Exploring chemical space for computational materials discovery

Abstract

Organic molecular crystals have potential uses for a wide range of applications. However, due to the vast number of possible molecules, exhaustively searching chemical space to find novel candidates for crystallisation, with promising solid-state properties, is prohibitively expensive. Therefore, more efficient approaches are required. Typically, such explorations of chemical space evaluate sampled molecules based upon related molecular properties to the targeted application of interest, ignoring the significant effect of the 3D crystal packing on the final properties. The main objective of this work is to address this by introducing the predicted crystal structures into the molecule’s evaluation.

For organic crystal structures, their stability is governed by weak intermolecular forces. Therefore, small modifications to sampled molecules, as part of a guided searches’ sampling, can have a substantial impact on the preferred crystal packings. This limits the effectiveness of using approaches such as templating crystal packings between different molecules. To rectify this, the effectiveness of incorporating crystal structure prediction (CSP) – a traditionally computationally expensive method which aims to find the preferred 3D crystal structure packings of molecules – into the fitness evaluations of an evolving population of candidate molecules is demonstrated with a genetic algorithm (GA). The most efficient implementation is tested, taking advantage of a minimum CSP sampling scheme strategy, allowing individual CSP calculations to be performed for each sampled molecule. This in turn can be seen to better guide the GA’s path, by more accurately informing the search using the property evaluations on the most likely predicted crystal structures.

This general CSP-GA approach has been demonstrated on the organic semiconductor chemical space, targeting molecules with the propensity to form crystals with high charge carrier mobilities, as well as a demonstration of a multi-objective fitness. Additionally, the improvements in automation of CSP workflows made in this work are shown to facilitate a large-scale study targeting porous materials.

More information

Identifiers

ORCID for Jay Johal: orcid.org/0000-0001-8489-4803

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

Catalogue record

Export record

Altmetrics