PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning-Enabled Wireless Power Transfer: A Task-Oriented Review of Structural Optimization, State and Parameter Estimation, and Adaptive Control

Denghao Luo, Yiheng Jiang, Jiaqi Zhao, Yuan Gao, Changsong Cai, Weiyi Huang, Yun Zhang

arXiv:2610.30101v1eess.SYcs.SY

TL;DR

WPT systems must deliver efficient, safe energy despite variations in geometry, coupling, load, alignment, and operating conditions. This review compares learning methods across design, estimation, and control tasks, finding complementary strengths alongside significant deployment constraints.

  • Problem

    The reviewed literature lacked a common task-oriented framework comparing learning methods across WPT design, estimation, control, and practical deployment dimensions.

    Source: §1.3 · p. 6

  • Method

    The review organizes WPT learning studies by structural design, parameter and state estimation, and control, comparing supervised, reinforcement, and hybrid or physics-informed approaches.

    Source: §6 · p. 32

  • Results

    The reviewed studies show complementary strengths: supervised learning supports surrogate modeling and fast inference, reinforcement learning supports sequential control and scheduling, and hybrid methods integrate data-driven learning with physical knowledge.

    Source: 3) Hybrid Intelligent Methods · p. 16, §6 · p. 32

  • Takeaways & Limitations

    Integrating physical knowledge through model inputs, architectural constraints, or physics-based losses is a route toward improved data efficiency, generalization, and physical consistency.

    Source: 2) Physics-Informed and Hybrid Learning · p. 31

    WPT data are costly to collect, and limited or imbalanced datasets may not generalize reliably to safety-critical conditions.

    Source: §5.1 · p. 29

Abstract

from arXiv · show

Wireless power transfer (WPT) systems must maintain efficient and safe energy delivery under variations in geometry, coupling, load, alignment, and operating conditions. Machine-learning methods are increasingly applied to accelerate structural design, estimate operating parameters and states, and support adaptive control and scheduling. This review develops a task-oriented framework organized around three interconnected domains: structural design and optimization, parameter and state estimation, and control and optimization. Supervised learning, reinforcement learning, and hybrid or physics-informed methods are compared in terms of data requirements, dependence on physical models, information sources, computational requirements, accuracy, online response, application scope, and optimization objectives. Supervised learning provides rapid surrogate modeling, regression, and classification when representative labeled data are available. Reinforcement learning addresses sequential control and scheduling but introduces challenges in reward design, safe exploration, and simulation-to-reality transfer. Hybrid methods integrate physical knowledge with data-driven models to improve data efficiency and physical consistency. The principal barriers to engineering deployment include limited and biased datasets, weak out-of-distribution generalization, insufficient physical consistency, unsafe online adaptation, embedded-resource constraints, and the absence of standardized benchmarks. Future research should prioritize uncertainty-aware and physics-informed learning, safe simulation-to-reality transfer, lightweight edge deployment, distributed coordination, reproducible evaluation, and digital-twin-assisted development.

1. Introduction

WPT faces coupled design, estimation, control, and integration challenges as systems move toward dynamic and demanding operating environments. This review organizes machine-learning approaches by task and compares their applicability, limitations, and deployment readiness.

1.1 Wireless Power Transfer Technology: Background and Application Requirements

  • Dynamic WPT must balance coupler efficiency, misalignment tolerance, electromagnetic compatibility, and size and weight as vehicle motion changes alignment.Position, mutual inductance, and load may require estimation from electrical measurements, while changing coupling and load call for timely operating adjustments.Source: §1.1 · p. 2
  • System-level performance depends on coordinated couplers, compensation networks, converters, sensing, and control, making isolated component optimization insufficient.Full system-level co-design remains an open research direction.Source: §1.1 · p. 3

1.2 Limitations of traditional design and control methods

  • Conventional workflows trade computational cost, physical-model dependence, and adaptability, with FEA sweeps and population-based optimization requiring repeated evaluations.Analytical models can be efficient under their assumptions, but coupling or load changes can affect impedance and efficiency.Source: §1.2 · p. 3

