Open Collections

Mokhtari, Narges

University of British Columbia

Compared to the finite element method (FEM), surrogate models enable more efficient structural analysis and optimization. Recent advances in deep learning have enabled neural networks to serve as surrogate models across various fields with remarkable success. However, their application in predicting stress distribution within 3D structures remains relatively unexplored. In this thesis, we propose a method to encode stiffened panels of varying geometries into 3D grid spaces suitable for processing by convolutional neural networks (CNNs). We compare the performance of these CNNs to multilayer perceptrons (MLPs) in predicting von Mises stress distribution in stiffened panels, employing principal component analysis (PCA) to reduce training complexity for the MLPs. Furthermore, we investigate the impact of skip connections by utilizing two different CNN architectures. Four case studies are conducted to assess how these neural network architectures predict stress distribution across various geometric configurations and loading conditions. The results reveal that CNNs, especially those with skip connections like U-Net, outperform MLPs by achieving less than 5% mean absolute percentage error compared to FEM results in all cases. While MLPs with PCA yield satisfactory outcomes for simpler problems, they falter in more complex tasks. CNNs demonstrate an exceptional ability to capture local stress variations, making them powerful surrogates to conventional finite element analysis (FEA). They also show great proficiency in predicting stress results within a limited amount of data. This underscores CNNs as a viable tool for real-world structural analysis.

Text

2024-12-23

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89981

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/