CMStatistics 2020: Start Registration
View Submission - CMStatistics
Title: Multi-fidelity statistical modelling for molecular crystal structure prediction Authors:  Olga Egorova - University of Southampton (United Kingdom) [presenting]
Roohollah Hafizi - University of Southampton (United Kingdom)
David Woods - University of Southampton (United Kingdom)
Graeme Day - University of Southampton (United Kingdom)
Abstract: Structural polymorphism occurs when crystallising the same molecule results in obtaining multiple (up to tens of thousands) of solid forms which vary in terms of the associated lattice energies. Obtaining reliable energy evaluations is of great importance, as energy differences are linked to the differences in physical and chemical properties of the structures which affect, for example, the suitability and safety of a pharmaceutical compound. Computational methods for crystal structure prediction (CSP), which would allow for reliable energy evaluations are highly demanding in terms of computational costs, making their direct application for all trial structures infeasible. We employ multi-fidelity Bayesian Gaussian process modelling to combine different levels of computational methods to obtain predictions for the highest level at much lower costs. We assess the uncertainties of the obtained energy predictions together with their propagations in energy rankings. The approach for energy surface optimisation is also considered.