Comparing Machine Learning Models for Reconstructing Microwave Thermal Fields

The investigation tested three compact coordinate-based regression families—ridge-regularized tensor-product B-splines, a Fourier-feature multilayer perceptron, and Chebyshev Elastic Net—using sparse samples from a published steady-state bioheat field. To prevent immediate-neighbor data leakage, the team employed five spatially blocked test folds, exclusion buffers, and limited training budgets of 1% or 5% of the available nodes.

According to the findings, the ridge-regularized tensor-product B-splines achieved the lowest mean held-out absolute errors in top optimizer configurations, alongside strong Dice scores in hot-bearing folds at the larger training budget. The project was supported by the KAU Endowment (WAQF) and the Deanship of Scientific Research at King Abdulaziz University in Jeddah, Saudi Arabia.