The prediction of donor number and acceptor number of electrolyte solvent molecules based on machine learning

Abstract

Electrolyte solvents have a critical impact on the design of high performance and safe batteries. Gutmann’s donor number (DN) and acceptor number (AN) values are two important parameters to screen and design superior electrolyte solvents. However, it is more time-consuming and expensive to obtain DN and AN values through experimental measurements. Therefore, it is essential to develop a method to predict DN and AN values. This paper presented the prediction models for DN and AN based on molecular structure descriptors of solvents, using four machine learning algorithms such as CatBoost (Categorical Boosting), GBRT (Gradient Boosting Regression Tree), RF (Random Forest) and RR (Ridge Regression). The results showed that the DN and AN prediction models based on CatBoost algorithm possesses satisfactory prediction ability, with R2 values of the testing set are 0.860 and 0.96. Moreover, the study analyzed the molecular structure parameters that impact DN and AN. The results indicated that TDB02m (3D Topological distance based descriptors−lag 2 weighted by mass) had a significant effect on DN, while HATS1s (leverage-weighted autocorrelation of lag 1/weighted by I-state) plays an important role in AN. The work provided an efficient approach for accurately predicting DN and AN values, which is useful for screening and designing electrolyte solvents.