Public defence in Engineering Physics, M.Sc.(Tech) Heikki Muhli

Title of the thesis: Dispersion interactions in machine learning potentials for large-scale atomistic simulations

Doctoral student: Heikki Muhli
Opponent: Head of Department, Professor Alexander Tkatchenko, University of Luxembourg, Luxembourg
Custos: Professor Tapio Ala-Nissilä, Aalto University School of Science, Department of Applied Physics

While machine learning (ML) has slowly started appearing in our everyday lives, ML approaches have already been widely adopted by different branches of science to accelerate research. In materials science, the main benefit of ML is its ability to predict new results from data that has been calculated with computationally expensive quantum-mechanical models. The ML models used for these calculations avoid these expensive calculations and predict quantum-mechanically accurate results in a fraction of the computational time.

Most of the ML models used in materials science are based on describing the surrounding environment of individual atoms using so called local descriptors. These descriptors are a mathematical tool to communicate the information of an atom and its environment, that is, its neighbors in the immediate vicinity of the central atom, to the ML framework such that it can make predictions on the changes in the observables related to the central atom when its neighbors move. The movement of these neighbors can cause bond breaking and changes in energy in the atomistic system.

As mentioned, these descriptors are often local, centered on an atom. However, some features of an atomistic system can be highly non-local, spanning effective distances much larger than what can be captured by these local descriptors. An example of this are dispersion interactions that are long-range interactions between dipoles induced on atoms due to quantum-mechanical superposition of electrons. These interactions are much weaker than covalent bonds between atoms but due to their long-ranged and cumulative nature, they are responsible for the stability of many materials, such as graphite and black phosphorus.

In this study, we addressed the challenge of implementing these non-local interactions using local descriptors in the ML framework. The main result of the study is the non-trivial formulation of the state-of-the-art many-body dispersion model in atom-centered fashion which had not been done before. The new ML potentials allow the user to simulate large atomistic systems whose structure heavily depends on accurate description of dispersion interactions with almost quantum-mechanical accuracy, in a fraction of the computational time required for the same calculations using traditional methods.

Keywords: dispersion interactions, machine learning, density-functional theory, many-body dispersion, molecular dynamics

Thesis available for public display 10 days prior to the defence at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/

Contact information:

Doctoral theses at the School of Science: https://aaltodoc.aalto.fi/handle/123456789/52