Quantitative Structure-Activity Relationship Of Dye-Sensitized Solar Cells Based On Deep Learning

Qye sensitizers are responsible for light collection and influence many key electron transfer processes of photovoltaic performance in dye-sensitized solar cells（DSSCs）.The quality of dye molecules affects the photovoltaic performance parameters of DSSCs,such as,photoelectric conversion efficiency（PCE）,short-circuit current density(J_sc)and open-circuit voltage(V_oc),etc.In order to further improve the photovoltaic performance of DSSCs,it is necessary to design and synthesize new efficient organic dyes.Quantitative structure-property relationship（QSPR）is used to study the relationship between the molecular structure of dye sensitizers and photovoltaic performance parameters,which can make up for the shortage of high experimental cost and long period to the maximum extent.Machine learning（ML）and deep neural networks（DNN）have emerged as a promising data-driven approach,which use self-learning and recognition of regularities and correlations in existing knowledge to predict the properties of different classes of dyes and determine optimal values to achieve the highest photovoltaic performance of DSSCs.In this paper,based on machine learning and deep neural network algorithm,the quantitative structure-activity relationship between different organic dyes and DSSCs photovoltaic performance parameters was studied.Firstly,partial least squares（PLS）,gradient lifting regression tree（GBRT）,random forest（RF）,support vector regression（SVR）and deep neural network algorithms were used to establish QSPR models between the molecular structure and photoelectric conversion efficiency of 229 triphenylamine DSSCs.The predictive performance of the model was evaluated by external test set validation and five-fold cross validation.The validation results show that the established models have high predictive ability and stability.Compared with other models mentioned above,DNN model has better performance;higher accuracy;better fitting and generalization ability.The importance of GBRT features indicates that the presence of Sdss C;F05[N-N];F06[N-O]descriptors will reduce the PCE properties of dyes and the presence of n N（CO）₂descriptor is beneficial to PCE values.The PCE value of dyes containing C-043 and C-038 fragments will be reduced.Secondly,RF;SVR and DNN algorithms were used to establish QSPR models between the molecular structure and photoelectric conversion efficiency of 281 porphyrin DSSCs.The predictive performance of the model was evaluated by external test set validation and five-fold cross validation.The validation results show that the established models have high predictive ability and stability.Compared with other models mentioned above,RF model has better performance and higher accuracy.The significance of model features indicates that the t statistics of F07[C-X];B05[N-S]and F10[N-O]are all very large and the t value is positive.This descriptor has made a positive contribution to PCE.The t value of Sds CH is negative and this descriptor makes a negative contribution to PCE.Therefore,fragments with a positive impact on PCE value should appear as far as possible in the design,while fragments with a negative impact on PCE value should be avoided as far as possible in the design to improve the photoelectric conversion efficiency of porphyrin dye sensitizers.Finally,the QSPR model of 1178 organic dye sensitizers and V_oc*J_scwas established by DNN algorithm by combining structure,quantum and experiment molecular descriptors.The established QSPR model successfully predicted the V_oc*J_scvalue of many kinds of DSSCs.The above results indicate that it is possible to introduce ML and DNN methods into the study of the relationship between the quantitative structural properties of the photovoltaic performance parameters of organic dye DSSCs.Meanwhile,DNN algorithm also shows its strong prediction ability,which provides an effective tool for predicting the photovoltaic performance parameters of dye molecules,thus saving the cost for the experimental synthesis.