Public final conference at DECHEMA

Which added value can artificial intelligence methods create for the process industry? After three years, the KEEN partners presented their results at a public final conference on May 22 and 23, 2023 at DECHEMA. Fourteen talks presented examples of successful applications of AI in collaboration between industry and research partners.

At the Laboratory of Engineering Thermodynamics (LTD) and the Machine Learning group of the Rheinland-Pfälzischen Technischen Universität (RPTU) Kaiserslautern-Landau, new methods for the prediction of substance data of mixtures were developed. These are based on matrix completion methods, as they are also used for recommendations for movies or music. In cooperation with Fraunhofer ITWM, this resulted in the software prototype "MatrixMole", which was tested with the partners Dortmund Data Bank (DDBST GmbH) and Covestro. As a result, the user receives statements about thermodynamic properties of binary substance mixtures that have not been measured. In addition, the uncertainty of the prediction is quantified. Based on the improved prediction capabilities for thermodynamic properties of mixtures, AI-based surrogates were used for process simulation in a second step. One focus was the development of online models with which the user can interactively investigate what-if scenarios. Hereby, optimization potential can be reliably and transparently identified.

In plant engineering, the use of AI has the potential to shorten the development time in basic as well as detail engineering. As part of KEEN, the Laboratory of Equipment Design at TU Dortmund University developed an AI-based algorithm that uses trained recurrent neural networks (RNNs) to suggest possible subsequent components for the user while creating R&I flow diagrams. Graph Neural Networks (GNNs) are used to perform a consistency check after the drawing. In the concept development phase, an AI algorithm suggests potentially successful separation operations for mixtures of substances. The developed tools are currently being tested by the partners in their software packages and will be available to other users in the near future.

As a consortium partner, Bayer provided a use case that involved the development of a column control system using data-based AI models. In perspective, plant operators might be supported by an increased plant autonomy via such an approach. In collaboration with TU Dortmund University, a rigorous dynamic column model was created and an ML-based control approach was tested to achieve optimal operation in terms of material and energy efficiency despite disturbances within the operating specifications. The developed solution was compared to a conventional PID control system and achieved a similar performance. "Despite the requirement for operating data, which is not always available at real plants, Bayer is pursuing AI-based methods, especially in modeling and process control," said Stefan Krämer (Bayer). Bayer is using AI in many areas of its business and operations to accelerate product development and to improve the way products are manufactured and delivered to customers. Bayer is committed to using AI - like all existing and new technologies - responsibly and in compliance with applicable laws and regulations.

An important basis of AI is a systematic data management, emphasized Ralph-Müller Pfefferkorn (TU Dresden, ZIH) in his presentation. He presented the modular metadata schema for describing process industry data ProMetaS (Process Engineering/Industry Metadata Schema) as well as the KEEN data platform for sharing and publishing data. During the project period, datasets from various academic and industrial KEEN partners were uploaded to the platform. The datasets are available for use by the scientific public (to be found at

In the final panel discussion entitled "Are we still KEEN on AI?", Marius Kloft (TU Kaiserslautern), Stefan Krämer (Bayer), Joschka Winz (TU Dortmund University, Process Dynamics and Operations Group), Armin Fricke (Capital-Gain Consultants), Tom Kraus (iit) and Sibylle Mutschler (Postfinance) discussed the value of data and the future of AI in the process industry. "KEEN gave me a realistic idea of what AI can and cannot do," Joschka Winz, a PhD student at TU Dortmund University, summarized his experience.

Keynote lecture:

Are we still keen on AI?
Sibylle Mutschler (Postfinance & KEEN Spiegelgremium)
The results to download:
From thermodynamics to processes: AI facing small data and big models
Michael Bortz (Fraunhofer ITWM)
Machine Learning Meets Physical Modeling: Hybrid Thermodynamic Models
Nicolas Hayer, Stephanie Peper, Fabian Jirasek & Hans Hasse (Universität Kaiserslautern)
Workflow for Integrating Surrogate Models into a Complex Flowsheet (Ammonia Plant)
Armin Fricke & Ivana Lukec (Capital-Gain Consultants) – Präsentation gern auf Anfrage
Explainable Optimization via explainable AI
Dominik Schack, Robin Schmidt, Vanessa Gepert (AirLiquide) & Marco Baldan, Patrick Ludl, Michael Bortz (Fraunhofer ITWM)
Smart Engineering – KI-Assistenz in Prozesssimulation und R&I-Planung
Norbert Kockmann (TU Dortmund) & Wolfgang Welscher (X-Visual Technologies)
Selbstoptimierende Anlage mit KI – wie weit sind wir gekommen?
Sebastian Engell (TU Dortmund)
Kognitive Sensoren als Voraussetzung für KI-Algorithmen
Laura Neuendorf (TU Dortmund) & Christiane Schlander (Merck)
Recipe optimization of batch distillation trajectories based on a data-driven model
Gerardo Brand Rihm, Erik Esche, Jens-Uwe Repke (TU Berlin) & Merlin Schüler, Corina Nentwich, Michael Kawohl (Evonik)
PID bis KI: optimale Kolonnenregelung in KEEN - was lernen wir daraus?
Stefan Krämer, Yak Ortmanns, Jörn Felix Hecht, Volker Roßmann (Bayer) & Mohamed Elsheikh, Sebastian Engell (TU Dortmund)
Development and utilization of a dynamic gray-box model for a fermentation process of a sporulating bacterium
Joschka Winz, Sebastian Engell (TU Dortmund) & Supasuda Assawajaruwan, Uwe Piechottka (Evonik)
Viele Daten, wenig Information – neue Wege zur Nutzung operativer Prozessdaten Leon Urbas (TU Dresden)
Systematisches Datenmanagement als Grundlage für KI
Ralph Müller-Pfefferkorn, Lincoln Sherpa, Valentin Khaydarov, Leon Urbas (TU Dresden) & Gregor Tolksdorf, Michael Kawohl, Michael Wiedau (Evonik) & Udo Enste (Leikon) & Marco Gaertler (ABB) & Martin Krawczyk-Becker (Krohne) & David Wagner-Stürz (Samson Group)
Deep Learning for Computer Vision in Process Industry
Valentin Khaydarov, Leon Urbas (TU Dresden)
Active learning and transfer learning for process analytics
Chen Song, Ruomu Tan, Marco Gärtler, Martin Hollender, Sylvia Maczey (ABB) & Franz Baehner (Bayer) & Bram Bamps (Covestro)

KEEN – was bleibt, was kommt? Kai Dadhe (Evonik) & Martin Hoffmann (ABB)