FrameGraph: A Scalable Performance Evaluation Method for Frame Structure Designs Using Graph Neural Network

Abstract Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN) encounter a critical scalability issue when applied to the performance evaluation task for frame structure designs. Specifically, a model of MLP or CNN is limited to structures of a particular topology type and fails immediately when applied to other topology types. In order to tackle this challenge, we propose a scalable performance evaluation method (called FrameGraph) for frame structure designs using Graph Neural Network (GNN), offering applicability to a wide range of topology types simultaneously. FrameGraph consists of two main parts: (1) Components and their connections in a frame structure are denoted as edges and vertices in a graph, respectively. Subsequently, A graph dataset for frame structure designs with different topologies is constructed. (2) A well-defined GNN design space is established with a general GNN layer, and a controlled random search approach is employed to derive the optimal GNN model for this performance evaluation task. In numerical experiments of car door frames and car body frames, FrameGraph achieved the highest prediction precisions (96.28% and 97.87%) across all structural topologies compared to a series of classical GNN algorithms. Furthermore, the comparison with MLP and FEM highlighted FrameGraph's significant efficiency advantage. This verifies the feasibility and optimality of FrameGraph for the performance evaluation task of frame structures with different topologies.