Fragment-based energy models and machine learning methods for the computational study of organic molecular crystals

Abstract

The focus of this work is to investigate approaches to improving crystal structure prediction (CSP) through the study of different energy models and the use of machine learning.Simple fragment-based models are introduced based on anisotropic force fields, and areobserved to consistently improve the ranking of known polymorphs in CSP landscapes.These models are found to be amenable to machine learning, especially when fragmentenergies are learnt, as opposed to total lattice energies. When applied to absolute energies, fragment-based models including two-body terms using Density Functional Theory(DFT) are found to be largely basis set agnostic. By contrast, models involving secondorder Møller-Plesset perturbation theory (MP2) depend heavily on the basis set used,with three-body terms becoming vital for some systems. Overall, these results highlightthe benefit of simple fragment-based models for CSP, providing a bridge between forcefield approaches and periodic DFT. Through the use of machine learning these models can remain competitive with the computational cost of force fields, allowing for theexploration of CSP landscapes at higher levels of theory.In addition to these models, Gaussian processes are used to learn cocrystal stabilisation energies, where CSP is used to estimate the stabilisation energies of hypotheticalcocrystals. Here CSP is found to be a viable tool for predicting stabilisation energies,though results from machine learning models remain modest.

More information

Identifiers

ORCID for Graeme Day: orcid.org/0000-0001-8396-2771

Catalogue record

Export record