Describing Chemical Simulation with a Bayesian Worldview

"This project relates to the Computational and Theoretical chemistry research theme of the EPSRC and aims to develop statistical methods by which to quantify the uncertainty and performance of chemical simulations.

Analysis of experimental techniques is becoming increasingly reliant on computational chemical simulations. However, these simulations often fail to cover experimental timescales, meaning that conclusions are extrapolated in the hope that simulations represent the real system. Limited computational resource requires sacrificing either system size/duration (ab-initio methods) or accuracy (classical force-fields) in order to make the computation tractable.

A Bayesian approach, where probabilistic models are proposed to describe our system (based on the underlying physics and chemistry) and then accepted or rejected based on their predictive power, should improve the comparison between simulation and experiment.

Recent developments in machine learning force-field (MLFF) methods may offer another promising avenue to the listed issues with molecular simulations by providing near ab-initio accuracy with classical force-field system sizes. MLFFs are developed by machine learning a potential energy surface that has been sampled with energies and forces from static quantum mechanical calculations. MLFFs have been widely employed to draw conclusions about system dynamics.

The analysis of experimental data, such as X-ray diffraction and neutron scattering, the generation of machine learning models and the use of quantum computing algorithms can all benefit from such probabilistic models.

Initially this project aims to apply the Bayesian method to Quasi Elastic Neutron Scattering (QENS) applying a similar methodology as within McCluskey, Coles and Morgan, (2024) to statistically efficiently quantify how the uncertainty from molecular dynamics simulations propagates when estimating important physical quantities such as the incoherent structure factor. This theoretical investigation, carrying out molecular simulations with conventional and MLFF will be combined with experimental QENS measurement of 3 model liquid systems, 3 model molecular systems, benzene, ethylene glycol and 1-butyl-3-methylimidazolium hexafluorophosphate (BMIM-PF6) conducted at the at the ISIS Neutron and Muon Source These systems display a range of molecular forces, (purely van der Waals, hydrogen-bonding, and ionic interactions, respectively), and so should provide an ideal benchmark producing valuable validation and fundamental understanding of the extensibility and limitations of these methods.

McCluskey, A.R., Coles, S.W. and Morgan, B.J. (2024) 'Accurate Estimation of Diffusion Coefficients and their Uncertainties from Computer Simulation'. arXiv. Available at: http://arxiv.org/abs/2305.18244 (Accessed: 17 October 2024). "