. https://doi.org/10.19206/ce-204049

Lee, P. C., Kim, S. Y., Ko, Y. K., Ha, J. U., Jeoung, S. K., Lee, J. Y., & Kim, M. (2022). Durability and service life prediction of fluorocarbon elastomer under thermal environments. Polymers, 14(10), 2047. https://doi.org/10.3390/polym14102047

Lee, J. C., Smaoui, S., Duffill, J., Marandi, B., & Varzakas, T. (2025). Forever chemicals PFAS global impact and activities, cascading consequences of colossal systems failure: Long-term health effects, food-systems, eco-systems. https://doi.org/10.20944/preprints202501.1698.v1

Mandlekar, N., Joshi, M., & Butola, B. S. (2022). A review on specialty elastomers based potential inflatable structures and applications. Advanced Industrial and Engineering Polymer Research, 5(1), 33-45. https://doi.org/10.1016/j.aiepr.2021.05.004

Masa, A., Hayeemasae, N., Ahmad, H. S., & Ismail, H. (2023). Nitrile glove waste as blending component for natural rubber and epoxidized natural rubber: functionality and thermal stability. MJCHEM, 25(4), 135-43. https://doi.org/10.55373/mjchem.v25i4.135

Moon, Y. I., Jung, J. K., Kim, G. H., & Chung, K. S. (2021). Observation of the relaxation process in fluoroelastomers by dielectric relaxation spectroscopy. Physica B: Condensed Matter, 608, 412870. https://doi.org/10.1016/j.physb.2021.412870

Muelaner, J. E. (2023). Decarbonized power options for non-road mobile machinery. https://doi.org/10.4271/epr2023002

Müller, M., Mishra, R. K., Šleger, V., Pexa, M., & Čedík, J. (2024). Elastomer-based sealing o-rings and their compatibility with methanol, ethanol, and hydrotreated vegetable oil for fueling internal combustion engines. Materials, 17(2), 430.
https://doi.org/10.3390/ma17020430

Nishina, H. (2008). Trends in automotive sealing materials technology. International Polymer Science and Technology, 35(3), 51-60. https://doi.org/10.1177/0307174×0803500315

Sahu, T. K., Shukla, P. C., Belgiorno, G., & Maurya, R. K. (2022). Alcohols as alternative fuels in compression ignition engines for sustainable transportation: A review. Energy Sources, Part A: Recovery, Utilization, and Environmental Effects, 44(4), 8736-8759. https://doi.org/10.1080/15567036.2022.2124326

Stevens, R. D. (2006). Fuel and Permeation Resistance of Fluoroelastomers to Ethanol Blends. In Rubber Mini Expo 06: 170 th Technical Meeting of the American Chemical Society, Rubber Division 2006.

U.S. Department of Energy (DOE). (n.d.). Biofuels: Energy for transportation. Energy.gov. https://www.energy.gov/cmei/fuels/biofuels-energy-transportation

US Department of Energy. (2019). Leveraging Biofuels for a cleaner mobility future. Energy.gov. https://www.energy.gov/cmei/articles/leveraging-biofuels-cleaner-mobility-future

Valentini, L., & Lopez-Manchado, M. A. (2020). Classification of rubbers and components for harsh environmental systems. High-Performance Elastomeric Materials Reinforced by Nano-Carbons, 1-14. https://doi.org/10.1016/b978-0-12-816198-2.00001-3

Vega Puga, E., Wachtendorf, V., Kömmling, A., Brendelberger, S., Jaunich, M., & Sattler, C. (2025). Lifetime prediction and degradation assessment of FKM and FFKM O-rings under high temperature thermo-oxidative ageing. Polymer Testing, 147, 108820. https://doi.org/10.1016/j.polymertesting.2025.108820

Xu, G. M., & Shuai, C. G. (2021). Axial and lateral stiffness of spherical self-balancing fiber reinforced rubber pipes under internal pressure. Science and Engineering of Composite Materials, 28(1), 96-106.

Yakın, A., Behcet, R., Solmaz, H., & Halis, S. (2022). Testing sodium borohydride as a fuel additive in internal combustion gasoline engine. Energy, 254, 124300. https://doi.org/10.1016/j.energy.2022.124300

Yang, G., Tong, L., You, Y., Lei, X., & Liu, X. (2019). A study on fluoroelastomer/MWCNTs-COOH dielectric composite with high temperature and acid resistance. Journal of Materials Science: Materials in Electronics, 30(17), 16359-16368.

Yılbaşı, Z. (2025). Biofuels, E-fuels, and waste-derived fuels: Advances, challenges, and future directions. Sustainability, 17(13), 6145. https://doi.org/10.3390/su17136145

About the author

DongHan Li

DongHan is currently Grade 11 student, studying in Beijing, China. During the course of high school, he has studied the A-level Curriculum in subjects such as physics and chemistry. DongHan has a strong interest in Engineering, especially in Chemical Engineering. Additionally, he has also found himself very intrigued by theories related to Nanotechnology along with fluid dynamics that have facilitated the process of completing this dissertation.

At the same time, DongHan’s interest in motorsports (Grand Touring 3 race cars ) has helped develop the general aim of this dissertation. As motorsports develop continuously over time, it is very promising and exciting to dig into and challenge the theories that already exist, hoping to offer alternative solutions to existing engineering challenges.

The post Fluoroelastomer Sealants and their Alternatives in Automotive Engines Supplied with E-fuels appeared first on Exploratio Journal.

