Learning atomic interactions for the thermodynamics of liquid mixtures

Marco Hoffmann, Sebastian Schmitt, Hans Hasse, Fabian JirasekarXiv 2026 · arXiv

TL;DR

COGNAC is a hybrid thermodynamic model that embeds a graph neural network (GNN) within the established UNIFAC framework to predict liquid mixture behavior with high accuracy. By learning interactions at the atomic level rather than using rigid functional groups, it outperforms state-of-the-art machine learning models while remaining fast enough for real-world chemical process simulations.

What Problem Does This Paper Address?

Designing and optimizing chemical processes like distillation or extraction requires precise knowledge of how different molecules interact in a liquid mixture. This "non-ideality" is captured by activity coefficients.

For decades, engineers have relied on Group-Contribution (GC) methods like UNIFAC. These are fast but limited: they break molecules into predefined "groups" (like -CH3 or -OH), which often fails to capture the nuances of molecular geometry, structural isomers, or complex neighborhoods. Conversely, modern Machine Learning (ML) models are highly accurate but computationally "heavy"—running a complex neural network at every step of a large-scale process simulation is often too slow to be practical.

The authors of this paper sought to bridge this gap, creating a model that offers ML-level accuracy with GC-level speed and thermodynamic consistency.

Methods

The authors introduce COGNAC (local-COmposition Graph Neural network for the prediction of Activity Coefficients). Its architecture is a clever hybrid:

  1. Graph Neural Network (GNN): A message-passing neural network takes SMILES strings (text representations of molecules) and encodes each atom based on its neighborhood (up to four bonds away). Unlike traditional methods, it treats polar hydrogens (H-bond donors) as explicit nodes to better capture hydrogen bonding.
  2. Parameter Generation: Instead of predicting activity coefficients directly, the GNN predicts the physical parameters ($R$, $Q$, and interaction coefficients $a, b, c$) required by the modified UNIFAC equations.
  3. Local-Composition Framework: These predicted parameters are plugged into the standard, physics-based UNIFAC equations. This ensures the model is "thermodynamically consistent"—meaning it obeys fundamental laws like the Gibbs–Duhem equation by construction.
  4. Differentiable Solver: To train on liquid-liquid equilibrium (LLE) data, the authors developed a differentiable phase-split solver, allowing the model to learn directly from complex phase-separation data.
Figure 1: Overview of the COGNAC model architecture and its application. a, Molecular graphs constructed from SMILES [30] using RDKit [31] are encoded by a message-passing neural network (MPNN). Heavy atoms and hydrogens bound to non-carbon atoms form explicit nodes, allowing hydrogen-bond donors and acceptors to be resolved. Atoms with identical molecular neighborhoods within the message-passing radius obtain identical embeddings. b, Two prediction heads map the context-aware atom embeddings ha to volume and surface-area parameters Ra and Qa, and pairs of embeddings to the symmetric interaction coefficients aab, bab, and cab of the exported parameter table (Methods). Summing the respective atom-level geometry parameters gives the molecular quantities ri and qi. c, The learned parameters and the state point (composition x and temperature T) are processed in a local-composition framework. The combinatorial contribution accounts for size and shape effects, while atomic surface areas and pair interactions determine the residual contribution. Together, the two contributions yield the activity coefficient of component i. d, For a specified component set, parameters need to be generated only once and can be reused across compositions and temperatures without further neural network evaluations. The obtained atom-level parameter vectors R and Q and interaction matrices a, b, and c enable, e.g., the prediction of vapor–liquid (VLE), liquid–liquid (LLE), and solid–liquid (SLE) equilibria, azeotropes, and residue curve maps for binary and multi-component mixtures, and can easily be implemented in process simulations.

Datasets

COGNAC was trained on a massive dataset of 822,780 experimental data points covering 48,432 binary mixtures from the Dortmund Data Bank (DDB). The data types include:

  • Vapor–liquid equilibria (VLE)
  • Liquid–liquid equilibria (LLE)
  • Solid–liquid equilibria (SLE)
  • Activity coefficients at infinite dilution (ACT)
  • Excess enthalpies (HE)

The model's geometry predictions were also regularized against quantum-chemistry-based COSMO cavity data from the CHAOS database to ensure the learned parameters remain physically meaningful.

Key Insights and Ideas

  • The "Once-and-Done" Advantage: The most significant practical insight is that the GNN only needs to be evaluated once for a given set of components to generate a parameter table. Once generated, these parameters can be reused across all temperatures and compositions in a simulation, making it nearly as fast as traditional UNIFAC.
  • Atoms Over Groups: The paper demonstrates that the limitation of traditional methods isn't the underlying equations, but the rigid definition of "groups." By allowing the GNN to define atom-level interactions based on the local environment, the model captures effects (like isomerism) that GC methods miss.
  • Interpretability: Because the model predicts physical parameters, researchers can inspect the results to see if they make chemical sense. For example, COGNAC correctly learns that water-oxygen and alkyl-group interactions are unfavorable, while amine-nitrogen and hydrogen interactions (H-bonding) are highly favorable.

Key Results

  • Superior Accuracy: COGNAC reduces median errors by 13–63% compared to HANNA ( a leading ML model) and significantly outperforms modified UNIFAC across all tested data types.
  • Zero-Shot Generalization: Although trained only on binary mixtures, COGNAC accurately predicts the behavior of ternary and quaternary mixtures, often outperforming UNIFAC even though UNIFAC was specifically fitted to that multi-component data.
  • Azeotrope Prediction: COGNAC correctly classifies 92.5% of binary azeotropes, a critical task for designing distillation columns.
  • Simulation Speed: In a batch distillation simulation, COGNAC was two orders of magnitude faster than the HANNA ensemble and only about 2x slower than the much less accurate UNIFAC.

Limitations and Future Work

  • Elemental Scope: The current model is limited to molecules composed of 12 specific elements (H, B, C, N, O, F, Si, P, S, Cl, Br, and I).
  • Complex Systems: Electrolytes and polymers are currently outside the model's domain. The authors suggest that future extensions could incorporate long-range electrostatic interactions and free-volume effects to address these categories.
  • Data Quality: Like all ML models, COGNAC is dependent on the quality of the underlying experimental data. While the authors performed manual inspections to filter outliers, the model's performance is ultimately bounded by the DDB's accuracy.

Summary updated

Abstract

Predictive models for activity coefficients, which describe the non-ideality of liquid mixtures, are foundational for process design and optimization. Group-contribution methods, such as UNIFAC, are computationally cheap and well established in process simulators. Modern machine learning models, such as HANNA, offer higher accuracy and broader applicability, but their integration into process simulation is expensive or even infeasible. In this work, we introduce COGNAC, a graph neural network that learns atom-level geometry and interaction parameters that depend on the molecular neighborhood and evaluates them in the UNIFAC framework. Trained on about 820,000 experimental data points, COGNAC achieves significantly higher accuracy than UNIFAC and HANNA for binary and multi-component mixtures, reduces error relative to HANNA by 13-63%, reliably predicts multi-phase equilibria and identifies azeotropes, yields interpretable parameters, and is strictly thermodynamically consistent. COGNAC evaluations run at a cost comparable to UNIFAC, thereby unlocking the potential of machine-learning models for process simulations.
BibTeX
@misc{hoffmann2026learning,
  author = {Marco Hoffmann and Sebastian Schmitt and Hans Hasse and Fabian Jirasek},
  title = {{Learning atomic interactions for the thermodynamics of liquid mixtures}},
  year = {2026},
  eprint = {2610.36440},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36440},
}
Figure