(142a) Machine Learning Collective Variable Discovery for Materials Design and Engineering

Ferguson, A., University of Illinois at Urbana-Champaign
Data-driven modeling and machine learning have opened new paradigms and opportunities in the understanding and design of soft and biological materials. The automated discovery of emergent collective variables within high-dimensional computational and experimental data sets provides a means to understand and predict materials behavior and engineer properties and function. In this tutorial, we will describe two machine learning techniques for collective variable identification – nonlinear manifold learning using diffusion maps, and nonlinear dimensionality reduction using autoencoding neural networks (“autoencoders”). We will provide illustrative applications of diffusion maps to recover low-dimensional assembly landscapes for self-assembling patchy colloids from simulated and experimental data, and inverse building block design by rational sculpting of the landscape to stabilize desired aggregates. We will provide illustrative applications of autoencoders to identify collective variables for protein folding and the use of these coordinates to accelerate molecular dynamics simulations.