Predicting Ternary Inorganic Photovoltaic Materials By Machine Learning

Solar energy is a clean and sustainable energy.The development of photovoltaic materials with first principles simulation calculation highly requires more computing cost and more time cost,which limits the development speed of researching photovoltaic materials.However,machine learning provides a new route of development of photovoltaic materials quickly.This paper uses machine learning to learn the existing photovoltaic material data and predicts four new photovoltaic materials.Combined with the first-principle density functional theory simulation calculation,machine learning analyzes the electronic structure,optical transition,photoelectric conversion efficiency and stability of these four materials.The training and evaluation datasets were constructed using the existing data of 2398 photovoltaic materials for training model and evaluating performance of model.The 241 features were calculated using the Voronoi method,and the 41 most important features were selected using recursive feature elimination and cross-validation(RFECV)methods.Comparing Gradient Boosting Regressor(GBR)and the other four model algorithms,we found that the prediction time cost of GBR model is comparable to that of the other four models,but the prediction accuracy is 10% or even 45% higher.We chose GBR algorithm,and the prediction accuracy of the model has reached 90.7% after training.Subsequently,using 29 materials as the basis,the prediction dataset containing 3,587 new materials was made by using the same master family element replacement method.Band gap was predicted by GBR model,and model selects out four materials with material preparation cost,enthalpy of formation(ΔH),Ba4Te12Ge4,Ba8P8Ge4,Sr8P8Sn4 and Y4Te4Se2.Further,when molecular dynamics simulation of the selected materials using first principle based density functional theory,we found that the four materials at 300 K always fluctuated little within 5 ps.The crystal structure of three candidates(Ba4Te12Ge4,Ba8P8Ge4 and Y4Te4Se2)remained stable,only the lattice of Sr8P8Sn4 remained stable after slightly tilted at 2 ps,which indicates that the four materials have good stability at room temperature.By calculating the band structure and transition rate of the carriers,it is found that Ba4Te12Ge4 and Y4Te4Se2 are direct band gap semiconductors,and the peaks of transition probability(P2)is at G point,which indicates that the optical transition of the two materials at the band edge is allowed and a light absorption coefficient.Although Ba8P8Ge4 and Sr8P8Sn4 are indirect band gap materials,they have good absorption properties.By calculating the spectroscopic limited maximum efficiency(SLME),it is found that the photoelectric conversion efficiency of these four materials is very close to that of CH3NH3 Pb I3.In order to improve the accuracy of energy prediction,this paper proposes a pooling representation for encoding the information in the shell electrons,and compares it with the Coulomb matrix.We find that the pooling representation is better performance on the neural network,and further improves the accuracy of energy prediction.