Towards Accurate and Efficient Polymorph Prediction Independently of External Potentials with FFLUX

Abstract

Polymorphism, where a molecule can adopt multiple crystal forms, poses a major problem in several industries due to the different physical properties that the different structures can exhibit. For example, in the pharmaceutical industry the sudden appearance of an unexpected polymorph can lead to large reformulation costs due to the new structure having undesirable properties. Having a good knowledge of the possible crystal forms can therefore increase the confidence in a product, ensuring that a metastable phase is not being crystallised. Crystal structure prediction (CSP) studies offer a way of identify possible polymorphs computationally, but generally require expensive methods to accurately predict the small energy differences between polymorphs. However, these studies tend to neglect thermal contributions to stability, relying on lattice energies rather than free energies to determine stability. This thesis moves towards using the machine learning potential, FFLUX, in efficient and accurate CSP studies. FFLUX uses Gaussian process regression (GPR) models trained on data from quantum chemical topology methods to predict atomic energies and multipole moments. Formamide is chosen as a test system in a series of proof-of-concept studies due to its small size and rigidity. The accuracy of FFLUX is demonstrated through simulations of clusters and the ambient and high-pressure crystal phases of formamide. Helmholtz free energies are calculated using lattice dynamics calculations within the harmonic approximation, making contact with the energy ranking in CSP studies. These calculations are performed $10^{5}$ times faster than equivalent density functional theory (DFT) calculations whilst maintaining a similar level of accuracy, showing that efficient CSP is feasible with FFLUX. Limitations in the representation of repulsion and dispersion through Lennard-Jones potentials are exposed in this work. To address these shortcomings GPR models that account for repulsive intermolecular interactions to be used in FFLUX simulations, further improving the accuracy of simulations. The work presented here showcases the potential of FFLUX in modelling molecular crystals, taking the first steps towards simulations free from external potentials, which will allow for more efficient CSP studies with the accuracy of state-of-the-art methods in the future.

Date of Award	10 Jul 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Jonathan Skelton (Co Supervisor) & Paul Popelier (Main Supervisor)