Abstract: Reliable prediction of solder joint microstructure evolution and mechanical degradation under thermal aging is essential for ensuring long-term reliability of electronic assemblies. However, existing approaches often require large datasets or fail to capture complex microstructure-property interactions, limiting their predictive accuracy. Here, we introduce a self-attention physics-informed neural network (SA-PINN), which embeds an Arrhenius-based diffusion model into the loss function, integrates a transformer self-attention mechanism to capture complex interactions between microstructure and performance, and employs a Squeeze-and-Excitation block for dynamic channel recalibration. Trained on 240 samples spanning widely used package types aged at 94, 120, and 150 °C, SA-PINN simultaneously predicts intermetallic compound (IMC) growth, Pb-phase coarsening, and shear strength, achieving R² values of 0.992, 0.958, and 0.932, respectively, while reducing mean absolute error (MAE) by up to 50 % compared with strong data-driven regressors. Ablation studies confirm that both attention and channel re-weighting contribute synergistically to performance gains, and Shapley analyses quantify the main effects of microstructural and geometric descriptors on each target, consistent with experiments; meanwhile, the model-learned activation energies for IMC growth are broadly consistent with values fitted from Arrhenius plots, supporting physical consistency. This data-efficient yet physics-consistent framework offers high interpretability and can be readily extended to fatigue-life forecasting and multi-field coupling problems, thereby providing a versatile tool for rapid and reliable assessment as well as design guidance in electronic packaging.

Key words: Machine learning, Physics-informed neural network (PINN), Reliability, Shapley additive explanations, Advanced packaging