In short
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.
It differs from a digital twin or a process simulator because it learns from the plant's own data, guided by its physics, and keeps learning as the plant changes.
In Orbital Plant Intelligence, the Orbital world model helps engineers detect, predict, optimise and simulate, so every answer about a unit tells one consistent story.
A reactor temperature is drifting, every alarm is quiet, and a shift log from last month mentions a similar drift after a feed change. Each of those facts lives in a different system. An experienced engineer connects them in their head. A world model for energy is built to connect them too.
In the first part of this series we set out why a language model on its own cannot do this. This part explains what a world model is, how it differs from a digital twin or an industrial foundation model, and what it gives engineers on shift.
What is a world model in AI?
A world model is an AI model that learns how a system behaves, so it can predict what happens next and what would happen if something changed. The idea comes from machine learning and robotics research, and it sits at the heart of what is now called physical AI. 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.
How is a world model different from a digital twin or a process simulator?
A digital twin is usually a model of one asset that engineers build and maintain largely by hand, and a process simulator solves first-principles models configured for a particular design case. A world model learns the plant's behaviour from its own data, guided by its physics, and keeps learning as the plant changes. It complements both rather than replacing them: rigorous simulators remain essential for design work.
Is a world model the same as an industrial foundation model?
The terms overlap. An industrial foundation model is a large model trained on industrial data that can be adapted to many tasks. Applied Computing uses the term world model to stress something more specific: the model represents how the plant behaves and responds to change, cause and effect, not only the patterns in its data.
How does the Orbital world model read a plant?
It reads the plant three ways, through its operating signals, its physics and its engineering documents, and draws on all three together.
How the world model reads a plant
Read the operating signals. Historian, control-system and laboratory data. The model tracks changing operating conditions, notices departures from expected behaviour and forecasts how the plant will perform as conditions evolve.
Model how the process behaves. First-principles physics: mass and energy balances, process relationships and equipment constraints. This is the basis for judging whether a change to the plant's operation makes sense.
Bring engineering context. Technical documents and operating knowledge, from P&IDs and procedures to work orders. This is what lets findings be explained in engineering terms.
Together they give Orbital an understanding of the unit that stays useful when a sensor drops out or when engineers ask something new.
Signals: from sensors to foresight
Real historian data is messy. Sensors fail, readings freeze, lab results arrive hours after the sample was taken. A model that only works on clean data will not survive contact with a real plant. So the Orbital world model is built to stay useful when data is missing, and to help distinguish a sensor fault from a real change in the process.
Physics: the rules the process has to follow
Much of the physics that governs a plant is already written down, in its design basis, heat and mass balances, equipment datasheets and operating manuals. Orbital builds on the engineering knowledge your plant already holds. What it relies on is visible to your engineers, who can review and correct it as they would any other engineering record. What is not written down, such as how a catalyst ages, is learned from the plant's own data.
The third part of this series explains why physics has to be part of how the model learns, not a check added at the end.
Language: knowledge your team can use
A shift log entry such as "raised hydrogen to hydrocarbon ratio after coke trend on the reactor" is valuable. It records a decision, its reason and its effect. Orbital links knowledge like this to the operating data it describes, so that past experience can inform what it says today, and so it can explain itself in the language engineers already use.
Why does one consistent story about the plant matter?
Because engineers cannot act on answers that contradict each other. The pattern in the data, the physics that explains it and the equipment it sits on tell one story about the plant, not three opinions to reconcile.
One plant, one story.
What can engineers do with a world model for energy?
In Orbital Plant Intelligence, the Orbital world model becomes four capabilities engineers use on every shift: detect, predict, optimise and simulate.
What the world model gives engineers
Detect. Compares how the plant is running with what it could achieve under the same conditions, so it can see a loss developing even while every tag stays inside its limits.
Predict. Connects a changing process to its most likely physical cause and forecasts where it is heading, including when an equipment constraint will start to bite.
Optimise. Compares possible responses against energy, throughput and equipment constraints, and brings the recommendation and its reasoning to the team.
Simulate. Tests a proposed change, such as a new feed or a different set point, against the model of the plant before anything changes on the unit.
Each capability draws on the same model of the plant, which is why a detection, a forecast and a recommendation about the same unit tell a consistent story. The final part of this series follows one operating question on a catalytic reformer from the moment it is asked to the moment an engineer decides what to do.
Frequently asked questions
What is a world model in AI?
A world model is an AI model that learns how a system behaves, so it can predict what happens next and what would happen if something changed. The idea comes from machine learning and robotics research and underpins much of physical AI. A world model for energy applies it to an industrial plant, learning from its operating data, its physics and its engineering knowledge.
How is a world model different from a digital twin?
A digital twin is usually a model of one asset that engineers build and maintain largely by hand. A world model learns the plant's behaviour from its own data, guided by its physics, and keeps learning as the plant changes. The two are complementary: a world model does not replace rigorous simulation for design, but it adapts to how the plant actually runs today.
Is a world model the same as an industrial foundation model?
The terms overlap. An industrial foundation model is a large model trained on industrial data that can be adapted to many tasks. Applied Computing uses the term world model to stress that the model represents how the plant behaves and responds to change, cause and effect, rather than only recognising patterns in its historical data.
What can the Orbital world model do?
In Orbital Plant Intelligence, the Orbital world model helps engineers detect losses developing before alarms, predict where the process is heading and why, optimise the operating point against energy, throughput and equipment limits, and simulate proposed changes before anything changes on the unit. All four draw on one consistent model of the plant, so their answers agree.
What happens when sensors fail or data is missing?
Real plant data is messy: sensors fail, readings freeze and lab results arrive hours late. The Orbital world model is built to stay useful when sensors drop out or readings are missing, and to help engineers distinguish a faulty instrument from a real change in the process, so a bad reading is less likely to be mistaken for a process problem.
Was Orbital previously described as a foundation model for energy?
Yes. Earlier descriptions called Orbital a foundation model for energy operations, a large model pre-trained on industrial time-series data. Since November 2026 Orbital is built on a world model for energy. It goes further: alongside the operating data it learns the plant's physics and engineering knowledge, so it can predict what happens next and what would happen if something changed.
About the author
Applied Computing research team
Applied Computing
The engineers, data scientists and AI researchers building the world model behind Orbital.
