Machine Learning Method to Fastly Screen Topological Magnetic Materials

Abstract

The setting up of the magnetic topological materials database provides the possibility of exploring in this field through the machine learning method. Here, we trained a support vector classification (SVC) model to predict the topological state of the magnetic materials. There are 551 of materials in the topological magnetic materials database, and for each material, the topological state is different for different U values as the date are the DFT results. So, different U values should be different data and finally, 1319 of data are used for the model training. Although the strongly imbalance of the data (only 87 of data are topological insulators (TIs)), the trained model still can reach high precision (0.99) for trivial materials, which means that almost all the un-trivial magnetic materials can be screened by our model. Then, the SVC model is used to predict the topological state of the magnetic materials in MAGNDATA database and select 65 materials as TIs and 15 materials as enforced semimetals (ESs) for all the four times prediction. Topological quantum chemistry method then used to test the topological state of the randomly selected 11 of the un-trivial materials and indicate that 5 of them are ES and 1 is TI. Then first principle method is applied to further inspect the topological state of the only TI. The atom unfilled orbitals and valence electrons are deemed as import features by our model, which is in consistence with the physical analysis results. This SVC model could help screen topological materials from large amount of magnetic materials.