Equivalent Modelling and Simulation Method for 2.5D Chips Based on Machine Learning and Multi-Physics Field Coupling

Aiming at the problems of low computational efficiency and high resource consumption of traditional finite element simulation owing to complex structures such as through-silicon vias (TSVs) and bumps in 2.5D chip packages, this paper proposes an intelligent equivalent modelling and simulation optimization method that integrates machine learning and multi-physics field coupling. By constructing a dynamic equivalent model adaptive mechanism and adjusting the material parameters in real time based on deep neural network to capture the temperature-stress coupling effect; combining with the multi-scale geometry simplification technology, the deep learning is used to identify the key regions and differentially assign the modelling accuracy, which reduces the overall mesh number by 50% while ensuring the refined simulation of key regions (e.g., heat-sensitive and stress-concentrated regions). Dynamic model reconstruction and real-time optimization under multi-physics field coupling are further achieved through the integration of sensor data and simulation feedback. The experimental results show that compared with the traditional finite element method, the method shortens the simulation time by more than 30%, reduces the memory consumption by 50%, reduces the root mean square error (RMSE) of the temperature field by 2.8$^\circ$C, and controls the maximum error of the stress field within 4.8%, which significantly improves the multi-physics simulation efficiency and accuracy of the complex 2.5D chip package and provides a highly efficient and reliable solution for the design optimization of high-density integrated chips. solution for the design optimization of high-density integrated chips.