Leveraging mixture representation design and chemical language model for property prediction of deep eutectic solvents

Deep eutectic solvents (DESs) have emerged as a promising class of solvents with exceptional structural tunability and excellent solvation capacity, whereas the vast chemical space and limited experimental data available make their rational screening extremely challenging. In this study, chemical language models pre‐trained on large‐scale unlabeled molecules are constructed to predict four key physicochemical properties of DESs. Multiple mixture representation strategies are designed to capture the interactive features between individual DES components, among which a novel CNN‐integrated mole‐weighted latent‐vector method proves the most effective. By hierarchically comparing the performance of models pre‐trained on subsets of varying scales, the optimal pre‐training data sizes are carefully determined for distinct DES property prediction tasks. Moreover, the influence of different data splitting strategies (random and mixtures out) on model performance is assessed. Based on these systematic comparisons, the optimal BERT models achieve notably superior performance compared to benchmark models in the literature.