Sustainable Design of Deep Eutectic Solvent Viscosity: A Structure–Property Machine Learning Framework with Structure-Disjoint Validation

Deep eutectic solvents (DES) have emerged as sustainable alternatives to conventional organic solvents, yet their industrial adoption remains hindered by limited predictive capability for transport properties, particularly viscosity. This work reports a machine learning framework for DES viscosity prediction built on 3440 experimental measurements spanning 237 systems (85 hydrogen bond acceptors, 86 hydrogen bond donors), with viscosities from 1.3 to 85,000 mPa·s over 278–378 K, in which component identity is defined by canonical isomeric SMILES rather than by name and structural assignments have been audited and corrected. A hierarchical feature engineering strategy generated 236 candidate descriptors, reduced to 128 modeling features by removing noncomputable, monotonically redundant and collinear terms, with no imputation of missing values at any stage. Ten algorithms spanning linear, kernel, neural and tree-ensemble families were benchmarked under identical tuning and evaluation; support vector regression achieved the lowest cross-validated error (R2 = 0.974, AARD = 12.2%) against 42% for linear models, establishing that nonlinearity accounts for two-thirds of the achievable error reduction. The central result is a quantified separation between interpolation and extrapolation obtained on identical data and a single model: AARD is 13% on a random hold-out, 75% for an unseen acceptor under structure-disjoint leave-one-HBA-out validation, 129% for an unseen donor, and 145% for an unseen chemical family. This 10-fold spread demonstrates that a single reported accuracy figure is uninterpretable for DES property models unless the splitting protocol is stated. Global SHAP attribution on the final model assigns 31.4% of total contribution to composition-weighted mixing descriptors and 23.7% to asymmetry descriptors quantifying HBA–HBD dissimilarity, establishing both families as substantial and nonredundant. The framework integrates gradient-free inverse design on a smooth surrogate, application-oriented clustering by viscosity and Arrhenius activation energy, and manufacturing robustness assessment, providing actionable tools for rational DES formulation together with explicit and empirically measured boundaries on their use.