POD-Based Machine Learning Approach for Coupled EM-Thermal Analysis in Microwave Heating

Microwave heating processes require accurate simulation of time-varying electromagnetic (EM) properties to ensure reliable temperature predictions. Conventional methods which require conducting full-wave EM simulations at every thermal step, impose substantial computational burdens. In this paper, we propose a machine learning-based approach for reduced-order modeling that efficiently predicts the specific absorption rate (SAR) distributions in coupled EM and thermal analyses. The method employs proper orthogonal decomposition (POD) to extract dominant spatial and temporal modes from simulation data, thereby reducing the parametric complexity even when training data are limited. A linear regression model is then established between time-variant material properties and temporal POD modes, which enables the accurate reconstruction of dynamic SAR distributions during the heating processes. Numerical validations using 3-D FDTD simulations of microwave heating with temperature-dependent Debye characteristics of the material properties show that the proposed method achieves a maximum temperature deviation of only 0.0095°C for a 32.65°C temperature rise in a 20-minute heating simulation. This corresponds to a 0.029% error while reducing the required number of EM simulations to $1/6$ compared to conventional methods. This significant reduction in computational cost and high prediction accuracy highlights the potential of the proposed method as an effective alternative for coupled EM-thermal simulations in microwave heating applications.