Test platform information for developed systems.

Thermal compression bonding (TCB) electrodes that initiate thermal fatigue cracks compromise reliability and takt time in electronic manufacturing, and accurate prediction of three-dimensional (3D) electrode cracks is a prerequisite for crack mitigation. This study developed a digital twin (DT) framework that combined physics-based simulation and artificial intelligence (AI). The framework used the extended finite element method (XFEM) to build a high-fidelity electrode DT and reproduced fatigue behavior under coupled electrical, thermal, and mechanical loading through adaptive updating. To alleviate the scarcity of crack data, a conditional variational autoencoder (CVAE) with a position attention (PA) mechanism was constructed, with an error of 0.7% to 1.3% relative to experimental results. Using the augmented data, the PA-RePointNet model was developed to predict 3D crack morphology. Results showed that PA-RePointNet surpassed PointNet++ and PointCNN in prediction accuracy and stability and achieved a mean absolute error (MAE) of 2.8, a root mean square error (RMSE) of 5.1, and a coefficient of determination (R²) of 0.9378, while the maximum relative error between the reconstructed 3D cracks and experimental measurements was 1.87%. This framework provides a high-precision solution for electrode crack prediction and opens a new pathway for intelligent maintenance of TCB electrodes in microelectronic manufacturing.