Machine Learning Accelerated Discovery of Mixed Anion Materials

Mixed-anion materials are interesting counterparts to their more widely studied single-anion compounds due to the increased flexibility of properties afforded by the presence of multiple anions. Here, we demonstrate how computational approaches, based on DFT datasets can be combined with materials informatics and machine learning (ML) models to accelerate materials discovery. We utilize a recently proposed improved crystal graph convolutional neural network (iCGCNN) model, and the Voronoi tessellation approach incorporated in the Materials-Agnostic Platform for Informatics and Exploration (MAGPIE). ML models are trained on several training datasets prior calculated in the Open Quantum Materials Database (OQMD) and evaluated on a same separate test set of 380 mixed-anion compounds. Surprisingly, the ML model trained on 3,460 unrelaxed mixed-anion compounds only dataset outperforms the other models with MAE 0.116 eV/atom. We therefore make predictions on a separate ~4,000 hypothetical mixed-anion compounds, which are subsequently validated by DFT calculations. We find 51 new (meta)stable mixed anion compounds using only 241 DFT calculations, a success rate of 21.2%, more than 3 times what is achieved in a typical high-throughput survey.