Low-data machine learning models for predicting thermodynamic properties of solid–solid phase transformations in plastic crystals

Plastic crystals, many of which are globular small molecules that exhibit transitions between rotationally ordered and rotationally disordered states, represent an important subclass of colossal barocaloric effect materials. The known set of plastic crystals is notably sparse, which presents a challenge to developing predictive thermodynamic models to describe new molecular structures. To predict the transformation entropy of plastic crystals, we developed a comprehensive database of tetrahedral plastic crystal molecules (neopentane analogs) and used several types of features, including chemical functional groups, molecular symmetry, DFT-calculated vibrational entropy, and energy decomposition analysis to train a machine learning model. To select the most relevant features, we used a correlation matrix to screen out highly correlated features and ran sure independence screening and sparsifying operator (SISSO) regression on the remaining features. The SISSO regression samples over combinatorial spaces, including operations and features, to find the relationship between material properties. Using a dataset of 49 plastic crystals and 37 non-plastic crystals based on a common tetrahedral geometry, we have demonstrated the effectiveness of this strategy. Furthermore, we applied this strategy to develop a regression model to predict transition entropy and enthalpy. The top 100 models from the operation space showed that the overall distribution of performance became narrower, sacrificing the top-performing model but avoiding the worst models. Using this approach, we identified the top-performing descriptors to further clarify the underlying mechanisms of the plastic crystal transformation.