Abstract

The field of bionic limb technology is currently undergoing a pivotal paradigm shift, transitioning from purely functional robotic design to highly integrated neuroprosthetic systems. While contemporary advancements in neural interfaces and powered robotics have improved user capabilities, clinical translation remains hampered by issues such as signal instability, limited sensory feedback, and poor biocompatibility of interfaces. This literature review synthesizes recent progress in the clinical translation of bionic limbs, focusing on three core pillars: advanced neural interfaces, osseointegration, and patient-centric design. We explore the role of bidirectional control, the application of reinforcement learning for robust intent decoding, and the integration of digital twins for personalised residuum management. By evaluating the co-evolution of surgical innovation and robotic hardware, this review highlights the critical need for a “clinical-push” methodology, in which design is driven by long-term residuum health rather than mere functional performance. The findings emphasise that successful clinical outcomes depend on treating the prosthetic and the user’s altered physiology as a single, integrated system, ultimately identifying key barriers to translation and proposing a framework for more equitable, accessible, and durable prosthetic care. 

Keywords: Neuroprosthesis, Bionic Limbs, Neural Interfaces, Osseointegration, Rehabilitation, Neuroplasticity, Bidirectional, Targetted Muscle Reinnervation (TMR), and Agonist-Antagonist Myoneural Interface (AMI)

1. Introduction: 

The cinematic image of a sophisticated, seamlessly integrated artificial limb, once a domain of science fiction, is slowly becoming a clinical reality. For centuries, the ambition to replace missing limbs with functional, anthropomorphic counterparts has driven immense innovation. Yet, the reality of many amputees remains defined by the limitations of traditional, socket-based prostheses. While these conventional devices provide basic mobility, they are frequently associated with chronic pain, skin breakdown, and osseous degeneration, leading to high rates of device abandonment. The core of this challenge lies in the “HMI paradox” whereby, as the level of amputation increases, the technical complexity required for control increases as well, while the available biological signals from the residual limb diminish. 

To overcome these barriers, the field has increasingly turned towards bionic limbs, which aim to replicate biological function through a marriage of robotics, surgery, and neuroscience. Unlike passive devices, modern bionic limbs leverage sophisticated neural interfaces, such as Targeted Muscle Reinnervation (TMR) and the Agonist-Antagonist Myoneural Interface (AMI), to create direct channels for motor intent and sensory feedback. Furthermore, the emergence of osseointegration, where prosthetics are directly skeletally attached, offers a stable, reliable foundation that bypasses the limitation of soft-tissue socket interfaces entirely. 

However, technology alone does not guarantee success. The history of prosthetic development has been dominated by a “technological-push” model, often neglecting the biological reality of the user. True clinical translation requires a transition toward a patient-centric approach that prioritises residuum health, embodiment, and long-term viability (Figure 1). This literature review examines the intersection of three fundamental domains: neural interfaces, which bridge the gap between brain and machine; osseointegration, which ensures skeletal and functional stability; and patient-centric design, which integrates digital monitoring and personalised rehabilitation (Figure 1). In the following sections, I will synthesise these advancements, starting with the technical evolution of bidirectional neural control, moving into the essential role of skeletal integration and residuum health, and concluding with a framework for how researchers may overcome current clinical and organisational barriers to make bionic limbs a sustainable, equitable standard of care. 

2. Neural Interfaces and Bidirectional Control

Establishing a reliable, bidirectional link with the peripheral nervous system serves as the foundational requirement for any functional bionic limb. Recent surgical breakthroughs seek to maximise the “information throughput” of this interface, ensuring that the human-machine interaction is not limited by a lack of viable control signals. 

Amputations affect hundreds of thousands of people each year in the United States and Europe, particularly. Therefore, there has been a growing interest in neuroprosthetics that can restore both movement and sensation. While today’s bionic limbs can perform increasingly precise movements, they still fall relatively short when replicating a natural limb. One of the biggest challenges regarding this is establishing a strong, long-term connection with the peripheral nervous system that can both monitor motor commands and deliver meaningful sensory feedback (Cho et al., 2023). Existing neural interfaces, whether engineered or biological solutions, often struggle with durability, biocompatibility, and limited signal variety. A new hybrid bionic interface has been established to bridge these gaps by combining a regenerative peripheral nerve interface (RPNI) with a traditional nerve interface with a single implant (Cho et al., 2023). The device uses a buckle-shaped design made from a shape memory polymer (SMP), allowing it to be easily implanted onto a severed nerve while also connecting to a muscle graft. Long-term in rabbits showed that the interface remained stable and biocompatible while reliably recording and stimulating neural signals (Cho et al., 2023). Most importantly, it was able to capture distinct motor signals during natural walking and use as a robotic leg. Overall, the hybrid interface offers a more flexible and information-rich way to connect nerves to prosthetics, bringing neuroprosthetic control closer to natural movement and sensation.

