Phase prediction and effect of intrinsic residual strain on phase stability in high-entropy alloys with machine learning

The phase formation and stability of high-entropy alloys (HEAs) are crucial to their properties, but the efficient prediction of them remains challenging due to the associated vastness of the composition space. In the present work, we study the formation and stability of solid-solution (SS) phases with a dataset consisting of 656 HEAs using machine learning (ML). We demonstrate the independence of nine physical parameters by self-organizing map (SOM) algorithm and rank them by feature importance methods, discovering that the root mean square residual strain (εrmsm s) is the most critical parameter. Based on the theoretical calculations and parallel coordinates plot (PCP) technique, we find that εrms can be used to quantitatively predict the stability of SS phases. Moreover, we apply the support vector machine, gradient boosting decision tree, multi-layer perceptron, and logistic regression algorithms to predict the formation of SS phases with testing accuracy values of 95.22 %, 94.78 %, 90.87%, and 89.57 %, respectively. Our results provide a perspective on the stability of SS phases from the viewpoint of intrinsic residual strain and show the support vector machine could be a better algorithm to predict the phase formation of HEAs. •The SOM algorithm demonstrates the independence of the nine parameters.•The feature importance methods suggest that ϵrms is the most critical parameter.•ϵrms can be used to quantitatively predict the stability of SS phases.•The SVM model provides the highest accuracy in SS phases prediction.