Machine Learning Based Electronic Structure Prediction: From Nanostructures to Complex Alloys

I will describe our work on using specialized first principles calculations with machine learning tools to enable prediction of the electronic structure of various nanomaterials and bulk systems. I will focus on two related but independent directions. First, I will show how helical and cyclic symmetry adapted density functional theory calculations may be used to train interpretable machine learning models of the electronic fields of quasi-one-dimensional materials. The descriptors in this framework are global geometry and strain parameters. Through examples involving distorted carbon nanotubes, I will show how the framework can be particularly accurate in its prediction, even with limited training data. I will discuss the use of this framework for automated materials discovery, and in multiscale modeling.<br/> <br/>Second, I will discuss the use of high-throughput first principles calculations to train machine learning models of bulk systems, possibly featuring some degree of disorder in atomic arrangements. The descriptors in this framework are local in nature and the prediction of electronic fields occurs in a pointwise manner spatially. I will describe how a combination of strategies, including transfer learning, thermalization, and the use of Bayesian Neural Networks can allow the development of machine learning models that are systematic, reliable, and very efficient (both in terms of training and prediction). I will discuss the use of this framework for calculation of the electronic structure of bulk solids with defects, and compositionally complex alloys, from which, other material properties of engineering interest may be inferred. I will end with a discussion of future research directions.