Industries / Mining and materials / Chemicals and petrochemicals / Refining
Process control
Process control keeps each unit at its targets minute by minute, and planning decides what those targets should be; most of both is now done by software that operators supervise.
4
Milestones achieved, from 1952 to 1999.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1952
Planning by linear programming
Abraham Charnes, William W. Cooper and B. Mellon published a study in the journal Econometrica of blending aviation gasolines in an integrated oil company, set up as a linear program. A linear program finds the cheapest plan that stays within a list of limits, such as stock on hand and each product's specification. In a 1990 history, Bodington and Baker said the industry's planning models built this way were well known and "often cited as a major triumph of applied operations research", and traced such methods from strategic planning through process control.
Credit
Abraham Charnes and William W. Cooper of the Carnegie Institute of Technology, with B. Mellon
1973
Control that plans ahead
During a 1973 strike, Charles R. Cutler of Shell Oil was stuck inside a Shell refinery for weeks and tried his untested control method on the running plant, according to a 2025 history by Sigurd Skogestad and Thomas Adams. The method, Dynamic Matrix Control (DMC), uses a model of how a unit responds to predict where it is heading, and plans moves for many valves at once; the family is now called model predictive control. Shell kept it a trade secret until 1980. By the time Cutler left in 1984 the company had more than 200 DMC applications, by Control Global's account.
Credit
Charles R. Cutler, Shell Oil
1984
Predictive control for sale
In June 1984 Cutler and his colleague Chuck Johnston left Shell and founded Dynamic Matrix Control Corp. In 1986 a DMC controller went onto a Sunoco hydrocracker, and Cutler later recalled that energy use fell 15% the first time it was switched on. Texaco engineers said in a technical presentation that DMC saved US$15,000 to 25,000 a day at their Port Arthur, Texas refinery. AspenTech bought the company in 1996.
Credit
Charles R. Cutler and Chuck Johnston
1999
Targets for alarm rates
The Engineering Equipment and Materials Users Association (EEMUA) published EEMUA 191, a guide to alarm systems written by their users with input from the UK Health and Safety Executive (HSE). In a 2000 information sheet the HSE noted that in the last 11 minutes before the 1994 explosion at Texaco's Milford Haven refinery, two operators had to recognize, acknowledge and act on 275 alarms. Citing the guide, it set targets: in normal running no more than one alarm every ten minutes on average, and no more than ten in the first ten minutes after a major upset.
LessonAn alarm system has to be sized to what an operator can read and act on.
Credit
EEMUA and the UK HSE
Watch for
Reinforcement learning on hydrocrackers
Imubit, a Houston software company, says it has over 100 closed-loop controllers running, which send new targets to a plant's existing control system minute by minute while operators supervise. It says one two-stage hydrocracker application changes reactor severity every 10 minutes and is worth US$2 million to 4 million a year per hydrocracker, with about US$0.40 a barrel of extra margin in one case. A 2026 trade article by an Imubit author describes a controller trained by reinforcement learning, meaning trial and error against a model of the unit, that weighs yield now against catalyst life later; it gives no figures. The results are anonymized, and no outside party has published a measurement of them.
Achieved whena refiner names the unit, and the refiner or an outside party publishes before-and-after measurements over a stated period.
Claimed by
Imubit, on its customer page as read in November 2026 and in PTQ's Catalysis 2026 issue
AI handles plant upsets
A preprint by Khalid Alhazmi reports that a general-purpose large reasoning model, an AI model that works through a problem in steps before it answers, ran a simulated plant through abnormal situations with no task-specific training and only the information an operator has. Acting through a limited set of actions that software checks before they run, it kept the plant within all hard limits in 39 of 39 abnormal situations and operating changes, where basic control failed in 15, and named the root-cause fault in 15 of 15 safety-critical cases. The plant is a standard simulated benchmark, not a refinery, and no person was in the loop.
Achieved whenanother group reproduces the result on a different plant model, such as a refinery unit, or a supervised plant trial is measured by someone outside.
Claimed by
Khalid Alhazmi, August 2026
Open
01
A refinery alarm-flood test
An open set of simulated alarm floods on a refinery unit, such as a fluid catalytic cracker, each labeled with its root cause, for scoring both classic alarm-analysis methods and AI operator assistants. An alarm flood is a burst of alarms faster than an operator can handle.
The briefCredit · unclaimed02
Control claims sorted by evidence
A public table of every published result for advanced control, real-time optimization and AI control at refineries, from vendor case studies, trade magazines and papers. Each row gives the unit, the baseline, the claimed benefit, the period measured, who measured it, and whether anyone outside the vendor and the refiner checked it.
The briefCredit · unclaimed
Open 01 · the brief
A refinery alarm-flood test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open set of simulated alarm floods on a refinery unit, such as a fluid catalytic cracker, each labeled with its root cause, for scoring both classic alarm-analysis methods and AI operator assistants. An alarm flood is a burst of alarms faster than an operator can handle.
Why it is still open
The open alarm datasets used in research come from simulated chemical plants: 300 abnormal situations from 8 root causes, with 7,343 alarm activations, on the Tennessee Eastman process, and alarm flood sequences from a vinyl acetate plant model, released by Ning Chen and Wenkai Hu in December 2025. A 2026 dataset on GitHub simulates five faults on an open fluid catalytic cracker model, but gives sensor readings, not alarms. The August 2026 reasoning-model test also used a standard simulated plant, not a refinery.
Built with what exists
The open fluid catalytic cracker and fractionator model that Santander, Kuppuraj, Harrison and Baldea published in 2022, written in MATLAB under the MIT license and based on McFarlane and colleagues' 1993 Model IV simulator; alarm limits set to EEMUA 191's rate targets; the Tennessee Eastman set's format as a template; a coding agent that tests its own work, such as one built on Inkling 2 Small (see the Note "Checked without being asked"); AI models run 20 times per scenario (see the Note "All twenty times").
First step
Run the open model, set alarm limits on its measured signals, and check that a day of normal running stays under one alarm per ten minutes.
Done looks like
At least 300 simulated abnormal situations from at least 8 root causes, released with alarm logs and labels under an open license; no more than one alarm per ten minutes in normal running; published scores for one classic method and two AI models, as the share of floods whose root cause each names correctly within the first ten minutes.
Open 02 · the brief
Control claims sorted by evidence
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every published result for advanced control, real-time optimization and AI control at refineries, from vendor case studies, trade magazines and papers. Each row gives the unit, the baseline, the claimed benefit, the period measured, who measured it, and whether anyone outside the vendor and the refiner checked it.
Why it is still open
Results are mostly told by vendors or recalled years later. Imubit anonymizes its refinery results; Cutler recalled a 15% energy cut on a 1986 hydrocracker; a compendium of over 40 refinery energy case studies on GitHub Pages repeats an unnamed US refinery's 8% energy and 3% throughput gains without saying who measured them. Bauer and Craig published a survey and framework for judging what advanced control earns in 2008, but a search found no public, case-by-case table graded by evidence.
imubit.comcontrolglobal.comamsamms.github.ioapi.crossref.org
Built with what exists
A web agent to collect case studies from vendor sites and trade-magazine archives such as PTQ's on digitalrefining.com; a language model that fills each row and quotes the sentence each field came from (see the Note "Findings that cite notes"); two people who check every row's evidence class; a spreadsheet and GitHub.
First step
Write the coding sheet and code 30 results from one trade magazine's archive.
Done looks like
At least 200 results coded; two people agree on the evidence class for at least 90% of a 50-row sample; the table reports the share of results checked by an outside party, by type of unit and by decade.