Ensemble Tree Machine Learning Prediction of Superhard Compositions

Superhard materials with a hardness > 40 GPa have various important applications. Here, we use machine learning (ML) simulations to find the elastic properties of chemical compounds. Following the work by Chen et al. [in npj Computational Materials 7, 114 (2021)], we use approximately 10,400 target compounds and 60 features, which are based on the properties of elements, orbital occupations, stoichiometric traits, and ionic bonding levels, to train ML models. Once the models are generated, they are utilized to predict the shear modulus, bulk modulus, and hardness of B-C-N compounds. We compare different ensemble models including random forests and gradient boosting trees. We also consider weighted samples in the training and compare the ML predictions to first-principles calculations. The results indicate that ML models can efficiently predict the mechanical properties with a reasonably good accuracy.