Industries / Mining and materials / Chemicals and petrochemicals
Process control
Process control keeps a plant's temperatures, pressures, flows and levels where they should be, using instruments, controllers and software.
5
Milestones achieved, from 1959 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1959
A computer closes the loop
At Texaco's Port Arthur refinery in Texas, an RW-300 computer ran closed-loop control of a polymerization process, reading the plant's measurements and adjusting it without a person in between.
Credit
Texaco and Ramo-Wooldridge
1975
Control spread across modules
Honeywell's TDC 2000 split control among many modules instead of one large central computer, so that one failure need not stop the whole process. Honeywell calls it the first distributed control system (DCS), and says other companies released such systems the same year.
Credit
Honeywell
1994
Too many alarms
A lightning strike at Texaco's Milford Haven refinery in Wales began a day of upsets that ended with about 20 tonnes of hydrocarbon released, an explosion and fires. The UK Health and Safety Executive found that an excessive number of alarms reduced the operators' response, and that the control screens did not give them an overview of the process.
LessonAlarms help only if operators can still read them in an emergency.
Credit
UK Health and Safety Executive
2023
AI runs a distillation column
After a 35-day field test in 2022, ENEOS Materials formally adopted Yokogawa's reinforcement-learning controller, FKDPP, to run a distillation column. Reinforcement learning trains software by trial and error. The two companies report that steam use, and the carbon dioxide that goes with it, fell 40% against the previous manual control.
Credit
Yokogawa and ENEOS Materials
2025
An open control system runs
ExxonMobil's resin finishing plant in Baton Rouge, Louisiana runs on an open, standards-based control system whose parts can come from different vendors. By September 2026 it had been live for about a year and a half: the site engineer reported fewer support calls and some early losses of console control, and a third-party assessment found one critical firmware issue. ExxonMobil estimates the system will cost at least 20% less over its life than a traditional one.
Credit
ExxonMobil, a founding member of the Open Process Automation Forum
Watch for
AI models at Evonik Oxeno
Evonik Oxeno says AI-based process models, built from over one million simulated plant states, give control recommendations within seconds in running operations at its Marl and Antwerp sites. It is working on sending the results straight into the plant's control system, and gives no energy or emission figures.
Achieved whenthe models adjust the plant without a person in between and Evonik or an outside party publishes measured energy savings.
Claimed by
Evonik Oxeno, November 2026
Evonik tests AI in Singapore
Evonik and Imubit started a pilot of data-driven process models at Evonik's Singapore facility. The models give operators "what-if" advice but do not control the plant, with a focus on catalyst life. The companies will set baselines and measurement methods, and give no targets.
Achieved whenEvonik publishes measured results from the pilot.
Claimed by
Evonik and Imubit, April 2026
Open
01
Startup procedures as tested code
A tool that turns a written startup procedure for a distillation column into an automated sequence of steps, following the ISA-106 models for procedure automation, and tests it in a dynamic simulator against people following the same written procedure.
The briefCredit · unclaimed02
AI control, fairly tested
A public comparison of a reinforcement-learning controller, a well-tuned conventional controller and model predictive control (MPC, which plans its moves ahead using a model of the process) on the same open distillation-column model, scored on steam use and time off specification across the same disturbances.
The briefCredit · unclaimed
Open 01 · the brief
Startup procedures as tested code
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that turns a written startup procedure for a distillation column into an automated sequence of steps, following the ISA-106 models for procedure automation, and tests it in a dynamic simulator against people following the same written procedure.
Why it is still open
Yokogawa writes that most incidents not caused by outright mechanical failure happen during abnormal situations such as startups and shutdowns, and that rarely used procedures are often left manual because companies judge them not worth the effort and old control-system tools make the work hard. The 2005 explosion at BP's Texas City refinery happened while a unit was being restarted.
Built with what exists
A free dynamic-modelling tool such as OpenModelica; the ISA-106 models; a coding agent that tests its own work, such as one built on Inkling 2 Small (Notes: "Checked without being asked").
First step
Write the startup procedure for one column in plain numbered steps and code the first three in the simulator.
Done looks like
For one published column design, 20 simulated startups run by the automated sequence and 20 by people following the written procedure, with time to reach specification, off-specification product and safety-limit breaches published for each run.
Open 02 · the brief
AI control, fairly tested
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public comparison of a reinforcement-learning controller, a well-tuned conventional controller and model predictive control (MPC, which plans its moves ahead using a model of the process) on the same open distillation-column model, scored on steam use and time off specification across the same disturbances.
Why it is still open
The one plant result, the 40% steam saving at ENEOS Materials, was measured against manual control, not against tuned conventional or model-based control.
Built with what exists
Sigurd Skogestad's Column A, a free 40-stage column model in MATLAB and Simulink with ready-made control loops; open reinforcement-learning libraries such as Stable-Baselines3; an open MPC toolbox such as do-mpc.
First step
Run Column A with its supplied control loops and record a baseline for ten disturbances.
Done looks like
Code and results published for 1,000 simulated disturbance runs per controller, with steam use and minutes off specification reported for each controller.