The laboratory announces one value, the in-line instrument announces another, and the meeting that follows runs two hours. In most cases neither instrument is at issue.
The scenario is always the same. A production line installs an in-line measurement, then compares it for a few weeks with the laboratory method. The gap is there, systematic. The sensor is suspected, the qualification redone, the probe changed, the supplier called back. A quarter goes by.
Or else: a calibration model works perfectly on the development spectrometer. It is loaded onto the workshop instrument, same type and same manufacturer, and the predicted values are offset, with no line in the schedule for it.
Neither story has an instrumental cause. Both have a cause of convention, and a cause of convention is settled with a document rather than with a technician.
In short
In most cases, the gap comes from an unwritten convention, not from a faulty instrument. A complete specification therefore names its conventions: for particle size, the basis of calculation, the detection angle and the algorithm; for a spectroscopic measurement, the spectral range, the resolution, the preprocessing and the reference instrument of the model. These are fixed at scoping.
First case: two particle sizers, eleven per cent apart, both right
A measured and published result shows the mechanism better than any explanation. On the same injectable lipid emulsion, two dynamic light scattering instruments return 332 nm and 369 nm. Both are Zav values — the intensity-weighted harmonic mean diameter, the quantity any DLS instrument reports by default. About 11 per cent apart, or 35 nm. In a common industrial specification, that is enough to open an investigation and occupy three departments.
Converted to a volume distribution, the same measurements give 415 ± 5 nm on both sides, with the same relative variance.
Published figures: T. Rooimans et al., “Development of a compounded propofol nanoemulsion using multiple non-invasive process analytical technologies”, International Journal of Pharmaceutics 640 (2023) 122960. The two instruments: a Zetasizer Nano S in visible light at 173°, and a NanoFlowSizer in the near infrared at 180°.
Nothing was wrong, and nothing was corrected. Both instruments were measuring the same emulsion, with the same physics. They were expressing their result in two different conventions.
And the cause is not the one usually given. The two starting values are the same quantity, returned by two instruments that do not illuminate at the same wavelength and do not collect at the same angle — 173° against 180°. While the droplets stay far smaller than the wavelength this changes nothing: scattered intensity then follows diameter to the sixth power, whatever the angle and the colour. Beyond that, the Mie regime begins, and the weight each size carries in the signal becomes specific to the instrument. At 350 nm in visible or near-infrared light, that is where we are.
The volume distribution depends on neither of those two choices. That is why both instruments recover the same value there. Two settings made in the factory, and rarely written into the specification. Particle size and granulometry.
One consequence runs against intuition, and it is the one that traps: outside the Rayleigh regime, the Zav can be smaller than the volume mean — here 332 and 369 against 415. The received idea that an intensity-weighted value is always the larger one holds in the Rayleigh regime only.
And the gap is not a detail. Simulations published by the makers of one of the two instruments — an interested party, but a checkable calculation — put a figure on it: between two instruments whose wavelengths differ by a factor of two, the gap between the two Zav values reaches 60 nm, or 20 %, around 300 nm for a broad population, and still exceeds 5 % for a narrow one. Comparison of Intensity and Volume-based size data for different DLS instruments, R. Besseling, M. Hermes & C. Schuurmans, InProcess-LSP.
The consequence is blunt: a particle size result is only comparable with a result expressed on the same basis. Comparing an intensity-based value with a volume-based one means comparing two different quantities that carry the same name and the same unit. See particle size.
Second case: the same model on two spectrometers
Spectroscopy plays the same tune. Two spectrometers of the same type, illuminating the same samples, return different spectra. The gaps are small and of no importance while a spectrum is read by eye, and decisive as soon as a multivariate model turns them into a number.
What differs between two instruments
Spectral resolution: two instruments cut the radiation with different fineness. Wavelength alignment, where a slight offset moves every band. The instrument line shape, the form the instrument gives to a very narrow line. And detector non-linearity.
Why that is enough to shift a value
A model calibrated on samples has learnt a numerical relation between spectral shapes and reference values rather than the chemistry. Move the shapes slightly and the relation applies differently. The model carries on returning a value, plausible and shifted — and residual monitoring is what flags it. That is what makes the point worth designing for rather than merely noting.
A pure-component model transfers better, and that is no accident
It is a structural advantage worth naming: it weighs heavily in a choice of architecture, and it is at scoping that it is quantified.
A pure-component model decomposes the measured spectrum onto pure component spectra. It rests on a physical property of the molecules rather than on a regression learnt on a population of samples. Changing instrument does not change the physics, so the model stays valid with a far lighter levelling. A model calibrated on samples learnt on one precise instrument, with its own characteristics, and it carries them with it.
That advantage sits alongside the conditions of the pure-component model
The pure-component model calls for three things: a composition that is known, finite and declarable, pure component spectra available in the real solvent, and a quantity that really is a concentration. Where your matrix is natural or variable, or where you are after a hardness or a particle size, the route is the model calibrated on samples, with the campaign that comes with it. It is a tie-breaker rather than a selection criterion. What each route costs.
The four families of transfer, in plain words
When a model calibrated on samples has to live on several instruments, four strategies exist. They cost differently and they arrive at different moments of the project.
| Family | Principle | What it assumes |
|---|---|---|
| Standardise the spectra | The spectra of the second instrument are corrected to resemble those of the reference instrument before being presented to the model | A sample set measured on both sides. The model is left untouched, a real advantage in a regulated environment |
| Correct the predictions | The model predicts, and its output is corrected by a relation established between the two instruments | The simplest and the most fragile: it assumes a stable bias over the whole domain |
| Resample onto a master instrument | One instrument is declared the reference. All the others are brought back to its wavelength grid and its response | A fleet discipline: the master instrument is maintained, requalified, and outlived at some point |
| Calibrate on several instruments | The calibration set holds spectra from several instruments from the start | The most robust and the cheapest, on condition of having thought of it before building the model |
None is superior in the absolute: the choice follows the fleet, the regulatory status of the model and the possibility of circulating samples.
The EMA NIR guideline frames these routes explicitly, in its § 7.3: a mathematical compensation inside the instrument software — bias, slope or vector correction — where that is enough, and otherwise calibration and validation repeated and confirmed on the additional instrument.
The transfer samples decide the result
All these methods rest on the same object: a small set of samples measured on both instruments, which establishes the correspondence. It is the load-bearing link, and the least prepared.
- They cover the domain rather than the centre. A set concentrated around the nominal value establishes a correspondence valid at the centre and shifted at the ends, which is where the decisions are taken.
- They stay stable over time. Between the measurement on the first instrument and the one on the second, the sample stays as it was. A product that takes up moisture invalidates the exercise silently.
- They are presented identically. Same packing, same thickness, same temperature, same probe geometry. Any difference of presentation ends up inside the correction and stays there.
- Few are needed, and well chosen. One of the rare situations in chemometrics where the number counts less than the spread, unlike the calibration itself. And the EMA NIR guideline gives the selection criterion: samples with good multivariate leverage — the ones that have a large influence on the calibration model. How many samples it really takes.
A transfer is planned at the design stage
This is the one recommendation on this page that costs money when it is left out, and almost nothing when it is applied. It consists of writing into the calibration protocol that the model will one day be carried elsewhere. Keep the transfer samples, measure on more than one instrument from the start, document the configuration of the development instrument, freeze the pre-processing chain.
And it carries a regulatory name. The EMA NIR guideline asks that a transfer be the subject of a comparability protocol, and provides for it to be filed with the initial application, as a post-approval change management protocol, where the transfer is foreseen. Planning it at the design stage is not an integrator’s precaution: it is the route the text opens.
A transfer improvised two years later is no longer a transfer
Because the samples of the initial campaign are gone. Because the development instrument has been repaired, recalibrated, sometimes replaced. Because the person who built the model has moved on. And because in a regulated environment, touching a validated model opens an impact assessment and a requalification. What follows is a new calibration rather than a transfer, under time pressure, on production batches.
A documentary problem before it is a technical one
The conclusion of both cases is identical. What is left out of the specification will be settled by the factory setting of an instrument that was never asked. A complete specification therefore goes beyond a target value and a tolerance. It names the conventions that give that value its meaning.
- For a particle size: the calculation basis, intensity, volume or number, the detection angle and the inversion algorithm.
- For a spectroscopic measurement: the spectral range, the resolution, the pre-processing chain and the reference of the instrument the model was established on.
- In both cases: the reference method, with its own conventions made explicit. With that in hand, a trial is judged on a real gap.
These elements are frozen at scoping, verified during integration and documented in the method validation file. None of those steps calls for a rare competence: simply for the question to have been asked before rather than during an investigation.
Frequently asked questions
How do we know whether a gap comes from a convention or from a real defect?
A difference of convention produces a systematic and structured gap: same sign, regular evolution with the measured value. An instrumental defect produces an erratic one. So the useful reflex is to plot the two series against one another before dismantling anything. A cloud aligned on a line that is not the first bisector points to a convention.
Does changing instrument mean redoing the whole calibration?
Not where the transfer was planned. With a preserved sample set and a documented configuration, the operation is short and bounded. Without it there is no transfer: the calibration is rebuilt, with a new reference value campaign, and the calendar becomes production’s again.
And if our supplier states that the instruments are interchangeable?
That is a statement that can be verified. Ask for the transfer protocol, the method used and the data supporting it on instruments comparable to yours. A declaration of interchangeability with an attached protocol is one you can use. The question belongs before the purchase, alongside what happens on the day the model drifts. Keeping a model alive.
Does a measurement with no multivariate model escape the problem?
It escapes the model transfer, and the convention question stays. The case of the 332 and 369 nm bears precisely on a measurement with no multivariate model. A quantity coming out of the physics still follows the choices of signal processing and weighting, made explicit with the same rigour. And there remains the question of which portion of material it was obtained on. How much material a sensor analyses.
You have a gap between two instruments. Tell us which ones, and on which quantity.
Forty-five minutes is enough to know whether the gap belongs to a calculation convention, to a model transfer to work up, or to a real measurement defect, and whether it justifies engaging the work.