Machine Learning Robust Classifications of Topological Materials

Tens of thousands of candidate topologically non-trivial materials had been predicted by large-scale studies combining ab-initio calculations, symmetry indicators, elementary band representations, and machine learning approaches [1-5]. However, practical requirements for high-throughput density functional theory (DFT) searches and inherent methodological classification biases introduce significant inconsistencies, such as the incompatible classifications in the topological databases [1-3] across subsets of experimentally observed materials cataloged in the Inorganic Crystal Structure Database (ICSD). In this talk, we present our machine-learning framework for classifying the topological states of crystalline materials, and share the key details of our curated training datasets and predictions. We extensively evaluate the accuracy, stability, and generalization of our models, and discuss how they can be useful to evaluate the reliability of previous predictions of individual materials' classification and overall statistical patterns and biases. [1] https://doi.org/10.1126/science.abg9094 [2] https://doi.org/10.1126/science.adf8458 [3] https://doi.org/10.1038/s41586-019-0937-5 [4] https://doi.org/10.1103/PhysRevB.101.245117 [5] https://doi.org/10.1063/5.0055035