Applied Computing

An AI brain for refineries and petrochemical plants, whose predictions also obey the laws of physics

Contact for pricing United Kingdom

A refinery has tens of thousands of sensors installed, but the data actually used for decisions is, per the company's account, less than 8%. Applied Computing's Orbital wants to put the remaining 92% to use — and its predictions must pass the tests of mass conservation and reaction kinetics, not just look statistically reasonable.

Features and application scenarios

Orbital stacks three models together. A time-series model connects to the control system and data historian, predicting trends and detecting anomalies; a physics model does the gatekeeping, ensuring predictions conform to mass balance, energy conservation, and reaction kinetics — this is its biggest difference from ordinary industrial AI; a language model, trained on chemical-engineering knowledge, can interpret plant documents and propose action suggestions. Target customers are energy operators managing critical infrastructure, especially oil, gas, refining, and petrochemical plants. The company is headquartered in London, UK, and is remote-first; the founder is an Imperial College alumnus, and former Shell AI-project lead Dan Jeavons serves as president, having completed a £9M seed round.

Suited to the process-engineering and operations departments of refining, petrochemical, and gas plants. Taiwan's petrochemical industry (the CPC and Formosa Plastics group systems) is exactly a typical scenario for this kind of technology — old equipment, lots of data, and experiential knowledge tied to senior engineers. This product's most noteworthy design is "the physics model as gatekeeper": purely data-driven models have long been distrusted in the process industry because they may give physics-violating suggestions engineers dare not follow. Constraining AI's output within a reasonable range with physics is the key to getting engineers to adopt it.

Main features

Time-series model: connects to the control system and data historian for trend prediction
Physics model: ensures predictions conform to mass balance, energy conservation, and reaction kinetics
Language model: trained on chemical-engineering knowledge, interprets plant documents and suggests actions
Anomaly detection and early warning
Dedicated design for refining, petrochemical, and oil-and-gas processes

Common uses

Parameter optimization of refining and petrochemical processes
Early detection of equipment anomalies
Knowledge retrieval of plant technical documents
Continuous improvement of operational safety and efficiency

Key Features

Time-series model: connects to the control system and data historian for trend prediction
Physics model: ensures predictions conform to mass balance, energy conservation, and reaction kinetics
Language model: trained on chemical-engineering knowledge, interprets plant documents and suggests actions
Anomaly detection and early warning
Dedicated design for refining, petrochemical, and oil-and-gas processes

Pros

The physics-constraint design lets engineers verify and dare to adopt it
The team has practical background at big energy companies like Shell
Handles both numerical data and technical documents, with complete coverage

Cons

Customer base limited to large process industries; not applicable to SMEs
Requires deep integration with the existing control system, so adoption is a long project
The company was founded in 2023, so long-term cases are still accumulating

Use Cases

Parameter optimization of refining and petrochemical processes
Early detection of equipment anomalies
Knowledge retrieval of plant technical documents
Continuous improvement of operational safety and efficiency

Editor's Note

Industrial AI has been hyped hard these past few years; the real barrier isn't how big the model is but whether engineers dare to follow it. Using physics as the gatekeeper is a smart move.

FAQ

What problem does the 'physics model gatekeeping' actually solve?

A purely data-driven model may give suggestions that are statistically reasonable but physically impossible, like operating parameters violating mass balance. Once engineers see such output, they never trust the system again. Adding physics constraints keeps the AI's suggestions at least within the physically feasible range — this is the watershed for whether it gets adopted.

Where is the difficulty of adopting it at Taiwan's petrochemical plants?

The biggest difficulty is usually not the technology but the data — whether the control system's historical data is fully preserved, whether labeling is consistent, whether it can be connected across plants. Most old plants get stuck at this stage. Doing a data inventory first rather than jumping straight to vendor selection is advised.