Physics-first AI for industrial energy and emissions
Every plant has the data. Almost none can act on it.
Twin Process AI models process equipment from first principles, checks whether your instruments can be believed, and turns what is left into operating moves priced in fuel, CO2 and money. Every recommendation is re-solved in the physics before it reaches the control room.
The problem
Industrial plants are full of data nobody can act on.
Billions have gone into sensors, historians and control systems. The measurement problem is solved. What is missing is a layer that turns any of it into a decision an operator can take this shift.
- 01
The data cannot be trusted
Plants running on assets twenty and thirty years old carry mis-mapped tags, dead transmitters, drifted instruments and lab results that never reconcile with the DCS. Machine learning applied to that data confidently produces the wrong answer.
- 02
The models have drifted
The simulation built at commissioning was right at commissioning. Feed changed, catalyst aged, exchangers fouled. What it predicts and what the plant does have been quietly separating ever since.
- 03
Optimisation happens once a year
An energy study takes weeks and describes a plant that has already moved on. Between studies, the operating window is set by habit and by whoever is on shift.
- 04
The losses are invisible
Excess air, fouled exchangers, hidden steam leaks and burners degrading behind a healthy-looking stack reading. None of it trips an alarm. All of it shows up in the fuel bill and the emissions figure.
Machine learning cannot fix any of this on its own. Applied to data that has not been reconciled, it learns the errors and reports them back with confidence.
No customer logos yet. Here is what we can show you instead.
Everything below is checkable. The engineering basis, the cross-validation method and the cases where the model failed are all published rather than summarised.
- 0.9934
Blocked-CV R2 on thermal efficiency. Nine of twelve targets clear 0.99
First-principles simulation of a 53.4 MW crude preheat heater driven through 120 days of hourly operation. Not a plant installation.
- 1,631x
Faster than a rigorous solve. 1.75 s becomes 1,073 microseconds
Measured on the same operating points, single core.
- 81.5 %
Of burner-fault hours caught that a stack O2 analyser cannot see, at zero false alarms
83 of 102 labelled hours across burner degradation and register imbalance. Zero flags across 2,499 healthy hours. Scored against the labels, built without them.
- 0.84
Of the predicted saving that physics actually delivered when we re-solved it
Ten recommended setpoints re-solved in the rigorous simulator. The surrogate predicted 1.52 GJ/h, physics delivered 1.27 GJ/h.
- 100 %
Of re-solved setpoints cleared every CO, tube metal, dew point and draught limit
The surrogate overstates the value of a move, not its safety. Every re-solved setpoint was valid in the rigorous model.
- ASME PTC 4
Full loss ledger. API 560 bands. NIST Shomate data
41 outputs, each with formula, assumptions, units, reference and validation.
First-principles simulation of a 53.4 MW crude preheat heater driven through 120 days of hourly operation. Not a plant installation. Calibration against a live heater is our next milestone. We publish the physics-corrected figures rather than the surrogate’s own estimate of its worth, because the surrogate is optimistic about the value of its recommendations and we would rather you heard that from us.
How it works
Model the equipment. Learn the plant. Price the move.
- 01
Model the equipment
We build the heater, not a correlation. Combustion stoichiometry, an enthalpy-based energy balance on NIST Shomate data, a Lobo-Evans radiant section and a rated finned convection bank. Stack temperature, the radiant and convection split, efficiency and tube metal temperature are outcomes of the geometry rather than numbers fitted to the answer.
- 02
Learn the plant
Your historian teaches the model what your unit actually does. Soft sensors infer the state nobody measures, fouling and true excess air among them. A surrogate then replaces the rigorous solve with a microsecond prediction, which is what makes optimising every hour of a year possible at all.
- 03
Price the move
A constrained optimiser finds the setpoint, prices it in fuel, CO2 and dollars, names the limit that stops it going further, and writes the reason in language an operator can act on at three in the morning.
The proof
We built it properly for one unit before claiming it for a plant.
Fired heaters burn the largest share of a refinery’s fuel, so that is where we started. Everything on this page that is marked as built, we built here first: the physics, the soft sensors, the optimiser and the checks that catch it lying.
- The unit
- 53.4 MW absorbed, 195,000 kg/h of 34 API crude
- Solved to
- 90.7 % LHV at 15 % excess air on a clean coil
- Against
- All inside the API 560 bands
The stack temperature, the radiant and convection split, the tube metal temperature and the efficiency all fall out of the geometry and the energy balance. None of them is a correlation fitted to a known answer, which is why the model still behaves when it is pushed somewhere the data never went.
First-principles simulation of a 53.4 MW crude preheat heater driven through 120 days of hourly operation. Not a plant installation.
- Efficiency
- 87.1 %
- Stack O2
- 3.73 %
- Stack temp
- 233 C
- CO
- 20 ppm
- Feed
- 189,546 kg/h
- Outlet
- 415 C
- Fouling
- 0.75
- Detected fault
- fuel_upset
Trim excess air 19.8 % to 9.9 % (stack O2 3.73 % to 2.08 %)
$45.44/h~$381,710/yr at this state
- +1.26 ptsefficiency
- −1.52 %fuel
- −0.617 t/hCO2
Watch Move in steps and hold. Stop if CO passes 100 ppm before the O2 target, or if the stack falls below 98 C.
Why +−
Every percent of excess air above what the burners need is nitrogen heated from 20 C to the 233 C stack and vented. Cutting stack O2 by 1.65 points removes that parasitic mass flow, which is why the stack falls 16 C and dry flue-gas loss drops with it. The duty is unchanged, so the whole of the recovered loss shows up as less fuel through the burners. The move stops where it does because CO 167 ppm at the 200 ppm limit.
