Neural network to predict 23Na NMR spectra of Nan clusters

Abstract

In order to understand the charging and discharging processes of sodium-ion batteries, we are interested in the relationship between the size of sodium clusters inserted into the hard carbon anode and the solid-state 23Na NMR chemical shifts. In this study, we investigated the predictability of the size dependence of 23Na NMR shielding constants by SchNet, a deep learning model that uses the distance between Na atoms without graph connection information. The data set required for training the neural network was constructed by density functional theory (DFT) calculations. This study shows that the neural network model, which only used structural data, achieved comparable accuracy in predicting the shielding constant to the Lasso model, which utilized gross orbital population predicted from DFT calculations. Moreover, by introducing a penalty term to the neural network's loss function, the neural network was able to reproduce the skewed distribution of the shielding constant without modifying its architecture.