Accelerated discovery of stable spinels in energy systems via machine learning

Discovery of new energy materials with thermal stability and special electro-optical properties has always been the goal and challenge of material science. As an important energy material, spinel has been widely used in the fields of photovoltaics, piezoelectric, catalysis, batteries, and thermoelectrics. However, there are many spinels with AB2X4 formula that have not been explored, especially for the ones with direct band gaps, which severely limit their applications. Here, we develop a target-driven method that uses machine learning (ML) to accelerate the ab initio predictions of unknown spinels from the periodic table of elements. Under this strategy, eight spinels with direct band gaps and thermal stabilities at room temperature are screened out successfully from 3880 unexplored spinels (CaAl2O4, CaGa2O4, SnGa2O4, CaAl2S4, CaGa2S4, CaAl2Se4, CaGa2Se4, CaAl2Te4). The screened spinels show good optoelectronic performance in the energy systems (thin-film solar cells, photocatalysts, etc.). Based on the XGBoost algorithm, a semiconductor classification model with strong structure-property relationship is established, with a high prediction accuracy of 91.2% and a low computational cost of a few milliseconds. The proposed target-driven approach shortens the research cycle of spinel screening by approximately 3.4 years and enables the discovery and design of a wide range of energy materials. Compared with traditional high-throughput material screening, the proposed method has potential applications in shortening the screening time and accelerating the development of material genomics.