2.1 Continuous Neural Control of a Bionic Limb

A 2024 study in Nature Medicine highlights a significant paradigm shift in prosthetic technology by prioritising surgical-neural integration through the Agonist-Antagonist Myoneural Interface (AMI) (Song et al., 2024). Unlike contemporary bionic limbs that rely on intrinsic pre-defined control frameworks to estimate gait phases, the AMI procedure restores a degree of natural proprioception by surgically reconnecting residual muscle pairs. This interface allows the human nervous system to directly and continuously neuromodulate the bionic limb (Figure 2), augmenting residual muscle afferents by 18% if biologically intact values (Song et al., 2024). This restoration of sensory feedback proved sufficient to enable a highly biomimetic gait, with participants achieving a 41% increase in maximum walking speed compared to matched control cohorts (Song et al., 2024). Furthermore, the study demonstrates that this level of afferent augmentation allows for seamless adaptation to complex, real-world environments such as stairs, slopes, and obstructed pathways without the need for robotic decision-making (Figure 2). Ultimately, these results suggest that even partial reinstatement of the biological feedback loop is sufficient for the human nervous system to regain versatile, biomimetic motor control, offering a potent alternative to autonomous robotic systems.

2.2 Improving Bionic Limb Control

Freitag et al. (2026) investigated the use of reinforced learning (RL) to improve the decoding of human motor intent for bionic limb control using electromyographic signals (EMG) (Freitag et al., 2026). Accurate and robust intent decoding remains a major challenge in prosthetics, and most existing approaches rely on supervised learning (SL) trained on static, labeled EMG datasets (Freitag et al., 2026). However, such data fails to reflect natural muscle activity during real-world use, limiting online performance and usability. The researchers propose a novel training framework that combines SL and RL. An initial control policy is first pretrained using supervised learning based on EMG recordings. This policy is fine-tuned using offline reinforcement learning with dynamic EMG data collected while participants interact with a custom-designed, Guitar Hero-inspired game. The game environment enables real-time, human-in-the-loop interaction and provides a structured reward signal based on movement accuracy and timing. The Advantage Weighted Actor Critic (AWAC) algorithm is used to improve the policy while remaining close to previously learned behaviours (Freitag et al., 2026). Experiments were conducted with nine able-bodied participants performing simultaneous finger movements. Results show a threefold increase in normalized cumulative reward and more than a two-fold improvement in decoding accuracy compared to the supervised baseline (Freitag et al, 2026). Improvements generalised to a separate motion test, demonstrating enhanced robustness. The study concluded that reinforcement learning on interactive, usage-based data can bridge the gap between offline training and real-world prosthetic control, supporting more intuitive and reliable bionic limb operation.  

The future of bionic limbs depends on bridging the gap between advanced engineering and the biological reality of the post-amputation physiology of patients (Paslausta et al., 2022). Engineers have developed sophisticated bidirectional systems capable of decoding motor intent and providing sensor feedback using surgical techniques like Targeted Muscle Reinnervation (TMR) and Agonist-Antagonist Myoneural Interface (AMI) (Paslausta et al., 2022) (Figure 3). Methods like TMR and AMI are highlighted for their ability to recreate biological feedback loops, such as proprioception, that the brain expects to receive (Figure 3). Moreover, researchers highlight the need for biomimetic encoding and moving beyond simple electrical pulses to natural sensory algorithms with a deeper focus on motor control theory and understanding how the brain manages postural balance and embodiment (Paslausta et al., 2022).

Restoration of sensorimotor function in upper limb amputees has historically been limited by the lack of reliable control, sensory feedback, and comfortable attachment (Ortiz-Catalan et al., 2023). Traditional socket-based prostheses often cause discomfort and rely on surface electrodes that are prone to signal interference. Ortiz-Catalan et al. detail the clinical implementation of a transradial neuromusculoskeletal prosthesis, which connects a bionic hand directly to the user’s nervous and skeletal systems (Figure 4). In a landmark case study, a patient with a below-elbow amputation received titanium implants in the radius and ulna for osseointegration, providing stable skeletal attachment (Figure 4). Surgically, the interface was enhanced by creating regenerative peripheral nerve interfaces (RPNIs) through the transfer of severed nerves into free muscle grafts (Ortiz-Catalan et al., 2023). These grafts, along with native muscles and the ulnar nerve, were instrumented with implanted electrodes (Figure 4). The resulting system is entirely self-contained, requiring no external batteries or processing units for daily use. 

Long-term results over three years demonstrated significant improvements:

1. Enhanced Function: Improved scores in the Southampton Hand Assessment Procedure (SHAP) and Assessment of Capacity for Myoelectric Control (ACMC).
2. Sensory Feedback: Consistent tactile sensations elicited via direct neural stimulation of the ulnar nerve.
3. Quality of Life: A marked reduction in phantom limb pain and a substantial increase in overall well-being. 

This integrated approach overcomes the “HMI paradox” by providing a high-bandwidth, stable, and bidirectional communication channel between the patient and the bionic limb (Ortiz-Catalan et al., 2023).

3. Neuroplasticity and User Adaptation

Once a physical neural connection is achieved, the focus must shift to the central nervous”s systems” innate ability to adapt to artificial input. Understanding the brain’s capacity for cortical reorganisation is essential, as the user must essentially re-learn how to control a limb that no longer communicates through the traditional biological pathways. 

3.1 Bidirectional Bionic Limbs

Although engineers have developed sophisticated bidirectional systems capable of decoding motor intent and providing sensory feedback, these technologies often operate in a vacuum, failing to account for how nervous systems reshape after losing a limb (Paslausta et al., 2022). Amputation triggers significant cortical reorganisation, where the brain’s sensory and motor maps may shrink or shift over time (Paslausta et al., 2022). To address this, a transdisciplinary approach is required to manage neuroplasticity. Thus, a solution that combines high-tech hardware with innovative surgical techniques must be implemented. Ultimately, restoring the true function of a missing limb requires treating the prosthesis and the patient’s altered physiology as a single, integrated system rather than separate components (Paslausta et al., 2022).  

