Abstract
High entropy alloys (HEAs), which are multi-component alloys having constituent elements in equi-atomic or near equi-atomic ratios, are receiving immense attention owing to their remarkable mechanical and physical properties. These unusual properties depend on one or more of the phases that these alloy systems constitute, namely solid solution (SS), intermetallic compound (IM), and amorphous (AM) phases. Therefore, phase prediction is crucial in selecting appropriate elements that lead to the formation of a HEA with desirable properties. In this work, machine learning (ML) models are used on a set of design parameters using multi-classification for HEA phase prediction. The ML models comprised ensemble-based (Random Forest and Stacked ensemble) and support vector machine (SVM) methods. To predict solid solution phase formation, researchers used atomic size difference and the parameter Omega = T-m Delta S-mix/vertical bar Delta H-mix vertical bar where T-m is average melting point, Delta S(mix )is entropy of mixing and Delta H-mix is mixing enthalpy. The features used in this study are 51, atomic size difference, average electron concentration, electronegativity difference, average melting point, mean atomic radius, mixing enthalpy, the entropy of mixing, and bulk modulus. The training dataset comprising of 601 as cast alloys is used with cross-validation for test data, and the phases SS, IM, AM, SS + IM, and IM + AM are predicted. The phase prediction accuracies are calculated using one-vs-rest, and precision-recall curves are plotted to determine the model performance. For phases AM, SS, and IM, the stacked ensemble displayed better accuracies when compared to SVM and Random Forest. The findings indicate that the stacked ensemble, which combines weak learners and meta-models, provides accuracy comparable to the neural network model accuracy reported in the literature. These findings provide insights to researchers and practitioners in selecting features and ML models while designing HEAs.
NSTL
Elsevier