ADASYN-assisted machine learning for phase prediction of high entropy carbides

Abstract

The lack of appropriate data and data imbalance hindered the development of ML models for identifying novel high-entropy ceramics. To circumvent data imbalance for ML-based ceramic design, we build a semi-synthetic database of high entropy carbides using literature data, atomic environment mapping-based structure plots and adaptive synthetic sampling (ADASYN) technique. A 5-fold cross-validated kNN classifier was trained on both original and balanced datasets. The kNN model trained on a balanced dataset has 95% testing accuracy while controlling for overfitting. SHAP describes the relationship between characteristics and goal variables. This paper shows a new ML approach with a decreased bias to predict high-entropy single-phase carbides preemptively.