On the generation of ab initio potential energy surfaces using machine learning techniques

A potential energy surface is a major pre-requisite to carrying out quantum molecular dynamics studies on chemical systems. These studies allow theoreticians to explore the behaviour of modern-day chemistry in ways that are not feasible in a lab, enabling predictions to be made about experiments performed at extreme temperatures and pressures, and even helping to reveal reaction pathways. To achieve this, a PES associates nuclear position and energy, constructing an n-dimensional surface from high-level ab initio calculations that are fit using physically motivated functions. However, this fitting is a painstakingly slow process, and must be tailored to individual systems. This thesis will explore three ways of speeding up the generation of ab initio PESs: simplifying the fitting process, reducing the number of fitting points, and reducing the computational cost of the ab initio calculations themselves. Machine learning (ML) algorithms offer a number of potential advantages for the construction of PESs: firstly, they represent more of a ``black-box'' approach to the fitting that promises an easier route to accurate surfaces; second, reducing the dimensionality of the problem holds the promise of constructing a surface from significantly fewer points. These algorithms also have access to active learning techniques that aim to reduce the size of machine learning datasets. As such, a particular subset of machine learning model, the neural network, will be used along side a novel application of a firefly inspired optimisation algorithm to speed up PES generation. While the development of new basis sets paired with correlation consistent effective core potentials will aim to speed up data generation.