2025 article
Predicting 3D magnetic topological insulators and semimetals with machine learning
Boulton, J. A., & Kim, K. W. (2025, August 22). Journal of Applied Physics.
open access via pubs.aip.org (publisher PDF)Machine learning (ML) models are used to identify topologically non-trivial magnetic insulators and semimetals among a group of candidate materials. The search for these topological materials is critical in the development of spintronic devices due to their strong magnetoelectric responses enabling low-power operations. In a conventional approach, such a task usually requires an extensive use of time-intensive density functional theory (DFT) calculations combined with a Hamiltonian model using maximally localized Wannier functions (MLWFs). It is hypothesized that a ML model may make the process more efficient. As a specific example, our investigation focuses on AxByCz compounds with A = Ni, V, Co, or Eu; B = Bi or Sb; and C = Te or Se, which have received much attention recently. Of the three ML models trained (Decision Tree, Random Forest, and XGBoost), XGBoost appears to provide the highest level of accuracy in the three relevant tasks (e.g., 87.6% in metal vs non-metal; 95.7% in trivial vs topological insulators; 84.1% in trivial vs topological metals). Subsequent application of these models to the AxByCz compounds leads to the identification of seven topological insulators (verified with DFT and MLWFs). On the other hand, only one of the predicted topological semimetal candidates turns out to have the desired properties. This challenge in identifying the topological semimetals appears to be due to the sensitive dependence on the Fermi level position, which is rather difficult to predict a priori. Nevertheless, the results illustrate that the ML models can indeed significantly enhance the efficiency in the search for magnetic topological materials.