Machine learning for phase selection in multi-principal element alloys

Abstract

Multi-principal element alloys (MPEAs) especially high entropy alloys have attracted significant attention and resulted in a novel concept of designing metal alloys via exploring the wide composition space. Abundant experimental data of MPEAs are available to show connections between elemental properties and the resulting phases such as single-phase solid solution, amorphous, intermetallic compounds. To gain insights of designing MPEAs, here we employ neural network (NN) in the machine learning framework to recognize the underlying data pattern using an experimental dataset to classify the corresponding phase selection in MPEAs. For the full dataset, our trained NN model reaches an accuracy of over 99%, meaning that more than 99% of the phases in the MPEAs are correctly labeled. Furthermore, the trained NN parameters suggest that the valence electron concentration plays the most dominant role in determining the ensuing phases. For the cross-validation training and testing datasets, we obtain an average generalization accuracy of higher than 80%. Our trained NN model can be extended to classify different phases in numerous other MPEAs.