1.3 Machine Learning-Enabled Wireless Power Transfer: Motivation and Current State

  • The reviewed literature has expanded across WPT tasks and learning categories, with supervised learning the most widely adopted and reinforcement and hybrid methods increasingly prominent.The review identifies a gap in task-oriented comparison across learning methods, operating robustness, safety, real-time capability, and deployment readiness.Source: §1.3 · pp. 5–6
  • The review compares supervised, reinforcement, and hybrid or physics-informed methods across structural design, state estimation, and control and scheduling.Its framework considers data needs, physical-model dependence, computational efficiency, real-time performance, robustness, and deployment readiness.Source: §1.3 · p. 6

1.4 Scope and Organizational Structure

  • The review proceeds from conventional methods to learning algorithms, task-specific applications, challenges, and future research directions.Highlighted directions include physics-informed learning, lightweight models, multi-agent coordination, and digital-twin-enabled WPT.Source: §1.4 · p. 7

2. Traditional design and control methods for wireless power transfer

Traditional WPT design and control combine physical models, numerical optimization, estimation, and feedback regulation. Their recurring trade-offs involve model fidelity, computational cost, and adaptability to changing conditions.

2.1 Traditional Analytical Design Methods

  • Conventional design uses analytical models, FEA, parameter sweeps, and population-based searches, but multiple geometric and material variables can still require costly repeated simulations.Population-based methods also depend on population diversity and settings, and penalty coefficients can be problem-dependent.Source: §2.1 · p. 8
  • Compensation networks can detune as components age or temperature changes, so active tuning can restore resonance or regulate power but requires additional system resources.Reported approaches include PWM-controlled equivalent inductance and switched-capacitor reactance adjustment.Source: §2.1 · p. 8

2.2 Traditional Control Methods

  • Conventional control combines state estimation with regulation, but controller performance depends on operating conditions, model adequacy, and available state information.PI/PID gains may degrade away from nominal conditions; SMC may chatter, while MPC depends on its prediction model and online computational resources.Source: §2.2 · pp. 9–10

2.3 Common Limitations in Traditional Methods

  • Repeated electromagnetic simulations and problem-dependent tuning constrain conventional design, while model fidelity, computational cost, and adaptability remain recurring trade-offs.Reduced-order models can omit geometric detail, edge effects, material nonlinearities, and high-frequency losses.Source: §2.1 · p. 9, §2.3 · p. 10

3. Machine Learning Algorithm Classification and Comparative Analysis

This section classifies WPT machine-learning research into supervised, reinforcement, and hybrid intelligent methods. It compares the categories to clarify their engineering strengths, limitations, and task fit.

3. Machine Learning Algorithm Classification and Comparative Analysis

  • Conventional WPT methods face high computational cost, dependence on predefined models, and limited adaptability, especially in high-dimensional or time-varying settings.These are trade-offs rather than universal failures; high-fidelity models can be costly, while reduced or linearized models may lose accuracy outside their assumptions.Source: §3 · p. 11
  • The comparison framework covers data and physical-model dependence, information sources, computation, accuracy, response speed, application scope, and optimization objectives.Differences across these dimensions complicate direct comparison and method selection for particular WPT tasks.Source: §3 · p. 12

3.1 Overall Algorithm Classification

  • The review organizes WPT machine-learning studies into supervised learning, reinforcement learning, and hybrid intelligent methods.It examines underlying principles, representative algorithms, typical applications, and engineering limitations.Source: §3.1 · p. 12

1) Supervised learning

Supervised learning maps labeled WPT inputs to design, performance, state, fault, or control targets. It offers rapid inference after training, but its practical value depends on representative data and task-specific constraints.

