Coupling Graph Neural Networks and Finite Element Analysis for Faster Mechanical Stress Prediction

Finite‑element analysis (FEA) is the de‑facto standard for structural assessment of automotive components, but repeated meshing, boundary‑condition setup, and solver runs make design exploration costly and time‑consuming. This paper investigates a data‑driven surrogate approach that couples graph neural networks (GNNs) with FEA to predict full‑field stress distributions on a tractor axle steering knuckle. Historical FEA results from multiple geometries, focusing on variations of the spindle fillet radius, were curated as a training corpus and ingested by Altair Physics AI, a geometric deep‑learning framework. The proposed workflow comprises dataset generation from validated FEA studies, GNN training with mean‑squared‑error loss, and out‑of‑sample validation on previously unseen geometries. The trained model delivers stress patterns and peak principal stress estimates within approximately ±10% of traditional FEA across the tested cases, while reducing prediction time from minutes–hours to seconds once trained. Learning curves indicate stable convergence with decreasing training and validation losses and mean absolute error (MAE) on the test set of ~6.6 MPa at the critical spindle radius. The results suggest that GNN‑based surrogates can accelerate early‑stage design iterations and enable broader parametric exploration, provided that new geometries remain inside the trained design space. We discuss dataset design, sources of error, and practical guidance for deploying GNN/FEA surrogates in industrial settings.