Fullerene-based organic thin-film solar cells (OSCs) with fullerene derivatives (PCBMs) as acceptor materials have been reported to provide a balanced short-circuit current density by appropriately selecting donor materials with different band gaps, as fullerene derivatives contribute little to the absorption of the mixed film.

However, due to the narrow absorption band of PCBMs, which cannot compensate for the broad region of the solar spectrum, resulting in a small photocurrent, fullerene-based OSCs have the disadvantage that they are largely dependent on the absorbance of the electron donor material. The search for donor materials is therefore key to producing high photoconversion efficiency. Therefore, an attempt was made to search for donor molecular materials for fullerene-based organic solar cells using machine learning, which allows prioritization of models for the construction of a dataset.

The acceptor material was a fullerene derivative (PCBM), while the donor material was selected based on the experimentally obtained photoconversion efficiency PCE (%), open circuit voltage V_OC, short circuit current density J_SC, curve factors FF, HOMO and LUMO, and 39 donor molecules with total energy data sets are used [1]. In this study, we used alvaDesc to perform molecular descriptor and fingerprint calculations, principal component analysis, correlation analysis and t-SNE analysis, alvaModel to create models and alvaBuilder to generate new molecules with specified properties from an arbitrary training set. First, calculations were performed with 5666 molecular descriptors and machine learning was performed using the language python. The molecular descriptors were used as explanatory variables and PCE, V_OC, J_SC and FF as objective variables, and features were predicted using Random Forest, followed by principal component analysis, correlation analysis and t-SNE analysis. The 39 donor molecules were divided into training and test data, and a regression model was created by machine learning with J_SC as the objective variable. Then, focusing on the structure of donor molecules with good J_SC values obtained from the experiments, new molecules were generated using the 39 donor molecules as the training set. The generated new donor molecules were evaluated by performing molecular descriptor and fingerprinting calculations and creating new regression models.