Computational Modeling of Indoor Organic Photovoltaics: Dataset Curation, Predictive Analysis, and Machine Learning Approaches

This study presents a comprehensive dataset that encompasses the indoor device performance of organic photovoltaic (OPV) materials, their corresponding SMILES codes, and frontier molecular orbital (FMO) energy levels. This dataset comprises a total of 128 subsets and features 64 pairs of donors and acceptors. We demonstrate that traditional models, such as the Shockley–Queisser limit and Scharber’s model, are insufficient for accurately predicting the behavior of indoor OPVs based on the molecular orbitals of these materials. In contrast, we explore the predictive capabilities of four machine learning (ML) models for estimating the power conversion efficiencies (PCEs) of indoor OPVs, utilizing molecular structure information and FMO data from the dataset we compiled. The trained ML models exhibit strong predictive performance with high correlation coefficients (r > 0.8) for indoor PCE values; notably, the support vector regression (SVR) model achieves the highest r of 0.878. The generalization capabilities of the models are also assessed using previously unseen materials, and the results demonstrate high accuracy rates. The SVR algorithm reaches the best average accuracy of 92.1%, underscoring its potential for efficiently screening materials for indoor applications. Our findings suggest that this dataset, with opportunities for future expansion, could significantly facilitate material design and accelerate computer-aided materials screening, reducing the need for extensive experimental testing in the development of indoor OPVs.