Evonik Oxeno, the German chemical company's C4 chemicals business, said Nov. 9 it has introduced AI-based process models into live operations at production sites in Marl, Germany, and Antwerp, Belgium. The models generate recommendations for plant operating conditions within seconds, helping engineers and operators adjust processes as feedstock composition and other operating conditions change.

Developed through the company's COMPAS project (Control and Optimization using AI Models for Process Advancement and Cost Saving), the models draw on more than 1 million pre-calculated process simulations. The resulting AI-based models, known as gray-box models, combine artificial intelligence with descriptions of process thermodynamics and kinetics to predict plant behavior across different operating conditions.

The models are designed to address limitations of conventional process simulations, which can be computationally intensive and time-consuming. By using previously calculated scenarios, COMPAS can quickly identify operating conditions intended to minimize energy consumption while maintaining product specifications.

Two applications involve separating 1-butene from a C4 feed stream through distillation and optimizing the oligomerization of butenes into higher-molecular-weight olefins. In a paper published Oct. 14 in Chemical Engineering & Technology, Evonik researchers detailed the methodology and its application in both processes. The study reported that use of a model-based operator dashboard in the 1-butene distillation process was associated with an average 6% reduction in daily steam consumption following its introduction in late 2024, while maintaining product specifications within defined limits.

Evonik said the models also have demonstrated close agreement with actual operating data and have helped reduce energy consumption and associated CO₂ emissions. Additional applications are being developed, including automated transfer of AI-generated results to plant control systems. The company is working toward a self-optimizing C4 production network, although its published methodology describes operators reviewing model recommendations before making operating adjustments.

The development follows other work by Evonik Oxeno to incorporate lower-carbon approaches into chemical production. In October 2025, the company and research partners at the Leibniz Institute for Catalysis and Ruhr University Bochum developed a catalyst system that uses carbon dioxide and green hydrogen in place of carbon monoxide for carbonylation reactions to produce esters used in fragrances and plastic precursors. The research also explored alternative feedstocks for industrial chemical production.

Why it Matters

The COMPAS project demonstrates how AI-based surrogate models can make complex process simulations more accessible for day-to-day plant operations. Rather than waiting for computationally intensive simulations, engineers can use the models to evaluate potential operating adjustments quickly, particularly when feedstock composition or other operating conditions change.

The development follows several recent efforts to bring AI into industrial operations. On Oct. 26, Honeywell announced plans to deploy an AI platform at Dangote's Nigerian refinery to identify production constraints and support operating decisions. On Oct. 21, Schneider Electric, NVIDIA and Hewlett Packard Enterprise announced plans for edge-based AI agents targeting control-system monitoring, cybersecurity and energy optimization. Earlier, on Oct. 16, Seeq and Applied Computing announced a partnership combining AI-based anomaly detection with industrial analytics to help investigate equipment and process problems.

Unlike some of these announced collaborations and planned deployments, Evonik reports that its models are already operating in production environments. The published study provides an example of measurable energy savings, while maintaining product specifications. The company is now working to expand the methodology to other chemical processes and further integrate model outputs with plant control systems.