Machine-learning models that infer a hard-to-measure variable — composition, quality, emissions — from cheaper, faster process signals already being collected.
AI-driven closed-loop advanced process control (APC 2.0)
Control Systems — DCS · PLC · ESD · HMI
Industry Acceptance
Near-term · 2–5 yrs
Evidence E3 · First production use
AI / ML
Machine learning and reinforcement learning applied to closed-loop process optimisation, alongside or above classical model predictive control. Specific commercial results exist; most of the evidence still comes from suppliers and their early customers.
What classical APC does well, and where it struggles
Advanced process control, usually model predictive control (MPC), has been standard on refinery and petrochemical units for decades. An MPC uses a dynamic model, identified from step tests on the plant, to move several manipulated variables at once and hold the unit close to its constraints. The benefits, typically more throughput, better yield and lower energy use, are well documented.
So are the weaknesses. The models are usually linear and degrade as the plant changes, so controllers need periodic re-testing and re-tuning. Highly nonlinear units, frequent grade changes and processes with poor measurements are hard to handle. When maintenance lapses, operators switch controllers off and the benefits erode, which is why service factor, the share of time a controller is on and active, is the number APC engineers watch most closely.
What the AI approaches change
The newer approaches use data-driven models, usually neural networks trained on years of plant history, and in some cases reinforcement learning, where a control policy is learned in simulation and then refined. They aim to capture nonlinear behaviour, adapt as the process changes, and reduce the step testing that classical APC needs.
Two published cases show the range. At a chemical plant of ENEOS Materials in Japan, Yokogawa's reinforcement-learning algorithm, Factorial Kernel Dynamic Policy Programming (FKDPP), controlled a distillation column autonomously for 35 days from January to February 2022. It held liquid levels and product quality while making maximum use of waste heat, on a column that had previously been adjusted by hand because PID control and APC could not cope. After a scheduled shutdown the trial continued for almost a year, and in March 2023 the AI was formally adopted for the plant. The companies report a 40 percent reduction in steam consumption and CO2 emissions, measured against the steam used to maintain liquid level under manual operation 1.
Separately, Imubit, a US supplier of what ARC Advisory Group has called closed-loop AI optimisation, said in September 2024 that its technology was used by seven of the ten largest US refiners, with more than 90 closed-loop applications deployed and refinery margin gains of $0.30 to $0.50 per barrel 2.
Both sets of figures are reported by the suppliers and their customers. They are specific and credible, but independent verification, consistent baselines and long-term service-factor data are not yet public.
| Classical MPC | AI-based optimisation | |
|---|---|---|
| Process model | Linear dynamic models from step tests | Nonlinear models learned from plant data, sometimes reinforcement learning |
| Upkeep | Periodic re-testing and re-tuning | Retraining, model monitoring and governance |
| Explaining a move | Traceable to model gains | Harder to explain for an individual move |
| Track record | Decades across refining and petrochemicals | Years, with a growing number of published cases |
Questions a control engineer should ask
The first design question is where the AI writes. In most current applications it writes setpoints or targets to existing controllers, often the MPC itself, rather than driving valves directly, and the regulatory and safety layers below are unchanged. That keeps the familiar limits, interlocks and operator controls in place and makes it easy to switch the application off safely.
The second is how the model is kept honest. A data-driven model learns how the plant was operated, including the habits of particular shifts, and can be unreliable outside that range. Applications need defined operating envelopes, monitoring of model performance, rules for retraining and a management-of-change process for model updates, just as configuration changes have.
The third is acceptance. Operators who cannot see why a controller made a move tend to take it out of service. Showing the model's intent, a clear handover procedure, and involving operators in commissioning matter as much as the algorithm.
How this differs from agentic AI
This entry covers optimisation within defined limits, usually as a supervisory layer. It is distinct from agentic AI for process control, elsewhere on the curve, in which software agents are given broader goals and the authority to plan and act. The reinforcement-learning case above shows that autonomous control of a single unit within tight bounds is possible today. It does not show that open-ended autonomous agents are ready for the control room.
Evidence timeline
- Jan–Feb 202235-day autonomous control of a distillation column by Yokogawa's FKDPP at ENEOS Materials. 1
- Mar 2023ENEOS Materials formally adopts the reinforcement-learning AI for the column; 40% steam and CO2 reduction reported. 1
- Sep 2024Imubit launches its reinforcement-learning-based optimisation; reports more than 90 closed-loop applications. 2
Why it sits at the Industry Acceptance stage
We keep this entry at the Industry Acceptance stage, but at its early edge, and have moved it earlier than in the September edition. Production applications exist and the architecture, AI writing targets to proven control layers, is sound. What is missing is independent, comparable evidence of results and of how service factor holds up over years rather than months.
Adoption signals to watch
- Independent or owner-published results with clear baselines and multi-year service-factor data.
- Adoption beyond refining into specialty chemicals, polymers and batch processes.
- Model-governance practices (validation, retraining, change management) written into owner standards.
Questions to ask before committing
- Exactly which variables does the application write to, within what limits, and what enforces those limits?
- What is the baseline for the claimed benefit, and who measured it?
- How is model performance monitored, and what triggers retraining?
- What happens when the application is switched off: does the unit return smoothly to existing control?
- How will operators see what the application intends to do?
Terms used on this page
- MPC
- Model predictive control, the algorithm at the heart of most advanced process control.
- Step test
- Deliberately moving an input to identify how the process responds.
- Service factor
- The percentage of time an advanced controller is in service and active.
Sources
- In a world first, Yokogawa's autonomous control AI is officially adopted for use at an ENEOS Materials chemical plant
- Imubit launches closed loop AI optimization solution powered by reinforcement learning
- Yokogawa's autonomous control AI adopted for use at an ENEOS Materials chemical plant
Related technologies
See this entry on the full maturity map · All Control Systems entries · How placements are made. Stage and time-horizon placements are editorial judgements based on the public sources listed above, as of October 2026. They are not engineering, safety or investment advice; see the Terms of Use.