Exploring the deviation from Nernst-Einstein conductivity in ionic liquids using machine learning

Abstract

Ionic liquids (ILs) are promising candidates for safer battery electrolytes due to their low flammability, but their low ionic conductivity constrains battery performance. Given the large chemical space of potential ILs, machine learning (ML) approaches are essential for accelerated screening. However, previous studies have shown that ML models trained on common computational descriptors to predict molar ionic conductivity have accuracies comparable to the Nernst–Einstein equation [Umaña, Cashen, Zavala and Gebbie, Digital Discovery, 2025, , 1423–1436]. Furthermore, experimental measurements show that the ionic conductivity of many ILs deviate substantially from that predicted by the Nernst-Einstein equation. While several mechanisms have been proposed, the structural origins of these deviations are not well understood. In this study, we develop ML models to predict the deviation of the ionic conductivity of ILs from Nernst-Einstein behavior using charge-based descriptors for individual ions. We observed that ML models trained using a smaller set of sigma profile-based descriptors had similar performance to those trained on a larger set of RDKit descriptors. Additionally, we found that models trained to predict this deviation could serve as a correction factor to the Nernst-Einstein equation and resulted in more accurate conductivity predictions compared to models that were designed to directly predict the molar ionic conductivity of ILs. We applied feature importance rankings to gain insight into model predictions and identified features relating to the cation alkyl chain length and the cation and anion polarity as being influential.