1) Supervised learning

  • Supervised learning spans regression, neural networks, ensembles, and deeper architectures for design prediction, diagnosis, performance modeling, and control-related tasks.Reported applications include coupler geometry prediction, coil quality-factor estimation, fault diagnosis, position and misalignment estimation, and power regulation.Source: 1) Supervised learning · p. 13
  • A random-forest coupler-geometry evaluation took a few seconds rather than approximately 5 min using FEA, an approximately 50-fold acceleration.The reported fault-identification accuracy for a hierarchical SVM using a single current sensor exceeded 95%.Source: 1) Supervised learning · p. 13
  • Supervised models approximate nonlinear input-output relationships from labeled data and provide rapid inference after training.Analytical models offer physical interpretability, whereas repeated FEA evaluation can be computationally costly.Source: 1) Supervised learning · p. 14
  • Supervised learning is a task-dependent complement to physical modeling, with effectiveness dependent on representative training data, uncertainty, generalization, physical consistency, and deployment computation.Conventional regression and deep architectures remain in use, with suitability depending on data availability, task complexity, interpretability, and deployment constraints.Source: 1) Supervised learning · pp. 13–14

2) Reinforcement learning

Reinforcement learning learns policies through interaction rather than labeled optimal actions, supporting sequential decisions in WPT. Its applications range from physical-system regulation to communication-oriented scheduling, but deployment faces training, safety, and sim-to-real challenges.

2) Reinforcement learning

  • RL methods include value-based, policy-gradient, and actor-critic approaches suited to different decision spaces and control problems.Value-based methods suit discrete decisions, policy gradients suit stochastic or continuous decisions, and actor-critic methods address high-dimensional continuous control.Source: 2) Reinforcement learning · p. 15
  • RL has regulated WPT voltage and output modes, tuned frequency, selected coils, and managed energy storage.Examples include compensation-capacitance selection under load variation, Q-learning that reduced UAV coil-activation steps by 92.2% versus exhaustive search, and offline TD3 tuning of PI parameters.Source: 2) Reinforcement learning · p. 15
  • Communication-oriented RL extends from converter control to network-level resource allocation and scheduling, where models, states, and objectives differ substantially.Applications include computation offloading, UAV trajectory and RIS configuration, and user scheduling and power allocation.Source: 2) Reinforcement learning · pp. 15–16
  • RL avoids labeled optimal actions, but training costs, reward sensitivity, unsafe exploration, limited guarantees, and simulation-to-reality degradation constrain deployment.Most reviewed WPT studies train policies offline in simulation and deploy the learned policy or optimized controller parameters.Source: 2) Reinforcement learning · p. 16

3) Hybrid Intelligent Methods

Hybrid WPT methods combine learning with physical knowledge, simulation, optimization, or conventional control. The review describes recurring integration patterns while noting application-specific designs and inconsistent evaluation.

3) Hybrid Intelligent Methods

  • Surrogate-assisted optimization uses learned performance mappings with simulation or optimizers to search WPT designs and settings.Examples include ANN-based biomedical coil design and offline phase-shift search followed by a lightweight ANN for real-time settings.Source: 3) Hybrid Intelligent Methods · pp. 16–17
  • Physics-guided learning incorporates governing relationships or mechanism-generated data to improve data efficiency and physical consistency.Analytical-model pretraining followed by limited FEA fine-tuning achieved coil-parameter estimation errors below 4%; underwater WPT identification errors were below 3% with response within 2.72 ms.Source: 3) Hybrid Intelligent Methods · p. 17
  • Learning-assisted conventional control retains established control structures while using learned estimates or surrogates to improve adaptation or controller design.An extreme learning machine estimated uncertainty bounds online in a fixed-time sliding-mode controller, reducing settling time by 45.70%–51.72% versus investigated PI baselines.Source: 3) Hybrid Intelligent Methods · p. 17
  • Multistage workflows coordinate offline optimization, surrogate modeling, and online adaptation across design, estimation, and control tasks.Examples include EV-WPT static optimization, ANN modeling, and PPO control, and assigning distinct algorithms to sensor-network deployment, clustering, and resource allocation.Source: 3) Hybrid Intelligent Methods · p. 17
  • Existing hybrid architectures remain application-specific and are evaluated on different datasets, objectives, and operating conditions.The review calls for standardized comparisons, component ablations, safety and stability guarantees, and embedded-hardware validation.Source: 3) Hybrid Intelligent Methods · pp. 17–18

3.2 Algorithm Comparison

  • For position-adaptive dynamic charging, candidate approaches include supervised position estimation with conventional control, direct RL policies, or physical-model and data-driven integration.The comparison frames method choice around data, physical-model assumptions, computation, deployment, and decision structure.Source: §3.2 · p. 18

1) Data dependency and information sources

The comparison distinguishes learning categories by their data and information sources, computational profiles, and decision structures. No category is universally superior; selection depends on task and deployment conditions.

