An Effective 3D Thermal Network Integrated with Deep Learning for Improved Prediction of the 3D Thermal Properties of Complex Packaging Patterns

As patterns on semiconductor packages become more complex and extend into three dimensions, it becomes increasingly difficult to accurately determine the thermal properties of the package. Heat transfer simulation based on the original resolution of the pattern is limited in practicality due to the high computational cost. In this study, we developed a compact model based on effective thermal conductivity (ETC) that can be applied to 3D packages with complex pattern. The model consists of a thermal network using ETC, and it is designed to reflect more diverse thermal properties by introducing diagonal ETC in addition to the existing orthogonal ETC. The proposed compact model accurately predicts the thermal properties of the 3D package with an error within 1% compared to the original detailed model, and the temperature distribution is also predicted with high accuracy. In addition, the total time from simulation file generation to simulation and post-processing was reduced by about 90% compared to the detailed model, demonstrating excellent computational efficiency while maintaining high prediction accuracy. Furthermore, experimental validation using Peltier devices confirmed the validity of the simulation results. We have demonstrated that the in-plane and through-plane ETCs of the unit cells can be predicted with high accuracy by utilizing material distribution image-based data, 2D convolution neural network (CNN), and 3D CNN, which enables us to quickly obtain thermal property maps of new packages without additional simulations. The proposed model can be used to quickly analyze the thermal behavior of 3D packages with complex patterns, and is expected to provide high efficiency in terms of time and computational cost during the design and test stage.