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

Abstract

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.

Introduction

This study presents a machine learning–based surrogate model to accelerate the Finite Element Analysis (FEA) process for designing automobile components, specifically a tractor steering knuckle. Traditional FEA is accurate but computationally expensive and time-consuming, especially during iterative design changes such as varying the spindle fillet radius. To overcome this limitation, the research proposes a Graph Neural Network (GNN) model that learns from historical FEA data to predict stress fields and peak principal stress values.

The steering knuckle, particularly the spindle radius (critical region), is analyzed since stress concentration in this area affects durability and strength. Instead of repeatedly running full FEA simulations, the trained GNN model maps geometric parameters and loading conditions directly to stress predictions, enabling rapid evaluation.

The dataset was generated from multiple FEA simulations with different spindle radii. The model was trained using mean-squared-error loss and validated using unseen geometries. Results show that the GNN predictions closely match traditional FEA results, with an average error of about 6.6 MPa and peak stress deviation within ±10%. Once trained, the model produces results within seconds, significantly reducing computation time.

Conclusion

A GNN?based surrogate trained on historical FEA of a steering knuckle can predict stress fields and peak principal stress with engineering?level accuracy while drastically reducing evaluation time. The approach supports accelerated early?stage design and broader parametric exploration, provided that new geometries remain within the trained design space. Future work includes multi?physics targets, uncertainty quantification, and active?learning schemes for automated dataset expansion.

References

[1] D. Nath, Ankit, D. R. Neog, et al., \"Application of Machine Learning and Deep Learning in Finite Element Analysis: A Comprehensive Review,\" Archives of Computational Methods in Engineering, 2024. https://doi.org/10.1007/s11831-024-10063-0 [2] E. Ababu, G. Markou, N. Bakas, \"Using Machine Learning and Finite Element Modelling to Develop a Formula to Determine the Deflection of Horizontally Curved Steel I-beams,\" in Proceedings of the 14th International Conference on Agents and Artificial Intelligence (ICAART 2022), pp. 958–963. DOI:10.5220/0010982400003116. [3] L. Liang, M. Liu, C. Martin, W. Sun, \"A deep learning approach to estimate stress distribution: a fast and accurate surrogate of finite-element analysis,\" J. R. Soc. Interface, vol. 15, 2018, 20170844. https://doi.org/10.1098/rsif.2017.0844 [4] J. Zhang, J. Zhao, Q. Rong, W. Yu, X. Li, R. D. K. Misra, \"Machine learning guided prediction of mechanical properties of TPMS structures based on finite element simulation for biomedical titanium,\" Materials Technology, vol. 37, no. 1, 2021, pp. 1–8. https://doi.org/10.1080/10667857.2021.1999558 [5] S. Rajasekaran, H. B. Khaniki, M. H. Ghayesh, \"On the mechanics of shear deformable micro beams under thermo-mechanical loads using finite element analysis and deep learning neural network,\" Mechanics Based Design of Structures and Machines, vol. 51, no. 12, 2022, pp. 6612–6656. https://doi.org/10.1080/15397734.2022.2047721 [6] A. Mujtaba, F. Islam, P. Kaeding, et al., \"Machine-learning based process monitoring for automated composites manufacturing,\" Journal of Intelligent Manufacturing, 2023. https://doi.org/10.1007/s10845-023-02282-2 [7] K. Naveen Kumar, M. Arshad, R. Sathpathy, \"Predicting von Mises stresses in I-sections using machine learning models,\" IRJMETS, 2022. DOI: https://www.doi.org/10.56726/IRJMETS31866 [8] H. Bolandi, X. Li, T. Salem, \"Bridging finite element and deep learning: High-resolution stress distribution prediction in structural components,\" Frontiers of Structural and Civil Engineering, 16, 2022, pp. 1365–1377. https://doi.org/10.1007/s11709-022-0882-5 [9] R. K. Maddipati, R. Agarwal, \"Machine learning based approach for brake pad pressure distribution,\" 2022. https://doi.org/10.22214/ijraset.2022.46611 [10] C. C. Harlapur, P. Kadiyala, S. Ramakrishna, \"Brake pad wear detection using machine learning,\" IJARIIT, vol. 5, no. 2, 2019. https://www.ijariit.com/manuscripts/v5i2/V5I2-1376.pdf [11] Q. Li, Z. Wang, L. Li, H. Hao, W. Chen, Y. Shao, \"Machine learning prediction of structural dynamic responses using graph neural networks,\" Computers & Structures, vol. 289, 2023, 107188. https://doi.org/10.1016/j.compstruc.2023.107188 [12] A. Srinivasan, B. Madhurya, S. S. Kangde, \"Machine Learning Based Approach for Prediction of Hood Oilcanning Performances,\" SAE Technical Paper 2023-01-0598. https://doi.org/10.4271/2023-01-0598 [13] Comprehensive introduction to Graph Neural Networks (web resource). https://www.datacamp.com/tutorial/comprehensive-introduction-graph-neural-networks-gnns-tutorial (accessed 13 Feb 2026) [14] Model training glossary (web resource). https://c3.ai/glossary/data-science/model-training/ (accessed 13 Feb 2026). [15] Model validation tutorial (web resource). https://www.tutorialspoint.com/model-validation-in-machine-learning (accessed 13 Feb 2026). [16] What is model validation (web resource). https://datatron.com/what-is-model-validation-and-why-is-it-important/ (accessed 13 Feb 2026). [17] Altair Physics AI Software (product page).

Copyright

Copyright © 2026 Parag Manohar Mahajan, Kumar Gaurav Gupta. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.