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.
Task-oriented framework for ML-enabled WPT
The review organizes the field around three core domains that together span the lifecycle of a WPT system from design through operation. Structural design and optimization deals with shaping coil geometries, optimizing coupling structures, and determining physical layouts that maximize power transfer efficiency given constraints on size, material, and placement. Parameter and state estimation focuses on using measured or inferred data to determine unknown operating conditions such as load impedance, coil alignment, or distance between transmitter and receiver. Control and optimization addresses the sequential decision-making problem of adjusting transmitter parameters in real time to maintain performance as conditions change. By separating these domains, the review can compare which machine-learning techniques are best suited to each stage and where coupling between domains creates additional challenges.
Supervised learning for surrogate modeling and estimation
Supervised learning methods, including neural networks and gradient-boosted trees, are used primarily for two tasks: building surrogate models that approximate expensive physical simulations, and regression or classification of operating parameters from sensor data. When a labeled dataset covering various geometries, loads, and alignments is available, a supervised model can predict output voltage, current, or efficiency with far less computation than a full electromagnetic simulation. The review notes that the main requirement is representative data that span the expected range of operating conditions; without such data, the surrogate may fail outside its training domain. Supervised approaches also appear in state estimation, where the model learns to map sensor readings to estimates of coil position, load power, or coupling coefficient. The authors compare these methods on data requirements, computational cost during training and inference, accuracy relative to physics-based models, and the ability to provide online responses suitable for real-time WPT control.
Reinforcement learning for adaptive control and scheduling
Reinforcement learning (RL) is examined for its capacity to handle sequential control and scheduling problems in WPT. In this setting, an agent learns a policy that maps system state to actions such as adjusting transmitter frequency, modifying duty cycle, or reconfiguring coil switches. The reward signal typically encodes the goal of maximizing transferred power while minimizing losses or avoiding unsafe operating regions. The review highlights both the potential and the difficulties of RL for WPT. On the positive side, RL can discover control strategies that are difficult to engineer by hand, especially when the system dynamics are complex or partially unknown. On the negative side, reward design is nontrivial; shaping a reward that captures all safety and efficiency desiderata without causing unstable behavior requires careful engineering. Safe exploration is another concern: an RL agent may explore actions that damage hardware or reduce efficiency during the learning phase. Simulation-to-reality transfer adds further complexity, as dynamics identified in simulation may not hold on physical hardware due to unmodeled effects. The review argues that RL methods must be combined with safety constraints, model-based initialization, or hybrid approaches to be viable in practical WPT systems.
Hybrid and physics-informed methods
Hybrid methods that combine physical models with data-driven learning are presented as a way to overcome some of the limitations of purely data-driven approaches. By embedding known physics—such as the relationship between coil geometry and mutual inductance—into the learning architecture, these methods can achieve greater data efficiency and stronger physical consistency. Physics-informed neural networks (PINNs) and similar formulations add terms that enforce differential equations or conservation laws as part of the loss function, guiding the model toward solutions that respect the underlying science. The review compares hybrid methods against purely supervised or RL-based baselines on several criteria: how much labeled data is needed to reach a target accuracy, whether the model extrapolates reasonably to unseen operating conditions, and the computational overhead of enforcing physics constraints. The consensus in the review is that hybrid methods offer a promising middle ground, especially when data are scarce or when the system must operate safely over long periods without retraining.
Barriers to engineering deployment
The review identifies six principal barriers that currently prevent widespread engineering deployment of ML-enabled WPT. First, limited and biased datasets: most publicly available datasets are collected under narrow operating conditions, so models trained on them may perform poorly in real-world scenarios with different loads, distances, or alignments. Second, weak out-of-distribution generalization: models that rely on statistical patterns learned from training data often fail when confronted with situations that differ in kind, not just in degree, from the training set. Third, insufficient physical consistency: a learned model may produce predictions that violate basic electromagnetic laws, leading to unreliable control actions or even hardware damage. Fourth, unsafe online adaptation: adapting a model's parameters in real time based on streaming sensor data can cause the system to enter unstable or unsafe regions if the adaptation logic is not rigorously verified. Fifth, embedded-resource constraints: WPT controllers often run on microcontrollers with limited memory, compute, and power, making it impractical to deploy large neural networks or complex RL policies without model compression or quantization. Sixth, the absence of standardized benchmarks: without common test sets, evaluation protocols, and metric definitions, it is difficult to compare methods across groups or to track progress over time. Each barrier is discussed in terms of its impact on real-world deployments and the types of solutions the literature is beginning to offer.
Future research directions
The review closes by outlining a research agenda that the authors believe should shape the next phase of work in ML-enabled WPT. Uncertainty-aware learning is highlighted as a priority: models that can quantify their own confidence, or that incorporate probabilistic representations of inputs and outputs, would enable safer online adaptation and more reliable system integration. Physics-informed learning should be extended beyond PINNs to include more flexible architectures such as normalizing flows, Gaussian processes with physics constraints, and multi-fidelity models that combine high-fidelity simulations with low-cost measurements. Safe simulation-to-reality transfer methods, including domain randomization, cycle-consistent adversarial networks, and system identification techniques that map simulated dynamics to physical systems, are needed to bridge the gap between laboratory benchmarks and field deployment. Lightweight edge deployment research should focus on model compression, spiking neural networks, and inference engines that run on microcontrollers with kilobytes of memory. Distributed coordination across multiple transmitter-receiver pairs, particularly for charging furniture or electric-vehicle pads, requires multi-agent RL or consensus algorithms that scale to many units. Reproducible evaluation protocols, including standardized datasets, evaluation metrics, and open-source code repositories, would enable fair comparisons and cumulative progress. Finally, digital-twin-assisted development, where a high-fidelity simulation of the WPT system is used to generate synthetic data, validate control policies, and explore design spaces before physical prototyping, is proposed as a way to reduce cost and accelerate the iteration cycle. The authors note that addressing these directions will require close collaboration between machine-learning researchers, electromagnetic engineers, and systems engineers.