Comprehensive Machine Learning Pipeline for Prediction of Power Conversion Efficiency in Perovskite Solar Cells

Abstract Perovskite solar cells (PSCs) have garnered considerable interest as a viable replacement for conventional silicon‐based solar cells, thanks to their high‐power conversion efficiency (PCE), low manufacturing costs, and ease of adjusting optoelectronic properties. To expedite the development of PSC devices, it is crucial to establish seamless connections between relevant parameters, thereby saving time and reducing the costs associated with actual experimentation. This capability can be achieved through advanced machine learning (ML) models, which are at the forefront of providing such interrelation capabilities. In this investigation, a comprehensive ML pipeline using scientific data from PSCs is established, which includes data processing methods and synthetic data generation. Various ML models such as linear, logic‐tree‐based, gradient‐based, discriminative‐based, and attention‐based neural network regressors are utilized to predict PCE of PSCs. The results indicate that the XG‐Boost ML algorithm displays a minimum mean absolute error (MAE) of 1.52, while the attention‐based neural network model TabNet also demonstrates an MAE as low as 1.62. Furthermore, the coefficient of determination for these models is as high as 0.919 and 0.915, respectively. The findings and developed models can aid in the study of training and evaluation processes and provide insights into correlations within perovskite data for required property prediction.