Can crystalline polymorphism be predicted from single-molecule properties?

Understanding the crystallization of molecules, and specifically the appearance of polymorphs, is a great challenge to modern chemistry, with both fundamental and practical aspects. Here, motivated by the established ability of Machine Learning (ML) algorithms to perform classification tasks, we harness ML-based tools and existing chemical datasets to ask whether the existence of polymorphs of a molecular crystal can be predicted based solely on properties of the corresponding single molecules. We implemented an algorithm that can predict the existence of polymorphism with an average accuracy of 65\%. Although this is not enough to generate a reliable “polymorph predicting” engine, our results reveal intriguing trends. We suggest that the limited performance of the model reflects the inherent bias of crystallographic data towards monomorphs. Namely, the fact that only one crystal form of a certain compound has been observed in experiments to date does not rule out the possibility that additional stable crystal structures could exist. Our results provide statistical support for a broadly held claim in the crystal engineering community that the proportion of possible polymorphs is, in fact, much larger than represented in the data.