Training and application of graph neural networks for predicting structural responses targeted at tall building structures

Finite element analysis (FEA) methods are typically computationally intensive and time-consuming, particularly in structural optimization tasks that involve multiple iterations. To improve efficiency, neural networks have been utilized for predicting structural response. However, the scarcity of datasets in engineering, the complexity of tall building data, and the challenges faced by existing graph neural networks (GNNs) in handling such complex data restrict the potential of GNN surrogate models. In order to solve these problems, this study proposes a GNN training and application framework specially designed for tall building structures. The objective is to replace FEA with GNN surrogate for structural analysis. Firstly, the complete information of tall building structures using graph is represented. Secondly, tall building structures dataset is quickly generated by parametric modeling. Finally, a floor feature enhancement strategy is presented and the loss function is modified to optimize the patterns of displacement curves. This leads to the development of the tall building graph neural network (TBGNN), which exhibits satisfactory performance on specific tasks with less computational resources based on less data and lightweight architecture. Numerical experimental results demonstrate that the improved model achieves an average increase of over 5 % in accuracy compared to other GNNs for predicting structural displacements, inter-story drifts and natural vibration period of reinforce concrete (RC) structures. Moreover, the TBGNN is extremely sensitive to both the wind loads and structural dimensional changes. The proposed GNN surrogate model saves about 90 % of time compared to the traditional finite element analysis methods. These findings highlight the feasibility of using GNN surrogate models to optimize tall building structural designs against wind loads, providing valuable insights for addressing this prominent challenge.