Project title: Machine Learning excited-state properties in materials
Supervisors: Johannes Lischner and Andrew Horsfield
Project description:
Many-body perturbation theory within the GW approach is one of the state-of-the-art methods for accurately calculating band structures of solids. However, to date, these calculations are limited to systems with a few hundred atoms. The key bottleneck of the GW method is the calculation of the dielectric matrix, which requires a large number of single-particle states for convergence. Model dielectric functions provide an efficient alternative to the exact dielectric matrix for many simple materials. However, most physical models have too few parameters to provide a good description of the dielectric matrix for complex materials systems.
In this project we aim at developing flexible machine learning models which can be used to reconstruct approximations of the dielectric matrix. To achieve this, it is necessary to condense the dielectric matrix into a small number of parameters, which can be predicted from atomic neighbourhood densities. Once a suitable set of parameters is found, we can extract them from ab initio calculations of the dielectric matrix and train machine learning models to predict them and reconstruct the full dielectric matrix. Initially we will test the approach on silicon clusters with up to 200 atoms, where the main aim is to establish whether it is possible to decompose the dielectric matrix into atomic contributions, where each atomic contribution only depends on the local chemical environment of an atom.