Can Machine Learning Predict the Space Group Preference of Organic Molecules?

Crystal structure prediction (CSP) is a valuable computational technique used to anticipate the likely crystal structures of a compound of interest. These methods have been proven useful in research and development of pharmaceutical solid forms and in guiding the discovery of materials with targeted properties. Despite success of CSP in these areas, its widespread application remains limited by computational cost. One approach to reduce the computational cost of CSP is to limit the search space of generated crystal structures; it is common practice to limit the search to a selection of the most frequently observed space groups, with the associated risk of excluding the space group of an observed crystal structure. As an attempt to reduce computational cost and ambiguity when choosing a set of space groups for CSP, we investigate the use of machine learning models to predict the most likely space group(s) of a given organic molecule. We find that both random forests and graph neural networks provide accuracies far above random, and better than what is achieved by selecting based on the overall space group frequencies observed for organic molecular crystals. The best model, using a graph neural network, achieves a maximum accuracy of 47.2% for single (top-1) space group prediction, which is an improvement of 8.2% above the reference. This model was trained with 3-dimensional molecular information, which improved accuracies compared to a model trained with only 2-dimensional bonding information. Furthermore, we found that random forest models performed best when both chemical and geometric molecular features are included in training, which indicates that both are important in defining a molecule's preferred space groups.