Machine learning model accelerates solvent optimization for liquid-liquid separation

Four-step workflow diagram for a machine-learning solvent selection method. Steps 1–2 (model development): training a model on property data via a feedback cycle, then identifying phases in solvent mixtures. Steps 3–4 (model validation): predicting partition coefficients for Arizona-series solvents and monomers, then selecting solvents using a 3D solubility-parameter plot.
Study overview: (1) feature engineering of solute and solvent representations; ML training to predict log10KP values, (2) phase behavior identification of solvent mixtures, (3) validation of ML model to extrapolate to solutes beyond those used in training, and (4) solvent selection in multiple case studies. Photo by: Courtesy of authors

Background/Objective

Solvent selection is a critical step in the design of chemical processes, influencing solubility, phase equilibrium, and reactivity. Solvents are often incorporated as mixtures, substantially increasing the size of the design space, particularly for liquid-phase separation of lignin-derived products due to similar physiochemical properties.

Approach

To address the challenge of selecting solvent mixtures for liquid-phase separation of lignin-derived products, researchers trained a machine learning model to predict the partition coefficient (log10Kp) of a solute as a thermodynamic descriptor of preferential partitioning between two immiscible liquid phases formed in a solvent mixture. They combined log10Kp predictions with computational evaluation of phase behavior and individual solvent sustainability to screen mixtures for two case studies: replacing mixtures containing halogenated solvents with more sustainable alternatives and identifying mixtures suitable for separating lignin products via centrifugal partition chromatography (CPC).

Results

By representing solvent mixture properties using Hansen solubility parameters and incorporating simple solute descriptors, the model can accurately predict log10Kp values for a range of chemically similar lignin-derived solutes obtained from two distinct experimental datasets in the literature. The model identified solvent mixtures predicted to achieve similar separation performance as known mixtures containing toxic halogenated solvents and solvents suitable for use with CPC as an emerging liquid-liquid separation method.

Significance/Impacts

Solvent selection impacts process economics and energy use in manufacturing. The data-driven computational tools can accelerate the discovery of green and other solvent mixtures and promote sustainable lignin valorization by identifying mixtures for separating lignin-derived products.

Kim, C., & Van Lehn, R. C. Computational Framework for Discovering Solvent Mixtures for Liquid–Liquid Separations. ACS Sustainable Chemistry & Engineering. (2026). [DOI:10.1021/acssuschemeng.6c06340]
 

Research Theme
Sustainable Biomass Conversion
Highlight - Slide