A Cascaded Model For Prediction Of Overall Power Conversion Efficiency Of All-Organic Dye-Sensitized Solar Cells

Dye-sensitized solar cells(DSSC) are the most promising low cost solar cell, which may be industrialized in the near future. The photovoltaic parameters include short-circuit photocurrent density(JSC), open circuit voltage(VOC), fill factor(FF) and overall power conversion efficiency(PCE). Among these parameters PCE is the most important one for the solar cell performance evaluation, so accurate prediction of PCE has great significance to improving the performance of dye sensitized solar cell. Besides, the organic dye molecules used in DSSCs have critical influence to PCE. In past two decades, quantum chemistry methods have made significant progress in basic theories and programs. It is based on first-principles in physics and they can explain the experiments, reaction mechanisms, and even predict new molecular properties before they are synthesized. However, the solar cell device too complicated to get the value of PCE directly from the molecular structures using quantum chemical methods. Machine learning methods can bypass complex reacting processes and find the effects of molecular structures and properties on these observed parameters directly. Thus, after using quantum chemical methods to calculate the physicochemical properties of molecules, machine learning methods can build the quantitative relationship between the molecules and the experimental values of PCE.Therefore, the main work is using the quantum chemical methods combined with machine learning methods to build quantitative structure-activity relationship(QSAR) between organic dye molecules and PCE. A density functional method, B3 LYP, with either STO-3G or 6-31G* is used to optimize 354 molecules’ structures and calculate their properties in both ground and excited state. After that, the support vector machine(SVM) in combination with six feature selection methods(multiple linear regression, genetic algorithms, mean impact value, forward selection, backward elimination and +n-m algorithm) is used to construct two kinds of regression models: a non-cascaded regression model and a cascaded regression model. Among them, the cascaded model is a two-level regression network: the inputs in the first level are molecular descriptors and outputs are the JSC, VOC and FF; the inputs of the second level are the predicting outputs of the first level and the ultimate end-point is the PCE. The experimental results show that the cascaded model is significantly superior to the non-cascaded model in the ability of predictivity, goodness-of-fit and stability. The best established cascaded model predicts the PCE of DSSCs with a mean absolute error(MAE) of 0.57(%), which is only about 10% of the mean value of PCE.This study demonstrates that the established cascaded model is able to effectively predict the PCE, and it maybe provides a useful tool for the PCE prediction and design of new dye sensitizers.