Graph neural networks as surrogate models for structural analysis: a study on buckling behavior

Date

2025-1-10

Author

Kurt, Ömer

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

120
views

80
downloads

Cite This

This thesis presents a novel approach to structural analysis using Graph Neural Networks (GNNs) as surrogate models, specifically focusing on predicting buckling behavior of thin-walled structures with and without stiffeners. The research addresses the computational challenges in traditional finite element analysis by developing an efficient machine learning framework that maintains accuracy while achieving computational speeds faster than conventional methods. To create a well balanced dataset, a comprehensive data generation pipeline is introduced, creating diverse structural geometries using Bezier curves and implementing systematic load case generation procedures. The study developed an enhanced graph representation system that effectively captures both local and global structural behaviors through innovative features such as super-nodes and virtual edges, while ensuring rotational and translational invariance through principal component analysis-based coordinate transformation. The framework demonstrates remarkable accuracy in buckling prediction across both non-stiffened and stiffened structures, with performance validated against finite element analysis results using multiple test datasets, including scaled geometries and complex loading scenarios. The framework reduces analysis time and enables rapid evaluation of multiple design variants. This reduction in computational time, combined with maintained prediction accuracy, demonstrates the potential of the framework to transform preliminary design processes. The research contributes to the growing field of machine learning in structural analysis by providing a robust methodology for creating efficient surrogate models and by demonstrating effectiveness of GNNs on a global property prediction like buckling.

Subject Keywords

Graph neural networks, Surrogate model, Machine learning, Buckling, Structural analysis

URI

https://hdl.handle.net/11511/113521

Collections

Graduate School of Natural and Applied Sciences, Thesis