Finite element analysis (FEA) can accurately predict the fatigue life of solder joints in IC packaging; however, the prolonged computational time required for analyzing different geometric parameter combinations limits design efficiency. In this study, a hybrid deep learning model combining physics-informed machine learning (PIML) with long short-term memory (LSTM) networks is proposed. FEA was performed on multiple geometric parameter combinations, and the Darveaux model was employed to calculate the fatigue life, thereby establishing a training database. The proposed dual-branch architecture processes static geometric parameters and temporal features of nonlinear plastic work simultaneously, with a physics-informed loss function ensuring predictions adhere to fatigue mechanics principles. The results demonstrate that the PIML-LSTM model outperforms traditional purely data-driven deep learning methods in solder joint fatigue life prediction. Physics constraints guide the model to converge to physically feasible solutions, improving the physical consistency and generalization capability of predictions. Once the model is established, the fatigue life for different parameter combinations can be rapidly predicted, providing a rapid prediction tool for packaging reliability assessment.
Shih et al. (Thu,) studied this question.