Black box
Learns from your history. Cannot tell you why.
What the model is
The Method
Physics first. Data second. Neither on its own.
A black-box model learns what your plant did. A first-principles simulator knows what a plant like yours should do. We run the physics and let your data calibrate it — which is why a fault the model has never seen still shows up on the first day, and why every number it produces has an engineering name rather than an index.
What we built
A grey box.
Physics decides the structure. Your data decides the condition. That is the whole idea.
White box
Textbook physics. Does not know your plant.
Grey box
Physics for the structure, your data for the condition.
Where the line sits
What is structural, and what your data is allowed to move.
This is the whole discipline of a grey box. The equations are fixed by engineering and never fitted. Only a short list of health parameters is fitted — and each one is a physical quantity an engineer can argue about.
Structural — fixed by physics, never fitted
- Kremser absorption factors across the absorber stages.
- Counter-current ε-NTU and LMTD for the exchangers.
- Manufacturer performance curves for the pumps.
- Energy and mass balances across the reboiler and the solvent loop.
- Vapour-liquid equilibrium for the solvent, from its own property data.
Fitted — health, and only health
- UA and cleanliness factor on each exchanger.
- Fouling resistance, tracked as a rate rather than a state.
- Stage efficiency in the absorber and the regenerator.
- Pump and machine degradation against the manufacturer curve.
- Nothing else. When one of these drifts, something physical is happening and an engineer can name it.
What that buys you
Two things this structure gives you that a fitted model cannot.
Faults it has never seen
Machine learning needs historical examples of each failure before it can recognise one. A residual against expected-healthy does not. That is the difference between a model that works on the first day and one that waits for you to break something first — and on a new-build unit, the second kind never becomes useful at all.
A drift you can act on
When a fitted weight in a neural network moves, you have learned nothing. When cleanliness factor on E-401 moves from 0.92 to 0.78, you have learned that the exchanger is fouling, roughly how fast, and what it is costing in reboiler duty. The parameter is the diagnosis.
The trade you are making
Grey-box is more work up front. We would rather say that now.
- It needs datasheets, and it needs an engineer who understands the unit. You cannot point it at a historian and walk away, and any vendor who tells you otherwise is selling you a black box.
- For a plant you intend to run for twenty years we think that is the right trade. But it is a trade, and you should know it before the first meeting rather than after the purchase order.
The method is easiest to judge from the output.
The platform runs on a real amine train with thirteen days of one-minute history behind it. Open a diagnosis and read the evidence for yourself.
Open the platform