School of Civil and Environmental Engineering
Ph.D. Thesis Defense Announcement
Deep learning characterization and mechanical ranking of microstructure features in geomaterials
By Daniel T. Chou
Advisor:
Chloé F. Arson (CEE/Cornell)
Committee Members:
Dr. Jorge Macedo (CEE /GT)
Dr. David Goldsman (ISYE/GT)
Dr. Richard Regueiro (CEAE/Colorado Boulder)
Dr. WaiChing (Steve) Sun (CEEM/ Columbia University)
Date and Time: Oct, 29, 2024. 1:00 pm ~ 4:00 pm, EST
Location: Mason 2119
Zoom Link: https://gatech.zoom.us/j/92616731798
ABSTRACT
Understanding the analytical relationship between fabric descriptors (i.e., microstructure descriptors) and stiffness tensors has been a long-standing challenge in geomechanics. This doctoral research aims to address this challenge by exploring three key research questions.
The first research question focuses on calculating 3D fabric descriptors from 2D images by using deep learning (DL) models. The performance of a pruned ACS-VGG19 network is assessed under different training metrics and configurations of trainable and fixed convolutional layers. The optimal model configuration utilizes the MSE loss function and fully trainable convolutional layers, achieving a Mean Absolute Percentage Error (MAPE) of 2 to 5\% for aggregate size, aspect ratios, and solidity. Computational costs increase linearly with the number of images extracted per direction, but performance improvements are marginal beyond a single image per direction.
The second research question is to understand the relative importance of microstructure features on local field variables, such as the stress field. Here, 2D composite materials made of a solid matrix and cracks are modeled numerically with the Finite Element Method (FEM) and cohesive zone models (CZM). A Non-Linear Variational Autoencoder (NLVAE) is proposed with a skew-normal distribution sampling and correlation penalties to improve latent feature disentanglement. The NLVAE effectively captures stress concentrations and reconstructs stress fields with consistent Normalized Mean Square Error (NMSE) across different stress components. Dynamic Time Warping (DTW) analysis reveals correlations between stress latent features and crack network descriptors, such as connectivity, path length, eccentricity, and centrality. Significant stress latent features and microstructure descriptors vary with boundary conditions, but certain latent variables consistently emerge as significant across different descriptors and loading paths. Notably, the distributional shape, tail behavior, and symmetry of microstructure descriptor distributions have more influence on the stress field than basic measures of central tendency and spread.
The final research question aims to explain why classic homogenization schemes break down in cracked solids. The information flow from microstructure variations to stiffness changes is measured. A support vector machine (SVM) with feature selection using mutual information (MI) and Analysis of Variance (ANOVA) highlights the influence of individual fabric descriptors on the breakdown of the Mori-Tanaka model.
This thesis explores the relative importance of statistical geometric descriptors in biphase composites for the reconstruction of 3D microstructure images, the estimation of the stress field, and the calculation of effective stiffness. Deep learning algorithms have been developed to discover microstructure self-organization patterns and to explain micromechanical phenomena. The results not only fill existing gaps in microstructure characterization, influential fabric recognition, and homogenization theory, but also offer numerous opportunities for future research, such as the effect of data preparation, extensions to 3D composites, metrics between sequences of statistical datasets, and extrapolation from DL.