Materials Informatics approach to design new high-entropy shape memory alloys

Abstract

The current study uses a materials informatics approach to design new high-entropy shape memory alloys (HESMAs). The two-stage Synthetic Minority Oversampling Technique (SMOTE) was used to enhance the dataset. Several machine learning models are trained and tested to predict martensite start temperature (Ts). The trained ML models are used to select new potential compositions of HESMAs. A modified Artificial Neural Network (ANN) model shows a testing R² score of 0.89 and predicted the new alloys transformation temperature comparable to the experimental data. The effect of material descriptors on the prediction of Ts is interpreted using SHAP and LIME. The global interpretation of the test dataset is done using SHAP and identifies that δ and VEC positively affect the mean SHAP value for the studied dataset. The DSC data shows the martensitic and austenite transformation, and the compression stress-strain diagram confirms the typical detwinning behaviour of shape memory alloys.