This paper uses random forests models to predict the elastic properties and mechanical hardness of compounds using only their chemical formula. The model, trained on over 10,000 compounds and 60 features, is used to create triangular graphs for B-C-N compounds, revealing that a 1:1 B-N ratio leads to various superhard compositions. These machine learning predictions are validated using evolutionary structure prediction and density functional theory (DFT), identifying BC₁₀N, B₄C₅N₃, and B₂C₃N as dynamically stable superhard materials (hardness > 40 GPa) potentially synthesizable via low-temperature plasma methods.

Wei-Chih Chen, Joanna N. Schmidt, Da Yan, Yogesh K. Vohra, Cheng-Chien Chen