Dynamic graph-based approach for prediction of spatiotemporal response of RC structure to impact loads

Abstract

Accurate prediction of concrete structure responses subjected to impact loads is crucial for effective structural designs and safety assessments against such loads. This study proposes the Dynamic Graph Auto-Regressive (DGAR) model, a novel machine learning approach for spatiotemporal response prediction and damage modelling of reinforced concrete (RC) structures subjected to impact loads. Leveraging graph neural networks (GNNs) as surrogates for computationally intensive numerical simulations, DGAR employs dynamic graph modelling with explicit element and edge erosion to capture localized damage evolution. By incorporating a virtual global element and a multi-task learning strategy, it predicts element-based responses, such as strain, stress, and displacement, as well as non-element-based parameters, such as impact force. DGAR’s auto-regressive mechanism supports iterative predictions, functioning as a data-driven simulator that accurately tracks dynamic responses across the entire structure and over time. Evaluation results highlight DGAR’s superior performance in capturing the complex spatiotemporal dynamic responses of RC structures subjected to various impact scenarios. By significantly improving the computational efficiency compared to conventional FE numerical models, enhancing damage prediction accuracy over existing GNN-based methods, DGAR establishes a robust and scalable framework for structural response simulation under impact loads.