A novel multi-task graph neural network model for cable force optimization in cable-stayed bridges

Abstract

This study proposes a novel multi-task graph neural network (MT-GNN) surrogate model for efficient, multi-objective cable force optimization in cable-stayed bridges. The MT-GNN is trained on heterogeneous graph data derived from finite element simulations conducted over a Latin Hypercube Sampled design space, leveraging a Huber loss function with uncertainty weighting to concurrently predict node displacements and element bending moments. Integrating this trained MT-GNN into the NSGA-II framework enables rapid optimization aimed at simultaneously minimizing total girder vertical displacements and total bridge bending moment energy. Case studies on two-dimensional (2D) and three-dimensional (3D) single-pylon cable-stayed bridges demonstrate that the proposed framework achieves prediction accuracies comparable to finite element (FE) analysis, while drastically reducing computational costs. The optimized designs exhibit superior structural performance over traditional strategies, particularly for complex 3D configurations. These results demonstrate that the MT-GNN-based framework offers a computationally efficient, robust, and practical tool for multi-objective cable force optimization in cable-stayed bridges.