Abstract:

Magnetocaloric performance is of vital importance for Mn-Fe-P-Si alloys. However, when processes and compositions are considered, designing alloys with large magnetic entropy changes (ΔS_m), low thermal hysteresis (ΔT_hys), and Curie temperatures (T_C) around room temperature become relatively complicated. In this study, we adopt machine learning methods to predict the magnetocaloric performance of Mn-Fe-P-Si compounds for the first time. To achieve this goal, 503, 465, and 660 data points for datasets with T_C, ΔT_hys, and ΔS_m are collected, respectively. The collected datasets contain parameters of compositions, preparations, heat treatment, and magnetic field changes. We search for the optimal configuration using various methods and also compare their mean squared errors (MSE) and allowable errors. Evaluation results show that the performance of neural networks (NNs) is better than other methods. Therefore, we select NN to explore the T_C, ΔT_hys, and ΔS_m values as a function of Mn, Si, metal/non-metal ratios, and B (Boron). We also propose to use the composition window with excellent magnetocaloric performance. These results not only help us gain deep insights into Mn-Fe-P-Si alloys but also accelerate the design process of alloys suitable for magnetocaloric materials. This work has the potential to solve the challenges and boost the research of Mn-Fe-P-Si alloys.

Key words: Mn-Fe-P-Si alloy, Material design, Magnetocaloric performance, Machine learning, Neural network