Predicting power conversion efficiency of dyes in dye-sensitised solar cells

Accurate prediction of power conversion efficiency (PCE) is essential for designing high-performance dye-sensitised solar cells (DSSCs). Current quantitative structure–property relationship (QSPR) models are often constrained by small datasets or narrow dye-class applicability. This study addresses these limitations by constructing a global random forest (RF) classification model for 1,186 diverse organic dyes across seven structural classes (coumarins, carbazoles, indolines, phenothiazines, porphyrins, triphenylamines, and diphenylamines). Employing a PCE threshold of 4.27% (Class 1: high-efficiency; Class 0: low-efficiency), the optimised RF model (29 descriptors, 300 trees, 5 mtry) achieved exceptional performance, with 99.68% accuracy on training data and 83.12% on independent test samples. Mechanistic analysis revealed that PCE improves with increased charge delocalisation (higher MATS4s/3s), enhanced dipole alignment (larger Eig05_AEA(dm)), optimal heavy atom periodicity (GATS8 m at 8-bond intervals), and strong donor–acceptor interactions (elevated B02[C–S]), while being limited by charge localisation (higher JGI8/JGI9) and excessive π-conjugation (increased SpMax2_Bh(s)). This global classification model predicts PCE across dye classes, broadening the design space for DSSCs.