In short
Data-only models know which answers look like the past. They do not know which answers the plant can physically deliver.
Physics-informed AI, which we call physics-grounded AI, makes mass and energy balances and equipment limits part of how the model learns, not a check added afterwards.
This matters most when the plant moves somewhere new, which is exactly when engineers need help and data-only models are least reliable.
A model trained on historian data learns which combinations of readings resemble the plant's past. That is useful, but it is not the same as knowing which combinations can actually exist. Historical plausibility is a statement about data. Physical feasibility is a statement about the plant.
What is physics-informed AI?
Physics-informed AI is machine learning that makes known physical laws, such as mass and energy balances and equipment limits, part of how the model learns, so its predictions respect what the plant can physically do. At Applied Computing we call it physics-grounded AI, and it is central to the Orbital world model. 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.
The idea has a substantial research history, including the physics-informed neural networks described by Raissi, Perdikaris and Karniadakis in 2019, and it is now used well beyond process plants, as its impact on fluid dynamics shows.
How do physics-informed models differ from data-driven models?
A data-driven model learns which answers look like the past. A physics-informed model also learns which answers are physically feasible. A data-only model can offer answers that look like the past but break a mass balance or push a heater past its limit.
Data shows what the plant has done. Physics sets what it can do.
Four gaps that physics closes
Where data alone falls short
Alarm versus loss. An alarm says a threshold was crossed. It cannot see a loss developing while every tag stays in range. A model that understands how the tags relate to one another can.
Correlation versus cause. When the feed or the operating conditions change, extending a familiar curve misses the mechanism. A forecast that follows the cause is far more likely to hold.
Plausible versus feasible. A statistically attractive answer can be physically impossible. A model trained on the physics is far less likely to suggest it.
Number versus reasoning. A number without its reasoning is hard to act on. Engineers need the evidence, assumptions and constraints behind a recommendation before they decide.
Where does the physics come from?
None of this requires inventing new physics. Much of it is already written down in the plant's design basis, heat and mass balances, datasheets and 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.
Not everything is written down. Some kinetics belong to the technology licensor, and behaviour such as how a catalyst ages is rarely documented at all. What is not written down is learned from the plant's own data, as the second part of this series describes.
Is physics-informed AI the same as hybrid modelling?
Not quite. Both combine physics with data. Hybrid, or grey-box, models often join a first-principles model and a machine learning model side by side, with one correcting the other. Physics-informed AI builds the physics into how the model learns.
Either way, there are two ways to bring physics in. The weaker way predicts first and attaches a warning when the answer breaks a balance. The stronger way makes the plant's physics part of how the model learns, so its predictions already respect the unit's mass and energy balances and its equipment limits.
A prediction that matches one sensor but breaks the unit's balance is still wrong. Over time, and allowing for inventory, flows in and out of each unit have to balance, energy has to be accounted for and equipment has to stay within its limits. Physics defines which answers are feasible. The operating objective decides which feasible answer is best.
Does physics-informed AI replace first-principles simulation?
No. Rigorous first-principles simulators remain essential for design and for studying cases far from today's operation. A physics-informed world model complements them in operations: it learns from live plant data, adapts as equipment ages and fouls, and answers operating questions in minutes.
Why does this matter on the plant?
Because the plant regularly moves somewhere it has not been before: a new feed, a higher rate, an unusual combination of conditions. That is exactly when engineers most need help, and exactly when a data-only model is least reliable, because it has nothing in its history to copy. A model that has learned the physics is better placed to answer, and to say how confident it is.
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 physics-informed AI?
Physics-informed AI is machine learning that makes known physical laws, such as mass and energy balances and equipment limits, part of how the model learns. Its predictions therefore respect what the plant can physically do. Applied Computing calls this physics-grounded AI, and it is central to the Orbital world model, its world model for energy.
What is the difference between physics-informed and data-driven models?
A data-driven model learns which answers look like the past. A physics-informed model also learns which answers are physically feasible. The difference matters most when the plant runs somewhere new, such as on a new feed or at a higher rate, because a data-only model then has nothing in its history to copy and may suggest answers that break a mass balance.
Is physics-informed AI the same as hybrid modelling?
Both combine physics with data. Hybrid, or grey-box, models often join a first-principles model and a machine learning model side by side, with one correcting the other. Physics-informed AI builds the physics into how the model learns, so its predictions respect balances and equipment limits from the start rather than being checked afterwards.
Does physics-informed AI replace first-principles simulators?
No. Rigorous first-principles simulators remain essential for design and for studying cases far from current operation. A physics-informed world model complements them in day-to-day operations: it learns from live plant data, adapts as equipment ages and fouls, and answers operating questions in minutes rather than requiring a new simulation study each time.
About the author
Applied Computing research team
Applied Computing
The engineers, data scientists and AI researchers building the world model behind Orbital.
