Elemental Frequency-Based Supervised Classification Approach for the Search of Novel Topological Materials

Abstract

The machine learning based approaches efficiently solve the goal of searching the best materials candidate for the targeted properties. The search for topological materials using traditional first-principles and symmetry-based methods often requires lots of computing power or is limited by the crystalline symmetries. In this study, we present frequency-based statistical descriptors for machine learning-driven topological material's classification that is independent of crystallographic symmetry of wave functions. This approach predicts the topological nature of a material based on its chemical formula. With a balanced dataset of 3910 materials, we have achieved classification accuracies of 82\% with the Support Vector Machine (SVM) model and 83\% with the Random Forest (RF) model, where both models have trained on common frequency based features. We have verified the performances of the models using $5-fold$ cross-validation approach. Further, we have validated the models on a dataset of unseen binary compounds and have efficiently identified 22 common materials using both the models. Next, we implemented the $first-principles$ approach to confirm the topological nature of these predicted materials and found the topological signatures of Dirac, Weyl, and nodal-line semimetallic phases. Therefore, we have demonstrated that the implications of frequency-based descriptors is a practical and less complex way to find novel topological materials with certain physical post-processing filters. This approach lays the groundwork for scalable, data-driven topological property screening of complex materials.