1) Data dependency and information sources

  • Supervised learning requires labeled input-output pairs, while reinforcement learning learns from interaction trajectories without target labels.Supervised labels may come from FEA, analytical calculations, or measurements; RL trajectories contain states or observations, actions, rewards, and subsequent states.Source: 1) Data dependency and information… · p. 18
  • Hybrid methods vary in data needs; physical constraints, mechanism-generated data, transfer learning, or simulation pretraining may reduce experimental-data requirements.The resulting performance depends on the validity of incorporated physical assumptions.Source: 1) Data dependency and information… · p. 19
  • Information sources can include measured data, simulations, governing equations, expert knowledge, optimization feedback, and control signals.Supervised and model-free RL inference need not require an explicit plant model, whereas model-based RL incorporates an environment model.Source: 1) Data dependency and information… · p. 19

2) Accuracy and Computational Characteristics

  • Learning can shift costly simulation, numerical inversion, or policy optimization offline, enabling rapid online inference when hardware and timing permit.This supports iterative structural design, online state estimation, and real-time control, subject to model complexity and deployment constraints.Source: 2) Accuracy and Computational… · p. 19
  • Supervised models primarily predict or classify, RL learns sequential policies through cumulative rewards, and hybrid frameworks coordinate prediction, optimization, estimation, and control.Hybrid approaches can span discrete and continuous variables or different time scales, but added components increase complexity and complicate attribution of improvements.Source: 3) Typical application scenarios and… · pp. 19–20
  • No method is universally superior; selection should reflect data availability, model fidelity, online latency, safety requirements, and task timing.The categories can be combined across design, estimation, and control, and common benchmarks and reporting standards are needed to assess added complexity.Source: 3) Typical application scenarios and… · p. 20

4. Specific applications of machine learning in wireless power transfer

The review organizes ML-enabled WPT research around structural design, parameter and state estimation, and control. It identifies practical deployment barriers and reviews learning-based approaches to structural optimization.

4. Specific applications of machine learning in wireless power transfer

  • WPT research spans structural design and optimization, parameter and state estimation, and operational control, with learning methods applied across all three domains.The review compares supervised learning, reinforcement learning, and hybrid or physics-informed approaches across these tasks.Source: §4 · p. 21
  • Practical deployment remains constrained by data scarcity and bias, limited physical consistency, online-adaptation safety, embedded resources, and explainability.Table 4 summarizes these challenges and potential mitigation strategies.Source: §4 · p. 21, Table 4 · p. 21

4.1 Structural Optimization

  • Magnetic-coupler optimization must address costly design-space evaluation as variables such as coil geometry, air gap, and magnetic-core properties affect performance.Parameter sweeps grow costly with design-variable count, while population-based optimizers may require many FEA evaluations and can converge prematurely.Source: §4.1 · p. 22
  • Learning-based structural methods use supervised surrogates, reinforcement-learning search, or hybrid learning and optimization.These approaches map design variables to performance metrics or learn policies for design-space exploration.Source: §4.1 · p. 22

1) Supervised Learning as an FEA Surrogate

Supervised models replace repeated FEA with rapid electromagnetic predictions, while the reviewed work also covers reinforcement-learning design and hybrid optimization. The section connects these approaches to parameter and state estimation under changing conditions.

