Deep eutectic solvent screening moves beyond quantum chemistry: TiNN splits the hydrogen-bond network into two steps and cuts measured solubility error from 23-fold to about 1.5-fold

Solubility prediction for deep eutectic solvents now follows the order in which the hydrogen-bond acceptor-donor environment forms and the solute is then accommodated, rather than concatenating acceptor, donor and solute as independent inputs at one level. Earlier machine-learning models typically concatenated constituent-wise molecular descriptors or combined independently encoded acceptors and donors at fixed molar ratios, treating temperature and pressure as ordinary features. On the carbohydrate solute subset, predictions improve progressively from the MLP baseline to the thermodynamics-informed model with concatenated input and then to the full two-level TiNN; the full TiNN reaches a test-set $R^2$ of 0.917, MAE of 0.126 and RMSE of 0.242.

A learned free-energy pathway links the molecular representation to the solubility response as temperature and pressure change. Mapping directly from the system representation to solubility treats temperature as an ordinary descriptor and leaves that response unconstrained. Training used 938 deep eutectic solvent-solute solubility records split by solute into roughly 72% training, 13% validation and 15% test; in leave-one-acceptor-donor-pair-out cross-validation over 81 distinct pairs, TiNN keeps a diagonal parity distribution while the MLP baseline develops horizontal bands of regression toward an averaged response.

The hydrogen-bond compatibility signal matters according to solute type, used where hydrogen bonding governs dissolution and largely disregarded where other mechanisms dominate. Zeroing that signal changes full-test-set $R^2$ by no more than 0.023 relative to keeping it, whereas shuffling it across solute pairs drops $R^2$ by 0.217. On the carbohydrate and lignin-derived subset, keeping the signal beats zeroing by 0.046 and beats shuffling by 0.412 in $R^2$; on the gas subset of CO2, SO2 and H2S, keeping and zeroing the signal are indistinguishable.

Without training on fractionation data, the same molecular solvation information still separates deep eutectic solvents that remove lignin and hemicellulose from corn stover well from those that remove them poorly. COSMO-RS gives a $\log_{10}$-scale MAE of at least 1.37 across nine deep eutectic solvent-solute systems, against 0.18 for TiNN. The nine systems are glucose, xylan and Kraft lignin measured in ChCl:lactic acid, TBAB:imidazole and ChCl:glycerol at 323.15 K, 101.325 kPa and a donor-to-acceptor molar ratio of 2:1; corn stover fractionation was tested at 373.15 K with a biomass-to-solvent mass ratio of 1:40 for five solvents, and the order of the intermediate solvents was not fully reproduced.

The demonstrated scope is training and testing on 938 deep eutectic solvent-solute solubility records, leave-one-acceptor-donor-pair-out cross-validation over 81 pairs, new measurements of glucose, xylan and Kraft lignin in three deep eutectic solvents of differing acidity and basicity, and fractionation of corn stover with five solvents; the intended readers are chemists and engineers working on solvent screening and biomass fractionation.

Absolute xylan solubility is underestimated, possibly partly because its polydispersity is represented by an idealized molecular structure and because relatively limited xylan data were available for training; in corn stover fractionation the order of the intermediate solvents is not fully reproduced, and the model correctly separates only the solvents with the highest and lowest fractionation capacity.