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

Area of Science:

• Materials Science
• Physics
• Computer Science

Background:

• Polymer-composite thermal interface materials (PC-TIMs) are critical for heat management in electronics and energy storage.
• Conventional methods for measuring their cross-plane thermal conductivity (κ) are often slow and complex.
• Accurate and rapid κ measurement is vital for optimizing PC-TIM performance.

Purpose of the Study:

• To develop a rapid and scalable diagnostic tool for predicting the thermal conductivity of PC-TIMs.
• To integrate infrared (IR) thermography with physics-driven machine learning (ML) for κ prediction.
• To enable accurate assessment of PC-TIMs in the low-conductivity regime (<5 W m^-1 K^-1).

Main Methods:

• A physics-driven ML framework was developed, combining IR thermography data with feature engineering.
• Over 200 thermal images from experiments and simulations were processed into structured feature vectors.
• Random forest regressors were trained on features including gradients, Laplacian variance, and thermal extrema.

Main Results:

• The ML framework achieved robust predictions of κ, with R² = 0.905 and MAE = 0.169 on experimental data.
• Random forest models outperformed linear and boosting models in predicting thermal conductivity.
• SHAP analysis confirmed the physical significance of density and gradient features for κ prediction.

Conclusions:

• Infrared thermography coupled with interpretable ML provides a rapid and scalable method for diagnosing PC-TIMs.
• This approach is suitable for PC-TIM development and manufacturing quality control.
• The framework enhances the ability to characterize materials with low thermal conductivity.