Machine learning study on magnetic structure of rare earth based magnetic materials

Abstract

Machine learning is playing an increasingly important role in discovery and design of new materials. In this work, 11 machine learning algorithms were trained to predict the material magnetic structure. Material composition and crystal structure are used to classify the dataset, and the relationship between multi-feature variables is constructed in a small sample space. The prediction accuracy of all models is above 0.73. Compared with non-decision tree models, optimized decision tree algorithms such as Gradient Boosting have greater advantages in binary classification. Neural Network has the best performance in predicting skyrmion structure, with accuracy and reliability of 0.93 and 97 %, respectively. Rare earth elements have a great influence on the material magnetic structure, and their proportion is negatively correlated with the generation of nonlinear or skyrmion structures. The material is more prone to nonlinear magnetic structure when the space group belongs to the cubic and the hexagonal crystal systems. Based on the Neural Network, the magnetic structures of several rare earth oxides are predicted. The skyrmion in SrR_xFe_12-x-yMg_yO₁₉ and La_xBa_1-xMnO₃ was observed by Neutron Powder Diffraction and magnetic force microscope, which effectively verified the model accuracy. This work provides new perspectives for machine learning in the discovery of nonlinear magnetic structures and rapid design of material compositions.