We present a highly efficient workflow for designing semiconductor structures with specific physical properties, which can be utilized for a range of applications, including photocatalytic water splitting. Our algorithm generates candidate structures composed of earth-abundant elements that exhibit optimal light-trapping, high efficiency in H2 and/or O2 production, and resistance to reduction and oxidation in aqueous media. To achieve this, we use an ionic translation model trained on the Inorganic Crystal Structure Database to predict over 30 000 undiscovered semiconductor compositions. These predictions are then screened for redox stability under hydrogen evolution reaction or oxygen evolution reaction conditions before generating thermodynamically stable crystal structures and calculating accurate bandgap values for the compounds. Our approach results in the identification of dozens of promising semiconductor candidates with ideal properties for artificial photosynthesis, offering significant advancement toward the conversion of sunlight into chemical fuels.
Read full abstract