Determining Temperature-Dependent Vickers Hardness with Machine Learning.

Assessing the hardness of structural materials at elevated temperatures is experimentally and computationally challenging, yet crucial for their success. In this work, a machine-learning method was developed to determine a material's temperature-dependent hardness based on its chemical composition and crystal structure. A total of 593 Vickers hardness data collected at various temperatures were extracted from the literature and used to train an extreme gradient boosting (XGBoost) machine-learning model. Applying a combination of composition descriptors and smooth overlap of atomic positions (SOAP) structural descriptors to represent these materials resulted in outstanding accuracy (R2 = 0.91; MAE = 2.52 GPa). The model's intrinsic variance was also measured by using a bootstrap aggregating (bagging) method, and the subsequent predictions showed strong agreement with the experimental data. The capability of the trained model was finally verified by demonstrating the model's ability to discriminate polymorphs, separate the properties of similar compositions, and reproduce the high-temperature hardness of several classic structural materials.