Physics-Informed Machine Learning for the Prediction of Thermal Conductivity from IR Images

Polymer-composite thermal interface materials (PC-TIMs) are essential for heat dissipation in electronics and energy storage systems, yet their cross-plane thermal conductivity (κ) is difficult to measure rapidly with conventional techniques. Here, we present a physics-driven machine learning (ML) framework that integrates infrared (IR) thermography with feature-engineered models to predict κ in the low-conductivity regime (<5 W m¹⁻ K⁻¹). More than 200 thermal images, obtained from experiments and Multiphysics simulations, were converted into universal temperature fields and transformed into structured feature vectors encoding gradients, Laplacian variance, and thermal extrema. Random forest regressors achieved robust predictions (R² = 0.905, MAE = 0.169) on experimental test sets, outperforming linear and boosting models. Domain adaptation via Gaussian-perturbed simulations enhanced transferability, while SHapley Additive exPlanations or SHAP analysis confirmed the physical relevance of density and gradient features. This work establishes IR thermography coupled with interpretable ML as a rapid and scalable diagnostic for PC-TIM development and manufacturing quality control.