Predicting metallicity and magnetic ground states with machine learning and high-throughput DFT

Many applications require screening for magnetic materials with specific magnetic orderings, but such information is only sparsely available in existing databases. Here we focus on metallic materials, because their magnetic properties are often well described by density functional theory (DFT) without any empirical corrections. First, we develop a machine-learning (ML) framework for classifying materials as metals or insulators, which is trained entirely on an experimentally verified input sets of metals and insulators. The model employs a tree-based gradient boosting classifier (XGBoost) with descriptors derived from elemental property statistics and features available from Materials Project and the SMACT library [1]. The trained model demonstrates strong separation between metals and insulators and shows physically meaningful feature importance. Using this ML classifier to preselect metallic compounds, we then use a high-throughput DFT-based workflow to map total energies of collinear magnetic configurations onto the Heisenberg model and determine the collinear magnetic ground states. The performance of this workflow is evaluated by comparing with available experimental data.

[1] D. W. Davies et al., J. Open Source Softw. 4, 1361 (2019).