Graph neural network-based 3D asphalt concrete pavement response modelling

This study proposed a Graph Neural Network (GNN) framework for modelling 3D asphalt concrete (AC) pavement responses. Based on the validated 3D finite element (FE) pavement model, the 3D finite element methods (FEM) pavement database of 840 FEM simulation cases was developed. The 3D FEM data were transformed into graph structures. The Graph Neural Network-based Pavement Simulator (GPS) was developed to model 3D pavement responses, comprising three main components: Encoder, Processor, and Decoder. Three variations of the GPS model were developed, utilising different Processor architectures: Graph Network and Deeper Graph Convolutional Network. Model performance was evaluated using two case studies: 'OneStep' for short-term predictions and 'Rollout' for long-term predictions. Results showed that using response differences between previous and target time steps as model output significantly reduced errors compared to using target response values. GPS models demonstrated accepted accuracy for both case studies; the long-term predictions ('Rollout') error was greater than that of short-term predictions ('OneStep'). GPS models achieved the rollout time of under 30 sec per FEM case, a dramatic improvement over the at least 12-hr runtime of traditional 3D FE models. The proposed GPS model could offer an accurate and computationally efficient alternative for predicting 3D pavement responses.