Na-ion solid-state electrolytes (SSE) is attracting spotlight as a key component of future power storage system due to their high energy density and low manufacturing cost. However, information on their mechanical properties are limited to find out optimal Na-SSE candidates, which is significant for sustaining structural stability at the electrolyte and electrode interface. Many previous studies prove that those properties can be discovered accurately through computational and experimental methods, but it is practically difficult to apply them to vast chemical space due to their expenses. In this study, machine learning based regression model, which is cost-efficient for material design, is developed for predicting mechanical properties of Na-SSEs. As a training set, 12,361 materials are obtained from well-known materials database (Materials Project) and they are described with chemical and structural descriptors. Then we construct model using LightGBM, along with sampling techniques considering distribution of database. Performance of developed model shows great accuracy (mean absolute error (MAE) value of 11.8 GPa and 15.3 GPa and R2 score of 0.72 and 0.87) for shear and bulk modulus, respectively. This model is then applied to 2,432 Na-SSE candidates whose mechanical properties are not explored in Materials Project. After obtaining ther mechanical properties by density functional theory(DFT) calculations, the accuracy of the constructed model is obtained as MAE of 8.75 GPa and 13.4 GPa and R2 score of 0.65 and 0.69 for shear and bulk modulus, respectively. Finally, the active learning process is conducted to propose the ideal materials with minimum dataset addition by lessening the prediction uncertainty (exploration). For evaluation, the performance from the exploration is compared to that from the random data addition. It shows that employing the optimization process exhibits superior improvement than random method as number of selected data are added to train set. We believe that constructed platform contributes to discovery of Na-SSEs with optimum mechanical properties in minimum cost.

Machine learning aided screening strategy for mechanically superior Na-based solid-state electrolytes

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.