Validation of a neural electromagnetic-thermal model for planar PCB coils in wireless power transfer using classical physics models

Wireless sensor nodes are increasingly powered inductively through planar PCB coils, whose design must balance transfer efficiency, received power and coil temperature. Neural network surrogate models can be used in place of finite-element simulation to explore this design space quickly, but their credibility is usually judged by test-set error, which cannot show whether the model has learned the underlying physics. This work proposes a validation method that treats established analytical models as reference standards. In ten one-factor sweeps, retuned to resonance at every point, a neural electromagnetic-thermal coil model is compared with seven analytical references for the electrical, magnetic and thermal quantities at identical geometry and operating point, with targeted finite-element solutions resolving disagreements. The model reproduces the references with median deviations of 1.3–10 % along the sweeps, recovers the expected power-law exponents and locates the optima correctly. The closest agreement belongs to self-checks built from the model’s own outputs, and level offsets of up to 20 % are the measured effect of the ferrite backing. Within the range covered by the training data, the method supports surrogate-based design of wireless sensor power supplies and transfers to other data-driven models in measurement and sensor design.