ControlIQ · interactive demo

ControlIQ · process control

The tuning was
never the problem.

Control algorithms are settled engineering — the backbone every plant already runs on, and they do their job. What decides whether a loop holds is the model underneath them: how well anyone actually knew the process when the controller was set up. Get that wrong and the controller is confidently wrong with it, and the plant gives you no sign. ControlIQ measures the loop, tells you whether the data determined a model, and commissions a controller only when it did.

FormulaIQStudio ProRoboBridgeControlIQ

01 · Commissioning

What it costs to do this properly.

Before anything else, the three questions an advanced-control evaluation actually turns on: how long will you be in my plant, what will you disturb, and what happens when it fails. The schedule scales with your loop’s time constant, so we can answer for a process we have never seen.

Your loop’s time constant

Time on your plant

Step test ≈3 τ, hold-out ≈5 τ. This is the only window that touches production. Identification, twin build and tuning are offline; the shadow period runs alongside your existing controller and changes nothing.

What gets disturbed

Supervised bumps of the manipulated variable, inside limits your engineers set, on a loop your operators choose. No new hardware and nothing of ours executing on your control system.

If the loop cannot be bumped at all — and many critical ones cannot — the route changes but the verdict does not. The model can come from setpoint moves under the controller already running, from history the plant has been recording all along, or from process design with no plant time at all (the greenfield path further down). What never changes is the last step: if the data was not enough to determine a model, we say so and do not tune from it.

What happens when it fails

Your controller keeps running. ControlIQ commissions the block that is already in your control system; there is nothing of ours in the loop to fail. In the supervisory mode on the roadmap, losing our link leaves that same tuned controller in charge, bumplessly.

calibration: the rig’s model came from 2.91 h of step test plus a 4.94 h hold-out and replayed a closed-loop run to 0.259 °C.

02 · Why this is the hard part

A step test that is too short does not look wrong.

Everything below runs on a model identified from real hardware — a physical temperature loop with a time constant just over an hour. Drag the slider to shorten the bump test, the way a busy plant would. Watch the model, and watch R².

Step test length rig τ ≈ 62 min · we budget 3 τ, and refuse below about 1

What you would have concluded

measured response fitted first-order model

—

R² of the fit — it stays this high all the way down, so goodness of fit cannot warn you

—

of the true time constant — the model is simply wrong

—

of the safety margin a proper test would have left you — this is what the short test actually costs

This is not a constructed example. It is the failure that happened on this hardware: a 900 s test against a plant with τ ≈ 3881 s returned a time constant four times wrong at R² = 0.993, and the number stood until the plant refused to behave. Our identifier now exits with an error rather than return it.

03 · The pipeline

Measure the loop. Decide whether to trust it. Then tune.

The same code on two plants that agree on almost nothing. Every rule it applies is published and cited — the value is in the acceptance step, not the arithmetic.

Plant

What it measured

Whether it may be used

Gains it produced

SIMC — Skogestad, J. Process Control 13(4), 2003.

04 · Controller choice

Most loops do not need MPC. We will tell you which.

One identification supports both. The step response that tuned the PID is exactly what Dynamic Matrix Control needs — DMC, Cutler & Ramaker 1980, the lineage behind every commercial MPC package. So we can just run both and look.

Ask the MPC for gentler valve tighter setpoint

This is the one dial an MPC has and a PID does not: it trades setpoint tracking against how much the valve moves. Slide it and watch both numbers move together.

Live

Slide left and the valve settles down while the error grows; slide right and the reverse. Every position is a real controller — there is no setting at which it beats the PID on both at once.

At matched actuator effort, across every scenario here

A properly tuned PID wins 6 of 7. That is the textbook result for a single unconstrained loop, and it is the most useful thing on this page: MPC earns its licence fee on hard constraints, interacting multivariable units, and feedforward from a measured disturbance slower than the control channel. A single loop exercises none of those, and we are not going to pretend it does.

This is the economic argument, not a modest one. Advanced process control is reserved for the few dozen loops per site whose value justifies a specialist building and maintaining a bespoke model. Every plant has hundreds more below that line, still running the tuning they were commissioned with. Those loops do not need a better algorithm — they need the model kept honest, at a cost per loop that makes reaching them worth doing.

05 · Deployment

Two ways in.

ControlIQ never replaces the controller in your control system. It either commissions the one already there, or it produces the loop before the plant exists.

Retrofit — your plant, today running here

live loop in your DCS, PLC or single-loop controller

↓ supervised bump

identify the plant as deployed — and accept or refuse it

↓

controller commissioned into the block already running

Nothing of ours runs on your plant. No new hardware, no safety re-certification, and if we vanish tomorrow your loop is unchanged.

Greenfield — before startup Studio Pro link in development

FormulaIQ — the formulation

↓

Studio Pro — the process model

↓

ControlIQ — commissioned controller, day one

No plant trial at all. Measured here: commissioning from a design model carrying 5 % parameter error costs 0.5 % against a real plant trial, averaged over every scenario on this page.

The unit is one loop. A refinery is one company with several sites and thousands of them, and they are commissioned, and drift, one at a time.

06 · Transfer

It does not need to know what your plant makes.

The two plants on this page share every line of the identification, acceptance and tuning code, and agree on nothing else. Nothing was written for either of them specifically — which is what makes the cost per loop low enough to matter.

Exothermic CSTR

Behaviour	reverse-acting
Manipulated	coolant temperature, K
Controlled	concentration, mol/L
Time constant	≈ 2 min
Nastiness	multiple steady states
Actuator	lag larger than the plant

Textbook reactor from the open-source pcgym package. A 10 % feed-temperature step crosses its ignition point, so the browser integrates it at 500 RK4 sub-steps per step — fewer than 100 and it diverges outright.

Thermal rig real hardware

Behaviour	direct-acting
Manipulated	heater duty, %
Controlled	water temperature, °C
Time constant	≈ 62 min
Nastiness	heat loss is a power law
Actuator	the heater is the lag

Not a textbook model — a physical loop, instrumented and driven through real step tests. Fitted to measured data and validated on a run it had never seen: 1.05 °C RMSE on hold-out data, and 0.259 °C replaying a closed-loop run that included 27 minutes of actuator saturation.

07 · Roadmap

What is not on this page yet.

Badged honestly, because the fastest way to lose a control engineer is to show them something that is not running.

OPC UA to your control system in dev

Supervisory setpoint writes into Honeywell, Emerson, Yokogawa, ABB, Siemens. Kill the link and your own controller keeps running, bumplessly.

Studio Pro handoff in dev

The design model arrives as an artefact instead of being re-measured, closing the greenfield path end to end.

Re-identification on drift research

Plants foul and catalysts age. Naive re-tuning of a drifted plant measurably makes things worse in our tests — we have the numbers, and we are not shipping a claim we cannot defend.

Interacting loops in dev

Everything measured here is one loop at a time. Units where loops fight each other are the case model predictive control was invented for, and they are next — not claimed today.

Learned control research

A policy that transfers across plants with modest fine-tuning, benchmarked continuously against a properly tuned PID — which, as the section above shows, is a harder opponent than it sounds.

The tuning wasnever the problem.