Dynamic process simulation
A plant simulator built to tune itself.
We are building Determyn, a first-principles dynamic simulator that models your plant from the records your site already keeps (historian, DCS configuration, operator graphics and P&IDs), fits the model to your plant’s own history, and reports in numbers how closely it matches. For operator training, startup and shutdown, and control testing.
- Historian
- Untuned model
- Tuned model
| Model | TT-104 RMS | Peak | TV-101 RMS |
|---|---|---|---|
| Untuned | 4.8 °C | +6.0 °C | 14 % |
| Tuned | 0.2 °C | +0.2 °C | 0.7 % |
Why simulators cost what they do
Most of a simulator project is tuning.
Laying out a model from drawings is quick. Matching it to the running plant takes weeks to months. An engineer adjusts valve characteristics, line resistances, heat-transfer coefficients and holdups one at a time, and in a coupled pressure-flow network one change moves matches elsewhere. When the plant changes, the model drifts and the work starts again.
COST
Services on a simulator project can run to $500,000 or more [source needed], and the software license is often the smaller part of the bill.
METHOD
Hand tuning is trial and error, one parameter at a time, in a network where the parameters interact.
UPKEEP
A DCS migration, a replaced valve or a fouled exchanger sends the simulator back for another round of services.
How it works
From the records you keep to a model you can check.
Each step shows where it stands today. We update these as the work lands.
-
Read what you have
Historian, DCS configuration (DeltaV first), operator graphics, P&IDs even when they are out of date, LIMS and datasheets. The software matches each instrument across every source.
Planned -
Build the network
A plant view that mirrors your P&IDs tag for tag, compiled into a solver model. First-principles mass and energy balances, with equation-of-state and activity-coefficient thermodynamics (Peng-Robinson, SRK, NRTL and others). The engine is built; building the plant view from your drawings is planned.
Engine built -
Tune against history
Hydraulics first, then heat transfer, then dynamics, fitted to trips, rate changes and upsets mined from years of history. We hold some events out, and the fit never sees them.
In progress -
Review a short list
Your engineer works through a queue ranked by impact on fidelity: suspected plant differences, unreliable instruments, and parameters the data cannot pin down.
Planned -
Keep it current
Drift and change detection against the historian are built into the tuning pipeline. Comparing each new DCS export with the last, and re-tuning only the sections that changed, are planned; the engineer will review only the differences.
In progress
Fidelity report
PlannedEvery model will ship with its fidelity measured.
We use the functional fidelity levels published by ISA, and are setting numeric tolerances for each. You will choose a target level for each unit and each use. The report scores the model against held-out events from your own history and says which level each unit reached, with the numbers behind it and its limits stated.
- Level 1
- Measurements match set points and move the right way when loop outputs change. Operator training.
- Level 2
- The same holds while valves open and close and pumps, fans and compressors start and stop. Training, startup and shutdown.
- Level 3
- Process gain, time constant and dead time are accurate enough to tune loops and see interactions. Control testing.
| Unit | Use | Target | Reached |
|---|---|---|---|
| V-101 Inlet separator | Training | L2 | L2 |
| K-101 Compressor | Control testing | L3 | L2, below target |
| E-101 Discharge cooler | Control testing | L3 | L3 |
| P-101 Separator liquid pump | Training | L2 | L2 |
- Steady state
- Dynamics
- Event replay
- Coverage
- Limits
K-101 is below target: its surge line rests on the vendor map because the history holds no surge event. The report names the test that would settle it.
Example layout, not results from a plant.
Findings
Built to find what the records get wrong.
When a section of the model will not fit, the cause is often a difference between the plant and its documents, or an instrument or record that cannot be trusted. The software reports it as a finding with its evidence instead of bending fitted coefficients to absorb it. Our test program plants each of these on purpose.
DETECTION BUILT · TESTED ON SYNTHETIC PLANTS
- Passing valve
- Undocumented bypass
- Stuck or drifting transmitter
- Fouled exchanger
- Valve replaced with a different Cv
- Historian clock off by an hour
PLANNED · NEEDS THE DOCUMENT READING OF STEP 1
- Mis-ranged transmitter
- Manual valve not where it is drawn
- Equipment removed but still drawn
- psig and psia mixed up
Uses
One model for training, startup and shutdown, and control testing.
Operator training
Operators will train on their own DCS screens, with the model connected to your vendor’s virtual controllers, so the simulator looks like the plant they run. We plan to draw exercises from real trips and upsets in your history.
Startup and shutdown
Empty vessels, dry lines, zero and reversing flow, and phases that appear and disappear. The engine is built to run through these states without failing, and tuning then makes it match. Trips and blowdown pass today on our reference plant; startup from rest and full shutdown are next.
Control testing and DCS migration
You will be able to check out new configuration against a model of your plant before cutover, and try a control strategy before it reaches the field, on units whose fidelity report shows Level 3.
Runs on your premises
The real-time model will run on your servers, next to your virtual controllers. Heavy tuning runs can go to the cloud if you choose.
PlannedWorks with your DCS
It will connect over OPC UA to your vendor’s virtual controllers. DeltaV configuration import comes first, with other systems after it.
PlannedYour model, your data
You own the model and your plant data. Every parameter records where its value came from: a drawing, a datasheet, a fit to history, or a default.
Bring your own AI agent
An MCP-capable agent will work on a sandbox copy of the model. It will propose changes for an engineer to approve, every call will be logged, and it will have no write path to the control system. The engine service already refuses agent writes to a live instance.
PlannedHow we test
Tested on virtual plants before it meets yours.
We never grade our engine with our own engine alone. Every test plant differs from the model being tuned, and only the scoring code can read the correct answers. Each plant hides faults that real plants have.
-
STAGE 1
In progressA structurally different plant
Our own engine running a more detailed plant than the model, with valve stiction, an undocumented bypass, exchanger fouling and a valve swapped mid-record, and a historian with stuck and drifting sensors.
-
STAGE 2
40 of 58 agreeAn independent engine
A plant section rebuilt separately in IDAES, the U.S. Department of Energy’s open process modeling framework, from published equations and without reading our engine’s code.
-
STAGE 3
PlannedReal equipment
A university unit operations lab with real sensors, where we will run our own step tests, then a design partner’s historical exports under NDA.
We are looking for one design partner.
A site with an upcoming DCS migration or a training gap, a well-kept historian, and willingness to share historical exports under NDA when the pilot starts. We want to hear from ethanol plants, refineries, specialty chemical sites and the integrators who run their migrations.
Tell us about your site and what is coming up. We reply within two business days.
Or write to [contact address]