Correlation between σ‑Profile Characteristics and Infinite Dilution Activity Coefficients of Choline Chloride-Urea DES: Experimental Determination and Machine Learning Interpretation

This study integrates inverse gas
chromatography (IGC) experiments
with machine learning (ML) to systematically investigate the thermodynamic
properties of choline chloride (ChCl)-urea (1:2) deep eutectic solvent
(DES) and its interaction mechanisms with organic solvents. IGC measurements
determined the infinite dilution activity coefficients (γ12∞) and related
thermodynamic parameters for 46 representative organic solvents within
the temperature range of 303.15–343.15 K. Results revealed
the hierarchy of solute–DES interaction strength: hydrocarbons
(increasing with chain length) > alkenes > ethers > aromatics
> ketones
> esters > alcohols (weakest due to hydrogen bonding). To enhance
γ12∞ prediction accuracy, a novel approach fused the quantized σ-profile
partitioning descriptors of the DES with temperature as input features,
constructing four ML models. Compared to the significant deviation
of the COSMO-SAC model prediction (R2 =
0.8224), the Extreme Gradient Boosting (XGBoost) model demonstrated
superior performance (test set R2 = 0.9979,
average absolute relative deviation (AARD) < 20%). Feature importance
analysis indicated that σ-profile regions corresponding to weak
hydrogen bond acceptor (HBAs) character [S3: −0.0084 ≤
σ ≤ 0 e/Å2] and weak hydrogen bond donor
character [S4, 0 ≤ σ ≤ 0.0084 e/Å2] contributed dominantly (42%) to the γ12∞ prediction. In contrast,
the strongly polar region [S5, 0.0084 ≤ σ ≤ 0.02
e/Å2] reduced γ12∞ by enhancing interactions, confirming
the “like dissolves like” principle. This framework
enables high-precision γ12∞ prediction solely from molecular structures
(Applicability Domain (AD) covers 93.85% of data), providing an efficient
and reliable theoretical tool for DES-based green solvent design and
optimization of industrial separation processes, such as benzene/methanol
systems.