Design guidance and band gap prediction of two-dimensional hybrid organic–inorganic perovskites by ensemble learning and graph convolutional neural networks

Abstract

Two-dimensional (2D) hybrid organic–inorganic perovskites (HOIPs) are promising materials for addressing the stability challenges in perovskite solar cells due to their exceptional environmental stability, exciton dynamics, and broad-band emission. However, research on the structure-directing role of organic cations in 2D HOIPs remains limited, and existing models for predicting their band gap lack sufficient accuracy. Here, we develop ensemble learning models and a site-attention-based graph convolutional neural network (SATGNN) to predict the dimensionality of lead iodide-based HOIPs and the band gap of 2D HOIPs, respectively. The ensemble learning models, leveraging molecular descriptors with eXtreme Gradient Boosting, achieve 88% cross-validation accuracy, and MaxAbsEStateIndex, Chi2n and Kappa2 are identified as critical features for dimensionality determination. The SATGNN model incorporates a convolution function tailored to the unique layered structure of 2D HOIPs and a novel site-attention mechanism to prioritize elemental contributions. The SATGNN significantly outperforms existing approaches by accurately capturing structural interactions and spatial configurations. Furthermore, the visualization of SATGNN confirms model's ability to identify structural features of 2D HOIPs and distinguishes the effects of different elemental types on material properties. By revealing interpretable molecular descriptors that govern 2D HOIP formation and integrating accurate band gap prediction, this two-stage framework offers both actionable design guidelines for organic cation selection and a scalable tool for the accelerated discovery of 2D HOIPs with targeted optoelectronic properties.