Artificial Intelligence for Chemical Processes

In the research group »KI-FOR«, we are working with our partners to investigate how Artificial Intelligence can be used for monitoring and analyzing chemical processes, even when only a limited amount of meaningful data is available. To achieve this, we combine mathematical models, simulations, and advanced AI methods to detect anomalies and critical conditions at an early stage.

In the first funding phase, we successfully demonstrated a »proof of principle« and showed that the combination of simulation and AI provides real added value. In the second funding phase, we are expanding the developed methods and investigating their transferability to other plants, processes, and operating conditions.

Why AI in the Chemical Industry Has to Overcome Particular Challenges

Chemical plants often generate only a limited amount of meaningful data for AI applications because their processes are usually highly consistent and anomalies are difficult to detect. Compared with applications such as image or speech processing, the chemical industry has significantly fewer suitable training data available. Although modern plants collect increasing amounts of measurement data, many processes operate under stable conditions for long periods of time. As a result, the variety of operating states required for AI models to learn reliably is often missing.

Another challenge is that measurements are frequently taken only at a few points within a plant. Sensors, for example, record temperatures, pressures, or concentrations at the inlet and outlet of a system. What happens between these measurement points often remains hidden. In addition, critical conditions or rare anomalies occur only occasionally. This makes it difficult for conventional learning approaches to identify patterns and distinguish between normal and unusual behavior.

In the first project phase, the team developed new methods for anomaly detection and provided extensive open datasets for chemical process data. The results show that modern AI methods can significantly improve the monitoring and analysis of chemical processes.

Mathematics and Simulation Close Data Gaps

However, AI alone is not sufficient to achieve optimal results. Our researchers therefore combine real measurement data with mathematical models and simulations. These models describe the physical and chemical processes within a plant and provide additional information about states that cannot be measured directly.

Based on this approach, realistic synthetic data are generated. They complement the available measurement data and help AI methods gain a better understanding of processes. This provides algorithms with significantly more information than can be captured by sensors alone. At the same time, mathematical models improve the traceability and interpretability of the results. The analysis is not based solely on data, but also takes known physical relationships into account.

An important component of the project is the use of Digital Twins. Put simply, we have developed digital representations of plants that can run "in parallel" with the real system in the future. The simulations continuously provide calculated states that are compared with the actual measurement data from the plant. Deviations between simulated and measured values can indicate unusual operating conditions or potential faults. This allows anomalies to be detected and assessed at an early stage.

An der Batch-Destillationsanlage der Forschungsgruppe »KI-FOR« untersuchen Forschende chemische Prozesse und erzeugen hochwertige Datensätze für Maschinelles Lernen und Digitale Zwillinge. — © RPTU Kaiserslautern-Landau
Die Batch-Destillationsanlage in Kaiserslautern liefert wertvolle Prozessdaten für die Forschung an KI-gestützten Methoden zur Analyse und Optimierung chemischer Prozesse.

Adaptive AI for New Plants and Processes

For complex industrial processes, simply collecting large amounts of data is not enough. Reliable analysis and evaluation of processes can only be achieved by combining expert knowledge, mathematical models, and simulated data.

The focus of the second funding phase is on transferring the developed methods to new applications. To this end, we are developing so-called »adaptive AI methods«. These methods can adjust to changing conditions. The team is investigating how the approaches can be applied to different plants, material mixtures, and operating conditions.

© Piotr Banczerowski / Fraunhofer-Gesellschaft
Prof. Dr. Michael Bortz giving a lecture on Applied Mathematics in Process Engineering.

On the Way to Autonomous Plants?

Fully autonomous chemical plants are still a vision of the future. Although automated systems already take over individual tasks, such as condition monitoring or predictive maintenance, humans remain responsible for safety-critical decisions.

In the long term, increasingly powerful Digital Twins and intelligent assistance systems offer great potential. They accompany processes, analyze operating conditions, and support the optimization of plants.

Bringing research and industry together

At Fraunhofer ITWM, we combine fundamental research with industrial practice. The insights gained from research projects like this one are directly incorporated into our collaborations with companies. At the same time, real-world applications help to further develop scientific methods.

By combining mathematics, Artificial Intelligence, and expertise from process engineering, we create solutions that are not only scientifically sound but also prove their value in industrial practice. The goal is to make production processes safer, more efficient, and more sustainable in the long term.

© KI-FOR
Gruppenbild bei Meeting der Forschungsgruppe »KI-FOR: Deep Learning auf dünnbesetzten, chemischen Prozessdaten«.

Our Project Partners

Rhineland-Palatinate Technical University of Kaiserslautern-Landau (RPTU), Prof. Dr. Marius Kloft, Research Group »Machine Learning« (Speaker)
Technical University of Munich (TUM)
TU Dortmund

Project Funding and Duration

The DFG research group »KI-FOR: Deep Learning on Sparse Chemical Process Data« has been investigating since 2022 how modern AI methods can be successfully applied to chemical processes, even when only limited data are available. Following successful results in the first funding phase, the German Research Foundation (DFG: Deutsche Forschungsgemeinschaft) approved a second funding phase and has been supporting the project for another five years since 2026.

Artificial Intelligence for Chemical Processes: Detecting Anomalies Early Despite Limited Data Availability

Why AI in the Chemical Industry Has to Overcome Particular Challenges

More Information