The concurrent TMR, targeted sensory reinnervation (TSR) for touch, and vibration-induced kinesthesia significantly advance bionic limb functionality toward able-bodied norms (Marasco et al., 2021) (Figure 5). By establishing a bidirectional neural-machine interface in participants with proximal neural-limb amputations, this approach moves beyond traditional myoelectric control to restore complex sensory-motor loops. The simultaneous delivery of tactile and kinesthetic feedback allows users to transition from speed-driven, compensatory behaviours to accuracy-maximising strategies, characterised by a marked reduction in visual dependency (Marasco et al., 2021). This restoration of feedback channels promotes a measurable shift in visuomotor behaviour whereby users redirect their gaze from the prosthetic hand toward the intended target while also enhancing limb ownership by aligning the device’s performance with the brain’s internal model (Marasco et al., 2021) (Figure 5). Ultimately, providing a multifaceted sensory experience is critical for reducing cognitive load and achieving “human-like” integration in prosthetic systems, effectively stratifying bionic performance away from conventional prosthetics and toward natural physiological function. 

3.2 Biometric Versus Arbitrary Motor Control

Contrary to the long-standing “biomimetic assumption” in neuroprosthetics, recent evidence from Nature Human Behaviour (2024) suggests that anthropomorphic control strategies, those that mimic biological movements, do not inherently provide superior long-term benefit for device learning or embodiment (Schone et al., 2024). By evaluating participants using a myoelectric bionic hand, researchers compared a biomimetic mapping against an arbitrary, non-biomimetic strategy to determine if recruiting pre-existing neural resources actually enhances performance. The findings revealed that while biomimetic control offered an initial “intuitive” advantage in speed and automaticity, these differences vanished as training progressed, with arbitrary users eventually matching their counterparts in dexterity and gesture switching (Marasco et al., 2021). The study further demonstrated that the arbitrary control group exhibited significantly higher levels of generalisation when they were tasked with a novel control mapping; arbitrary users maintained their proficiency, whereas biomimetic users experienced a sharp decline in performance (Marasco et al., 2021). Interestingly, both groups reported significant and comparable increases in a sense of embodiment, specifically regarding agency and body ownership, thus indicating that the bian’s integration of the tool depends more on the reliability of the human-machine interface than on anatomical mimicry. These results suggest that by moving away from strict biomimicry, designers can leverage the nervous system’s capacity for reinforcement learning to create more flexible, functional, and adaptable prosthetic solutions. 

4. Osseointegration and Skeletal Integration

While neural signals provide the intent, the physical stability of the limb is what allows that intent to be translated into meaningful action. Osseointegration represents a landmark advancement in this regard, moving away from adverse effects and issues of traditional sockets toward a permanent, bone-anchored skeletal attachment.

Traditional socket-based prostheses frequently cause discomfort, skin damage, and limited mobility, often resulting in high rates of device abandonment (Frossard and Lloyd, 2021). In contrast, osseointegrated bionic limbs are directly anchored to the bone, offering significantly improved stability and mobility (Frossard and Lloyd, 2021). This direct connection facilitates “osseoperception”, a form of sensory feedback that allows users to walk more naturally and experience an enhanced sense of embodiment (Frossard and Lloyd, 2021). By providing a stable mechanical foundation, osseointegration overcomes the functional limitations of the soft-tissue interface.

Azocar et al. researched the design, implementation, and clinical testing of the Open Source Leg (OSL), an integrated robotic knee-ankle prosthesis developed to advance research and clinical evaluation of powered lower limb prostheses in real-world environments (Figure 6). While powered prostheses can improve walking speed, stability, and energy efficiency for individuals, progress has been limited by high costs, proprietary designs, and the lack of standardized hardware for testing control strategies. The OSL addresses these challenges by providing a low-cost, portable, customisable, and fully open-source platform (Azocar et al., 2020). The prosthesis integrates high-torque electric motors, low-ratio belt transmissions, embedded sensing, and low-level control software, thus enabling independent operation of the knee and ankle joints (Azocar et al., 2020). A configurable series elastic actuator allows researchers to adjust joint stiffness and study different control options. Extensive benchtop testing demonstrated high bandwidth, accurate position and current control, reliable torque output, and safe thermal performance (Azocar et al., 2020). Clinical testing with 3 individuals with lower limb amputations showed that OSL supports stable ambulation across level ground, ramps, and stairs, using impedance-based control whereby participants achieved kinematics and kinetics comparable to able-bodied walking (Azocar et al., 2020) (Figure 6). By releasing hardware designs, software, and benchmark datasets, the OSL lowers barriers to prosthetic research and enables fair comparison of control strategies, supporting future innovation in bionic limb technology. 

