A novel prediction approach of three-dimensional thermal fatigue cracks in thermal compression bonding electrodes based on digital twin.

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.