Physics-informed graph neural network for 3D spatiotemporal structural response modeling of flexible pavements

Abstract

Quantifying pavement damage is crucial for roadway agencies' maintenance planning. This study proposed a Physics-informed Graph Neural Network-based Pavement Simulator (PhyGPS) to predict three-dimensional (3D) asphalt concrete pavement responses, building upon an established data-driven Graph Neural Network-based Pavement Simulator (GPS) model. The key innovation lies in integrating knowledge graphs and mechanics equations to create a physics loss function, distinguishing it from its data-driven counterpart. The physics loss function comprises strain-displacement and stress loss components derived from 3D strain-displacement relations and stress equilibrium principles. A thorough 3D finite element (FE) pavement database supported the model development. The 3D FE pavement data was transformed into graph format where nodes and edges represent 3D FE pavement models’ nodes and node connections, respectively. Performance evaluation employed two case studies: “OneStep” for assessing short-term predictive capabilities and “Rollout” for examining long-term prediction accuracy under practical conditions. Results demonstrated that the physics-informed GPS model showed superior long-term predictive capability and robustness while maintaining excellent short-term accuracy compared to the data-driven model. Both models achieve rollout time under 8 s per FE simulation case, a dramatic improvement over the 12-h runtime of traditional 3D FE pavement models. The PhyGPS model successfully integrates physics principles, spatial relationships between structural components, temporal correlations in structural data, and complex material properties, offering an accurate, robust, and computationally efficient solution for predicting 3D pavement responses.