Using Graph Attention Networks for Accurate Pavement Structural Response Prediction

Abstract: Mechanistic analysis is crucial in pavement engineering for predicting how pavements behave under traffic loading. Traditional methods, such as Layered Elastic Theory (LET) and Finite Element Analysis (FEA), are widely used. This paper explores the application of Graph Neural Networks (GNNs), specifically the Graph Attention Network (GAT), to predict pavement structural responses based on LET. GAT offers the ability to selectively focus on critical layer interactions, enhancing its accuracy in predicting strain propagation and deformation. By representing pavement systems as graph structures—where layers are nodes and interactions are edges—GAT enables efficient and accurate predictions without requiring mechanistic pre-analysis, which typically involves running 1-2 time steps of FEA to compute initial physical responses. The primary objective of this study is to demonstrate that the GAT model can function independently of mechanistic pre-analysis, addressing a key limitation in current methods. Additionally, we compare the GAT model with conventional Artificial Neural Networks (ANNs), showcasing its superior ability to capture complex spatial relationships and generalize across different pavement configurations. By naturally capturing these spatial and physical interactions, GNNs provide more accurate predictions than ANNs, which often struggle with generalization due to their lack of physical boundary knowledge. On evaluation, the GAT model achieved mean square errors of 18.40, 2.74, and 2.76 for the vertical, radial, and tangential strains, respectively, with corresponding mean absolute errors of 1.76, 0.78, and 0.77. These results illustrate GAT’s potential to reduce computational burdens and open new possibilities for replacing FEA in large-scale simulations.