HybridGamma
- Author(s)
- Ulderico Di Caprio, Jan Degrève, Peter Hellinckx, Steffen Waldherr, M. Enis Leblebici
- Abstract
Predicting molecular interactions is a crucial step for chemical process modeling. It requires the full knowledge of the analyzed system, however, this is often impossible in complex real-world cases. Machine learning (ML) techniques overcome this bottleneck and enhance systems predictability using data. Hybrid modeling (HM) is an established technique combining first-principle information and ML techniques. This work introduces a mathematical framework to predict activity coefficients employing HM approach. The obtained models are physically consistent and can handle systems with unknown components or external sources of deviation. The framework is validated on experimental and in-silico cases employing different training approaches. In all the tested cases, the HM showed remarkable prediction capabilities with coefficients of determination R2 above 0.98 for the predicted variables. This work proposes and develops a novel way to approach the HM of molecular interactions by embedding physical laws within the model structure. We encountered three main benefits in applying thermodynamically consistent HMs for activity coefficients: the reduction of tuneable parameters, the increased prediction capabilities, and the physically-consistent behavior of the model.
- Organisation(s)
- Functional and Evolutionary Ecology
- External organisation(s)
- Katholieke Universiteit Leuven, University of Antwerp
- Journal
- Chemical engineering journal
- Volume
- 475
- ISSN
- 1385-8947
- DOI
- https://doi.org/10.1016/j.cej.2023.146104
- Publication date
- 11-2023
- Peer reviewed
- Yes
- Austrian Fields of Science 2012
- 204003 Chemical process engineering
- Keywords
- ASJC Scopus subject areas
- General Chemistry, Environmental Chemistry, General Chemical Engineering, Industrial and Manufacturing Engineering
- Portal url
- https://ucrisportal.univie.ac.at/en/publications/dfc3b1c8-76ee-45fb-87d6-2e16da218498