Large chemical language models for property prediction and high-throughput screening of ionic liquids

Abstract

Ionic liquids (ILs) possess unique physicochemical properties and exceptional tunability, making them versatile materials for a wide range of applications. However, their immense design flexibility also poses significant challenges in efficiently identifying outstanding ILs for specific tasks within the vast chemical space. In this study, we introduce ILBERT, a large-scale chemical language model designed to predict twelve key physicochemical and thermodynamic properties of ILs. By leveraging pre-training on over 31 million unlabeled IL-like molecules and employing data augmentation techniques, ILBERT achieves superior performance compared to existing machine learning methods across all twelve benchmark datasets. As a case study, we highlight ILBERT's ability to screen ILs as potential electrolytes from a database of 8 333 096 synthetically feasible ILs, demonstrating its reliability and computational efficiency. With its robust performance, ILBERT serves as a powerful tool for guiding the rational discovery of ILs, driving innovation in their practical applications.