Machine Learning Potentials for Perovskite Science - Applications and Development

First principles, or ab initio, modelling is an essential tool in modern chemistry and materials science. The main limitation of ab initio methods, however, has always been their high computational cost. Electronic structure methods such as density functional theory or coupled cluster are simply too slow to directly simulate the complex processes that most practical research in materials science focuses on. This is rapidly changing, however, with the advent of machine learning interatomic potentials (MLIPs). MLIPs can learn to reproduce the behaviour of ab initio calculations but at a fraction of the computational cost. MLIPs are already transforming science by enabling large, accurate simulations of complex systems, thereby providing insight which is inaccessible by other means. This thesis presents two main themes of research. Firstly, MLIPs are used to tackle two problems in the study of perovskite photovoltaics. Perovskites are a class of materials which have emerged as excellent functional materials for building solar panels and other optoelectronic devices. Over the last 15 years the development of perovskite photovoltaics has been meteoric, and there is intense effort focussed on understanding and improving the properties of these materials. However, perovskites are structurally complex at both the nano- and micro-scale. There is a large degree of structural diversity and the nanostructure is highly dynamic. As a result, there are multiple open questions about the nanostructre of these materials, which limits our understanding of how the photovoltaic properties arise from the underlying chemistry. This thesis uses large scale, accurate simulations (made possible by MLIPs) to give a detailed, real-space picture of the local structure in the prototypical perovskite CsPbI3. It is shown that the nanostructure can be characterised by two-dimensional local regions of correlated structure and motion. We study in detail how these regions are arranged, including the characteristic time and length scales. Following this, we address the more general problem of crystal structure prediction in arbitrary perovskites. One reason that so many researchers are interested in perovskites is that the design space is extremely large: One can build these materials using a wide variety of different chemical sub-units, and it is hoped that outstanding materials live somewhere among the many possible perovskite. The problem is that to explore this space, one has to synthesise and characterise many different candidate materials. This thesis presents a solution in the form of a MLIP-based structure prediction tool. Using our procedure, one can accurately predict the crystal structure of any, hypothetical, two-dimensional organic perovskite given only the chemical composition. The accuracy of the method is demonstrating by predicting and then synthesising a previously unreported material. Experiments confirmed that the crystal structure of the new perovskite indeed matched the prediction. Finally, the last part of thesis addresses a fundamental limitation in all of MLIP development. Almost all widely used MLIPs systematically neglect long-range electrostatic interactions. This is problem not only perovskites, but for numerous applications throughout chemistry and materials science. To address this, a unifying framework of electrostatic and self-consistent MLIPs is presented. Using this framework, we are able to express existing solutions to long-range MLIPs in a common language, and make clear connections to both density functional theory and classical charge models. This allows one to understand the limitations of a broad range of previous approaches. Following, we implement two new architectures which are the natural conclusion of this theoretical work. A simple test system is then introduced to probe relevant behaviour of electrostatic MLIPs: Conducting and insulating behaviour at a metal-water interface. It is shown that previously proposed models fail to reproduce certain aspects of this system, whereas our models naturally learn all of the physics even in extrapolative tests. It is hoped that these architectures can soon be applied to many exciting problems in computational chemistry and materials science.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Baldwin, William

Advisor dc:contributor.advisor

• Csanyi, Gabor

• machine learning
• machine learning interatomic potentials
• molecular dynamics
• perovskite photovoltaics
• perovskites

Licence

• https://www.repository.cam.ac.uk/bitstreams/79ea7202-3d95-4f03-8aa3-f5692378a856/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0009-0005-9863-6422
0000-0002-8180-2034

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/397726