Deep-learning-enabled high-throughput Screening of MXene photocatalysts for hydrogen production

Abstract

The challenge of photocatalytic hydrogen production has motivated a targeted search for MXenes as a promising class of materials for this transformation because of their high mobility and high light absorption. High-throughput screening has been widely used to discover new materials, but the relatively high cost limits the chemical space for searching MXenes. We developed a deep-learning-enabled high-throughput screening approach that identified 14 stable candidates with suitable band alignment for water splitting from 23 857 MXenes. Through the deep learning framework utilizing the layered structure of two-dimensional materials trained on the C2DB, we predicted the properties, including formation energy, convex hull energy, and bandgap with mean absolute errors of 0.06 eV per atom, 0.06 eV per atom, and 0.14 eV, respectively. Through further density functional theory and non-adiabatic molecular dynamics calculations, we identified a series of descriptors that demonstrate the photocatalytic potential of the candidate MXenes in solar-to-hydrogen efficiency, light absorption, and carrier separation. Using symbolic regression, we proposed a descriptor that captures the relationship between the built-in electric field and the nonadiabatic electron–hole coupling. Notably, the presence of built-in electric fields in Janus MXenes can suppress recombination rates due to spontaneous separation of photogenerated carriers. Our work demonstrates an efficient computational strategy for inversely designing 2D photocatalysts and provides a strategy for designing high-performance photocatalysts for hydrogen production.