Autoregressive Graph Transformers with hybrid local–global attention for progressive damage evolution in composite laminates

Finite Element Analysis (FEA) offers trusted solutions for progressive failure in composite structures, but its computational cost becomes prohibitive in design-space exploration. Graph Neural Networks (GNNs) are therefore emerging as effective surrogates for mesh-based mechanics, representing unstructured meshes as graphs. However, standard GNNs exhibit over-smoothing and limited long-range information propagation, and extending mesh-based surrogates from static field prediction to progressive failure remains an open problem: damage evolution is autoregressive, with compounding rollout errors under initiation and irreversibility constraints.We present the first Graph Transformer methodology for autoregressive prediction of progressive failure in composite laminates, combining a frequency-controlled local–global attention backbone with a two-stage decoder that predicts the failure index and, gated by the failure-initiation criterion, the associated damage variable.The framework integrates with commercial FE workflows through an automated FEA-to-GNN pipeline and supports multi-GPU data-parallel training. We verify the methodology on progressive failure in CFRP open-hole plates, from a single 90° ply to a full 8-ply laminate in which ply damage evolves together with delamination at cohesive interfaces, achieving R² = 0.980 for the failure index and R² = 0.977 for damage on the single ply, with a speedup of approximately 5,900× over the corresponding FEA simulation for the laminate configuration and with standard message passing GNN architectures without attention mechanisms showing a 48–52% higher mean absolute rollout error for the failure-index and damage fields. The methodology further demonstrates strong extrapolation capabilities, predicting failure evolution in two-hole plates with no measurable loss of accuracy despite training on single-hole plates only.