Machine Learning-Assisted Screening of Hexagonal 2D van der Waals Bilayers for Photocatalysis

From the photoelectric effect to the development of contemporary photocatalysts, the hunt for efficient energy conversion has been a tale of invention and discovery. Since 1972, water-splitting photocatalysis has evolved as an effective strategy for advancing toward sustainable energy. The van der Waals (vdW) heterostructures, formed by the vertical stacking of two distinct 2D materials, exhibit unique electronic properties. The weak vdW interactions tend to provide the benefit of efficient carrier separation, thus making them promising candidates for photocatalysis. However, analyzing all possible combinations of 2D materials is impractical through traditional approaches, necessitating the development of predictive models to automate and accelerate the quest. Herein, we propose a hybrid approach using machine learning (ML) in conjugation with first-principles calculations to predict the properties of hexagonal vdW bilayers for application in photocatalysis. Our ML workflow comprises the following major steps: (1) constructing a vast material space of bilayers and their descriptors using a 2D material database, (2) labeling a diverse set of bilayers using density functional theory (DFT) calculations, (3) training the supervised ML models on a labeled data set for binding energy, interlayer distance, band gap, work function, and band edges of heterostructures, (4) evaluating the performance of models on the validation set, and (5) predicting the properties of the unlabeled data set and screening the bilayers feasible for overall water-splitting (OWS) photocatalysis. The computational framework presented here tends to establish the relationship between 2D monolayers and vdW bilayers. Our findings highlight the potential of this approach in accelerating the search for novel photocatalysts by efficiently and accurately predicting their properties, thereby contributing to the broader goal of sustainable energy production.