1) Supervised Learning as an FEA Surrogate

  • R^2=0.9977 for receiver-coil radius prediction, while the predicted optimum differed from FEA by 5.43%.Other reviewed supervised surrogates report faster FEA-like evaluation, improved coil performance, and low errors with fewer than 1,000 training samples.Source: 1) Supervised Learning as an FEA… · p. 22
  • Neural-network surrogates also support multiphase and 3D couplers, coil and ferrite geometry, magnetic-field mitigation, and underwater or implantable WPT designs.For underwater and implantable applications, the surrogate is coupled with a multi-objective optimizer.Source: 1) Supervised Learning as an FEA… · p. 22

2) Reinforcement Learning for Design-Space Exploration

  • With 300 episodes—about 2.3% of 12,870 possible structures—RL found a core with 7% higher measured coupling than the reference at 98 cm.An improved DD-coupler design reduced receiver-side core volume by 28% while increasing mutual inductance by 2.3%.Source: 2) Reinforcement Learning for… · p. 23
  • RL also addresses discrete, sequential infrastructure decisions, including charging-system design and charger placement for EVs, UAVs, and sensor networks.These applications have objectives and constraints distinct from electromagnetic component optimization.Source: 2) Reinforcement Learning for… · p. 23

3) Hybrid Method

  • The reviewed hybrid approaches combine learning-assisted metaheuristics, iterative data refinement, or offline design optimization with online adaptive control.These are presented as three representative patterns for combining complementary methods.Source: 3) Hybrid Method · p. 23
  • The filter-enhanced fuzzy recurrent neural-network controller’s experimentally validated system achieved a peak transfer efficiency of 95% at a 135 mm air gap.Bennia et al. coupled a neural-network surrogate with genetic and coyote optimization, reporting up to 76% efficiency for implantable WPT coils [55]. Fawad et al. reported acceleration factors of approximately 4,300, 16,960, and 10,138 for three respective models compared with ANSYS Maxwell simulations.Source: 3) Hybrid Method · p. 24
  • For an undersea coupler, an electromagnetics-embedded BPNN used less than half the training data and cut optimization time by at least six orders of magnitude versus a purely data-driven model.Other application-specific hybrids report validated designs spanning mobile-device, resonant-coil, and 2 kW LCC-S systems; the added components increase implementation and validation complexity.Source: 3) Hybrid Method · p. 24

4.2 Parameter and State Estimation

  • Key parameters and states, including mutual inductance, load, coupling, and receiver position, vary with misalignment and operating conditions and may require inference from electrical measurements.The reviewed approaches include direct data-driven, hybrid model-and-data-driven, and receiver-position or misalignment estimation.Source: §4.2 · p. 24

1) Direct Data-Driven Parameter Estimation

The reviewed studies estimate WPT parameters and receiver position from electrical responses, sometimes incorporating physical relationships or auxiliary sensing. They also use learned mappings for control, with reliability dependent on data coverage and operating conditions.

1) Direct Data-Driven Parameter Estimation

  • Transmitter-side electrical features supported estimates of mutual inductance and load resistance with maximum errors of 2.93% and 7.4%, respectively.Other reviewed studies reported relative mutual-inductance errors below approximately 1.2% for most evaluated cases and used transmitter-side sensing without receiver communication.Source: 1) Direct Data-Driven Parameter… · p. 25

2) Hybrid Model- and Data-Driven Estimation

  • On a 10 kW platform, a physics-informed network reported relative errors of 7.7%, 2.1%, 2.0%, and 1.8% for displacement, mutual inductance, leakage inductance, and load resistance.Mechanism-generated data with experimental adaptation yielded errors below 3% under three underwater conditions and response time within 2.72 ms.Source: 2) Hybrid Model- and Data-Driven… · p. 25

3) Receiver Position and Misalignment Estimation

  • Receiver-position studies reported average errors below 2.40 mm and a test-set correlation coefficient of approximately 0.9994.Another study reported approximately 5% average relative error and maintained efficiency close to 90% under misalignment using adaptive transmitter switching.Source: 3) Receiver Position and Misalignment… · p. 25
  • A two-dimensional workspace study reported 1.83 cm RMS positioning error under electromagnetic interference and measurement noise.Position estimates were also combined with reinforcement-learning charging-path scheduling, which achieved approximately 3.5 cm positioning error and a 20.9% average charging-efficiency improvement.Source: 3) Receiver Position and Misalignment… · pp. 25–26
  • Position classification reached 94.05% and 95.99% accuracy for random forest and DNN, respectively, in a 16-class case; efficiency-prediction error was below 1%.The reviewed work also includes joint misalignment and foreign-object detection, camera-based mutual-inductance estimation, and symbolic regression.Source: 3) Receiver Position and Misalignment… · p. 26
  • Estimator reliability depends on measurement quality, training-data coverage, uncertainty quantification, and generalization across alignments, loads, temperatures, media, and topologies.The review describes rapid inference after offline training, but these factors condition its practical reliability.Source: 3) Receiver Position and Misalignment… · p. 26

