Study On Prediction Models For Overall Power Conversion Efficiency Of Organic Solar Cells Based On Ensemble Learning

Solar cells can directly convert solar energy into electricity,which is the most effective way to utilize solar energy.Organic solar cells have attracted much attention because of their low cost,light weight and large area flexible devices.Among them,Power Conversion Efficiency(PCE)is a crucial parameter for evaluating the performance of Organic Solar Cells(OSCs),and the accuracy of its prediction directly affects the performance of Solar Cells.However,the relationship between molecular structure properties and PCE is complicated due to the complex power conversion process.The overall power conversion efficiency of organic solar cells is difficult to obtain by either calculations or experiments.As a branch of machine learning,ensemble learning can effectively bypass the complex experimental process and directly construct the Quantitative structure-activity Relationship(QSAR)between molecular Structure properties and photoelectric conversion efficiency of solar cells,which breaks the bottleneck of weak learners and is usually significantly more accurate than base learners and is usually significantly more accurate than base learners.Therefore,in order to achieve high accurate models,various ensemble learning methods are investigated.On the one hand,from the perspective of global modeling,three types of global ensemble models is constructed,including GBDT of Boosting and RF of Bagging in the way of homogeneous ensemble,as well as the SVM-KNN-WMA of heterogeneous ensemble.On the other hand,this paper studies a scheme of "clustering first,modeling later",and establishes a local heterogeneous ensemble model L-SVM-KNN-WMA.The global heterogeneous ensemble SVM-KNN-WMA uses weighted majority algorithm(WMA)with combination of regression support vector machine(SVM),K nearest neighbor(KNN),which outperformed the best single base learner,a support vector machine model,and it has better generalization ability than the other two ensemble methods GBDT and RF for QSAR models.According to the structural similarity of molecules,the local heterogeneous ensemble model L-SVM-KNN-WMA established by applying k-means clustering method to the training set achieved high accuracy and generalization.This study shows that the QSAR model based on ensemble learning to construct molecular structure properties and photoelectric conversion efficiency can predict new organic solar materials with high PCE,which solves the problem that traditional quantum chemical calculation methods consume a lot of computational resources,reduces the experimental cost and the experimental time.It is of great significance for practical application in the future.