4.1 Bionic Limb Replacement 

Amin (2022) examined the evolving role of bionic limb replacement in the context of lower extremity reconstruction. Advances in lower extremity reconstruction, including microvascular techniques, nerve and tendon transfers, and complex bone reconstruction, have expanded limb salvage options, but outcomes are still often limited by infection, unstable soft tissues, prolonged rehabilitation, chronic pain, and donor site morbidity. When salvage is unsuccessful, amputation with prosthetic fitting remains a common solution; however, conventional prosthetic fitting frequently fails to provide comfortable, efficient, and intuitive function, particularly due to issues with socket fit, increased energy expenditure, and lack of sensory feedback. Emerging bionic limb technologies offer a potential alternative by aiming to restore bidirectional communication between the nervous system and prosthetic devices (Amin, 2022). Approaches such as myoelectric control, targeted muscle reinnervation, and direct neural interfacing through extraneural, intraneural, and regenerative electrodes have demonstrated promise in improving motor control and sensory perception (Amin, 2020). Nevertheless, significant challenges remain, including signal instability, biological compatibility, invasiveness, and long-term reliability (Amin, 2020). Collectively, this body of work suggests that future progress in lower limb reconstruction will increasingly rely on interdisciplinary advances in neural interfaces, robotics, and tissue engineering to develop more functional and intuitive bionic limb replacements. 

4.2 Self-Contained Bidirectional Bionic Limbs

Surgical neural engineering and human-machine interfacing (HMI) have reached a milestone with the development of fully self-contained, bidirectional bionic limbs. A primary challenge in prosthetics is the “HMI paradox” whereby, as amputation levels rise, the complexity of the required prosthesis increases while the number of available biological control signals decreases (Dosen, 2023). To overcome this, a comprehensive HMI pipeline integrates advanced surgical techniques, such as muscle grafts and targeted muscle reinnervation, to create additional signal sources (Ortiz-Catalan et al., 2023). The system implemented by Ortiz – Catalan et al. utilises osseointegration, which is the permanent attachment of the prosthesis to the skeleton via titanium implants. This provides a stable gateway for bidirectional communication. Motor control is achieved through intramuscular and epimysial electrodes that record from native muscles and grafts, while sensory feedback is restored by delivering electrical stimulation to the ulnar nerves via extraneural cuff electrodes (Ortiz-Catalan et al., 2023). Long-term clinical results over three years demonstrate that this approach enables stable, dexterous control of individual digits and provides intuitive tactile sensations (Ortiz-Catalan et al., 2023). By combining skeletal attachment with implanted neural interfaces (Figure 7), this research moves closer to the ideal reality of bionic replacement: a limb that offers effortless, simultaneous control and natural somatosensory feedback, significantly enhancing the user’s daily quality of life.

5. Sensory Feedback and Embodiment

The ultimate success of a bionic limb is often defined by the user’s sense of ownership over the device, referred to as a psychological state known as embodiment. This state is achieved only when the brain receives consistent, multi-modal sensory feedback, such as heat, pressure, and kinesthesia, that aligns with its internal motor expectations. 

5.1 Thermally Sentient Bionic Limbs

Ortiz-Catalan argued for exploring thermal sensation as an added sensory modality in bionic limbs through sensory reinnervation. When nerves from an amputated limb naturally or surgically reinnervate the skin of the residual limb, they create “phantom maps”. By stimulating these specific areas with compact thermoelectric devices, researchers successfully triggered sensations of heat and cold, localised by the brain to the missing upper limb (Ortiz-Catalan, 2024). Moreover, in clinical trials, participants were able to distinguish between different materials, such as copper or glass, and accurately identify the temperature of objects they made contact with while wearing the bionic limb (Ortiz-Catalan, 2024). Unlike electrical nerve stimulation, which often may feel artificial, thermal stimulation of the phantom maps feels natural because it activates biological sensors as intended (Ortiz-Catalan, 2024). However, these maps are often somatopically disorganized or incomplete, though remaining stable for 48 weeks (Ortiz-Catalan, 2024). Therefore, it does still offer a promising avenue for long-term research and daily use. Ultimately, while thermal sensing may enhance the bionic limb user’s sense of “embodiment”, its primary value lies in improving the functional and social capabilities of artificial limbs.

5.2 Bionic E-Skin

Xu et al. (2024)  introduced a self-powered bionic droplet electronic skin (DES), designed to help bionic limbs establish a sense of touch to liquids (Figure 8). Generally, e-skins focus on solid pressure or temperatures, but most fail to perceive the movement of liquids. The proposed e-skin uses the triboelectric effect, meaning it generates its own electricity from the friction created when a water droplet slides across the surface, requiring no external power (Xu et al., 2024). This technological breakthrough is an interlaced electrode network and “overpass” connection design (Xu et al., 2024). This allows the skin to track the precise 2D trajectory, velocity, and acceleration of a droplet without the signal getting “crossed” or blurry. It’s so sensitive that it can even distinguish between different liquids such as tap water, rainwater, and seawater, based on their specific electrical signatures (Xu et al., 2024). Beyond just sensing, researchers demonstrated autonomous regulation similar to human neuromodulation (Xu et al., 2024). For example, when the skin detects a liquid leak, it can identify the direction of the flow and trigger an intelligent closed-loop system to shut a valve and stop the leak (Xu et al., 2024). Moreover, the e-skin is made of flexible, waterproof materials, and the AI allows for navigation and reaction to wet, unpredictable environments like rain or industrial spills as intuitively as a human would (Xu et al., 2024).

6. Patient-Centric Design and Residuum Health 

A sustainable bionic solution must prioritize the long-term health of the residual limb, or residuum, above mere mechanical performance. By utilising predictive tools such as digital twins, clinicians can now proactively manage the health of the patient’s tissue and bone, preventing the complications that historically led to prosthetic abandonment. 

