Research on heat conduction performance evaluation of thermal grease based on physical constraints of 3D CNN

Abstract To address the challenges in real-time heat conduction performance evaluation for thermal grease at chip-heat sink interfaces and overcome the limitations of conventional deep learning models that may produce predictions violating thermal conduction principles, this study proposes an evaluation method based on physical constraint 3D convolutional neural network (CNN). The method employs real-time thermal imaging sequences of the heat sink surface as input, utilizes a 3D CNN to extract spatiotemporal variations, and infers the heat conduction performance for thermal grease. Meanwhile, to ensure that the classification results comply with thermal conduction principles, this paper proposes a constraint method based on gradient order, embedding it into the loss function to achieve physical constraints on the model prediction. Under high-load dynamic heating conditions, the proposed model achieves 92.86% classification accuracy in distinguishing the heat conduction performance for thermal grease, with all the classification results adhering to thermal conduction principles. This outcome demonstrates that the methodology presented in this study can be used for real-time, reliable decision support for industrial thermal management applications.