Machine Learning Assisted Design of Type-II Two-Dimensional Heterostructures for Photocatalytic Water Splitting

Abstract

Type-II two-dimensional (2D) heterostructures are promising for photocatalytic water splitting but face exploration challenges due to high experimental/computational costs. Here, we propose an efficient data-driven approach for the rapid discovery of type-II van der Waals heterostructures (vdWHs) without the need for preoptimization of structures or precise stacking information. To meet this end, a specially designed matrix descriptor is developed to capture the important interlayer interactions. Coupled with a one-dimensional convolutional neural network, this descriptor can well describe weak interlayer interactions in heterostructures, allowing direct prediction of bandgap and band edge positions of arbitrary 2D heterostructures. 800 potential candidates are successfully screened out of nearly 10⁵ heterostructures for type-II vdWHs, and further comprehensive band structure and optical absorption spectra calculations reveal the potential of WS₂/Rh₂Br₆ and Al₂S₂/PtS₂ as water splitting photocatalysts. This work provides a data-driven approach to energy materials discovery and offers a cost-effective alternative to traditional methods.