6.1 Advances in Clinical and Prosthetic Care

Frossard et al. (2022) highlighted the necessity of a critical shift in this field of medicine, moving beyond just building better AI to focus on the health of the residuum. Generally, prosthetics fail because of the adverse effects, such as skin issues, pain, and bone loss, caused by the interface between the body and the device. This leads to an increase in limb abandonment. To fix this, researchers are pushing “Bionic 4.0 solutions”: direct skeletal attachment, osseointegration, and advanced neural interfaces such as TMR to get rid of pain caused by sockets and to improve control while reducing phantom pain (Frossard et al., 2022). After surgery, rehab blueprints should be established with examples such as botulinum toxin to stop skin movement around implants and haptic sleeves that give users a sense of touch (Frossard et al., 2022). Furthermore, a healthy residuum is a multifaceted goal whereby the patient’s age, weight, and specific alignment of the prosthetic joint are all key factors, not just the surgery. Consequently, new technology such as the Computer Assisted Limb Assessment (CALA) allows clinicians to objectively map out patients’ phantom pain, thus making it easier to treat (Frossard et al., 2022). Overall, the goal should be to stop oscillation of patients between satisfactory mobility and bedridden pain by threatening the prosthesis and the patient’s biology as a single, connected system.

To address the risks associated with bone-anchored systems, such as infection, implant failure, and fractures, a novel and non-invasive diagnostic approach using a “digital twin” of the residuum was proposed by Frossard and Lloyd (2021). This high-fidelity, physics-based virtual model integrates real-time data from wearable sensors to monitor internal tissue response during prosthetic use (Frossard and Lloyd, 2021). By visualising stresses and movements within the limb, clinicians and patients can better manage loading patterns and reduce injury risk (Figure 8). This transformative step toward intelligent bionic limbs supports personalized rehabilitation and improves long-term patient safety. 

7. Clinical Translation Challenges and Barriers

While laboratory milestones of the past decade are impressive, the transition from an experimental prototype to a standard-of-care medical device remains the most significant hurdle in prosthetic medicine. Evaluating bionic systems through a lens of translation-readiness reveals that high-performance metrics in a controlled setting do not always correlate with daily clinical utility. It is important to recognise that an ideal device in a lab may still fail in the real world if the surrounding infrastructure is not ready. 

7.1 Technical Barriers: Reliability and Biotic-Abiotic Integration

The first layer of obstruction involves the physical and electrical longevity of the system. For a bionic limb to be truly translation-ready, it must move beyond “benchtop” success and demonstrate resilience against the unpredictability of human life. Biocompatibility remains a major concern; even the most advanced electrodes face the risk of signal instability and decay over time as the body’s foreign-body response encapsulates sensors in fibrous tissue (Amin, 2022). Furthermore, achieving a truly permanent solution requires overcoming the risks of chronic infection at the percutaneous interface, which is a significant hurdle for the long-term reliability of fully implantable systems (Bumbaširević, 2020). Additionally, maintaining stable closed-loop control is technically taxing outside of controlled laboratory environments, where external interference and extraneous variables may disrupt the delicate neural signals required for intuitive motor execution. 

7.2 Clinical Organisation Barriers: The Infrastructure of Care

Advanced surgical interventions such as TMR or AMI require interdisciplinary teams that many regional clinics cannot support, leading to a disconnect between technological potential and actual clinical integration (Frossard et al., 2022). This disconnect is exacerbated by a reimbursement paradox: national health services and insurance providers are frequently hesitant to cover the high upfront costs of bionic limbs without longitudinal data proving they significantly reduce long-term secondary healthcare expenditures, such as caregiver reliance or mental health intervention. Without standardised clinical training and clear financial pathways, these innovations risk remaining research papers rather than becoming accessible medical standards. 

7.3 Ethical and Equity Considerations: Democratising Innovation  

Many current innovations are hindered by proprietary designs and “closed” software ecosystems that prevent interoperability between different manufacturers. This underscores a desperate need for standardised platforms that could lower production costs and facilitate broader distribution (Azocar et al., 2020).It also raises a critical question of equity, whereby if bionic limbs are only accessible to a small cohort of patients with premium insurance or specific military status, the field may inadvertently widen existing health disparities. An ethical transition toward “translation-readiness” must involve a shift from high-cost, exclusive prototypes toward inclusive, robust, and globally accessible bionic solutions.

8. Future Directions

While the preceding sections have outlined the state of the art and the barriers to implementation, the path forward requires a transition from demonstrating feasibility to ensuring longitudinal reliability. The next generation of research must prioritise the convergence of artificial intelligence, materials science, and standardised clinical protocols to transform bionic limbs from laboratory prototypes into lifelong biological extensions. 

8.1 Technical Future: Intelligence and Sensation:

The technical evolution of bionics will likely focus on the development of more robust, long-term neural interfaces. Current evidence suggests a critical need for long-term human data to validate the stability of bidirectional communication over decades (Cho et al., 2023). Furthermore, the integration of ‘Reinforcement Learning’ (RL) must move beyond virtual simulations into real world settings, allowing devices to adapt dynamically to a user’s unique environment (Freitag et. al, 2026)Beyond motor control, the future lies in multi-modal feedback, incorporating thermally sentient capabilities and bionic e-skins to provide a nuanced sense of touch, temperature, and pressure (Ortiz-Catalan, 2024).

8.2 Clinical Future: Trials and Long-Term Outcomes

The clinical trajectory must shift toward larger trials and standardised long-term outcome measures. To ensure patient safety, future research should integrate residuum-centered protocols that monitor the health of the host tissue alongside device performance (Amin, 2022).This involves moving toward fully implantable systems that minimise the risk of infection while maximising signal fidelity.

