Data-Driven Approach to Accelerate the Design of Halide Perovskite for Photovoltaic Application Using Electronic Properties As Descriptors

Perovskite solar cells (PSCs) have garnered remarkable attention for their efficiency and tunable optoelectronic properties. However, their instability to moisture and heat poses challenges. Traditional trial-and-error approaches for finding stable halide perovskites are inefficient due to the vast compositional possibilities within ABX3 perovskite structure. This study uses machine learning (ML) to predict bandgaps and photovoltaic parameters of PSCs, using a dataset of 447 data points containing chemical composition, bandgaps, and photovoltaic parameters. Various ML models including support vector regressor, random forest, gradient boost regressor, XGBoost, extratree Regressor, and AdaBoost Regressor have been used herein. Positional average elemental property (PAEP) approach is introduced to featurize the data. As ABX3 perovskite involves distinct A, B, and X sites, the proposed PAEP model captures site-specific effects, enhancing model accuracy. The best model exhibits impressive r-values of 0.98 for bandgap prediction and 0.86 for power conversion efficiency of PSCs. Elemental properties of B and X sites, such as ionization energy, electron affinity, and electronegativity, are found to be crucial features in the analysis by Shapley additive explanation. This study underscores the potential of ML in designing novel, stable, and efficient PSCs, offering a more efficient alternative to conventional trial-and-error methods. This study employs machine learning (ML) to design novel halide perovskites, using the positional average elemental property approach for featurization. Various ML models are trained, and the best-performing model achieves impressive accuracy. Shapley additive explanation is used to analyze the best trained model and a set of plausible compositions is predicted.image (c) 2024 WILEY-VCH GmbH