Tuning analyzer and composition loops
An analyzer loop is not a slow temperature loop and it is not a pH loop. The measurement arrives in bursts — a chromatograph cycle, a sample transport delay, a conductivity cell that sees yesterday's fluid — and the controller is blind for most of the time between updates. Treating it as a continuous PV with a large time constant produces a tune that fights a number that has not changed yet.
pH is its own nonlinearity. Composition, conductivity, moisture and GC loops share the sample-system problem instead: dead time you cannot tune out, a sample that may not represent the stream, and a cycle time that sets the fastest reset you can justify.
The sample system is the loop
Time from process to reading is the performance limit. Fast loops, short fast loops, and a cell that is actually in the stream all beat a clever PID. If the chromatograph cycle is four minutes, integral action faster than that is guessing between samples.
A sample that is unrepresentative — lagging, condensing, leaking air — gives a controller a PV that is not the process. No constant fixes that. Fix the sample, then talk about the PID.
What the controller can do
PI, slow, often with the proportional and integral terms acting on a filtered or held value rather than the raw analyzer spike. Derivative is almost always harmful: the PV is a step every cycle, and derivative turns that step into a kick. Gain scheduling helps when the process gain changes with composition, which it often does, but only after the sample is honest.
Do not borrow the pH playbook
pH is dominated by the titration curve. Analyzer loops are dominated by delay. Mixing, reagent valves and dual-base/acid trains are pH fixes. They are not the first things to reach for on a GC or a conductivity loop in a clean stream.
Questions that come up
Can I use Smith predictor or dead-time compensation?
Only if the dead time is known and stable and the model is kept. On an analyzer whose cycle time drifts, a predictor that assumes a fixed delay will be wrong in the same way a fast reset is wrong. Most plants are better served by a slower PI and a faster sample.
Related
Last reviewed 2026-09-10.