AI-aided search for new materials

Symbolic regression methods to predict hardness of materials (arXiv:2304.12880)
Here, we introduce a low-dimensional physical descriptor for Vickers hardness derived from a symbolic-regression artificial intelligence approach to data analysis. This descriptor is a mathematical combination of materials’ properties that can be evaluated much more easily than hardness itself through the atomistic simulations, therefore suitable for a high-throughput screening. The artificial intelligence model was developed and trained using the experimental hardness values and high-throughput screening was performed on 635 compounds, including binary, ternary, and quaternary transition-metal borides, carbides, nitrides, carbonitrides, carboborides, and boronitrides to identify the optimal superhard material. The proposed descriptor is a physically interpretable analytic formula that provides insight into the multiscale relationship between atomic structure (micro) and hardness (macro). We discovered that hardness is proportional to the Voigt-averaged bulk modulus and inversely proportional to the Poisson’s ratio and Reuss-averaged shear modulus. Results of high-throughput search suggest the enhancement of material hardness through mixing with harder, yet metastable structures (e.g., metastable VN, TaN, ReN2, Cr3N4, and ZrB6, all of them exhibit high hardness).