A machine learning approach for in silico prediction of the photovoltaic properties of perovskite solar cells based on dopant-free hole-transport materials

Herein, we report a novel and powerful approach developed to lower the manufacturing costs of PSCs and make them powerful and cost-effective technology through identifying more efficient and low-cost dopant-free hole-transport materials (HTMs) for PSCs by avoiding the shotgun approach of synthesizing hundreds or thousands of HTMs, fabricate devices and characterize them to identify the best performing candidate(s). With these in mind, the primary objective of this study is to develop a machine learning model extracted from automated-quantitative structure-property relationship (autoQSPR) models that could accurately predict various photovoltaic properties, including power conversion efficiency (PCE), open-circuit photovoltage (VOC), and short-circuit photocurrent (JSC), of dopant-free HTM-based PSCs. High-efficacy AutoQSPR models capable of accurately predicting photovoltaic properties were developed by considering experimental photovoltaic property data sets from dopant-free HTMs used in the fabrication of PSCs with different architectures, e.g., conventional (n-i-p) and inverted (p-i-n), which are fabricated using methyl ammonium lead halides (MALHs) or mixed cation lead halides, were used to develop high efficacy autoQSPR models capable of accurate predictions of photovoltaic properties. The Schrodinger Suite was utilized to build autoQSPR models to predict PCE, VOC, and JSC utilizing different kinds of molecular descriptors such as 2D binary fingerprints, one dimensional (1D), two dimensional (2D), and three dimensional (3D). Notably, 2D binary fingerprint descriptors generate models that outperform 1D, 2D, and 3D molecular descriptors calculated using the well-known PaDEL calculator. The developed autoQSPR models were improved when the PSC configurations were considered and a significant predictive ability (test set Q2 > 0.5) was achieved for all autoQSPR models that involved the use of 2D binary fingerprint descriptors. The model confidence was validated by the utilization of dopant-free HTMs that were not present in the original dataset used to build the models. Furthermore, the most efficient models were used to propose potential HTM candidates to benefit the scientific community and scholars.