Generalizing property prediction of ionic liquids from limited labeled data: a one-stop framework empowered by transfer learning

Abstract

Ionic liquids (ILs) could find use in almost every chemical process due to their wide spectrum of unique properties. The crux of the matter lies in whether a task-specific IL selection from enormous chemical space can be achieved by property prediction, for which limited labeled data represents a major obstacle. Here, we propose a one-stop ILTransR (IL transfer learning of representations) that employs large-scale unlabeled data for generalizing IL property prediction from limited labeled data. By first pre-training on ∼10 million IL-like molecules, IL representations are derived from the encoder state of a transformer model. Employing the pre-trained IL representations, convolutional neural network (CNN) models for IL property prediction are trained and tested on eleven datasets of different IL properties. The obtained ILTransR presents superior performance as opposed to state-of-the-art models in all benchmarks. The application of ILTransR is exemplified by extensive screening of CO₂ absorbent from a huge database of 8 333 096 synthetically-feasible ILs.