Abstract
Accurate prediction of blast loads on structures is essential for blast-resistant design. Traditional high-fidelity numerical simulations remain computationally expensive, while existing data-driven approaches are often limited to fixed and structured meshes. To address these challenges, this study proposes a graph neural network–based method for blast load prediction on dynamic unstructured meshes. A Dynamic Mesh Temporal Graph Network (DMTG-Net) is developed, integrating a cross-mesh temporal alignment mechanism and physics-informed learning strategies. The cross-mesh alignment, implemented via a k-nearest neighbor-based mapping, enables consistent supervision under time-varying mesh topologies. In addition, dynamic graph construction incorporates propagation-time-aware edge features to reflect physical information transfer, and residual constraints derived from momentum and energy considerations are introduced to enhance physical consistency. Thereafter, a multi-scenario dataset is constructed based on experimentally validated computational fluid dynamics (CFD) simulations, including single-, double-, and triple-source detonation cases with adaptive mesh refinement. The proposed model achieves high prediction accuracy, with coefficient of determination (R²) values exceeding 0.99 in most cases, and effectively captures shock wave propagation, reflection, and interaction processes. Furthermore, the model demonstrates robust generalization capability under variations in detonation locations and moderate extrapolation of charge mass, accurately predicting key engineering quantities such as peak overpressure and arrival time. The results indicate that the proposed method provides an efficient and reliable surrogate modeling framework for blast load prediction in complex engineering scenarios involving dynamic meshes.

This publication has 19 references indexed in Scilit: