Edge attention-based graph neural network for truss structures elastic analysis and optimization

This study proposes a graph edge-attention network (GEAN) to include the topological features and physical information of structures using a graphic representation in the model training process. Unlike conventional graph attention network-based models, which compute message-passing weights solely from node features, GEAN introduces an edge-attention mechanism to jointly model structural node and edge features. The message-passing weights are determined based on stiffness information of structures, enabling the model to capture the influence of element stiffness on force transmission and enhancing physical interpretability. Three numerical examples with different dimensions are used to evaluate the performance of the proposed method. The results demonstrate that GEAN achieves significantly competitive prediction results compared to traditional surrogate and graph neural network-based models. As the dimensionality of the problems increases, GEAN with affordable samples consistently keeps the prediction error around 3%, far smaller than that of the other surrogate-based approaches.