Efficient graph neural networks for predicting the responses of truss structures

Precisely predicting structural responses is essential for evaluating the safety and reliability of structural systems. This study explores the performance of graph neural networks (GNNs), including GraphSAGE and ChebNet, as effective surrogate models for predicting nodal displacements and bar internal forces of truss structures. The structural information, including node coordinates, external loads, boundary conditions (node features), cross-sectional areas (edge features), and node connectivity, is represented as graph data and used to train the GNNs. Notably, edge-to-node feature transformation is applied to enhance node feature information, and knowledge-based boundary conditions are accurately integrated into the GNNs, thereby improving the model accuracy. The results show that the nodal displacements and internal forces obtained from the developed GNNs closely align with those from the traditional finite element method. In addition, five widely used machine learning algorithms, including Decision Tree (DT), Random Forest (RF), AdaBoost (AB), Gradient Boosting (GB), and eXtreme Gradient Boosting (XGB), are adopted to assess the efficacy of the developed GNNs. The comparative results demonstrate that the developed GNNs outperform DT, RF, AB, GB, and XGB models for both nodal displacement and internal force predictions. This study reveals that the GNNs hold significant promise for improving the accuracy and efficiency of structural analysis.