The last few per cent
Well-run plants are not short of competent people or good control systems. Advanced process control keeps units stable, and experienced operators know their levers. Yet most units leave value on the table every day, because conditions keep moving and there are too many interacting variables for anyone to balance consistently, shift after shift.
Feed changes. Catalysts age. Exchangers foul. Ambient temperature swings between night and day. Quality is measured a few times a day, and alarms arrive late. Decisions rest on individual know-how, so their quality varies by shift, by person and by plant.
Orbital Plant Intelligence is built for that gap. It works on the outcomes an operations team is measured on: yield, efficiency, reliability, anomalies and maintenance. It does four jobs that operators and engineers already do, earlier and with the reasons attached: detect, predict, optimise and simulate. And through Workspace, it puts the tools in engineers' hands.
Detect: catch problems early, with the cause
Orbital notices when equipment or a process starts behaving differently from what the physics and the unit's own history say it should. It then traces the likely cause upstream, so the warning reads as an explanation rather than a symptom.
Drift flagged hours or days before an alarm.
The root cause, not just the tag that moved.
Anomalies ranked by their effect on yield and reliability, so the team knows what to look at first.
Predict: know now, not in six hours
Lab samples arrive a few times a day. By the time a result comes back, the unit has moved on. Orbital estimates quality every minute from the signals the plant already measures, with a confidence band, and checks each estimate against lab results as they arrive.
It also looks ahead: for the moment a constraint will bind, and for drift that will become a problem. A typical warning reads: purity is heading for its limit in about seven hours, and reactor inlet temperature has drifted below its usual band. That gives the team time to act, and a place to start.
Optimise: find today's best way to run
Orbital searches for the operating point that balances energy, yield and quality under today's conditions, and recommends set-points with the expected gain. A recommendation on its own is not enough, so each comes with its reasoning: the constraints that are active, the margin left on each, and what is holding the unit back.
That last point is often the most valuable. Knowing which constraint is binding, and why, turns optimisation into debottlenecking: the team can see what it would take to remove the limit, not just how to live with it.
Engineers do not need a black box that says turn this up. They need to see which constraint is in the way, and what it costs.
Simulate: try it before you touch the plant
Engineers ask what-if questions every day. What happens if we raise feed by five per cent? What if the next cargo is a leaner crude? Orbital runs the change against the physics of the unit and its operating limits, shows the expected effect on the constraints that matter, and says how confident it is. The plant is run exactly as it is today until the team decides.
Grounded in the physics of the unit
Statistically likely is not the same as physically possible. All four jobs draw on the world model of the plant, which learns from the unit's own history and is shaped by the physics of the process, such as mass and energy balance. Futures the unit cannot reach are ruled out, and outside the range the plant has seen before, Orbital is more careful and says how confident it is.
Built around how the plant is actually run
Every plant has its own operating philosophy: preferred levers, soft limits the team will not cross, equipment it treats carefully. Recommendations that ignore those will not be used, however good the numbers look. Part of every deployment is capturing that philosophy, so Orbital's suggestions fit how the unit is run.
The output follows the rhythm of the working day. Before the morning load decision, the team sees the same picture: the recommended operating point, the binding constraints, the quality outlook and anything that drifted overnight. Whether a recommendation was taken, and what followed, feeds back so the next one is better.
Workspace: your ideas, your tools
When we put AI in front of plant engineers, the lesson came quickly. Accurate models are expected. What engineers want is to use that intelligence flexibly, inside the way they already work, and very few of them want to chat. Fixed dashboards, on the other hand, are quick to read but rigid.
Workspace sits between the two. Engineers describe what they need in their own words, and Orbital builds it from the plant's live data.
Build dashboards by describing them. Orbital proposes the cards, you adjust the proposal, and it builds them from live data.
Turn repeated work into apps. Describe a task once, such as a daily shift comparison, and run it whenever you like.
Ask anything about the plant. Ask in plain words, and Orbital works the figures out from the data, then explains how it got there.
Figures are calculated from the plant's data, not written by a language model, and every one is checked against its source. When Orbital cannot answer from the data, it says so. Workspace draws on the same plant knowledge, calculations and limits as detect, predict, optimise and simulate, so a figure on a dashboard always matches the figure behind a forecast or a recommendation.
A dashboard that is fast to build but wrong is worse than no dashboard at all. The checking is the product.
What it does not do
Orbital Plant Intelligence does not run the plant. It does not replace advanced process control, the lab or the people who supervise the unit. It reads from the plant and never writes to it. It recommends, explains and learns. Engineers decide.
Where to start
We start with one unit and one question, once its data is connected through Orbital Data Connect. Forecasting quality is often the first step, because the value is easy to see and easy to check against the lab. Optimisation follows once the team trusts the system behind it.
About the author
Orbital product team
Applied Computing
The engineers, data specialists and AI researchers who design and build Orbital, working alongside the operators who use it.
