Machine Learning for Condensed Matter Physics

Abstract

This thesis is about the application of machine learning (ML) methods to a variety of problems in condensed matter physics. As condensed matter physics has large computational and experimental datasets readily available - or in the computational case it is often easy to generate them - it is an ideal field for ML. Here ML can be used to speed up existing calculations and routines, or more ambitiously, make new discoveries. Specifically, in this thesis, I look at using ML to: efficiently solve lattice models, predict the Curie temperature of ferromagnets and discover new molecules. Solving lattices models exactly, is generally only possible for small systems, as the dimensionality of the wavefunction increases exponentially with the system size. In this context, we apply ML to: construct exact density functional theory maps, which bypass the calculation of the wavefunction, and learn the propagator, which evolves the system in time. For a ferromagnet to have practical applications it must have a large Curie temperature, however it is very difficult to predict the Curie temperature from first principles. Using a dataset of experimental measurements, we create a Curie temperature predictor, with a mean absolute error of around 50 K. Finally, we look at using ML for the task of discovering new molecules that have certain properties. We demonstrate a ML architecture, capable of learning molecular distributions.