What surface characteristics truly affect thermal contact resistance -- An interpretability study based on deep learning and convolutional neural networks

he advancement of technology presents new challenges for heat dissipation. Thermal contact resistance is a critical factor in efficient thermal management. To gain a deeper understanding of thermal contact resistance, we developed a deep learning model based on convolutional neural networks for its prediction, which also serves as a reference for determining the actual contact area between two surfaces. The model utilizes the DenseNet121 architecture, and the dataset is generated according to surface fractal theory and multi-point contact mechanics. It was trained on the training set and demonstrated strong predictive performance on the testing set. Two specimen sets, produced through grinding and turning processes, provided both surface morphology and measured contact thermal resistance, which were used to evaluate the models accuracy, yielding comparable results. Guided Backpropagation and Class Activation Mapping were utilized for the interpretability study of the models visualizations. The results indicate that the contact and non-contact areas of the two surfaces each influence the prediction outcomes, thereby validating the models effectiveness. The surface features impacting contact thermal resistance were also directly visualized. This approach offers a new methodology to interpret the effects of surface characteristics on thermal contact resistance, advancing understanding and prediction capabilities in this field.