Towards a generalized surrogate model for truss structure analysis using graph learning

Truss analysis has been well investigated by researchers and engineers with a large number of structural analysis software that can quickly and reliably provide analysis results. However, these methods require either expensive commercial software or self-developed in-house codes based on structural expertise along with advanced programming skills. Thus, incorporating this software into other complex applications, such as an online structural analysis framework, multiple-objectives truss optimization, structural reliability, etc., is highly challenging. This study proposes a novel and efficient surrogate model for performing truss analysis based on graph theory and deep learning algorithms, which is applicable for various truss topologies with different load scenarios without requiring retraining as other Deep Learning (DL)-based counterparts. The truss connectivityinformation is expressed through adjacency matrices, while material properties, external loads, and boundary conditions are considered node features. The performance and efficiency of the proposed methods are successfully validated with numerous unseen 2D trusses, providing highly similar results compared to the conventional finite element method.