Open‐circuit voltage loss and dielectric constants as new descriptors in machine learning study on organic photovoltaics

• Published in: International journal of quantum chemistry 2023-03, Vol.123 (5), p.n/a
• Main Authors: Yang, Bing, Zhang, Cai‐Rong, Wang, Yu, Zhao, Miao, Yu, Hai‐Yuan, Liu, Zi‐Jiang, Liu, Xiao‐Meng, Chen, Yu‐Hong, Wu, You‐Zhi, Chen, Hong‐Shan
• Format: Article
• Language: English
• Subjects: Algorithms; Chemistry; Circuits; Correlation coefficients; dielectric constant; Electric potential; Machine learning; molecular descriptors; open‐circuit voltage loss; organic photovoltaics; Parameters; Permittivity; Photovoltaic cells; Physical chemistry; Quantum physics; Voltage
• ISSN: 0020-7608; 1097-461X
• DOI: 10.1002/qua.27039

Molecular descriptors are critical for determining the accuracy of machine learning (ML) study on organic photovoltaics (OPV). To unravel the complex relationship between molecular properties and device performance, on the basis of 510 donor‐acceptor pairs in OPV active layer, the open‐circuit voltage loss (VOC‐loss), dielectric constants of donor and acceptor (ε‐D and ε‐A) were firstly implemented into property descriptor set that includes 41 quantities totally. Then, the five ML algorithms were applied to compare the property descriptor sets with and without VOC‐loss, ε‐D and ε‐A (coded as new and old sets) in the prediction of photovoltaic parameters. The ML results of Pearson's correlation coefficient and the slope of regression lines indicate the performances of new molecular descriptor set are prevailing to that of old set. Furthermore, the Gini important analysis indicates that the ε‐D, ε‐A and VOC‐loss are very important parameters for determining device performance. Higher dielectric constants and lower VOC‐loss will be more beneficial to the performance of OPV devices. The open‐circuit voltage loss (VOC‐loss), dielectric constants of donor and acceptor materials (ε‐D and ε‐A, respectively) were firstly implemented into molecular property descriptor set (MPDS). The machine‐learning algorithms random forest, extra trees regressor, gradient boosting regression tree, adaptive boosting and extreme gradient boosting were applied to compare the MPDS with and without VOC‐loss, ε‐D and ε‐A in photovoltaic parameters prediction.