Expanding the chemical space of ionic liquids using conditional variational autoencoders

Abstract

Ionic liquids (ILs) are salts set apart by their low melting points and can act as highly tuneable solvents with broad application potential, for example as catalysts, in batteries, and for drug delivery. The potential chemical space of ILs is vast, with only a very small region having been explored to date. Machine learning offers a promising approach to advance into this vast space of unexplored ILs; however, existing IL databases contain limited ion diversity, constraining the performance of generative models. To address this, we introduce conditional variational autoencoders (CVAEs) and a novel ion scoring method as a conditioning factor. The ion score prioritises ions with a higher likelihood of forming low-melting-point ILs. Our CVAEs effectively generate novel and diverse cations and anions. Furthermore, we constructed a melting point prediction model to identify cation-anion pairs that are likely to yield ILs with low melting points. Visualisation of the generated ILs alongside existing ones reveals that our approach effectively expands the chemical space of ILs with novel structures. Molecular dynamics simulations further validate that 13/15 of the generated ILs possess desirable low melting points (<373 K). The associated code is available at github.com/fate1997/ILGen-ion.