Inferring Low-Dimensional Representations of Microstructure Using Convolutional Neural Networks
This event is part of the PhD Final Oral Exams.
Examining Committee: William Klein, Pankaj Mehta, Rama Bansil, Kevin Black, and Claudio Chamon
Materials microstructure effects are crucial to modern high-performance materials. Recent highly successful approaches to materials design use machine learning algorithms for optimization; however, these approaches have not incorporated microstructure. One requires a faithful, compact quantification of microstructure to integrate it into the machine learning workflow. Motivated by statistical physics, we establish a machine learning-driven method that envisions microstructure variations as existing near a low-dimensional manifold. We show how recent work on texture image synthesis using Convolutional Neural Networks (CNNs) can be used to define a faithful representation for microstructures. This allows us to extract distances between texture images by leveraging CNNs that were originally trained for computer vision tasks. We then use manifold learning algorithms to embed these distances into a low-dimensional space, thereby constructing faithful, compact microstructure manifolds.