Accelerated design for elastocaloric performance in NiTi-based alloys through machine learning

Abstract

NiTi-based shape memory alloys (SMAs) are regarded as one of the most promising materials for engineering applications of elastocaloric refrigeration. A critical mission is to efficiently explore the new NiTi-based SMAs with large adiabatic temperature change (ΔTad). We proposed a new material design method that combines highly correlated microscale physical information (volume change, ΔV) into machine learning to predict ΔTad of NiTi-based alloys. First, we tightly coupled machine learning with first-principles calculations to accelerate receiving lattice parameters before and after the phase transformation and predict the ΔV, which shows excellent performance with the coefficient of determination R2 > 0.99. Then, relevant features, including the ΔV, are considered to predict the ΔTad in NiTi-based SMAs. Moreover, due to the small dataset, the principal component analysis and the independent component analysis are added. We evaluate the performance of three machine learning models [Lasso regression, support vector regression, and decision tree regression (DTR)]. Finally, the DTR model exhibits a high accuracy for predicting ΔTad (R2 > 0.9). Introducing the feature of ΔV into the machine learning process can improve the accuracy and efficiency of model design. Further, this work paves the way to accelerate the discovery of new excellent materials for practical applications of elastocaloric refrigeration.