Machine learning based approach for phase prediction in high entropy borides

The current investigation uses a novel data-driven machine learning (ML) technique to provide the appropriate phase structures of multicomponent high entropy diborides. In the absence of sufficient experimental data for higher order diborides, the semi-empirical technique was first used to create the dataset using structure plots and develop a k-Nearest Neighbour (kNN) algorithm based classification model based on many structural and thermodynamic characteristics. Correlation analysis shows that the input feature space is suitable to model HEC compositions for the current dataset. Furthermore, testing accuracy of similar to 90% and excellent performance of other model evaluation metrics showed that the trained kNN model could further predict both quaternary and quinary high entropy diborides. A high F1 score (= 0.941) denotes low false positives and low false negatives for the trained model, which reciprocates how well the model predicts true values. The examination of the predicted compositions indicates that the ML framework developed can use the trend learned from the structure plot along with the available data from the literature to make reasonable predictions. ML with new data can be a reliable design method for the phase prediction in high entropy ceramics without requiring any prior experiments.