Advancing Thermodynamic Group-Contribution Methods by Machine Learning - UNIFAC 2.0 - Enhanced prediction accuracy and expanded parameter coverage

Aug 4, 2026

Accurate thermodynamic-property prediction is crucial for chemical engineering, yet traditional physical group-contribution (GC) models suffer from historically incomplete parameterizations that limit both accuracy and usable chemical scope. A new framework combines GC with machine-learning matrix completion to infer a complete set of UNIFAC pair-interaction parameters, yielding UNIFAC 2.0. Training and validation on over 224,000 experimental points significantly reduce mean squared error and eliminate gaps in the original parameter table, while enabling straightforward updates with new data or application-specific tailoring.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages32

Reading Time1 h 21 min

AccessFree to download

Abstract

Motivation: limits of classical GC models

Method: combining GC with matrix completion for UNIFAC

UNIFAC 2.0 training, validation, and performance

Updating and extending the model with new data

arXiv :2408 .05220v1 [physics .chem-ph] 25 Jul 2024 Advancing Thermodynamic Group-Contribution Methods by Machine Learning: UNIFAC 2.0 Nicolas Hayer,† Thorsten Wendel,† Stephan Mandt,‡ Hans Hasse,† and Fabian Jirasek ∗ ,† †Laboratory of Engineering Thermodynamics, RPTU Kaiserslautern, Erwin-Schrödinger-Str. 44, 67663 Kaiserslautern, Germany ‡Department of Computer Science, University of California, Irvine, CA 92617, USA E-mail: [fabian.jirasek@rptu.de](fabian.jirasek@rptu.de) 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 pairinteraction 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. Main Understanding the thermodynamic properties of mixtures is indispensable in chemical engineering and various related disciplines. However, the vast combinatorial diversity of mixtures makes it impossible to study each relevant mixture experimentally, necessitating reliable prediction methods. Group-contribution (GC) methods address this challenge by deconstructing components into structural groups, significantly reducing the number of parameters since the number of structural groups is much smaller than those of individual components. These methods rely on modeling pair interactions between these structural groups to describe mixture behavior. The effectiveness of GC methods hinges on selecting suitable groups and accurately determining their interaction parameters, both of which depend crucially on the database used for method development and parameterization. Among GC methods, UNIFAC stands out as the most sophisticated and widely adopted approach for predicting activity coefficients in liquid mixtures. Since its introduction in 1975, 1 UNIFAC has undergone continuous refinement and improvement, 2–7 becoming integral to industrial process simulations. Available in both public 7 and commercial 8 formats, UNIFAC supports diverse applications, including variants like UNIFAC LLE 9 for predicting liquid-liquid equilibria. All UNIFAC variants rely on the same equations but differ in the number and type of groups considered and their parameterization. The process of finding suitable UNIFAC parameters was, in the past, sequential and based on a stepwise extension whenever data became available. This tedious process makes it very difficult to modify decisions taken at early steps. This study addresses the challenges of updating and improving UNIFAC by leveraging modern computational techniques, aiming to enhance prediction accuracy and expand its applicability across a broader range of components and mixtures. Throughout this work, we reference the latest published version of UNIFAC. It was trained on a broad data basis focusing on vapor-liquid equilibrium data to develop a widely applicable model, not one for some specific purpose. 7 It is astonishing that, despite the importance of UNIFAC, this version is about 20 years old. The leading developers of UNIFAC have updated the method since then, but they have not disclosed these updates – they are only available for members of th

FAQs

What problem does UNIFAC 2.0 address in classical group-contribution methods?

Classical GC methods, including earlier UNIFAC versions, contain historically incomplete pair-interaction parameter tables, which restrict applicability and reduce prediction accuracy.

How does the proposed approach generate the missing UNIFAC parameters?

It combines group-contribution modeling with machine-learning matrix completion to predict a complete set of pair-interaction parameters for UNIFAC.

What evidence is provided for the performance improvement of UNIFAC 2.0?

UNIFAC 2.0 is trained and validated on more than 224,000 experimental data points, showing significantly enhanced accuracy, including roughly halving mean squared error, and improved coverage by filling parameter gaps.

Aug 4, 2026

Upload by Clementine