Classifying crystalline phases in complex systems: a machine learning approach with order parameters.

Crystal polymorphism is a common phenomenon in many atomic and molecular systems. Identification and accurate classification of crystal polymorphs during molecular simulations is of scientific interest while studying crystallization, nucleation, and growth. Classification of polymorphs during simulations requires the use of well-designed order parameters, which are not obvious a priori. In this work, we use the XGBoost machine learning algorithm to identify a minimal set of Steinhardt bond order parameters to accurately classify polymorphs in a wide variety of crystal systems. Our success demonstrates the wide applicability of the method developed in this manuscript.