Physics-Guided FE Surrogate CDNN With Grouped Attention and Monotonicity Constraints for Solder Joint Thermal Fatigue Life Prediction

Abstract

The prediction of thermal fatigue life for solder joints is a key aspect of reliability design in electronic packaging. Existing methods face challenges such as high barriers to traditional finite element (FE) simulation, low efficiency, and poor reusability, while purely data-driven models suffer from issues like the “limited transparency” nature, overfitting in small-sample conditions, and limited physical interpretability. This article proposes a physics-guided FE surrogate convolutional deep neural network (CDNN) model that integrates a physical grouped attention mechanism with dynamic monotonicity constraints, specifically designed for high-accuracy and interpretable life prediction under small-sample conditions. Through fivefold stratified cross-validation, the results show that the model achieves an <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> of <inline-formula> <tex-math notation="LaTeX">$0.8604~\pm ~0.0181$ </tex-math></inline-formula> and a root mean square error (RMSE) of <inline-formula> <tex-math notation="LaTeX">$0.0825~\pm ~0.0039$ </tex-math></inline-formula>, demonstrating prediction accuracy comparable to random forest (RF) and higher than extreme gradient boosting (XGBoost) and multilayer perceptron (MLP), while providing improved physical interpretability. Additionally, the parameters learned by the model align closely with classical thermal fatigue failure theories, providing clear physical interpretability. These findings indicate that the proposed method offers a novel technical approach to the thermal fatigue life prediction of solder joints in electronic packaging, with significant engineering potential.