Machine Learning Assisted Evaluation And Design Of Organic Photovoltaic Materials

Due to limited fossil fuel reserves and increasingly severe environmental problems,the demand for renewable energy has increased significantly in the past few decades.The efficient use of solar energy is expected to become a key breakthrough in solving energy and environmental problems.Among the many forms of solar energy utilization,solar cells provide an economical way to directly convert solar energy into electrical energy.Traditional silicon-based solar cells require complex processes and expensive cost in manufacturing,and the rigidity limits their application in wearable and flexible devices.Among the new generation of solar cells,organic photovoltaics have the advantages of light weight,flexibility,and can be fabricated by a continuous solution method,attracting widespread attention.In the field of organic photovoltaics,researchers focus on developing novel organic molecules with high power conversion efficiency.However,the efficiency of designing new molecules has been limited by the conventional way in material discovery,i.e.the trial-and-error method.This method requires significant amount of time and resources,leading to a tedious and time-consuming process to design new materials for organic photovoltaics.Since the first report of organic photovoltaics in 1973,less than 2000 donor materials have been tested in solar cells.In addition,at present the design of new materials still needs to be carried out according to previous experience,leading to a certain blindness.The molecular structure and important units related to high performance are not clear.To improve the efficiency of material development and make full use of the existing experimental data,we use machine/deep learning algorithms to establish the relationship between the chemical structures of the donor materials and the power conversion efficiencies,thus realizing the rapid evaluation of new materials.Meanwhile,the machine learning method is in conjunction with experiment,assisting or even directly guiding the design of new molecules.In previous studies,high-throughput and high-precision first-principles calculations that need expensive calculation cost were employed to obtain the microscopic properties of a material as the input of the machine learning model.In chapter two,we explore the feasibility of using images of chemical structures to establish a structure-property relationship.Deep learning has been proved to be able to rapid evaluate organic photovoltaic materials.The model in this paper allows direct use of images of chemical structures as input,exhibiting excellent nonlinear analysis capabilities and low demand for computing power.After training the binary classification model with the Harvard Clean Energy Project Database,a prediction accuracy of 91.02% can be reached by the optimal model on an independent verification dataset.By visualizing the data processing process of the model,it is found that deep learning can extract features in the chemical structures,such as specific elements and chemical bonds.The virtual molecules in the Harvard Clean Energy Project Database have not been synthesized by experiments,and their molecular structures are simpler than that of the molecules reported in the literatures.In chapter three,we collect the published literatures in the field of organic photovoltaic,and record the structures of the donor materials as well as the photovoltaic parameters of the corresponding devices to establish an organic photovoltaics donor materials database.In order to provide some guidance and suggestions when applying machine learning methods in material discovery,especially using small database,we studied the influence of programming language expression for molecule structures(images,ASCII strings,descriptors and fingerprints)on the performance of models.It is found fingerprints with length over 1000 bits can obtain high prediction accuracy.The easy availability makes them a suitable form to be used as the input of machine learning model.In addition,the reliability of machine learning is further verified by screening 10 newly-designed donor materials.Good consistency between model predictions and experimental results indicates machine learning is a powerful tool to pre-screen new materials,thus accelerating the development of organic photovoltaic field.In chapter four,we continue to explore the advanced functions of machine learning as well as the potential value of big data,and attempts to use machine learning to achieve the automatic design of high-performance donor materials.Through importance analysis with machine learning and statistics,we identify the key building units correlated to the high efficiency of the donor molecules.Then,these building units with different functions are combined onto two types of backbones,and a virtual library consisting of 18960 new small molecule donor materials is generated.After screening by a machine learning model,the theoretically predicted efficiency of 6337 candidates are over 8%.After the compatibility of Y6 as acceptor material,the best performed ones exhibit predicted efficiency values over 15%.Their calculated electronic properties by DFT show exceptional potential for high-efficiency carrier transport,making them promising candidates for high-performance organic photovoltaics.The result indicates the framework we developed is able to design new materials based on existing big data.It is believed that this framework also suits other materials,thus accelerating the discovery of new materials.In a word,machine learning method has a bright future in material development.It can not only predict the properties of materials to realize the rapid evaluation and screening of new materials,but identify important units to design new materials as well.