Deep Learning-based Stress Prediction for Short Fiber Reinforced Composites Using TabNet

Abstract

This thesis investigates how deep learning models can predict stress uncertainties in short fiber-reinforced composites (SFRCs). The central question is whether machine learning models, particularly TabNet, can accurately predict stress variations within SFRCs under different fiber distributions. To address this, full-field data were generated using Digimat and expanded through data augmentation to train the TabNet model. Model performance was evaluated using root mean square error (RMSE), and the results show that the TabNet model effectively predicts stress variations across different realizations of representative volume elements (RVEs). The model captures the stress uncertainties arising from the microstructural variability of the material while maintaining high accuracy. This study demonstrates that combining TabNet with data augmentation significantly reduces the computational resources required for traditional full-field simulations while providing accurate predictions of stress uncertainties in SFRCs, highlighting its potential applications in composite material design and manufacturing.