8.3 Design and Policy Future: Equity and Standardisation 

From a policy perspective, the field must address the “siloing” of technology. Moving away from proprietary designs toward standardised, adaptable platforms is essential for lowering costs and ensuring equitable access (Azocar et al., 2020). Future regulatory pathways must be streamlined to accelerate the transition from the “bench to the bedside”, ensuring that breakthroughs in the lab actually reach the diverse global population of limb loss patients.

8.4 Core Strategic Objectives:

• Small Learning Curves:

Future prosthetic design must prioritise intuitive control to minimise the cognitive burden on the patient. By aligning the device’s behaviour with the brain’s existing internal motor models, researchers can ensure that users do not require months of intensive training, making the technology accessible to a wider demographic regardless of technical proficiency.

• Simple Procedures for Insurance Coverage:

Clinical translation depends on moving toward surgical techniques that are standardised enough for general hospitals to perform. By simplifying the implementation of interfaces like TMR or AMI, the procedure becomes more “reimbursable” by insurance, shifting advanced bionics from an experimental luxury to a standard, covered medical necessity for all patients.

• Clinical Translation Focus:

The field must move away from “technological push” models toward a “clinical pull” strategy. This means prioritising the practical needs of the patient, such as the device weight, battery life, and skin comfort, over adding high-tech features that have little impact on daily quality of life or long-term functional independence in home environments. 

• Integration of AI (System Thinking):

Rather than using AI for isolated tasks, future systems must adopt “system thinking” AI. This allows the limb to simultaneously manage motor intent, sensory feedback, and environmental adaptation. Instead of a one-size-fits-all functional design, AI will personalise the device’s response in real time, matching the specific physiological signatures of each individual user.

• Approval and Acceleration of Clinical Trials:

The primary bottleneck is no longer the technology itself, but the speed of regulatory approval. Accelerating the transition to large-scale clinical trials is the key to proving long-term safety and efficacy. Without these trials, even the most revolutionary designs will remain stuck in a “proof of concept” phase without reaching the public.

• Integration of Digital Twins:

 Digital twins allow researchers to design and test prosthetics in a high-fidelity virtual environment before any surgery occurs. By simulating the interaction between the device and the patient’s unique bone and tissue structure, clinicians can perfect the design, predict potential failures, and minimise the risks of post-operative complications or revisions.

9. Conclusion

The evolution of bionic limb technology represents one of the most ambitious frontiers in modern medical engineering. As this review has demonstrated, the field has successfully moved beyond simple mechanical substitution toward a sophisticated model of neural and skeletal integration. However, the true measure of success for these innovations is not found in laboratory benchmarks or peak performance metrics, but in their capacity for meaningful clinical translation. For a bionic system to be effective, its design must be clinically driven, thus originating from the actual physiological and psychological needs of the patient rather than being a product of technological push, where features are added simply because they are technically possible.

The trajectory of a successful bionic innovation must follow a clear path: from the research bench, where fundamental science is born, to the clinical bedside, where the technology is safely and intuitively integrated into a patient’s life; and finally to the boardroom, where standardised manufacturing and insurance reimbursement models ensure the device is economically viable and globally accessible. By prioritising “system thinking” through AI and digital twins, and by lowering the cognitive and financial barriers to entry, the field can ensure that bionic restoration is no longer an experimental luxury/ Ultimately, the goal is to transform these devices into seamless biological extensions, restoring not just function, but a genuine sense of embodiment and independence to those living with limb loss. 

Bibliography

Amin, K. R. (2022). Bionic limb replacement: an evolving concept in lower extremity reconstruction. Plastic and Aesthetic Research, 9(3), 24. https://doi.org/10.20517/2347-9264.2021.119

Azocar, A. F., Mooney, L. M., Duval, J.-F., Simon, A. M., Hargrove, L. J., & Rouse, E. J. (2020). Design and clinical implementation of an open-source bionic leg. Nature Biomedical Engineering, 4(10), 941–953. https://doi.org/10.1038/s41551-020-00619-3

Bumbaširević, M., Lesic, A., Palibrk, T., Milovanovic, D., Zoka, M., Kravić-Stevović, T., & Raspopovic, S. (2020). The current state of bionic limbs from the surgeon’s viewpoint. EFORT Open Reviews, 5(2), 65–72. https://doi.org/10.1302/2058-5241.5.180038

Cho, Y., Jeong, H. H., Shin, H., Pak, C. J., Cho, J., Kim, Y., Kim, D., Kim, T., Kim, H., Kim, S., Kwon, S., Hong, J. P., Suh, H. P., & Lee, S. (2023). Hybrid Bionic Nerve Interface for Application in Bionic Limbs. Advanced Science, 10(35). https://doi.org/10.1002/advs.202303728

Dosen, S. (2023). Toward self-contained bidirectional bionic limbs with high information throughput. Science Robotics, 8(83). https://doi.org/10.1126/scirobotics.adk6086

Freitag, K., Karayiannidis, Y., Zbinden, J., & Laezza, R. (2026). Fine-Tuning Myoelectric Control Through Reinforcement Learning in a Game Environment. IEEE Transactions on Biomedical Engineering, 73(1), 168–179. https://doi.org/10.1109/tbme.2025.3578855

Frossard, L., Conforto, S., & Aszmann, O. C. (2022). Editorial: Bionics limb prostheses: Advances in clinical and prosthetic care. Frontiers in Rehabilitation Sciences, 3. https://doi.org/10.3389/fresc.2022.950481

