Graph neural network-based algorithm for design optimization of truss structures

A novel Graph Neural Network (GNN)-based algorithm for design optimization of truss structures is proposed in this paper presents through a combination of Multi-head Attention (MHA) mechanism and a newly developed hybrid adaptive sampling strategy. The algorithm involves the generation of high-fidelity training datasets employing Finite Element Analysis (FEA) to establish a Graph Attention Network (GAT) surrogate model, the approximation of objective or constraint functions by the trained GAT model, and the solution of structural optimization problems using the Particle Swarm Optimization (PSO) algorithm. As a result, the GAT surrogate model dramatically reduces the number of required high-fidelity training datasets, and shows a strong ability to approximate the objective or constraint functions, thereby enabling accurate estimation of the optimum solution. It is shown through two numerical examples on truss structures that the proposed algorithm provides accurate and computationally efficient estimates of the solutions of structural optimization problems.