A model that admits what it cannot know.
Two things every vendor in this space could publish and does not.
We correct our own numbers
Ten recommended setpoints were re-solved in the rigorous simulator. The surrogate predicted 1.52 GJ/h of saving. Physics delivered 1.27. Every limit still held, so the safety was sound and the value was overstated.
- Surrogate predicted
- 1.52 GJ/h
- Physics delivered
- 1.27 GJ/h
- Delivered ratio
- 0.84
Every figure quoted on this site is the corrected one. The surrogate sits furthest from its training data on the largest trims, which is exactly where it is most optimistic.
We switch off the sensors that fail
Three latent states were tested against the tags a historian already carries. Two are good enough to ship. One is not, so nothing downstream reads it.
- Fouling index
- Ships0.971
- True excess air
- Ships0.981
- Fuel hydrogen
- Gated off0.050
Fuel hydrogen cannot be inferred from the available tags, so it is switched off and nothing downstream reads it. A refinery that wants it online needs a Wobbe meter or a fast GC on the fuel header. More modelling will not recover a signal the tags do not carry.
The number
Fuel saved and carbon avoided are the same move.
A heater trimmed to the right excess air burns less gas and emits less CO2 in exactly the same proportion. There is no trade-off to manage here, which is why energy optimisation is the cheapest decarbonisation a refinery has available.
805 t/yr
CO2 avoided, per heater, per year
Physics-corrected. A mid-size refinery runs dozens of fired heaters, and none of this needs new capital equipment.
$96k
Saved per heater, per year
Averaged across every hour of a 120-day history, then scaled by the fraction of predicted saving that physics actually delivered.
up to $380k
At the hours furthest off optimum
A single operating state, not a year at that rate. Shown because those are the hours worth catching, and never on its own.
- 0.63 %
- Less fuel fired
- +0.65 points
- Mean thermal efficiency gain
- 2,675 of 2,783
- Hours where a move was available
The same result reported the way a plant reports it: energy per tonne of feed against budget. Optimisation moves that line, and it is the line a general manager already tracks every month, which makes the case without translation.
Assumptions: Fuel at 6 USD/GJ, carbon at 40 USD/t, 8,400 operating hours per year. Physics-corrected throughout. Physics-corrected. First-principles simulation of a 53.4 MW crude preheat heater driven through 120 days of hourly operation. Not a plant installation.
Where this goes
One plant, six layers. Three of them exist today.
The end state is every unit on a site modelled, reconciled and optimised continuously. We are building it in the order that makes each stage useful on its own, and we mark honestly which stages are finished.
- 01In progress
Heat and material balance
Reconcile flows, temperatures and pressures until mass and energy close. Nothing downstream is trustworthy until this does, and in most plants it has never been done on live data.
- 02Planned
Tag validation
Find the instruments that are drifted, frozen, mis-mapped or simply wrong, and score the confidence of every tag before a model is allowed to read it.
- 03Built
First-principles simulation
Build the equipment properly: combustion, heat transfer, hydraulics and an efficiency ledger that a process engineer can audit line by line.
- 04Built
Constrained optimisation
Find the best safe operating point against real limits, price the move in fuel, CO2 and money, and name the constraint that stops it going further.
- 05Built
Machine learning layer
Soft sensors for the state nobody measures, a surrogate fast enough to search a year of operation, and anomaly detection that does not depend on fault labels.
- 06Planned
Operator copilot
Every unit on the site, answering questions in plain language and carrying the reasoning of engineers who have retired.
3 of 6 phases built and tested, for one unit type. Everything marked in progress or planned is a statement of direction, not a capability you can buy today.
Where we fit
We are not replacing anything you already run.
Your simulator, your control system and your historian all do things we do not and will not. The gap is between them, and it is where the operating decision actually gets made.
Process simulators
Does well
The reference calculation. Decades of validated thermodynamics and unit models.
The gap
A model built at commissioning drifts as feed, catalyst and fouling change. Keeping it current is manual work that few plants have the staff for.
Advanced process control
Does well
Holds a unit steady against constraints, continuously and automatically.
The gap
It controls to a target somebody else chose. Deciding whether the target is the right one is a separate question, and it is usually answered once a year.
Industrial data platforms
Does well
Collect, store and visualise plant data at scale, reliably.
The gap
A historian will show you a tag that has been reading the same value for six weeks without telling you it is broken. Storage is not validation.
Engineering consultancies
Does well
Deep expertise, and a study that answers the question properly.
The gap
It arrives weeks later and describes a plant that has since moved. The knowledge leaves with the consultant.
We sit between them: reading what the plant recorded, checking whether it can be believed, and turning it into a move with a price on it.
Who reads the output.
- Energy and plant managers
- A number against a target, and the hours where a saving was available but not taken. Fuel, CO2 and cost per tonne of feed, tracked between shutdowns rather than reported after them.
- Process engineers
- The engineering basis for all 41 outputs, with formula, assumptions, units, reference and validation for each. If you disagree with a number you can find out exactly where it came from. Nothing is a black box.
- Reliability and sustainability leads
- Peak tube metal temperature, acid dew point margin, fouling trend and emissions, all inferred continuously from tags you already record rather than from a survey once a year.
Bring us a heater and its historian.
Send us 90 days of tags and we will tell you what the unit is leaving on the table, which of those hours we would have caught, and where our model disagrees with your plant. The last one is the useful part.