Explainable optimized 3D-MoRSE descriptors for the power conversion efficiency prediction of molecular passivated perovskite solar cells through machine learning

Interface molecular passivation is widely utilized to improve the performance and stability of perovskite solar cells (PSCs). However, designing efficient passivation molecules is still challenging. Owing to the fast development of machine learning (ML) methods, screening passivation molecules from molecular libraries with ML models becomes promising. Herein, 3D-MoRSE descriptor sets were introduced to predict the device power conversion efficiency (PCE) by machine learning with automatic relevance determination regression. By fine tuning the scale factor (sL), we found that sL from 0.04 to 0.50 could achieve satisfactory prediction results, and convergence is achieved at s x sL approximate to 1.40. Among all investigated atomic properties, atomic electronegativity and ionization potential revealed a strong correlation with PCE. We identified that molecules with abundant carbon-nitride single or partial-double bonds may achieve good surface passivation and realize high PCE. The highest coefficient of determination (R2) of 0.87 was achieved, demonstrating an improvement of approximately 0.11 compared to existing models. Using our optimal models, we predicted the PCEs of the devices with passivated molecules from the PubChem database, and three molecules in the top 15 candidates show good passivation ability in reported experiments. This work provides a simple and efficient method for molecule description for highly accurate ML predictions, which could accelerate the discovery of new molecules for PSC passivation. The 3D-MoRSE descriptor is optimized and introduced to predict the device power conversion efficiency of perovskite solar cells by machine learning.