(474c) Machine Learning-Assisted Design of Deep Eutectic Solvents Based on Uncovered Hydrogen Bond Patterns

Non-ionic deep eutectic solvents (DESs) have emerged as designer solvents for applications such as catalysis, extraction, and carbon capture and conversion. A major challenge is the lack of an efficient tool to discover DES candidates. Currently, the search relies heavily on the researchers intuition or a trial-and-error process, which leads to a low success rate or bypassing of promising candidates. Recognizing the central role that hydrogen bonds play in the DES formation, this work aims to decipher the hydrogen bond features for DESs and develop machine learning models to predict the potential of a system to be DES based on the hydrogen bond-based descriptors. We first analyze the hydrogen bond properties of 38 known DES and 111 known non-DES systems using their molecular dynamics simulation trajectories. The analysis reveals two features for DES compared to non-DES: (a) the imbalance between the numbers of the two intra-component hydrogen bonds, and (b) more and stronger inter-component hydrogen bonds. Based on the analysis results, we developed 30 machine learning models using ten algorithms and three types of hydrogen bond-based descriptors. The model performance is first benchmarked using their average and minimal ROC-AUC values. We also analyze the importance of individual features in the models and the results are consistent with the simulation-based statistical analysis. Finally, we validate the prediction ability of the models using the experimental results of 34 systems. The Extra Trees outperforms the others in the validation with an ROC-AUC of 0.88. Our work iterates the importance of hydrogen bond in DES formation and shows the potential of machine learning in discovering new DESs.