Machine learning-based phase prediction in high-entropy alloys: further optimization of feature engineering

Abstract

Understanding and predicting phase transitions in high-entropy alloys (HEAs) are pivotal for alloy design and performance optimization. This study aims to utilize artificial feature parameters for the predictive modeling of phase transitions in HEAs. Four elementary algorithms and four ensemble algorithms were employed for model selection. An innovative classification approach was introduced, classifying HEAs into eight distinct structural categories: IMS, AM, FCC + IM, BCC + IM, FCC, SS, BCC, and IM. To enhance feature selection, a three-phase feature screening method was devised, resulting in the identification of seven highly representative features from an initial pool of 64. These selected features were then used for model training. A comparative analysis was conducted against feature sets obtained from six peer-reviewed publications to validate the efficacy of the chosen features. In addition, various oversampling techniques were incorporated to further optimize model performance. Upon examination, factors such as electronegativity differences, heat of vaporization, and melting point temperatures play a decisive role in distinguishing alloy phase structures. Interactions between important characteristics exhibit significant differential impacts on phase structures.