| Feb 17, 2026 |
A physics informed machine learning model predicts thermal conductivity from infrared images in milliseconds, enabling fast, non contact screening for electronics and energy systems.
|
|
(Nanowerk News) Measuring how well a material conducts heat is essential for designing next-generation electronics, batteries, and power systems. But traditional techniques for determining thermal conductivity are slow, equipment-intensive, and poorly suited for high-throughput manufacturing environments.
|
|
Researchers at Clemson University have now demonstrated a physics-informed machine learning framework that can predict thermal conductivity directly from infrared (IR) images. The approach combines controlled thermography experiments with multiphysics simulations and interpretable machine learning to deliver rapid, non-contact conductivity estimates in polymer-composite thermal interface materials (PC-TIMs).
|
|
The findings have been published in ACS Applied Materials & Interfaces (“Physics-Informed Machine Learning for the Prediction of Thermal Conductivity from IR Images”).
|
 |
| Figure 1: Integrated workflow combining multiphysics simulation, infrared thermography, feature engineering, and physics-informed machine learning to predict the thermal conductivity of polymer–composite thermal interface materials. (Image: Reproduced with permission from American Chemical Society) (click on image to enlarge)
|
|
Rather than solving complex inverse heat transfer equations, the team converted more than 200 IR thermographs into structured temperature fields and extracted physically meaningful features such as thermal gradients, Laplacian variance, and extrema. These descriptors were used to train a Random Forest regression model optimized through cross-validation.
|
|
“Thermal conductivity measurement is traditionally cumbersome and not easily scalable,” said Ramakrishna Podila, professor of physics at Clemson University and the principal investigator of this study. “Our goal was to reframe the problem. Instead of solving the heat equation numerically from boundary conditions, we let physics-informed machine learning learn the mapping directly from temperature fields. The result is a fast, non-contact diagnostic tool that could fundamentally change how thermal materials are screened and certified.”
|
|
Crucially, the researchers addressed the common “domain shift” problem between idealized simulations and noisy experimental data by introducing Gaussian-perturbed thermal fields during training. This hybrid dataset significantly improved model transferability from simulations to real measurements.
|
|
The resulting model achieved strong predictive performance (R² ≈ 0.90 on independent test data), with mean absolute errors well within practical screening tolerances for low- to mid-conductivity polymer composites.
|
 |
| Figure 2. Comparison of simulated (COMSOL) and experimental IR temperature fields for a 20% BN–TPU composite, with resized and temperature-normalized images enabling direct, unified feature extraction across datasets. (Image: Reproduced with permission from American Chemical Society)
|
|
To ensure the model was not operating as a black box, the team employed SHAP (SHapley Additive exPlanations) analysis. The results confirmed that density and spatial temperature gradients, physically meaningful heat transport indicators were the dominant predictors, validating that the algorithm was learning interpretable thermal physics rather than spurious correlations.
|
|
“We were particularly excited to see that the model’s most important features aligned with physical intuition,” said Shinto M. Francis, a graduate student of physics and astronomy at Clemson University and also the lead author of the study. “The SHAP analysis showed that temperature gradients and density-related descriptors were driving the predictions, which tells us the algorithm is not just fitting noise — it is capturing real heat transport behavior.”
|
|
Because IR imaging is fast and non-contact, the framework could be integrated into in-line quality control systems for battery modules, printed circuit boards, or power electronics assemblies. With prediction times on the order of milliseconds per image, the method offers a pathway toward real-time thermal diagnostics in manufacturing.
|
|
By bridging experimental thermography, COMSOL multiphysics simulation, and interpretable machine learning, this work establishes a scalable route for rapid thermal metrology, potentially transforming how thermal interface materials are developed and certified for next-generation energy and electronic systems.
|
|
Source: Clemson University (Note: Content may be edited for style and length)
|
