 |
Center for Nonlinear Studies |
Modern materials design harnesses complex microstructure effects to develop high-performance materials. Recent highly successful approaches to materials design use machine learning algorithms for optimization. In order to incorporate microstructure into these approaches, a compact description of microstructure is required. However, faithful quantization of microstructure is an unsolved problem. We establish a method motivated by statistical physics which envisions microstructure variations as a low-dimensional manifold. We construct this manifold by leveraging multiple machine learning techniques including transfer learning, dimensionality reduction, and computer vision breakthroughs with convolutional neural networks.
Host: Chris Neale