1373579 - Graph-based Simulator to Predict Airport Flexible Pavement Responses from 3-D Finite Element Analysis

An extensive experimental, analytical, and numerical investigation of airfield asphalt pavement responses due to airplane gear loading is critical to validate pavement design approaches. The empirical method to estimate structural responses of flexible pavement using layer elastic and two-dimensional finite element (FE) analysis has been refined by using three-dimensional FE method. Although 3D FE methods have been proved to accurately capture various controlling parameters (e.g., non-uniform tire contact stresses, multiple-wheel loading, etc.), the time required to properly model the pavement system and the computational expensive process have hampered the implementation of this approach. This study proposes a graph network-based simulator (GNS), built by graph neural network (GNN), to simulate complex physical responses (strain, stress, and displacement) from small sets of FE data. It imposes strong inductive biases, where rich physical states are represented by the graph of interacting FE elements, and complex dynamics are approximated by learning message-passing among nodes of graphs. This GNN model will be trained using the FE database developed by the Illinois Center for Transportation (ICT), which contains more than 800 cases of FE results for flexible highway pavements. Although model training is time consuming, the simulation time could be reduced from days/hours to minutes for each case as long as the model is trained successfully. Ultimately, the GNN model of highway pavement could be adapted to simulate airfield pavement by transfer learning and adding a few airfield pavement FE cases.