Graph Neural Network Based Surrogate Model for Design Informed Structural Optimization

Abstract

Structural optimization of skeletal forms is crucial in weight-sensitive applications. Optimizing such structures often involves iterative, computationally intensive methods, which are inefficient under varying design parameters and constraints. This paper introduces a novel surrogate model based on Graph Neural Network (GNN) for real-time structural optimization, aimed at significantly reducing computational costs. In our approach, trusses composed of pin joints and connecting members are represented as graphs, where joints correspond to vertices and members to edges. This correspondence forms the use of Graph Neural Networks (GNNs) to predict topology and size-optimized truss structures. The GNN models the truss as a graph, with edges denoting member cross-sectional areas and nodes representing truss joints, based on input parameters such as geometry, load combinations, and boundary conditions. The resulting predicted structure reflects the optimized topology and member sizes. The proposed model bypasses the need for iterative computations by learning from a dataset comprising various problem definitions and their corresponding optimized results. This GNN-based optimization holds substantial promise for design scenarios requiring rapid and reliable optimization, demonstrating the potential for significant computational time savings while maintaining high accuracy in predicting near-optimal truss layouts. This is particularly significant in the context of sustainability, where industrial users can produce optimally designed structures with minimal material usage within a fraction of the computational power and time required for different applications. Testing results indicate that the model effectively generalizes across various design scenarios, providing near-optimal solutions with minimal computational effort. Specifically, the predicted structures exhibited a normalized root mean square error (NRMSE) of less than 10−3 and R2 values approaching unity. Additionally, predictions were made in under 0.01 s, demonstrating both accuracy and efficiency.