Search

This work evaluated the phase prediction capability of high entropy alloys using four supervised machine learning models K-Nearest Neighbors (KNN), Multinomial Regression, Extreme Gradient Boosting (XGBoost), and Random Forest. The study addresses the challenge of predicting multicomponent alloys by considering the overlapping of multicategorical stability parameters. Eight prediction classes (FCC, BCC, FCC+BCC, FCC+Im, BCC+Im, FCC+BCC+Im, Im and AM) were used. Finally, the predicted results were compared with those of two new alloys fabricated by induction melting in a controlled atmosphere using X-ray diffraction (XRD). The results indicate that with a robust database, appropriate data treatment, and training, satisfactory and competitive prediction indicators can be obtained with traditional machine learning predictions based on four prediction classes: Solid Solution (SS), Solid Solution with Intermetallic (SS+Im), intermetallic (Im), and amorphous (AM). The best predictive model obtained from the four evaluated models was Random Forest, with an accuracy of 72.8% and ROC AUC of 93.1%. • Machine learning was applied to 2434 experimental data for phase prediction. • The best evaluated model is Random Forest with 72.8% Accuracy and 93.1% ROC AUC. • An adequate prediction was found for two new HEAs obtained experimentally. • The packing factor may be the key to the increase in HEA prediction performance. [ABSTRACT FROM AUTHOR]