Machine learning for electronic structure

Machine learning (ML) methods have recently experienced rising popularity in quantum chemistry as a means to bypass expensive electronic structure calculations, which are used to calculate quantum mechanical properties for systems of atoms. This has led to advances in a broad range of applications, largely due to the fact that ML approximations have a much lower computational cost in comparison to electronic structure calculations, while maintaining comparable levels of accuracy. However, most of these models are designed to predict a predetermined set of quantum chemical properties and need to be retrained if an application requires properties outside of this set. Therefore, the focus of this thesis is constructing ML models that are capable of directly predicting the electronic structure of a system of atoms, since any other electronic property can be derived from it. Electronic structure is usually expressed in the form of a wave function or electron density, which are continuous multidimensional functions, making the learning problem non-trivial. In the course of this thesis, we explore appropriate compact representations for electronic structure and develop ML methods capable of learning and accurately reproducing these representations. By incorporating prior knowledge about the electronic structure problem into the machine learning models, we are able to construct models that also preserve the symmetries of electronic structure under various transformations, leading to high data-efficiency. We demonstrate the utility of these machine learning models for electronic structure by applying them to a range of quantum chemical problems, such as learning a density functional mapping between different levels of electronic structure calculations to enable accurate and fast approximations, reconstruction of highly accurate electron densities and infrared spectra and prediction of electronic wave functions via representations of the Hamiltonian operator.

Author dc:creator

• Bogojeski, Mihail

Advisor dc:contributor.advisor

• Müller, Klaus-Robert

Licence dc:rights.uri

• https://creativecommons.org/licenses/by/4.0/

Language dc:language.iso

en

Identifier URI

https://doi.org/10.14279/depositonce-18286

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/19488