A Physics Regularized Neural Network (PRNN) Model for Fan-Out Package Junction-to-Case Thermal Resistance Prediction

Abstract

The thermal resistance of Fan-Out Package Junction-to-Case interfaces is a critical parameter that guides modern thermal design and reliability optimization in chip packaging. Conventional simulation-based thermal resistance evaluation is a time-consuming and computationally intensive process, which limits the efficiency of design optimization iterations. Physics-Regularized Neural Network (PRNN) represent a novel machine learning framework that incorporates physical laws into the training architecture, thereby demonstrating unique advantages in multiphysics problems involving complex structures. In this study, a PRNN model was developed for predicting die-to-package thermal resistance. The incorporation of the physical laws governing heat transfer processes led to a substantial enhancement in the training speed and prediction accuracy of the neural network. By embedding the physical laws governing heat transfer processes, we effectively improved the training speed and prediction accuracy of the neural network, demonstrating exceptional generalization capabilities even when working with a limited training dataset.