Machine learning based material selection: quaternary chalcogenides for energy applications

Abstract

Photovoltaics (PV) and Water Splitting (WS) are critical target areas for material development. Machine Learning (ML) is the only viable method for finding suitable candidates from the immense number of possible compounds. As a class of materials, Quaternary Chalcogenides appear quite promising for potential use in these applications. However, only a small subset of them have been studied. This work is the first ML-based material selection analysis for these materials, focusing on A2BCX4-type compounds. The band gap was the primary property targeted in the study due to its fundamental role in determining a material’s suitability for these applications. The band gap prediction performance achieved by ML models was competitive with those reported for other classes of compounds in the existing literature. The relationship between the total energy and space group of Quaternary Chalcogenides was also explored, including a proposed protocol for using our ML models for total energy to determine the space group given only the composition of the compound. In addition, the relationship between various elemental properties and band gap was investigated via SISSO. For this purpose, we developed our own implementation of this algorithm in Python. The heuristic formulae obtained allow computationally inexpensive estimation of the band gap. An extremely simple but versatile model was used to understand the band gap change Quaternary Chalcogenides and justify the SISSO models. The understanding gained via these formulae has potential applications in band gap tuning. These results were used to identify candidate compounds with possible use in 3 areas - PV Solar Cells, Photocatalytic WS, and Photoelectrochemical WS.