Graph neural networks for the prediction of infinite dilution activity coefficients

Edgar Iván Sánchez Medina, Steffen Linke, Martin Stoll, Kai SundmacherDigital 2022 · DOI

TL;DR

Researchers have developed a Graph Neural Network (GNN) framework to predict infinite dilution activity coefficients ($\gamma_{ij}^\infty$), a fundamental thermodynamic property used in chemical separation processes. The model outperforms traditional state-of-the-art mechanistic methods like UNIFAC and COSMO-RS in accuracy while providing 100% coverage of the organic chemical space, effectively bypassing decades of manual "group-contribution" parameter estimation.

What Problem Does This Paper Address?

In chemical engineering, designing processes like distillation or liquid-liquid extraction requires knowing how different molecules interact in a mixture. The "infinite dilution activity coefficient" ($\gamma_{ij}^\infty$) is a critical value that describes how a single solute molecule behaves when surrounded entirely by a solvent.

The challenge is that the "chemical space"—the number of possible solute-solvent combinations—is effectively infinite. Experimentally measuring every pair is impossible. For decades, engineers have relied on mechanistic models (like UNIFAC) that break molecules into functional groups. However, these models require experts to manually define groups and interaction parameters, which is slow and often leaves "blind spots" where the model simply cannot make a prediction because a specific group hasn't been characterized yet.

Methods

The authors propose an end-to-end deep learning approach using Graph Neural Networks (GNNs).

  1. Molecular Representation: Molecules are converted from SMILES strings into graphs where atoms are nodes and bonds are edges. Features like atom type, hybridization, and bond type are encoded via one-hot encoding.
  2. The "Two-Tower" Architecture: The system uses two separate GNNs—one to extract a "fingerprint" for the solvent and another for the solute.
  3. Message Passing & Pooling: The model uses edge-conditioned convolutions and a "Jumping Knowledge" scheme (averaging information across different layers) to capture both local and global molecular structures. A Set2Set pooling layer then converts these graphs into fixed-length vectorial representations.
  4. Prediction: The solvent and solute vectors are concatenated and passed through a Multi-Layer Perceptron (MLP) to predict $\ln(\gamma_{ij}^\infty)$.
  5. Hybrid Modeling: Beyond a pure GNN, the authors tested "parallel residual hybrid models." In this setup, a traditional mechanistic model makes a baseline prediction, and a GNN is trained specifically to predict the residual error of that model, combining physical theory with data-driven correction.

Datasets

The study utilized a comprehensive database compiled by Brouwer et al. (2019).

  • Scope: 4,460 experimental data points at 298.15 K.
  • Diversity: 156 unique solutes and 262 unique solvents, covering 2,810 unique binary systems.
  • Cleaning: The authors averaged repeated entries for the same systems to ensure a clean training set of unique binary pairs.

Key Insights and Ideas

  • Automated Feature Engineering: Unlike UNIFAC, which requires humans to decide which "groups" matter, the GNN automatically learns the most relevant structural features for thermodynamic non-ideality through backpropagation.
  • Universal Coverage: Specialized mechanistic models like MOSCED or Abraham often have higher accuracy but only cover ~45% of the dataset because they lack parameters for certain molecules. The GNN provides 100% coverage without sacrificing accuracy.
  • The Power of Ensembles: By training an ensemble of 30 GNNs (bagging), the researchers significantly reduced prediction error, proving that model averaging is highly effective for noisy thermodynamic data.
  • Theory + ML: The hybrid approach showed that GNNs can "fix" the known biases in mechanistic models (like the positive error bias in UNIFAC), suggesting a path forward where ML complements rather than replaces classical thermodynamics.

Key Results

  • Superior Accuracy: The GNN ensemble achieved a Mean Absolute Error (MAE) of 3.91 on the test set, outperforming the widely used UNIFAC-Dortmund (MAE 8.54) and COSMO-RS (MAE 10.78).
  • Robustness: Even the "single" GNN model performed competitively with state-of-the-art mechanistic methods while maintaining much broader applicability.
  • Hybrid Performance: The hybrid GNN-COSMO-RS model significantly tightened the error distribution compared to the standalone COSMO-RS model, effectively acting as a data-based correction layer for quantum mechanical calculations.

Limitations and Future Work

  • Temperature Constraints: The current model is trained only for data at 298.15 K. Real-world industrial processes operate across a wide range of temperatures, so incorporating temperature dependency is a necessary next step.
  • Chemical Scope: While excellent for organic systems, the model has not yet been tested on ionic liquids or inorganic salts, which are increasingly important in green chemistry.
  • Composition Dependency: This work focuses strictly on infinite dilution. Extending the GNN architecture to predict activity coefficients at any concentration (the entire composition range) remains a significant open challenge for the field.

Summary updated

Abstract

Graph neural networks were trained for the prediction of infinite dilution activity coefficients.
BibTeX
@inproceedings{medina2022graph,
  author = {Edgar Iván Sánchez Medina and Steffen Linke and Martin Stoll and Kai Sundmacher},
  title = {{Graph neural networks for the prediction of infinite dilution activity coefficients}},
  booktitle = {Digital},
  year = {2022},
  doi = {10.1039/d1dd00037c},
}
Figure