Transfer Learning for Misaligned WPT Coils

Misalignment is the enemy of reliable wireless power transfer. Analytical models are fast but they start to drift as lateral offsets grow. Pure ML models learn the drift but need a lot of data. This note covers a blended approach I used for my M.Tech thesis to keep estimation error below 4%.

Dataset & Preprocessing

• Litz-wire and ferrite-core rectangular coils with lateral offsets up to 80 mm.
• Base parameters from analytical model: self/mutual inductance, coupling, AC resistance.
• Augmentations: Gaussian noise on measured voltages/currents and randomized coil spacing.
• Targets: corrected inductance and resistance values validated against lab measurements.

Model Stack

Start with the analytical outputs as features, then fine-tune a shallow regressor:

• Backbone: pretrained MLP on synthetic coil data.
• Heads: two regression heads for L and R to decouple error surfaces.
• Loss: Huber with per-parameter weighting to prioritize mutual inductance accuracy.

Results

• Error under 4% across offsets; improved efficiency by ~2% over purely analytical baselines.
• Transfer learning cut the required measured samples by ~60% compared to training from scratch.
• Model stays lightweight enough for rapid sweeps across coil geometries.

Takeaways

Combining physics-informed priors with ML keeps the model grounded and reduces data needs. The analytical model is the launchpad; the learner cleans up the edge cases caused by misalignment.