CAL‐Crack Net: A Deep Learning Framework for Predicting Nonlinear Multicrack Propagation With Mutual Interference

Mutual interference among multiple cracks introduces strong nonlinearity and uncertainty into their propagation, challenging for traditional methods. This study proposes CAL‐Crack Net, a data‐driven deep learning framework for multicrack path prediction based on phase‐field simulations. A boundary‐constrained algorithm generates diverse geometric models (single to triple cracks), whereas phase‐field simulations provide high‐fidelity path data (~30,000 points per case). An efficient gradient‐priority sampling strategy compresses this data to 3000 points per path, preserving essential features for model training. CAL‐Crack Net combines CNN, attention, and LSTM modules to process crack tip sequences, extracting spatial features, capturing temporal evolution, and focusing on critical regions. Validation on 30 independent test samples shows that the predicted paths achieve an average Dice similarity of approximately 87% and an average Hausdorff‐based contour similarity of approximately 93%, accurately capturing nonlinear details. SHAP‐based post hoc attribution identifies correlations between crack‐coordinate inputs and predicted phase‐field values, providing qualitative interpretability support, whereas physical consistency is assessed separately through energy and convergence analyses.