Advancing Thermodynamic Group-Contribution Methods by Machine Learning: UNIFAC 2.0

Nicolas Hayer, Thorsten Wendel, Stephan Mandt, Hans Hasse, Fabian JirasekarXiv 2024 · DOI · OpenReview

TL;DR

This paper introduces UNIFAC 2.0, a machine learning-enhanced version of the UNIFAC group-contribution method for predicting thermodynamic properties of liquid mixtures. By leveraging matrix completion, UNIFAC 2.0 fills critical parameter gaps in the original UNIFAC model, significantly improving prediction accuracy and expanding its applicability to a much wider range of chemical systems.

What Problem Does This Paper Address?

Accurate prediction of thermodynamic properties is crucial for designing and optimizing chemical processes. Group-contribution (GC) methods, like UNIFAC, are widely used for this purpose because they break down complex molecules into smaller, recurring structural groups, reducing the number of parameters needed. However, the traditional UNIFAC method suffers from incomplete parameterization, meaning it cannot predict properties for many mixtures due to missing interaction parameters between these groups. This limitation restricts its applicability and can lead to inaccurate predictions.

Methods

The core innovation of this work is the integration of machine learning, specifically matrix completion methods (MCM), into the UNIFAC framework. Instead of relying on sequential, manual fitting of parameters, UNIFAC 2.0 treats the pair-interaction parameters as entries in a matrix. MCM is then used to learn "group features" from experimental data and predict all missing entries in this interaction parameter matrix. This approach allows for an end-to-end training process, where the entire model is optimized based on experimental data, rather than relying on a piecewise, intuition-guided parameterization. The authors employ a Bayesian probabilistic model within the MCM framework, treating parameters and predictions as random variables to capture uncertainty.

Datasets

The primary dataset used for training and validation is the Dortmund Data Bank (DDB), described as the most extensive database for thermodynamic properties. The authors utilized over 224,000 experimental data points from this database. They specifically focused on binary mixtures whose components could be decomposed into UNIFAC subgroups.

Key Insights and Ideas

  • Bridging the Gap: The most significant insight is that machine learning, particularly matrix completion, can effectively "fill in the blanks" of incomplete parameter tables in established physical models like UNIFAC. This directly addresses a long-standing limitation of GC methods.
  • End-to-End Learning: Moving from sequential parameter fitting to end-to-end training of the entire model (including group features and interaction parameters) leads to a more robust and accurate prediction system.
  • Robustness to Missing Data: The MCM approach inherently handles missing data by learning underlying patterns, making UNIFAC 2.0 less susceptible to the limitations imposed by incomplete experimental datasets for specific parameter pairs.
  • UNIFAC 1.0 is Outdated: The paper highlights that the widely used UNIFAC 1.0 is about 20 years old and that its developers have not publicly released updates, creating a need for accessible improvements.

Key Results

  • Enhanced Accuracy: UNIFAC 2.0 significantly improves prediction accuracy, nearly halving the mean squared error (MSE) compared to UNIFAC 1.0.
  • Expanded Applicability: UNIFAC 2.0 can be applied to all mixtures in the DDB, whereas UNIFAC 1.0 is limited to predicting only 58% of binary mixtures due to missing parameters. This applicability extends dramatically to multi-component mixtures.
  • Extrapolation Capability: UNIFAC 2.0 demonstrates strong predictive performance even for components or pair-interactions not explicitly seen during training, highlighting its robustness and generalization ability.
  • Simplified Implementation: The hybrid model retains the UNIFAC framework, making it easy to implement by simply replacing the parameter table in existing process simulators.

Limitations and Future Work

  • Symmetric Model Performance: While a symmetric UNIFAC model (UNIFAC 2.0 sym) was developed for greater physical consistency, it showed slightly reduced predictive accuracy compared to the asymmetric UNIFAC 2.0. The authors focused on the asymmetric version due to its better performance.
  • Future Work: The authors suggest that the end-to-end training approach allows for easy updates if new data become available or for tailoring the model to specific needs. The developed methodology could be applied to other group-contribution methods.

Summary updated

Abstract

Accurate prediction of thermodynamic properties is pivotal in chemical engineering for optimizing process efficiency and sustainability. Physical group-contribution (GC) methods are widely employed for this purpose but suffer from historically grown, incomplete parameterizations, limiting their applicability and accuracy. In this work, we overcome these limitations by combining GC with matrix completion methods (MCM) from machine learning. We use the novel approach to predict a complete set of pair-interaction parameters for the most successful GC method: UNIFAC, the workhorse for predicting activity coefficients in liquid mixtures. The resulting new method, UNIFAC 2.0, is trained and validated on more than 224,000 experimental data points, showcasing significantly enhanced prediction accuracy (e.g., nearly halving the mean squared error) and increased scope by eliminating gaps in the original model's parameter table. Moreover, the generic nature of the approach facilitates updating the method with new data or tailoring it to specific applications.
BibTeX
@misc{hayer2024advancing,
  author = {Nicolas Hayer and Thorsten Wendel and Stephan Mandt and Hans Hasse and Fabian Jirasek},
  title = {{Advancing Thermodynamic Group-Contribution Methods by Machine Learning: UNIFAC 2.0}},
  year = {2024},
  eprint = {2408.05220},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.05220},
}
Figure