Anyone operating large chemical plants constantly has to make thousands of decisions. Which temperatures are optimal? How should a column be operated? What impact do fluctuating raw material qualities have? And how can an optimal product be achieved using as little energy as possible?
In many plants, questions like these are anything but trivial. The individual process steps influence one another, the composition of the raw materials is constantly changing, and there is often not just one right answer, but numerous ways to operate the plant properly. Accordingly, the challenge lies in finding the best option among countless possibilities.
Evonik Oxeno is working on this. With the COMPAS project, the company has developed a technology that combines decades of process engineering knowledge with methods of artificial intelligence.
The goal: to operate production plants even more efficiently, react faster to changing conditions, and unlock energy-saving potential. The technology is already in use at the Marl, Germany, and Antwerp, Belgium, sites.
ELEMENTS-Newsletter
Get exciting insights into Evonik’s research and its relevance to society—conveniently delivered via email.
When Experience Alone Is No Longer Enough
Chemical production plants are among the most complex technical systems. Many processes taking place inside cannot be directly observed. At the same time, a vast number of variables interact with one another. Even minor changes in temperature, pressure, or material flows can have a significant impact on the overall process.
For this reason, experts have relied on process simulations for decades. Such models mathematically describe chemical and physical relationships and make it possible to predict a plant's behavior. However, these calculations are usually computationally intensive. As a result, they are of limited use for making rapid decisions during ongoing operations.
Learning with Physics in the Background
The COMPAS team therefore took a different approach. Instead of recalculating every conceivable plant state over and over again, it first simulated more than one million possible operating scenarios. Based on this data, AI models were developed that can reproduce the results of these complex simulations in fractions of a second.
"We combine our many years of knowledge about chemical processes with the capabilities of artificial intelligence. This allows us to make complex process simulations directly usable for plant operations and support the most efficient control strategy possible, even under changing operating conditions," says Dr. Niklas Paul, Head of Engineering at Evonik Oxeno.
The term artificial intelligence often evokes the image of algorithms that learn exclusively from data. In an industrial environment, however, this approach quickly reaches its limits. Historical operating data only captures a fraction of all possible plant states. Some situations occur rarely, while others have never been observed at all.
Grey-box models
That is why COMPAS relies on so-called grey-box models. They combine the strengths of traditional process simulations with the capabilities of artificial intelligence. In the first step, the process is described using classic simulations based on physical models and experimentally determined material data, and then validated against real operating data. This serves as the foundation for creating the most accurate digital representation of the process possible. Using the validated simulation, the researchers subsequently generate a comprehensive training dataset that goes far beyond the volume of data available in actual practice. Finally, this dataset serves as the basis for training specially developed, custom-programmed AI models.
The result is a kind of digital twin that understands the most important laws governing the process while reacting fast enough to be deployed during ongoing operations. Instead of calculations taking hours, the system delivers a recommendation within seconds on which operating mode is particularly efficient under the current conditions.
What are grey-box models?
Grey-box models connect two worlds: traditional scientific process models and artificial intelligence. The physical models provide the fundamental understanding of a process. The AI uses this knowledge to recognize patterns and make very rapid predictions. This results in models that are both robust and practical for real-world operations.
Millions of Simulations for Better Separation
The benefits of the technology are particularly impressive in the production of 1-butene. This important chemical raw material is derived from complex C4 hydrocarbon mixtures. The challenge: many of the molecules it contains share similar properties and can only be separated from one another with a significant input of energy.
Traditionally, such plants are often operated with safety margins to maintain the required product quality under all circumstances. However, this increases energy demand. The COMPAS models help identify the exact operating points at which product specification and energy consumption are optimally balanced.
To do this, the system continuously processes current process data and provides the results to plant operators via a dashboard. The final decision remains with the human operator, while the software delivers a scientifically sound recommendation.
The effect is measurable: in the initial applications, specific steam consumption dropped by around six percent. At the same time, CO₂ emissions were reduced by more than 8.8 percent. In the long term, the team even sees savings potential of between 9 and 18 percent. These savings have also partially offset the sharp rise in energy prices seen recently.
Optimization at the Heart of the C4 Verbund
Another field of application is olefin oligomerization. In this process, smaller molecules are combined into larger building blocks that are later used in a wide variety of chemical products. For example, butenes are turned into octenes, dodecenes, or hexadecenes.
This process, too, is governed by numerous influencing factors. Raw material compositions change, catalysts age, and various process parameters are closely interconnected. Optimizing it is therefore like solving a highly complex puzzle.
The AI models make it possible to analyze such relationships much faster and more comprehensively than with conventional methods. In historical operating data, COMPAS identified several operating modes that would noticeably reduce energy demand without jeopardizing production targets or safety requirements. In some cases, even minor adjustments to individual parameters were enough to achieve a significantly more efficient mode of operation.
The Human Factor Remains Crucial
No matter how powerful the models are, responsibility deliberately remains with the people working in the plants. The recommendations are not implemented automatically, but are evaluated and approved by experienced specialists. The developers see precisely this combination of human experience and data-driven optimization as a key success factor. The software is not meant to replace employee expertise, but to expand it. At the same time, it provides new insights into the interrelationships of complex production processes.
"COMPAS shows that the knowledge jointly built up over many years by Operations, Engineering, and Research creates new value in process optimization when combined with artificial intelligence. Based on this technology, we are working step by step toward a C4 Verbund that will optimize itself in the future," says Professor Robert Franke, Head of Simulation & Kinetics at Evonik Oxeno.
The Path to the Self-Optimizing Chemical Plant
For now, COMPAS only provides recommendations. Yet the technology already shows what the future of chemical production could look like. Looking ahead, the models developed are set to work directly with digital control systems. Plants could then continuously analyze which operating mode is optimal under current conditions and make corresponding adjustments in near real time.
The long-term vision even extends beyond individual plants. In the future, entire integrated production networks could independently coordinate and optimize their operations. Raw material fluctuations, energy input, and product yields would all be taken into account simultaneously. What still seems like a future scenario today is already taking concrete shape with technologies like COMPAS.