A model is not delivered once, it lives. The material changes, a supplier changes, the instrument drifts. So the question that decides the cost of ownership is who owns the model, and it has two good answers.
The first answer is your own team, and it is the one the Sustain phase is built for. It assumes someone at your site whose trade this is. Many plants run without a chemometrician: a sorting platform, a multi-product food site, a CDMO that changes client every two years. For them the second answer is the right one. We build the model and we maintain it.
What decides between the two is rarely technical. An installation goes in, the first months are good, and then the person who built the model moves to another post. The installations still measuring three years later are the ones where that role was named at purchase, with a maintenance budget behind it.
A model with a named owner is a model that keeps its value. That owner sits at your site, or with us, and both routes are written the same way: who reads the residuals, on what criterion, and who recalibrates.
In short
Three routes: a turnkey calibration, checked on your batches before it is used; a model built on your material; a model we maintain over time. In all three, the same answers are written down: who reads the residuals, on what criterion, and who recalibrates.
Three ways to have a model, at three different costs
A ready-made calibration
On common matrices a model already exists — and there are two kinds, better not confused. Quantitative calibrations, which return a content: those of ZEISS on agricultural and food products, and those of the CRA-W in Gembloux, which we work with on any NIR instrument. Identification libraries, which return a name: the ones Enwave ships with its handheld Raman instruments. The fastest route, and the lightest to start with — bought or rented.
It was built on their sample population, so it is checked on your own batches before it is used, and it usually takes a bias and slope adjustment. Three checks settle it before any commitment: does its range cover your usual values, is its reference method yours, and does the gap measured on twenty of your batches sit inside what you accept. Three times yes, and you have saved a sample campaign.
A model built on your material
Where the matrix is particular, the common case, the model is built from your reference values and your real drifts.
Deliverable: the model, its written range of validity, the pre-processing chosen and the reason for each choice. What a calibration costs, route by route.
A model we maintain
Watching the residuals, detecting samples outside the range, recalibrating when a supplier or a season changes, transferring to a second instrument. A commitment over time.
It is the answer for a site that runs without a chemometrician, and it is the one that keeps an installation measuring for ten years. What keeping a model alive means.
In which software? Yours, preferably
A model lives in a software environment, and you often already have one. Moving it is costly, in habits as much as in documentation. So we work in yours when it exists. Blend endpoint criteria run in the instrument’s own software.
Vektor Direktor
The platform on which we develop and maintain when the choice is ours. Model versioning, history and change log, which makes it traceable which version produced which result.
Sensologic
The chemometrics software used with ZEISS spectrometers. When your installed base is ZEISS, the model stays in the instrument’s environment, and operations do not change tools.
Unscrambler
Long practice, in spectral library development and in training. If your models are already there, they stay there: we work inside it rather than imposing a migration on you.
Training on these tools is part of what we carry, and it is handled with the rest of the move to autonomy. The Sustain phase.
What “maintaining a model” means, in concrete terms
The word covers six actions, and none of them is automatic.
- Reading the residuals. A control chart on the gap between prediction and reference. It is the first sign of a drift, and it comes well before the result goes out of specification.
- Detecting entry into unknown territory. Hotelling’s T², Q residuals, distance to the training population. A model always predicts something, including on a sample it has never seen. That is exactly what these indicators make visible.
- Deciding when to recalibrate. A change of supplier, a new campaign, a recipe modification. The question is not to recalibrate often, it is to know on what criterion.
- Topping up the calibration set. Taking the missing samples, obtaining their reference values, readjusting without destroying what worked.
- Transferring to another instrument. A second analyser, a replacement after a breakdown, a second line. Why two instruments do not give the same figure.
- Tracing. Which version of the model produced which result, and on what date. Without that, a result from last year can no longer be defended.
Software that carries versioning and a change log makes these actions verifiable. It does not, however, make your system compliant as a result: overall compliance depends on your authorisations, your procedures and your audit trail review. What a declaration of conformity never covers.
