Graph neural networks with hybrid local-global attention for effective prediction of mechanical response in structures

Graph Neural Networks (GNNs) are emerging as a transformative approach for predicting mechanical response in structures by naturally encoding unstructured finite element meshes as graphs. While traditional finite element analysis (FEA) provides trusted solutions for mechanics problems, it encounters significant computational and scalability bottlenecks. Existing machine learning approaches using regular grids fail to capture the irregular nature of real-world meshes, whereas standard GNNs suffer from over-smoothing and limited long-range information propagation that compromise accuracy. To overcome these limitations, we present a Graph Transformer methodology that implements an encoder-processor-decoder framework augmented with a frequency-controlled hybrid local-global attention mechanism, which systematically bridges local mesh connectivity with global information awareness across the computational domain. Our approach integrates seamlessly with real-world FEA workflows through an automated FEM-to-GNN pipeline that converts FE simulations to graph representations and generates training datasets directly from commercial solvers, enabling rapid deployment across various engineering applications without manual preprocessing bottlenecks. We validate our approach on open holes Carbon Fibre Reinforced Polymer (CFRP) laminate plates under various loading conditions and geometric configurations, and extend validation to nonlinear woven composites exhibiting plasticity and progressive damage under shear-dominated loadings, employing systematic hyperparameter optimisation with Tree-structured Parzen Estimator (TPE) sampling and mesh convergence studies. The linear case demonstrates excellent predictive accuracy (R2 = 0.98, RMSE = 0.00028) with 70× speedup over equivalent-accuracy FEA and 58 − 64% reduction in training and validation loss compared to standard message-passing GNN architectures without attention mechanisms, while the nonlinear case achieves R2 = 0.97 (RMSE = 0.0013) with 660× speedup.