Classification and Prediction of Skyrmion Material Based on Machine Learning.

ABSTRACT

The discovery and study of skyrmion materials play an important role in basic frontier physics research and future information technology. The database of 196 materials, including 64 skyrmions, was established and predicted based on machine learning. A variety of intrinsic features are classified to optimize the model, and more than a dozen methods had been used to estimate the existence of skyrmion in magnetic materials, such as support vector machines, k-nearest neighbor, and ensembles of trees. It is found that magnetic materials can be more accurately divided into skyrmion and non-skyrmion classes by using the classification of electronic layer. Note that the rare earths are the key elements affecting the production of skyrmion. The accuracy and reliability of random undersampling bagged trees were 87.5% and 0.89, respectively, which have the potential to build a reliable machine learning model from small data. The existence of skyrmions in LaBaMnO is predicted by the trained model and verified by micromagnetic theory and experiments.