Research On Materials Screening And Ionic Conductivity Prediction Of Solid Electrolyte Based On Machine Learning Methods

In order to accelerate the development of lithium-ion batteries to meet the growing demand,designing and discovering solid-state electrolyte materials with good performance has become a research hotspot for the development of lithium-ion batteries.The key of research lies in how to quickly screen out excellent solid electrolyte candidate materials and predict ionic conductivity accurately.In traditional solid electrolyte materials screening and ionic conductivity prediction,the experimental measurement method and computational simulation method are mainly used.However,these two methods have problems such as low time efficiency and dependence on material microstructure.With the establishment of Materials Genome Initiative(MGI)and the development of material informatics,machine learning has been widely used in the research of lithium-ion batteries due to its superior performance in terms of prediction accuracy,time efficiency,applicability and so on.However,the application of machine learning to solid-state electrolyte materials is still at a preliminary stage of development.In the research,there are problems such as the difficulty in selecting the optimal subset of features,the unclear basis for the evaluation of the performance,and the low accuracy of ion conductivity prediction.Therefore,aiming at these problems,the feature selection method for analyzing the influence factors of solid electrolyte performance,the clustering method for evaluating the performance and the regression integration method for predicting the ion conductivity are studied in this paper.The main research contents and innovations of this paper are as follows:1)An expert experience integrated multi-layer feature analysis method is proposed and used to handle sparseness,irrelevance and redundancy issues in solidstate electrolyte data.From three aspects: sparseness assessment,correlation assessment and redundancy assessment,this method effectively combines multiple feature selection methods,and hierarchically selects features of the original feature set.Meanwhile,scoring method is used to quantify the expert experience,and expert experience and feature selection results are combined by establishing the feature importance score indicator,achieving a coordinated analysis of feature selection.The experimental results show that this method can effectively screen out the optimal feature subsets while fully integrating expert experience and improve model prediction accuracy.2)A cloud model based uncertainty clustering method is proposed and used for solid electrolyte material screening.This method uses the backward cloud generator to construct cloud models for each cluster,calculates the degree of certainty of each data in each cluster by the forward cloud generator to achieve soft clustering,and utilizes digital features to extract qualitative rules from clusters.The experimental results show that this method can extract qualitative rules from clusters and realize the effective division of solid electrolyte materials.3)A cloud model based regression ensemble method is proposed and used for the prediction of ionic conductivity.This method introduces ensemble learning to the prediction of ionic conductivity,uses a variety of heterogeneous linear regression algorithms as individual learners.Then,it integrates the regression coefficient equations obtained by the learners with the function cloud generator,and constructs cloud regression relationship between description factors and ionic conductivities.The experimental results show that this method can obtain the regression equation that characterizes the structure-activity relationship between the description factors and the ionic conductivity,and predict the ionic conductivity with high precision,providing a new solution for predicting ionic conductivity.