1) Supervised-Learning-Based Control

  • A neural impedance-matching method maintained approximately 90% efficiency over 10–25 cm, while a simulated vehicle-position method achieved up to 174% higher efficiency than lane-center motion-control baselines.Supervised control avoids online exploration but remains constrained by the operating conditions represented in training data.Source: 1) Supervised-Learning-Based Control · p. 26

2) Reinforcement-Learning-Based Control

Reinforcement learning is reviewed for WPT regulation, frequency and transmitter control, and system-level scheduling, while hybrid methods combine prediction with optimization or adaptive decisions. The methods offer varied operating results but differ in training, implementation, and validation demands.

2) Reinforcement-Learning-Based Control

  • In simulation, Q-learning recovered from a load disturbance within approximately 4 ms; a separate regulation strategy reported maximum and average output errors of 1.29% and 0.93%.The latter study reported negligible influence on system efficiency.Source: 2) Reinforcement-Learning-Based Control · p. 27
  • Deep RL reduced transmitter-coil activation steps by 92.2% versus exhaustive search on an 8 × 8 array, with about five activations required on average.Other reviewed applications include UAV charging and offline PI-parameter optimization for frequency regulation.Source: 2) Reinforcement-Learning-Based Control · p. 27
  • For lunar multi-energy WPT, PPO reduced energy-supply interruptions by 54.6% and improved beam-alignment efficiency by 41.3% versus static scheduling.The reviewed system-level work also covers adaptive beam scheduling, multimodal UAV charging, joint resource allocation, and underwater trajectory optimization.Source: 2) Reinforcement-Learning-Based Control · p. 27
  • RL control faces offline training costs, sensitivity to state and reward definitions, exploration-related safety risks, and simulation-to-reality transfer challenges.These are identified as challenges despite RL’s adaptability.Source: 2) Reinforcement-Learning-Based Control · p. 27

3) Hybrid and Surrogate-Assisted Control

  • A hybrid 430 W/85 kHz prototype reported prediction MAPE below 3.5% and maximum system efficiency of 93.5%.The approach combined an MLP with PSO; another reviewed framework integrated static optimization, an ANN surrogate, and PPO-based dynamic control.Source: 3) Hybrid and Surrogate-Assisted Control · pp. 27–28

4) Comparative Assessment

  • Hybrid prediction plus optimization or adaptive decision-making can reduce repeated costly model evaluations, but typically requires more implementation and validation than a single model or conventional controller.PI/PID controllers retain low, predictable online computational burden; numerical results across studies are not standardized cross-platform rankings because tasks, operating ranges, hardware, and metrics differ.Source: 3) Hybrid and Surrogate-Assisted Control · p. 28, 4) Comparative Assessment · p. 28

5. Challenges and Future Directions

The review identifies deployment barriers across data, generalization, computation, physical consistency, and evaluation, and outlines directions for transferable, data-efficient learning. It emphasizes that models must be assessed within their operating boundaries and task-specific deployment constraints.

5. Challenges and Future Directions

  • The review groups shared deployment barriers as data scarcity, distribution shift, limited physical consistency, embedded-resource constraints, safety requirements, and absent standardized evaluation protocols.It identifies six research directions, including data-efficient transfer, physics-informed modeling, lightweight edge deployment, coordination, benchmarks, and digital-twin-assisted development.Source: §5 · p. 29

