Machine learning of heavy rare earth-based low-temperature magnetocaloric materials under complex classification methods

Abstract

Low-temperature magnetic refrigeration materials have important application value in the fields of aerospace, military defense, medical and health care, and low-temperature physics. Compared with the “stir-fry” material R&D model, machine learning can accelerate the data analysis and R&D of new materials. However, at present, the research on machine learning is limited to the simulation and prediction of a single type of material, and the research object is mainly focused on materials with high phase transition temperature. Obviously, the generalizability of this kind of research is low. In order to solve the above scientific problems, this paper uses the gradient boosted regression tree (GBRT) model to train the dataset of heavy rare earth-based alloys under different classification methods, taking the chemical composition as the characteristic attributes, and the Curie temperature and maximum magnetic entropy as the target attributes. It was found that the performance of heavy rare earth-based alloys was different under different classification methods. The prediction of Curie temperature under each classification method is very good, but the prediction effect of magnetic entropy change with non-intrinsic properties is completely different. Among them, the prediction effect of magnetic entropy change of the whole dataset is poor，and the training results of the hexagonal structure and ternary alloys dataset are excellent. The correlation coefficient (R²) of the hexagonal structure on the training set and the test set is 0.99 and 0.77, respectively. The fitting results for the ternary alloys also reached 0.97 and 0.80, respectively. However, the orthogonal structure and binary structure datasets performed extremely poor, and the R² value of the test set was only 0.31 and 0.17, respectively. For the first time, we obtained the good prediction results of magnetocaloric effect materials under complex classification methods through the GBRT machine learning model, and revealed the influence of different classification methods on the target and feature attributes of the dataset from the physical level. These results are helpful to accelerate the development of low-temperature magnetocaloric materials.