Study Of Interfacial Molecular Stacking Patterns And Machine Learning For Organic Photovoltaics

With the massive consumption of fossil energy,the accompanying pollution of the environment,and the reserves of fossil energy are declining sharply while the cost of collection is gradually increasing.Therefore,the energy crisis of the 21st century,caused by energy shortage and environmental pollution,has led to the urgent need to find a new energy source with huge reserves and clean and no pollution,which is a great challenge for mankind.Organic solar cells（OSCs）are favored for their advantages such as light weight,flexibility,non-toxicity,non-pollution,abundant material sources,easy of synthesis and other merits.Since the active layer material has a decisive influence on the photovoltaic performance as a key component of OSCs devices,they have received a lot of attention from researchers.Firstly,in order to investigate the effect of different molecular stacking patterns in bulk heterojunctions on the photovoltaic performance,D18 and Y6 were selected as the donor and acceptor materials to investigate the stacking model at the heterojunction interface by quantum chemical calculations.Secondly,since the intrinsic relationships between OSCs device performance are very complex,machine learning（ML）methods are used to explore them in order to more accurately describe the intrinsic connections between device parameters.The trained ML models were trained on the basis of a database with 379 groups of donor-acceptor pairs,and the trained ML models were used for OSC device performance measurement and new molecule design.The main research of this thesis is as follows.Different intermolecular stacking patterns lead to different heterojunction morphologies,which in turn have an important impact on the photovoltaic performance of OSCs.Here,in order to investigate how molecular stacking pattern in OSCs heterojunction impact on electronic structures and properties,and then influence photovoltaic performance,this work selected D18:Y6 OSC as a representative system.On the basis of semi-empirical quantum chemistry and density functional theory calculations,we studied the geometries,electronic structures and excitation properties,as well as electron and hole couplings of D18and Y6 molecules,Y6 dimers,D18/Y6 and D18/Y6-dimer complexes which are heterojunction interface models.The results indicate that the molecular stacking effects on open-circuit voltage(V_OC)can be attributed to the influences on charge transfer（CT）excitation energies because molecular stacking modes at donor/acceptor heterojunction interface can significantly change excitation energies.Furthermore,the molecular stacking modes can cause drastically different electron and hole couplings,and then affect charge transfer/transport rates.Hence,the molecular stacking effects on short-circuit current density(J_SC)can be understood from the effects on electron and hole couplings.In the development of OSCs,machine learning（ML）is an effective method that can enable molecular prediction as well as molecular design.Firstly a database was created in this work,which was composed of 397 groups of donor-acceptor pairs selected from 193 published experimental articles.In the input data of ML,since different molecular descriptors have different effects on ML prediction accuracy,the selection of suitable molecular descriptors becomes crucial.Molecular fingerprint,as a kind of descriptor,can provide an accurate and effective description of the structure,properties,substructure and other information of molecules by binary encoding.In order to reveal the effect of Morgan fingerprints of 1024 bits of the donor and acceptor as well as the effects of root mean square roughness（RMS）and donor-acceptor weight ratio（D/A ratio）on the predictive performance of the photovoltaic device parameters,four ML algorithms were selected under supervised learning,namely random forest（RF）,adaptive boosting（Adaboost）,extra trees regressor（ETR）,gradient boosting regressor trees（GBRT）.Also,in order to effectively evaluate the prediction performance of the selected ML model,four different evaluation metrics were selected and evaluated by scoring.Among the evaluation metrics are the coefficient of determination（r²）,Pearson correlation coefficient（r）,root mean square error（RMSE）,and mean absolute error（MAE）.The evaluation results showed that the RF model exhibited the highest prediction accuracy,with the coefficient of determination r² reaching 0.47 and the Pearson correlation coefficient r reaching 0.68.Meanwhile,11 new acceptor molecules were designed,and the device performance parameters were predicted by RF model for 11 of them.The predicted results showed that two of the molecular structures,A1 and A2,showed better potential.The design and performance prediction of high-performance molecules by ML can effectively accelerate the molecular prediction and molecular design of high-performance donor and acceptor materials.