Screening for shape memory alloys with narrow thermal hysteresis using combined XGBoost and DFT calculation

Shape memory alloys (SMAs) are desirable candidates for elastocaloric effect materials, but they all suffer from large thermal hysteresis (Thys). This study analyzes multicomponent TiNi-based SMAs dataset by machine learning (ML) to explore new SMAs with narrow Thys. The second-largest eigenvalue λ2 of the stretch transformation matrix U is added to the original dataset to guide the ML process as a feature. Firstly, λ2 is obtained by first-principles calculations combined with ML. XGBoost Regressor (XGBR) combined with Leave-One-Out Cross-Validation (LOO-CV) is selected from four algorithms for modeling with the highest coefficient of determination R2 of 0.87. The introduction of λ2 improves the performance of the model. The dataset is divided into 15 groups based on different doping elements (such as Hf, Cu, Zr, etc.), among which TiNiCu is the most predictive component with the R2 of 0.89. Over 500 TiNiCu components are randomly generated and predicted Thys. Based on the contour maps created from the prediction results, it is found that Thys is likely to decrease with the increase of Cu doping in general, and minimum Thys occurs when the Cu is about 15 at. %, which is consistent with the existing experimental results. Eventually, a potential Thys minimum (1.2 K) region of TixNiyCuz (58.3%≤x ≤ 58.5%, 26.5%≤y ≤ 27%, 14.8%≤z ≤ 15.3%, x + y + z = 100%) SMA composition is predicted. Our study not only provides a potential selection of narrow Thys TiNi-based SMAs but also indicates combining of XGBoost and DFT calculation is an effective strategy for materials design.