5.1 Common Limitations of Current Research

  • High-quality labeled WPT data are costly and difficult to obtain; simulation differs from hardware, experiments cover limited configurations, and abnormal-condition data are particularly scarce.Limited or imbalanced training data may not generalize reliably to safety-critical conditions.Source: §5.1 · p. 29
  • Surrogate models often need additional sampling or adaptation when topology, frequency, materials, or transmission medium changes, and their effectiveness depends on physical-model validity and representative adaptation data.Mechanism-generated pretraining, online updates, and low-fidelity analytical priors are reviewed as approaches that reduce exclusive dependence on costly data.Source: §5.1 · p. 29
  • Most studies test a specific topology, configuration, frequency, load, or platform, while errors can increase under out-of-distribution conditions.Evaluation should distinguish simulation, laboratory prototype, hardware-in-the-loop, and field deployment, and report the operating boundaries for acceptable accuracy and stability.Source: §5.1 · pp. 29–30
  • Deployment requires balancing model accuracy against inference time, memory, computation, and hardware cost, with timing assessed against each task’s requirements.Offline lookup tables and symbolic regression can reduce online computation, but systematic trade-off methods remain needed.Source: §5.1 · p. 30
  • Data-driven predictions or control actions may violate engineering constraints, and physics-informed, symbolic-regression, and prior-input methods do not automatically guarantee safety, stability, or out-of-training generalization.The review names constraints including energy conservation, circuit relationships, exposure limits, device ratings, and soft-switching conditions.Source: §5.1 · p. 30

1) Data-Efficient and Transferable Learning

  • Transfer learning, domain adaptation, few-shot learning, and adaptive sampling are proposed to reduce additional data needs when WPT geometry, topology, or operating range changes.The review also highlights iterative data refinement with electromagnetic-simulation feedback as a strategy to investigate for lowering data-generation cost while maintaining accuracy.Source: 1) Data-Efficient and Transferable… · p. 30

2) Physics-Informed and Hybrid Learning

The review presents physics-informed learning, lightweight deployment, distributed coordination, reproducible evaluation, and digital twins as directions for addressing WPT engineering challenges. Its conclusion connects these approaches to broader deployment constraints and research priorities.

2) Physics-Informed and Hybrid Learning

  • Integrating physical knowledge through model inputs, architectural constraints, or physics-based losses may improve data efficiency, generalization, and physical consistency.Hybrid frameworks can combine mechanism-generated data with limited experimental measurements; future work should extend them to structural optimization and closed-loop control with safety constraints.Source: 2) Physics-Informed and Hybrid Learning · p. 31

3) Lightweight Models and Edge Deployment

  • Model compression, quantization, distillation, and compact architectures may reduce memory use and computation for embedded WPT applications.Evaluation should report model size, inference time, hardware platform, and control-period requirements, and compare learning implementations with conventional alternatives.Source: 3) Lightweight Models and Edge… · p. 31

4) Multi-Agent and Distributed Coordination

  • Multi-agent reinforcement learning could coordinate transmitter activation, charging schedules, beam coordination, and power allocation, but existing cited single-agent studies do not demonstrate distributed coordination.Future work should address scalability, communication overhead, partial observability, and coordination stability; federated learning may support collaboration without direct exchange of raw operational data.Source: 4) Multi-Agent and Distributed… · p. 31

5) Standardized Datasets and Evaluation Protocols

  • Common benchmark datasets and evaluation protocols could support reproducible comparison across representative operating variations and measures of accuracy, efficiency, robustness, and deployment cost.The review recommends common training-test partitions and cross-device validation.Source: 5) Standardized Datasets and… · p. 31

6) Digital Twins and Safe Simulation-to-Reality Transfer

  • Digital twins can support data generation, rare-condition evaluation, and controller testing before deployment, but should connect simulation and verification rather than replace experimental validation.The review calls for improved calibration, online updating, uncertainty representation, and validation against experimental systems.Source: 6) Digital Twins and Safe… · pp. 31–32

6. Conclusion

  • The review finds supervised learning effective for surrogate modeling and inference with representative labeled data, reinforcement learning suited to sequential control, and hybrid methods promising for physical consistency.Deployment remains constrained by dataset diversity, cross-device generalization, safety, interpretability, and the gap between simulation and hardware.Source: §6 · p. 32
Loading 2610.30101v1…