Thermodynamics-Informed Learning Captures Hierarchical Hydrogen-Bond Networks in Deep Eutectic Solvents

Yanpeng Ye, Chaoran Duan, Mian Xu, S. Mani SarathyarXiv 2026 · arXiv

TL;DR

Researchers have developed TiNN, a thermodynamics-informed neural network that predicts how complex biomass components dissolve in Deep Eutectic Solvents (DESs). By mimicking the hierarchical nature of molecular interactions and enforcing thermodynamic consistency, TiNN achieves significantly higher accuracy than traditional physics-based models and generalizes effectively to entirely new solvent combinations.

What Problem Does This Paper Address?

Deep Eutectic Solvents (DESs) are the "green" darlings of the biorefinery world. They are tunable, non-toxic, and excellent at breaking down lignocellulose (the tough structural material in plants) into useful fuels and chemicals. However, a DES isn't a single molecule; it’s a mixture of a Hydrogen Bond Acceptor (HBA) and a Donor (HBD) that forms a complex, cooperative network.

The challenge is that the solubility of biomass in these solvents isn't just the sum of the individual parts. It depends on a hierarchical hydrogen-bond network: first, the HBA and HBD must associate to form the solvent environment, and then that environment must accommodate the solute (like lignin or cellulose). Existing machine learning models often treat these components as a simple "bag of features," failing to capture this ordered interaction. Furthermore, traditional physics-based models like COSMO-RS often struggle with the non-ideal chemical environments found in DESs, leading to massive prediction errors.

Methods

The authors introduced two primary innovations in the TiNN architecture to bridge the gap between "black-box" ML and chemical reality:

  1. Hierarchical Representation: Instead of concatenating all molecular descriptors at once, TiNN uses a multi-stage approach. It first encodes the HBA and HBD into a "DES environment" representation, incorporating their molar ratios and internal compatibility. Only then is the solute embedding integrated into this pre-formed environment.
  2. Thermodynamic Inductive Bias: To ensure the model doesn't make unphysical predictions, the authors constrained the output using a learned free-energy pathway. The model is forced to estimate latent enthalpy ($\Delta H$) and entropy ($\Delta S$) terms, ensuring that the predicted solubility follows the fundamental relationship: $\Delta G \approx \Delta H - T\Delta S$.

The molecular structures themselves were processed using Equivariant Graph Neural Networks (EGNN), which allow the model to understand the 3D geometry of the molecules rather than just their 2D connectivity.

Figure 2. Molecular representation and predictive performance of TiNN. (a) Representation strategies for DES–solute systems. Left: direct concatenation of independently encoded HBA, HBD, and solute embeddings. Right: TiNN first combines the HBA and HBD into a DES representation, incorporating composition and internal hydrogen-bond compatibility, and then integrates the solute through DES–solute hydrogen-bond compatibility. (b–d) Parity plots on the carbohydrate subset for (b) MLP with concatenated HBA, HBD, and solute descriptors, (c) TiNN with concatenated input, and (d) TiNN with the hierarchical three-molecule representation. (e) Training- and test-set predictions of the full TiNN model; the reported test-set metrics are calculated on held-out samples.

Datasets

The model was trained and evaluated on a compiled dataset of 938 experimental records covering:

  • Solutes: Carbohydrates (cellulose/hemicellulose derivatives), lignin-derived compounds, and gases (CO₂, SO₂, H₂S).
  • Conditions: Varying temperatures, pressures, and HBD:HBA molar ratios.
  • Validation: The authors also performed new experimental measurements on nine DES-solute systems (acidic, basic, and neutral) and tested the model's ability to rank solvents for real corn stover fractionation.

Key Insights and Ideas

  • Order Matters: The hierarchical architecture (forming the solvent environment before adding the solute) significantly outperformed flat concatenation. This confirms that the sequence of molecular association is a critical "hidden" feature in solvation chemistry.
  • Physics Prevents Overfitting: By enforcing the $\Delta G$ relationship, TiNN remained robust in "sparse" regions of the dataset (e.g., very high temperatures) where standard ML models like XGBoost or simple MLPs saw their error rates spike.
  • Mechanism-Aware Learning: Through ablation studies, the authors found that a "hydrogen-bond compatibility" descriptor was vital for predicting biomass solubility (which relies on O-H and N-H bonds) but was correctly ignored by the model for gases like CO₂, which dissolve via different mechanisms like cavity formation.

Key Results

  • High Accuracy: TiNN achieved a test-set $R^2$ of 0.917, far outperforming standard descriptors and graph-based baselines.
  • Crushing the Competition: In head-to-head experimental tests, TiNN reduced the geometric mean fold error from 23-fold (with COSMO-RS) to just 1.5-fold.
  • Zero-Shot Transfer: Remarkably, even though TiNN was only trained on the solubility of isolated molecules, it successfully distinguished between DES formulations that were effective vs. ineffective at fractionating real corn stover. This suggests the model captured the fundamental molecular determinants of biomass deconstruction.
Figure 3. TiNN maintains low error even in data-sparse regions (a, b) and captures temperature trends more accurately than standard MLPs (c, d).

Limitations and Future Work

While TiNN is a major step forward, the authors noted a few areas for improvement:

  • Xylan Complexity: The model slightly underestimated xylan solubility, likely due to the polydisperse (varied chain length) nature of real xylan compared to the idealized structures used in training.
  • Kinetics vs. Thermodynamics: The model focuses on solubility (thermodynamics), but real-world biomass refining also involves reaction kinetics (breaking chemical bonds), which TiNN does not yet explicitly model.
  • Data Scarcity: While 938 records is good for this niche, more diverse data on exotic HBA/HBD combinations would further improve the model's "chemical imagination."

Summary updated

Abstract

Deep eutectic solvents (DESs) offer a tunable platform for lignocellulose refining, but their solvation behavior depends on multicomponent hydrogen-bond networks rather than individual constituent properties. We develop TiNN, a thermodynamics-informed neural network that hierarchically represents HBA-HBD association and solute accommodation through a learned free-energy pathway. Trained on 938 records, TiNN achieves a test-set $R^2$ of 0.917 and generalizes to new HBA-HBD combinations and sparse conditions. Ablation identifies solute-specific stabilization as the largest contributor to prediction accuracy, while the importance of hydrogen-bond compatibility depends on the dissolution mechanism. Experimental measurements of glucose, xylan, and Kraft lignin in acidic, basic, and neutral DESs show that TiNN reduces the geometric mean solubility error from more than 23-fold with COSMO-RS to approximately 1.5-fold. Without training on fractionation data, TiNN also distinguishes DES formulations with high and low lignin and hemicellulose removal from corn stover. These results support chemically grounded solvent screening for lignocellulose fractionation.
BibTeX
@misc{ye2026thermodynamics,
  author = {Yanpeng Ye and Chaoran Duan and Mian Xu and S. Mani Sarathy},
  title = {{Thermodynamics-Informed Learning Captures Hierarchical Hydrogen-Bond Networks in Deep Eutectic Solvents}},
  year = {2026},
  eprint = {2611.06590},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06590},
}
Figure