TSO-GCN: A Graph Convolutional Network approach for real-time and generalizable truss structural optimization

The truss structural optimization is a major research topic in the field of structural, civil, aerospace engineering, etc. Conventionally, the truss structural optimization methods are often inefficient because they run in iterations and are computationally intensive during each iteration. In this paper, we propose a generative design framework based on deep learning networks to predict three dimensional structural topologies without iterative computations while achieving acceptable accuracy. Different from most commonly used deep learning driven structural optimization approaches that transform structural geometries into images, our innovation lies in solving optimization problems based on geometric analysis from the perspective of graphs, and efficiently predict the near optimal truss structure with a negligible computational time. Therefore, it shows potential significance in design scenarios when the structure is described by connections and massive number of optimization computations are required. The proposed generative design framework is called TSO-GCN (Truss Structural Optimization - Graph Convolutional Network), which is an encoder–decoder based graph convolution network designed to map the problem definition and the desired truss layout. Once trained, it is expected to directly predict truss layouts by feeding into the encoded optimization problem definitions. To train the TSO-GCN, a dataset consisting of different number of problem definitions and their corresponding minimum volume results generated by conventional methods is constructed and fed into the network. The experiments show that TSO-GCN can predict results with the near optimal accuracy compared with conventional approaches while costing an average time of only 1 s. Besides, extra experiments with totally unseen dataset are performed to demonstrate the generalizability of the proposed method.