In short
Orbital, built on the Orbital world model, works a question through the way a careful engineer would, and answers only when the evidence is sufficient.
If Orbital cannot answer with confidence, it says so. It reads from the plant and never writes to the control system.
In a hypothetical reformer example, a heavier feed is flagged days before any alarm, with set points an engineer can review.
An engineer's question contains no tag names, no time window and no definition of enough evidence. "Are we on track to meet our production plan?" is a perfectly clear question to a person on shift, and a very vague one to a computer. Turning it into something a model can answer is most of the work.
How is AI used for refinery optimisation in practice?
In practice, AI for refinery optimisation helps engineers spot losses developing before alarms, forecast how a unit will respond and compare options against energy, throughput and equipment limits. Engineers then review the recommendation and decide.
Orbital works a question through the way a careful engineer would. It works out what needs to be known, gathers it from the right sources, and answers only when the evidence is sufficient, showing where each piece came from. Underneath it is the Orbital world model, described in the second part of this series. A world model for energy is an AI model that learns how an industrial plant behaves, from its operating data, its physics and its engineering knowledge, so it can predict what happens next and what would happen if something changed.
Does Orbital control the plant?
No. Orbital reads from the plant and never writes to the control system. If it cannot answer with confidence, it says so rather than guessing. Engineers decide, and any change goes through the plant's normal operating procedures.
What does this look like on a catalytic reformer?
This is a hypothetical example, not a customer result. It uses a continuous catalyst regeneration (CCR) reformer with simplified numbers to show the shape of an interaction.
It is day five of the run. A production engineer asks mid-shift whether the unit is on track to meet its production plan. Feed is 120 tonnes an hour against an on-spec product target of 110 tonnes an hour. Every monitored signal is inside its alarm limits.
Following the answer
The evidence. The latest lab result shows a heavier feed: the final boiling point has risen from 170 to 180 °C since the day before. Reactor inlet temperature and recycle hydrogen purity are where they were. The changed feed is the new information.
The forecast. If nothing changes, coke on the catalyst is forecast to pass the 5 wt% operating limit, and on-spec product to fall to around 100 tonnes an hour within about five days. The shortfall is not missing mass: it leaves the unit as lighter, lower-value products.
The options. Raising the hydrogen to hydrocarbon ratio from 2.5 to 2.8 and catalyst circulation from 1,000 to 1,080 kg an hour keeps both inside their equipment limits, with feed unchanged. The higher ratio adds load on the recycle gas compressor.
The expected result. Coke settles back to about 4 wt% and product returns to around 110 tonnes an hour. The extra compressor energy and regeneration costs are shown separately, so the trade-off is clear.
The decision. The two proposed set points and the expected response go to the engineer, who reviews them and decides. Nothing changes on the plant until they do.
None of the monitored signals had crossed an alarm. Without a model that understood how a heavier feed affects coke laydown, the first sign of trouble would have been lost production several days later.
How does this relate to APC and real-time optimisation?
Orbital works alongside both. Advanced process control (APC) holds a unit at its targets, and real-time optimisation (RTO) sets those targets from a process model. Orbital helps engineers understand why performance is changing and which targets are worth revisiting, and leaves the change itself to the engineer and the plant's existing control layer.
Why does it take both a world model and a system around it?
Because each half fails without the other, as the first part of this series argued. A grounded world model behind a dashboard still leaves the work of reaching a decision with the engineer: the problem was never a shortage of charts, it was the effort needed to get from charts to a defensible action. And a capable system wrapped around an ungrounded model can run many steps and still be wrong.
The world model gives trustworthy evidence. The system around it decides how much evidence a decision needs.
Together they let an engineer ask a question in plain words and get back an answer with the evidence, the reasoning and the trade-offs laid out, ready to review. To see this on real plants, read how Indorama Ventures engineers verified Orbital on their own unit, how KBR brings Orbital to its customers, and what Orbital Plant Intelligence does.
To see the Orbital world model on one of your own units, book a technical briefing with our team.
Frequently asked questions
How is AI used for refinery optimisation?
AI for refinery optimisation helps engineers spot losses developing before any alarm, forecast how a unit will respond to a change in feed or conditions, and compare options against energy, throughput and equipment limits. Engineers then review the recommendation and its evidence and decide. In Orbital, this work draws on the Orbital world model, a world model for energy built for industrial plants.
Does Orbital control the plant?
No. Orbital reads from the plant and never writes to the control system. It brings engineers a recommendation with its evidence, reasoning and trade-offs. Engineers decide whether to act, and any change goes through the plant's normal operating procedures and existing control systems, so responsibility for the plant stays with the people who run it.
How is Orbital different from APC and RTO?
Advanced process control (APC) holds a unit at its targets, and real-time optimisation (RTO) sets those targets from a process model. Orbital works alongside both rather than replacing them. It helps engineers understand why performance is changing, and which targets are worth revisiting, using a world model that learns from the plant's own data and physics.
What happens if Orbital is not confident in an answer?
It says so rather than guessing. Orbital answers only when the evidence is sufficient, and shows where each piece of evidence came from. When the evidence is incomplete, it makes clear what is missing, so engineers can judge how far to rely on the answer or decide what further information they need before acting.
Is the catalytic reformer example a real result?
No. It is a hypothetical example on a continuous catalyst regeneration (CCR) reformer, with simplified numbers. It shows the shape of an interaction: a heavier feed flagged days before any alarm, a forecast of rising coke on the catalyst, and proposed set points that an engineer reviews before anything changes on the plant. It is not a customer result.
About the author
Applied Computing research team
Applied Computing
The engineers, data scientists and AI researchers building the world model behind Orbital.
