Combining CALPHAD and Machine Learning to Design Single-phase High Entropy Alloys

Abstract: Although extensive experiments and computations have been performed for many years, the phase selection rules and prediction of single-phase HEAs currently remain elusive [1]. An underlying reason is that many of these phase selection rules were proposed or developed based on experimental data of limited composition spaces and/or insufficient data, and thus they are often not robust for the phase selection of HEAs. Machine learning (ML) methods have recently been used to predict the phase formations/selections of HEAs and achieved a certain degree of success. However, they still suffer the problems of small datasets and irrational selection of suitable physical descriptors. Moreover, ML models are often perceived as a ‘black box’ to the materials research community, thus often lack of clear physical meaning/understanding [2]. To effectively guide the design of HEAs, it is important and necessary to develop convenient and yet effective phase selection rules. In the talk, we report our research work on combining CALPHAD machine learn to design single-phase high entropy alloys. First, we used Thermo-Cal software together with the database TCHEA3 to generate a large dataset with more than 300,000 quinary data formed by Al, Co, Cr, Cu, Fe, Mn, Ni, and Ti. Next, we selected initial 15 features and employed a machine learning model to rank their importance. Based on the ranking, we then further performed feature reduction and identified most important features that governed the formation of single phases of FCC, BCC and other phases. Our study showed that at least 5 important features were needed to best discriminate the three classes of phases. These 5 features are equilibrium temperature, average atomic radius, average valence electron, difference in electronegativity, and difference in valence electron. The inclusion of equilibrium temperature, which was often neglected in previous studies, highlights its importance in the phase selection of HEAs. Our ML model was tested on 155 experimental data and reached a high accuracy of 81%. Based on the 5 important features and the large dataset, we established new phase selection rules for HEAs of single-phase FCC and BCC and achieved the accuracy of 93% and 92%, respectively. In addition, we also clarified the controversy concerning the phase selection rules in HEAs and offered in-depth insights into the relationships among composition-feature-phase of HEAs. Finally, we designed 222 new single-phases HEAs of BCC and FCC structures for experimental exploration [3].

Keywords: High entropy alloy, CALPHAD, Machine learning; Phase selection rule.
Full Article: https://proceedings.iaamonline.org/ar...

Join us: https://www.iaamonline.org/