Interpretable phase structure and hardness prediction of multi-principal element alloys through ensemble learning

Abstract

Optimizing the phase structure is critical for enhancing the mechanical properties of multi-principal element alloys (MPEAs). This study employed a stacking strategy within machine learning to build an ensemble model aimed at improving the accuracy of MPEA phase structure prediction, with an emphasis on the interpretability of the results. By utilizing Pearson correlation coefficients and mutual information scores, the importance of five key features was analyzed: valence electron concentration, difference in electronegativity, difference in atomic radius, mixing entropy, and mixing enthalpy, and weights were assigned accordingly. These features were used as the input variables to train the ensemble learning models. After comparing various models, it was found that an ensemble comprising Random Forest, XGBoost, CatBoost, and logistic regression performed optimally, achieving an accuracy of 0.875 and F1 score of 0.8731. Experimental validation confirmed the reliability of the ensemble model’s predictions. Furthermore, to demonstrate the applicability of the proposed ensemble model to continuous datasets, experiments were conducted to predict the MPEA hardness. The results show that the model also predicted the MPEA hardness, indicating that ensemble learning algorithms can effectively handle different types of data in material property predictions. In summary, this study highlights the potential value of ensemble learning in material science. Finally, the method of the ensemble learning model guiding material composition design is discussed in detail, which provides technical support for MPEA design and broadens the application scope of such algorithms.