@Article{cmes.2026.083687,
AUTHOR = {Shamsh Parveen, Mehtab Alam, Ashraf Ali, Abdullah Alourani},
TITLE = {Deep Learning-Accelerated Extended Finite Element Method for Crack Propagation Analysis in Composite Structures},
JOURNAL = {Computer Modeling in Engineering \& Sciences},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/CMES/online/detail/27756},
ISSN = {1526-1506},
ABSTRACT = {Accurate prediction of complex failure modes in anisotropic composite structures—specifically matrix cracking, fiber rupture, and delamination (stratification)—remains a central challenge in computational fracture mechanics. The primary goal of this work is to bridge the gap between high-fidelity physical modeling and computational efficiency. While the extended finite element method (XFEM) enables mesh-independent crack modeling, its computational cost limits scalability. This work proposes a deep learning–accelerated extended finite element framework (DL-XFEM) that couples physically admissible XFEM fields with a neural network surrogate to predict incremental crack growth. XFEM is employed to generate stress-intensity factors and fracture-consistent state variables, which are used to train the network to replace the most computationally intensive crack-update steps while preserving fracture-mechanics constraints. The proposed framework achieves an 88.3% reduction in runtime relative to standard XFEM. Quantitatively, the method delivers an average trajectory prediction error of RMSE = 0.083 MPa√m and MAE = 0.071 mm, reducing prediction error by 61.2% compared to standard Physics-Informed Neural Network (PINN) baselines. Generalization is demonstrated across multiple stacking sequences and loading scenarios, and predictions are validated against experimental crack-growth data for aluminum, titanium, and carbon/epoxy laminate specimens. The results indicate that DL-XFEM provides a scalable and physically consistent approach for accelerating fracture simulations in composite structures without sacrificing predictive reliability.},
DOI = {10.32604/cmes.2026.083687}
}