Chemometrics, machine learning and Annex 22
Under the draft EU GMP Annex 22 on artificial intelligence, a PLS model is a machine learning model: its parameters are optimised from data. Whether the final text confirms this remains to be seen. Released for consultation in 2025, the draft had not been adopted as of September 2026, and it covers only the manufacture of medicinal products and active substances.
For critical applications, it accepts frozen, deterministic models that always return the same result for the same spectrum. It excludes models that keep learning in service, generative AI and LLMs.
Its requirements overlap with familiar chemometric practice:
- an independent test set, never seen during development;
- a review of the variables driving each decision;
- a confidence indicator, and an “undecided” result when it is too low, the role Hotelling’s T² and Q residuals already play;
- monitoring of performance and of input data drift, under change control.
One requirement is new: whoever has seen the test data should not train the same model alone. It is best organised at the scoping stage. Regulatory texts.
What stays with you, whichever route you take
Three things stay yours on all three routes, and they are better known at scoping than at commissioning.
The reference values
A model is built against a laboratory method, and the scatter of that method sets the accuracy you can reach. Knowing that figure early tells you exactly what the model can deliver, and that is a useful conclusion.
The operating decision
When the measurement leaves its range, someone decides what happens to the batch. That decision belongs to your operations and your quality assurance. We provide the signal and the criterion.
The documentation
Specifications, qualification protocols, change control. They live in your document system. We write them with you, and they stay with you.
What a turnkey calibration covers
A calibration delivered and a calibration verified on your material are two steps
A manufacturer’s calibration is a real saving of time, and we use them whenever they fit. It was built on a population of samples that is not yours, with a reference method that is not necessarily your laboratory’s.
Three checks settle it. Does the range the calibration covers contain your usual values. Is its reference method yours, or does an adjustment follow. And does the gap measured on twenty of your batches stay within what you can accept. Where all three answer yes, you have saved a sample campaign. Where they do not, you know exactly what is missing.
It is the same logic as everywhere else on this site: a result announced by a supplier is verified on your material before it is relied on. What validating a procedure means.
A partner calibration is not bought, it is installed
A purchased model does not run on your spectrometer on day one, and it ages. Those are the only two serious objections to the ready-made route, and they decide whether the saved sample campaign is real. Both are handled, and both are part of the service.
Transfer onto your instrument
A model lives on the instrument that built it. Moving it onto yours is a calibration transfer, not a file copy: the problem is the same as between two spectrometers in one fleet.
The ring test
Your batches are measured at your site and at the model holder’s, and the results are compared. It is the defensible form of the twenty-batch check described above: a gap established rather than declared.
Updating, under annual contract
A material changes from one season to the next. The model is revisited at an agreed interval on recent batches, and the version in service is traced. Without that, a model degrades quietly.
Bought or rented. Rented, the model becomes an annual operating line rather than an investment — which changes nothing technically, and a great deal about the decision to start.
Where this applies. For quantification: on receipt and in the laboratory, on agricultural and food materials, with the calibration sets of the CRA-W and ZEISS — and with ZEISS the same service approach also exists in-line, on the process. For identification: on plastics and polymers, with the libraries shipped on Enwave‘s handheld Raman instruments — more than 45 plastics on the PolyMax, more than 145 entries on the PolyLab, and your own spectra can be added. Sentronic publishes a raw-material identification database of its own, validated against chapter 2.2.40 of the European Pharmacopoeia. Elsewhere we develop your models, or take over and maintain the ones you already have.
Tell us who looks after your models today. That is where the answer starts.
Forty-five minutes is enough to see whether an existing calibration saves you a sample campaign, what a model built on your material would ask for, and who keeps it alive afterwards. Bring your constraints: the software you run, how many models you hold, how often your materials change. If your team can carry it alone, we will tell you what it needs to hold up over time.