Optimizing Photoelectrochemical Photovoltage and Stability of Molecular Interlayer-Modified Halide Perovskite in Water: Insights from Interpretable Machine Learning and Symbolic Regression

Interpretable machine learning models are desired for materials and chemical design processes, while the stable optoelectronic properties of the halide perovskite materials in hostile conditions such as in water are prerequisites for their wider industrial deployment. In this study, we demonstrate an experimentally verified interpretable machine learning pipeline coupled with a symbolic regression method to optimize and understand the stability and photovoltage of the molecular interfacial layer-modified perovskite film in aqueous solution. An accurate machine learning model is achieved via the random forest algorithm; this leads to the successful experimental validation of a champion CH₃NH₃PbI₃/bimolecule/TiO₂ system giving a stable photovoltage output in water with two interlayer molecules, basic violet 7 and rhodamine B. The experimental stability and photovoltage outputs of the machine learning-predicted sample in the aqueous solution are enhanced by 4 and 16.7% compared with those of the untreated CH₃NH₃PbI₃ film under the same aqueous conditions; the atomic adsorption structures are then investigated via density functional theory calculations. A t-SR symbolic regression method is developed to design relevant molecular descriptors, which employs a traversal algorithm comparing 8.95 million expressions and obtains a descriptor with 18% improvement. The present study provides a machine learning platform to accelerate the design of stable and high-performance perovskite films in extreme conditions, and the method can be elaborated to other surface systems.