Graph attention networks enhanced predictive modeling for penetration-explosion damage in concrete structures

Rapid assessment of damage under combined penetration–explosion loading remains challenging due to the high cost of conventional simulations and experiments. Data-driven models improve efficiency but often lack physical consistency and struggle to capture the damage evolution process of concrete under extreme loading. This study introduces GAT-ImpactNet—an attention-enhanced graph neural network designed to advance engineering informatics by enabling real-time, physics-guided damage assessment for defense-related structural analysis. The model leverages high-fidelity numerical simulations, validated through projectile penetration and explosion tests on ultra-high-performance concrete targets, to establish a reliable dataset. GAT-ImpactNet incorporates physical priors via the loss function, allowing the graph attention mechanism to effectively capture complex structural interactions while enhancing prediction accuracy and computational efficiency. Validation against LS-DYNA simulations shows relative errors of 4.84% for penetration depth and 3.92–4.19% for blast cavity dimensions. Attention visualizations further demonstrate the model’s interpretability by highlighting critical structural interactions. This work contributes a novel, physics-guided deep learning framework that integrates fundamental engineering principles with neural networks, offering a robust informatics tool for real-time safety evaluation under extreme dynamic loads.