Machine Learning Assisted Modelling of Atomic-Resolution Electron Microscopy

The rapid advancements in nanomaterial science, particularly in the context of renewable energy solutions, have brought Transmission Electron Microscopy (TEM) to the forefront as a crucial analytical tool. The interaction of high-energy electrons in a microscope with the sample cause direct damage to the sample, induce diffusion, and trigger chemical reactions. The focus of this thesis is on a detailed understanding and quantification of the primary mechanisms by which the electron beam influences the sample, leading to such damage. It particularly focuses on gold nanoparticles supported by hexagonal Boron Nitride (hBN)/Titanium Dioxide (TiO₂) and the two-dimensional material Molybdenum Disulfide (MoS₂).
At the core of this research is the development of advanced computational methodologies, combining Machine Learning, Kinetic Monte Carlo (KMC), Molecular Dynamics (MD) simulations, and Density Functional Theory (DFT), to analyze and predict the interactions between the electron beam and nanomaterial samples. These methods aim to distinguish between intrinsic properties of the materials and alterations caused by the electron beam, a challenge highlighted by the necessity for high-resolution imaging in TEM.
The thesis is structured into five comprehensive chapters, beginning with a foundational overview of TEM and beam-induced damage. It then progresses into a hybrid KMC-MD approach, shedding light on the atomic-level dynamics and damage mechanisms in gold nanoparticles. This method demonstrates significant potential in understanding and quantifying the processes of diffusion and atomic displacement under electron irradiation.
Further, the study dives in into the behavior of MoS₂ under electron beams, using KMC to investigate the formation of sulfur vacancies and the influence of variables like charge states and coordination numbers on defect formation.
A significant part of the thesis is dedicated to exploring Machine Learning and Graph Neural Networks (GNNs) applications in material science. It focuses on the development and implementation of Machine Learning interatomic potentials, particularly Neural Equivariant Interatomic Potentials (NequIP), showcasing their ability to predict heat transport dynamics with accuracy comparable to DFT, but with greater computational efficiency.
The thesis also highlights the efficacy of active learning strategies, specifically the "query-by-committee" technique, in refining machine learning models for molecular dynamics simulations. This approach enhances the predictive capabilities of the models, particularly in large molecular systems, and is exemplified in the comparative analysis of heat transport in gold nanoparticles on TiO₂/hBN interfaces.