Design And Application Of Solid Electrolyte Material Screening And Conductivity Prediction Software Based On Machine Learning

In the field of new energy batteries,solid-state electrolytes are expected to be used in the next generation of all-solid-state batteries due to their high safety,long cycle life,wide electrochemical window and high energy density.Although solid electrolyte is an effective substitute for liquid electrolyte,its ionic conductivity is not as good as that of liquid electrolyte.The traditional trial-and-error method is often time-and resource-intensive.With the improvement of database quality,the increase of data volume,and the rapid development of computing power and algorithms,the"fourth science paradigm"based on first-principles computing and data-driven has gradually penetrated into various research fields.Machine learning has been gradually applied to the research of battery and other new energy materials for its high efficiency,applicability and good prediction accuracy.At present,there are still many problems in machine learning applications in materials field,such as the difficulty to select the optimal feature subset and the single performance evaluation index,which make the model easily fall into overfitting and result in poor generalization performance.To solve these problems,this paper designed a visual machine learning software-MLearning,developed a global optimal feature subset selection method and proposed an effective feature selection method based on high dimensional small samples.Based on this software,the potential fast ionic conductor structure screening and ionic conductivity prediction can be realized efficiently and accurately.The detailed research contents and innovations are as follows:（1）A relatively comprehensive,universal and automatic machine learning software MLearning is designed.Its main functions include:MP database and OQMD database structure and information batch acquisition,structure descriptor extraction,feature selection,algorithm construction,comprehensive performance evaluation and automatic machine learning interface.In particular,a variety of cross-validation evaluation schemes,hyperparameter automatic search schemes and a feature selection method based on high dimensional small samples are developed.（2）The factors affecting the ionic conductivity were analyzed in depth,several new correlation descriptors were proposed,and an interpretable logistic regression model was constructed based on MLearning software to evaluate the model performance,and then some intuitive explanations of influencing the ionic conductivity were given.Furthermore,the trained optimal model was applied to all the data containing lithium structure in MP database to screen out potential fast ion conductors.Finally,the model shows a high LOOCV classification accuracy of nearly 90%.Finally,bond-valence energy analysis was performed on more than 60 selected structures to roughly evaluate the ion migration energy barrier.The results show that there are relatively low ion migration energy barrier.（3）Some simple molecular and electronic descriptors are proposed to predict the room temperature Li⁺conductivity of garnet-type solid electrolyte.Based on the feature selection module,algorithm construction and evaluation module of MLearning software,the explainable logistic regression and random forest regression algorithm models were constructed respectively to evaluate the performance of the models and give some intuitive explanations affecting the ionic conductivity to guide the theoretical design and experimental preparation.Finally,high classification accuracy（close to 0.9for cross validation）and mean absolute error(0.24 log₁₀（σ）)for LOOCV are demonstrated.