Two suppliers, the same quantity to measure. The first announces a sample campaign and a development. The second announces that there is nothing to develop. Both are sincere. They are talking about different calibration burdens.
The calibration burden is what the measurement asks you to supply before a usable result: samples, laboratory analyses, line time, or nothing. It depends on your product, and it decides the calendar and the cost of ownership.
Three levels, and what each asks of you
Full calibration
A model is regressed on your samples, analysed by a reference method. The route of partial least squares and principal component regression.
What you supply: batches covering the future variability, laboratory values, and upkeep of the model over time.
Lean chemometrics
The spectrum is used with no reference value: moving block standard deviation, comparison of variances, unsupervised principal component analysis.
What you supply: spectra. What you get: a criterion of stabilisation or of relative conformity, rather than a content.
Calibration-free
Either the spectrum is decomposed onto pure component spectra, or the quantity comes straight out of the physics: a length, a distribution, a diffusion coefficient.
What you supply: the exhaustive composition, or nothing. What is undeclared is not seen.
This is an axis rather than an opposition, one the literature describes too: the work of an integrator is to place your process on that axis, at the point that fits your question.
The six degrees the literature distinguishes
Our three levels are a working simplification. The reference text distinguishes six, and that granularity earns its place as soon as the scope of a campaign is being negotiated:
- Opportunistic training set — unstructured data accumulated in production or drawn from historical records.
- Full calibration — every source of variance in the operating range is covered.
- Efficient calibration — the same range, described with fewer datapoints.
- Partial calibration — only the variance directly relevant to the target analyte is covered.
- Minimal calibration — datapoints reduced to the strict minimum, sometimes a single one.
- Calibration-free — no reference samples at all.
The three middle degrees are the ones that get negotiated. That is where the calendar differences of a project sit, and it is the conversation that “all or nothing” makes impossible.
Calibration-free is still chemometrics
The phrase circulates widely. A few pages further on, the same documents describe the same method as lean chemometrics and quantitative spectral decomposition. Both statements sit awkwardly together.
A decomposition onto pure components is a chemometric method. It is the oldest of the multivariate calibration methods. What disappears is the process sample campaign. The chemometrics stays, and the entry price moves rather than vanishing.
The wording matters beyond vocabulary. In front of a validation manager or an inspector, announcing a measurement with no chemometrics announces a method with no domain of validity and no anomaly detection. That is not what the method does, and the accurate wording is the one that carries the discussion.
What each level costs, and where it reaches its edge
| Level | What you supply | Time to first result | Where it reaches its edge |
|---|---|---|---|
| Full | A population of samples covering batches, seasons, suppliers, with laboratory values | The longest: it follows the production calendar and the variability to be covered | Outside the domain learned, the model returns a plausible value, which is why out-of-domain detection is part of the design |
| Lean | Spectra, and a stopping criterion to define | Short. The parameters are set at instrument qualification, the threshold is determined on the product | It says the blend is stabilised rather than the content is 5.0 per cent |
| Calibration-free | The exhaustive composition and pure component spectra, or nothing at all | Short, where the composition really is known and declarable | An undeclared constituent is not seen, and an interaction between constituents shifts the accuracy |
One point rarely made: decomposition onto pure components assumes that absorbances add. At high concentrations that additivity moves, through interactions between solutes, changes in the structure of water, band saturation. Serious implementations add a corrective term built on a dilution series. In other words there is a calibration, and it bears on standards rather than on your batches.
Four questions decide the level that applies
- Is your composition known, finite and declarable? A formulated medium, a buffer, a culture medium, a solution of excipients: yes. An agricultural raw material varies by nature, and its exhaustive composition stays out of reach.
- Are you after a value or a state? The content is 5.0 per cent calls for a full calibration. The blend is stabilised is obtained with no reference value at all.
- Is the quantity a concentration? A particle size, a thickness, a phase distribution are not concentrations. They come from direct physical measurements, or from correlations to calibrate.
- What will you have to demonstrate, and to whom? A requirement for a formalised domain of validity and anomaly detection applies whichever level is chosen. It stays in place when there is no regression.
The diagnosis stays compulsory in all three cases
Hotelling’s T squared, spectral residual, Mahalanobis distance: these indicators say whether the sample presented resembles the ones the method was established on. A decomposition onto pure components produces its own residual, and a residual that rises signals material the library has yet to describe.
What changes from one level to the next is the regression. Not the monitoring. What a method demonstrates before it is declared valid.
What the literature says
The subject is treated in the peer-reviewed literature, and the works below are the ones we cite in a meeting when the discussion stalls.
- Rish A. J., Henson S., Drennen J. K., