Every measurement technology has a domain where it excels, and that domain is knowable before installing. Establishing it early is what makes an installation useful for years — and what points to the right route when another one serves your case better.
Three questions settle it, and all three are asked before the purchase. How many samples did the demonstration cover, and did they span the variability of your line? Does the available nozzle look at a zone where the product moves? Will the planned model get to see your changes of raw material?
Those three questions have one thing in common: they are answered before the purchase. What follows is organised by cause rather than by technology, because it is the cause that points to the route. The same family of instruments excels on one process and gives way to another next door.
In short
The limits fall into five families of causes: the matrix calls for a different physics; the quantity sought can only be followed by correlation; the physical conditions of the measurement are still to be met; the decision rate sets what can be controlled; the organisation is still getting ready. Each one shows up during a trial, sometimes in a single meeting, and is dealt with upstream.
Five families of causes, and how to recognise them early
Each is recognised during a trial, sometimes in a single meeting. And each is handled upstream, where it is really settled: the measurement point, the sampling, the choice of physics.
1. The matrix calls for another physics
A frequent cause, and a physical one first. The signal exists, and the material shapes it before it reaches the detector.
A product that is heterogeneous at the scale of the probed volume gives an exact measurement of a place that represents that place: a spectroscopic measurement analyses a few milligrams, and where the composition varies on a larger scale, each acquisition describes its own spot. A fluorescent matrix sits a RAMAN signal on a background well above the useful bands. A strongly aqueous medium closes MIR, since water absorbs there. A heavily pigmented coating keeps optical coherence tomography from the interface it is asked to measure, since without light returning from the lower layer there is no thickness.
TERAHERTZ has its own frontiers, and the manufacturer writes them: polar liquids such as water are the demanding case, the lines of solids are broadened, and through a pack there is strong spectral distortion from water absorption and scattering. Limits of principle rather than instrument defects. How much material a sensor really analyses.
2. The quantity sought is followed by correlation
There is no pure spectrum of a hardness. Nor of a viscosity, nor of a particle size. Those quantities are not chemical species, so nothing in the radiation corresponds to them directly.
They can be followed, though, because they vary alongside something that does carry a signature. What happens then is a correlation rather than a measurement, which calls for a model calibrated on samples, a population of reference samples, an explicit domain of validity, and upkeep.
And that type of model is the first to move. A change of raw material supplier, a batch of different origin: the correlation shifts, and the out-of-domain diagnostic is what flags it. A correlative model holds well, on condition that who watches it and updates it has been decided. Keeping a model alive.
3. The physical conditions are still to be met
The measurement is possible, and the place decides. The most mechanical cause: it turns on the installation rather than the chemistry of the product — and an installation is something you can design.
- Turbulent flow, or air bubbles. For an in-line particle size measurement the manufacturer writes it in black and white: the flow stays laminar and bubbles are governing. A bypass loop drawn for the measurement — bore, slope, purge — brings both conditions within reach.
- Shear that dominates diffusion. Classical DLS theory assumes purely random motion: flow must be avoided, or corrected for. Depth-resolved instruments measure the flow profile and derive the shear rate from it, which separates the two contributions — up to a point. Beyond it, the motion of the fluid covers the Brownian motion the size is drawn from, and it covers it the more as the process is agitated, which is where you wanted to measure. The correction has its own conditions: it fails on pulsating flow and on bubbles.
- A working distance that cannot be held. Some contact-free measurements call for a probe to product distance controlled to within a few tens of microns. Where the mechanical eccentricity of the equipment is larger, the answer is mechanical rather than optical.
- The absence of optical access. No free nozzle, no window. And a new penetration on qualified equipment brings a requalification. The question is handled at scoping, with the engineering team.
4. The pace of decision sets what is steerable
A measurement can be accurate, robust, perfectly calibrated. And arrive on a different time scale from the decision.
Two simple rules, and they settle many projects. An acquisition time longer than an event averages that event with the rest. And an at-line measurement returning its result in ten minutes documents a forty-minute operation rather than steering it.
So the first question of a project is at what pace you have to decide, and the technology follows from the answer. Where the pace asked for exceeds what the envisaged point allows, it is the measurement point that moves — and the question reopens at once.
That pace can be put in figures: the ratio of the decision window to the duration of the method meant to answer inside it. Below 2 the laboratory no longer keeps up, and that is where in-line measurement becomes at once the most needed and the most demanding. The table, operation by operation, on a bioproduction process.
The guideline states the same rule: ICH Q13, adopted on 16 November 2022, asks that data-averaging time intervals take into account the relevance of the PAT measurement frequency to the residence time distribution and to the process response time (§ 4.2, Control Strategy). What continuous manufacturing asks of the measurement.
5. The organisation is still getting ready
This one carries no technical dimension, and it is the one that carries furthest — because it is prepared before the installation.
A process contact able to say which drifts really exist. Line time for the trials, so the work goes beyond the easy batches. A stabilised reference method, since where the laboratory value varies, no model will be better than it. And an owner for the model, named for the rest of the lifecycle.
These points are settled before the installation, and they decide what the project will return. We give them a page of their own.
Two method limits of their own
There are cases where everything is in place, a favourable matrix, an available access, a compatible rate, and the method itself sets a boundary. Two of them come back often enough to be announced up front.
The width of the distribution sets what can be read
In light scattering granulometry a field rule holds that beyond a polydispersity index of 0.3 the default interpretation no longer applies and further analysis becomes necessary. It is not a normative threshold: general chapter USP <430> defines the index without setting a limit on it, and the only published threshold we could attribute is Malvern Panalytical’s, above 0.7. It is a limit of the mathematical model that turns the signal into a size, not a limit of the instrument. On a very spread suspension the displayed value stays stable and reads as a trend.
Two close populations are reported together
Two populations are resolved separately where their size ratio exceeds a factor of about three. Below that, one population is seen, whose mean moves. Where your question is whether you have agglomerates, and those agglomerates are less than three times the primary particles, another measurement answers it. A backscatter configuration improves the resolution in some cases without changing the nature of the method: dynamic light scattering remains a low-resolution technique, and the manufacturer writes so itself.
Both boundaries are published by the manufacturer, in its own technical documentation. That is where they are looked up, and where they are checked before a specification is written.
What we do when the answer is another route
Four routes open up then, and one of them is almost always practicable.
- Reformulate the question. The most productive route. Many requests are put as a value, I want to know my content, when the decision bears on a state: is it homogeneous, is it stabilised, has it changed since the last batch. Measuring a state often removes the need for a calibrated model, and with it the reference campaign and the upkeep.
- Move the measurement point. A signal absent in the vessel can be there at the pump outlet, on a bypass, at the entry of the next operation. Changing place changes the homogeneity the sensor sees, the geometry and the access.
- Change technology family. A matrix that closes one route opens another: a medium opaque to optical waves often stays accessible to electrical tomography. That is the role of a multi-brand integrator.
- Keep the control in the laboratory, and write down why. Sometimes the soundest decision is to install nothing for now, and, documented, that is a decision that keeps its value: which conditions would reopen the question, which equipment change would make the access possible.
The central idea: a documented conclusion is a deliverable. A report explaining how the signal separates, what was tested and which alternative is worth the next look saves the same study from being relaunched in three years by someone else.
It is also what separates a proof of concept from a commercial demonstration. A demonstration sets out to succeed. A proof of concept sets out to know.
Frequently asked questions
How do we know early which way it will go?
Through a trial on real samples, covering the variability you want to detect. The first question is separation, before accuracy: do two states you want to distinguish give signals that differ beyond the noise? Where the signals sit close at that stage, the most productive move is to reformulate the question — that is where the most ground is gained.
Is a laboratory demonstration enough?
It is a good start and a different situation. A demonstration runs on stabilised product, at controlled temperature, with no vibration, no flow, on chosen samples. Production brings all four at once. Hence the value of qualifying the instrument in motion, in the very conditions where it will do its work. Validating an analytical procedure.
How do you follow a quantity with no signature of its own?
It is correlated with a quantity that does, and that works in many industrial cases. The condition is to own what it implies: a declared domain of validity, residual monitoring, out-of-domain detection, and someone whose responsibility it is. With those in place, a correlative model earns its keep.
What does a report that points to another route hold?
What was measured, on which samples, with which settings. The decision criterion retained and the result obtained. The analysis of the cause, which of the families above is in play. And the alternative routes. It is a document that reads again in five years, when the instrument has changed.
Tell us what you are trying to measure. We will tell you which route to take.
Forty-five minutes is enough to identify which of these five causes bears on your project, what to verify first, and whether a trial is justified.