Open Challenges to Unlock Deep Eutectic Solvent Discovery

Anastasia K. Lavrinenko, George I. Puskov, Rishat Rafikov, Andrei Dmitrenko

Keywords: deep eutectic solvents, machine-learned interatomic potentials, molecular dynamics, AI-guided design

TL;DR: Machine-learned interatomic potentials could enable rational DES design, but require accurate treatment of electrostatics and hydrogen bonding, robust transport-property convergence, and broad chemical-space transferability.

Abstract: Deep eutectic solvents (DESs) have attracted considerable attention as tunable, often low-toxicity alternatives to conventional organic solvents with a wide range of applications in synthetic chemistry, biomass processing, CO$_2$ capture, drug delivery, and electrochemistry. They can be described as multicomponent mixtures with a melting point significantly lower than that of the neat components due to complex hydrogen-bond network. The discovery of new DESs remains mostly empirical and is limited by the infinite number of combinations of suitable components. Existing computational approaches, from thermodynamic models and molecular dynamics (MD) to direct machine learning prediction of macroscopic properties, suffer from limited transferability, low accuracy of transport properties, and lack of mechanistic interpretability. Machine-learned interatomic potentials (MLIPs), which enable MD simulations at near ab initio accuracy and low computational cost, represent a largely untapped approach for DESs. Only a few dedicated studies have been published to date. We outline key scientific and methodological challenges, including long-range electrostatics, a complex hydrogen-bond network, slow convergence of transport properties, and chemical-space transferability, and discuss research directions to develop MLIPs that can accelerate the rational design of next-generation green solvents.

Submission Track: Findings, Tools, & Open Challenges (Tiny Paper)

Submission Category: AI-Guided Design

Submission Number: 70