Frossard, L., & Lloyd, D. (2021). The future of bionic limbs. Research Features, 134. https://doi.org/10.26904/rf-134-7477

Guo, K., Lu, J., Wu, Y., Hu, X., & Yang, H. (2024). The Latest Research Progress on Bionic Artificial Hands: A Systematic Review. Micromachines, 15(7), 891. https://doi.org/10.3390/mi15070891

Long, Z., Qiu, X., Chan, C. L. J., Sun, Z., Yuan, Z., Poddar, S., Zhang, Y., Ding, Y., Gu, L., Zhou, Y., Tang, W., Srivastava, A. K., Yu, C., Zou, X., Shen, G., & Fan, Z. (2023). A neuromorphic bionic eye with filter-free color vision using hemispherical perovskite nanowire array retina. Nature Communications, 14(1). https://doi.org/10.1038/s41467-023-37581-y

Marasco, P. D., Hebert, J. S., Sensinger, J. W., Beckler, D. T., Thumser, Z. C., Shehata, A. W., Williams, H. E., & Wilson, K. R. (2021). Neurorobotic fusion of prosthetic touch, kinesthesia, and movement in bionic upper limbs promotes intrinsic brain behaviors. Science Robotics, 6(58). https://doi.org/10.1126/scirobotics.abf3368

Ortiz-Catalan, M. (2024). Thermally sentient bionic limbs. Nature Biomedical Engineering, 8(8), 938–940. https://doi.org/10.1038/s41551-023-01174-3

Ortiz-Catalan, M., Zbinden, J., Millenaar, J., D’Accolti, D., Controzzi, M., Clemente, F., Cappello, L., Earley, E. J., Mastinu, E., Kolankowska, J., Munoz-Novoa, M., Jönsson, S., Cipriani, C., Sassu, P., & Brånemark, R. (2023). A highly integrated bionic hand with neural control and feedback for use in daily life. Science Robotics, 8(83). https://doi.org/10.1126/scirobotics.adf7360

Pasluosta, C., Kiele, P., Čvančara, P., Micera, S., Aszmann, O. C., & Stieglitz, T. (2022). Bidirectional bionic limbs: a perspective bridging technology and physiology. Journal of Neural Engineering, 19(1), 013001. https://doi.org/10.1088/1741-2552/ac4bff

Schone, H. R., Udeozor, M., Moninghoff, M., Rispoli, B., Vandersea, J., Lock, B., Hargrove, L., Makin, T. R., & Baker, C. I. (2023). Should bionic limb control mimic the human body? Impact of control strategy on bionic hand skill learning. https://doi.org/10.1101/2023.02.07.525548

Schone, H. R., Udeozor, M., Moninghoff, M., Rispoli, B., Vandersea, J., Lock, B., Hargrove, L., Makin, T. R., & Baker, C. I. (2024). Biomimetic versus arbitrary motor control strategies for bionic hand skill learning. Nature Human Behaviour, 8(6), 1108–1123. https://doi.org/10.1038/s41562-023-01811-6

Song, H., Hsieh, T.-H., Yeon, S. H., Shu, T., Nawrot, M., Landis, C. F., Friedman, G. N., Israel, E. A., Gutierrez-Arango, S., Carty, M. J., Freed, L. E., & Herr, H. M. (2024). Continuous neural control of a bionic limb restores biomimetic gait after amputation. Nature Medicine, 1–10. https://doi.org/10.1038/s41591-024-02994-9

Xu, Y., Sun, Z., Bai, Z., Shen, H., Wen, R., Wang, F., Xu, G., & Lee, C. (2024). Bionic e-skin with precise multi-directional droplet sliding sensing for enhanced robotic perception. Nature Communications, 15(1). https://doi.org/10.1038/s41467-024-50270-8

Zhang, H., & Lee, S. (2022). Robot Bionic Vision Technologies: A Review. Applied Sciences, 12(16), 7970. https://doi.org/10.3390/app12167970

About the author

Sharanya Seth

Sharanya Seth is a Grade 11 student at United World College South East Asia in Singapore with a strong interest in medicine, particularly in surgical fields such as Obstetrics and Gynaecology and Plastic Surgery. She has gained early clinical exposure through internships in both Singapore and India, where she observed procedures including hysterectomies and fertility treatments, and assisted with patient interactions in a general practice setting. These experiences provided her with insight into clinical environments, patient care, and the doctor-patient relationship.

Sharanya is actively engaged in scientific research, having submitted a CREST Award project investigating the inhibitory effects of natural disinfectants such as turmeric and garlic on E. coli growth. She is currently conducting an Extended Essay exploring glucose diffusion and its relevance to diabetes, further developing her ability to apply scientific principles to real-world medical challenges.

In addition to her academic pursuits, she leads a service initiative supporting individuals with intellectual disabilities, where she organises cognitively stimulating activities to promote engagement and social interaction, as well as fosters an environment of respect and inclusivity. She also volunteers regularly at the National Kidney Foundation in Singapore, where she helps in the Dialysis Centre and Enrichment sector with patients. She has been recognised on the High Honour Roll award for three consecutive years and has achieved distinction in piano examinations, along with athletic awards.

The post Clinical Translation of Bionic Limbs: Neural Interfaces, Osseointegration, and Patient‑Centric Design appeared first on Exploratio Journal.

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.