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 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 machine-learning 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.

Introduction

In chemical engineering, understanding the thermodynamic behavior of chemical mixtures is critical for designing industrial processes like distillation and pharmaceutical formulation. However, researchers face a "combinatorial explosion": with millions of known chemicals, the number of possible binary (two-component) mixtures is astronomical, making exhaustive experimental testing impossible.

Traditional physical models, such as UNIQUAC, excel at predicting how a specific mixture's properties change with temperature or concentration. However, they typically require experimental data for every specific pair of chemicals to determine interaction parameters. Conversely, Group-Contribution (GC) methods like UNIFAC generalize by analyzing molecular fragments, but they often struggle with rare fragments or complex molecular architectures that lack sufficient experimental coverage. This leaves significant gaps in our ability to predict the behavior of many chemical systems.

Methodology: The MCM-UNIQUAC Hybrid

To bridge these gaps, researchers developed MCM-UNIQUAC, a hybrid model that integrates a Matrix Completion Method (MCM)—an algorithm similar to those used in recommendation systems—directly into the physical UNIQUAC framework.

The MCM-UNIQUAC framework combines physical modeling with matrix completion to generalize across binary systems and thermodynamic conditions.

The approach follows three key steps:

  1. Latent Feature Learning: The MCM learns "latent features" (mathematical fingerprints) for 1,146 different chemicals based on existing data.
  2. Predicting Interaction Energies: These features are used to predict the pair-interaction energies ($U_{ij}$) between any two molecules, including those never tested together in a laboratory setting.
  3. Physical Integration: These predicted energies are plugged into classical UNIQUAC equations. This ensures the model adheres to thermodynamic laws, allowing it to generalize across continuous variables like temperature and pressure.

Notably, the authors found that using a single symmetric interaction energy ($U_{ij} = U_{ji}$) maintained high accuracy while improving physical consistency and simplifying the machine learning task compared to standard asymmetric parameters.

Data and Training

The model was trained using the Dortmund Data Bank (DDB), a premier repository of thermodynamic data. The dataset included 363,181 experimental measurements of activity coefficients across 1,146 distinct chemical components. Despite the large number of data points, the experimental coverage is sparse: data exists for only about 12,200 (less than 2%) of the 650,000+ possible binary combinations of these components.

Key Results and Insights

The MCM-UNIQUAC model demonstrated several significant advantages:

  • Superior Accuracy: The hybrid model achieved a lower Mean Absolute Error (MAE) on unseen test data than the modified UNIFAC (Dortmund) model, the current industry gold standard.
  • Comprehensive Parameterization: For the first time, a complete set of UNIQUAC parameters was generated for all 656,085 possible binary systems among the studied components.
  • Zero-Shot Extrapolation: By focusing on pair interactions, the model successfully predicted the vapor-liquid equilibria of ternary (three-component) mixtures without having seen data for those specific complex systems during training.
  • Qualitative Success: The model accurately captured complex phase behaviors, such as high-boiling azeotropes and heteroazeotropes, in systems entirely absent from the training set.

Conclusion and Future Directions

By predicting intermediate physical parameters (interaction energies) rather than final properties (activity coefficients), the model gains a "physical common sense" that simple black-box ML models lack. It effectively maps chemicals into a multi-dimensional "latent space" where molecules with similar interaction profiles are clustered together.

While the model relies on the mean-field lattice assumptions of UNIQUAC—which might be challenged by extremely non-ideal systems—its success suggests a new paradigm for thermodynamic modeling. Future work may extend this hybrid MCM approach to other frameworks, such as Equations of State (EoS) or the NRTL model, to predict properties like excess enthalpies and heat capacities.

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 <i>pair-interaction energies</i>, 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 machine-learning 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 <i>complete</i> 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