Hierarchical graph-based machine learning model for optimization of three-dimensional braced steel frame

This paper proposes a method for topology optimization of three-dimensional braced steel frames under static seismic loads, aiming to minimize structural volume while satisfying response constraints. The method consists of a proposed hierarchical graph-based machine learning model that sequentially embeds graph representations of structural elements to create a comprehensive lumped mass model suitable for seismic analysis and optimization of building frames. The proposed model is utilized as a reinforcement learning agent, a class of machine learning, to observe the current building frame configurations, modify the frame to improve its performance, and adjust the model parameters using the obtained reward function. Numerical results demonstrate that the proposed graph-based model can improve its performance better than benchmark of a graph neural network utilized in previous research. When applied to three cases of unseen large three-dimensional building frames, the trained agent with the proposed model outperforms the genetic algorithm and simulated annealing in the optimization task with 11–46% lower objective function while utilizing the same computational cost. The generality, performance, and computational efficiency of the agent indicate its applicability to applied to the optimization of three-dimensional frames.