Machine learning-assisted prediction of organic solar cell efficiency from device reflectance spectra

Organic solar cells (OSCs) have gained significant attention due to their cost-effectiveness, lightweight, and flexibility. However, traditional optimization methods are time-consuming and labor-intensive. This study introduces a novel approach to rapidly predict the power conversion efficiency (PCE) of OSCs by leveraging their device reflectance spectra and machine learning techniques. A dataset comprising 606 reflectance spectra and corresponding PCE values for six types of OSCs was constructed using the Finite-Difference Time-Domain method. Various preprocessing techniques (STD, SG, D1, D2) and feature extraction methods (PCA, CARS, SPA) were applied, followed by training with traditional machine learning models (PLSR, KNNR, RFR, SVR, ELM). The STD-CARS-RFR model achieved a prediction accuracy (R 2 ) of 0.987. Further improvements were made by integrating a multilayer perceptron regression (MLPR) model, raising the accuracy to 0.992. The introduction of an attention mechanism in the Attention-MLPR model further optimized the prediction, achieving an R 2 of 0.998 and a root mean square error of 0.253. These results demonstrate that combining reflectance spectra with machine learning techniques can efficiently predict OSC efficiency, offering a non-contact alternative to traditional photoelectric measurement methods and holding significant potential for practical applications.