Roles and opportunities for machine learning in organic molecular crystal structure prediction and its applications

Abstract

The field of crystal structure prediction (CSP) has changed dramatically over the past decade and methods now exist that will strongly influence the way that new materials are discovered, in areas such as pharmaceutical materials and the discovery of new, functional molecular materials with targeted properties. Machine learning (ML) methods, which are being applied in many areas of chemistry, are starting to be explored for CSP. This article discusses the areas where ML is expected to have the greatest impact on CSP and its applications: improving the evaluation of energies; analyzing the landscapes of predicted structures and for the identification of promising molecules for a target property. Graphical abstract

Frontiers of molecular crystal structure prediction for pharmaceuticals and functional organic materials

“…surrogate models 43 ) that more effectively filter structures to reduce the number of final structures for which DFT-D calculations are needed, or even adopting entirely new data-driven topological approaches for generating short-lists of candidate structures quickly, without extensive hierarchical filtering algorithms. 44,341,342…”

“…48 Machine learning potentials represent a natural extension of this idea. 44,100,101 Low-cost semi-empirical methods are similarly promising for intermediate renement of crystal structures and lattice energies. [102][103][104][105][106][107][108][109] These can be further combined with D-ML, in which an ML model is trained to correct a simpler model up toward the quality of a more expensive one.…”

“…Rapid improvements in machine learning potentials will likely also increase the accuracy with which those free energy simulations can be performed by enabling dynamics-based approaches to be used on potentials that approach quantum mechanical accuracy. 44,101 Improved understanding of how the strengths and weaknesses of widely-used DFT-D methods (e.g. delocalization and other systematic errors) impact crystal energetics will make it easier to identify when crystal energy rankings are likely to be problematic.…”

“…For further reading, readers are also referred to several excellent earlier reviews that focus on CSP methods and applications in greater technical detail. 22,23,34–44…”

“…For further reading, readers are also referred to several excellent earlier reviews that focus on CSP methods and applications in greater technical detail. 22,23,[34][35][36][37][38][39][40][41][42][43][44] 2 The crystal structure prediction challenge…”