Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions

Fabian Jirasek, Robert Bamler, Sophie Fellenz, Michael Bortz, Marius Kloft, Stephan Mandt, Hans Hasse

Abstract

Predictive models for the thermodynamic properties of mixtures are essential in chemical engineering and chemistry. While classical models excel at generalizing over continuous conditions like temperature and concentration, they often struggle with discrete binary systems where experimental data is missing. Conversely, machine learning matrix completion methods (MCMs) are adept at generalizing across discrete systems but lack physical grounding. This paper presents MCM-UNIQUAC, a hybrid approach that embeds an MCM within the UNIQUAC physical model. By predicting pair-interaction energies, the model achieves complete coverage for over 650,000 binary systems and outperforms the industry-standard modified UNIFAC model.

The Challenge of Mixture Complexity

In chemical engineering, designing processes such as distillation or drug formulation requires precise knowledge of how different chemical components interact, typically quantified by activity coefficients. However, the chemical space is vast. With over $10^8$ known components, the number of possible binary mixtures is astronomical. Currently, experimental data exists for less than 2% of the most common binary systems.

Traditional physical models, such as UNIQUAC, are effective at interpolating across temperatures and concentrations but require specific experimental data to tune parameters for every unique pair of chemicals. Group-contribution methods like UNIFAC attempt to circumvent this by breaking molecules into functional groups, but they fail when specific group-interaction parameters are missing from their databases.

The MCM-UNIQUAC Hybrid Approach

To address these limitations, researchers developed MCM-UNIQUAC. This model bridges the gap between two distinct methodologies:

  1. Matrix Completion Methods (MCM): Inspired by recommender systems used in digital platforms, MCMs learn "latent features" of chemicals. These features allow the model to predict interactions between two chemicals even if they have never been tested together experimentally.
  2. UNIQUAC Framework: A well-established physical lattice model that describes the Gibbs excess energy of a mixture.

The innovation lies in the integration: the machine learning component does not predict the final thermodynamic property directly. Instead, it predicts the pair-interaction energies ($U_{ij}$), which are then used as inputs for the physical UNIQUAC equations. This ensures that the model's outputs remain thermodynamically consistent while benefiting from the predictive power of machine learning. The model was trained using a Bayesian framework (Variational Inference) to rigorously account for uncertainty.

Data and Training

The model was trained and validated using the Dortmund Data Bank (DDB), one of the most comprehensive repositories of thermodynamic data:

  • Scope: 1,146 distinct chemical components.
  • Dataset: 363,181 experimental data points for activity coefficients across 12,199 unique binary systems.
  • Conditions: A wide temperature range from 183 K to 638 K.

Despite the large number of data points, the 12,199 systems represent a tiny fraction of the potential combinations, highlighting the need for robust predictive capabilities.

Key Results and Insights

The hybrid model demonstrated several significant advantages over existing methods:

  • Superior Accuracy: MCM-UNIQUAC outperformed the modified UNIFAC (Dortmund) model—the current "gold standard"—in predicting activity coefficients on held-out test sets.
  • Complete Parameter Coverage: For the 1,146 components studied, the model provides a complete set of parameters for all 656,085 possible binary systems. This enables modeling for millions of mixtures that previously lacked experimental or group-contribution data.
  • Zero-Shot Generalization: By learning underlying molecular features, the model can perform "zero-shot" predictions for binary pairs entirely absent from the training set.
  • Multicomponent Extrapolation: Because the model is built on fundamental pair interactions, it can predict the behavior of ternary (3-component) or higher mixtures using only the parameters learned from binary data. It successfully predicted complex phase behaviors, such as azeotropes, in both binary and ternary systems.
  • Physical Consistency: While standard UNIQUAC often uses asymmetric parameters ($Δ U_{ij} \neq \u0394 U_{ji}$) to fit data, the authors found that a single, physically consistent symmetric interaction energy ($U_{ij} = U_{ji}$) predicted by the MCM achieved comparable accuracy while maintaining theoretical integrity.

Conclusion and Future Outlook

MCM-UNIQUAC represents a significant step forward in making thermodynamic modeling truly predictive. By embedding machine learning within a physical framework, the model overcomes the data scarcity that plagues traditional methods.

However, some limitations remain. The model relies on the "mean-field" approximation of lattice theory, which may be overly simplistic for highly non-ideal or extremely complex mixtures. Future work could involve extending this framework to other thermodynamic models, such as Equations of State (EoS) or the NRTL model, to handle high-pressure systems or properties like excess enthalpies. This hybrid approach provides a scalable template for integrating machine learning into classical physical sciences.

Summary updated

Abstract

Predictive models of thermodynamic properties of mixtures are paramount in chemical engineering and chemistry. Classical thermodynamic models are successful in generalizing over (continuous) conditions like temperature and concentration. On the other hand, matrix completion methods (MCMs) from machine learning successfully generalize over (discrete) binary systems; these MCMs can make predictions without any data for a given binary system by implicitly learning commonalities across systems. In the present work, we combine the strengths from both worlds in a hybrid approach. The underlying idea is to predict the pair-interaction energies, as they are used in basically all physical models of liquid mixtures, by an MCM. As an example, we embed an MCM into UNIQUAC, a widely-used physical model for the Gibbs excess energy. We train the resulting hybrid model in a Bayesian machinelearning framework on experimental data for activity coefficients in binary systems of 1146 components from the Dortmund Data Bank. We thereby obtain, for the first time, a complete set of UNIQUAC parameters for all binary systems of these components, which allows us to predict, in principle, activity coefficients at arbitrary temperature and composition for any combination of these components, not only for binary but also for multicomponent systems. The hybrid model even outperforms the best available physical model for predicting activity coefficients, the modified UNIFAC (Dortmund) model.
BibTeX
@inproceedings{jirasek2022making,
  author = {Fabian Jirasek and Robert Bamler and Sophie Fellenz and Michael Bortz and Marius Kloft and Stephan Mandt and Hans Hasse},
  title = {{Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions}},
  booktitle = {Chemical},
  year = {2022},
  doi = {10.1039/d1sc07210b},
}
Figure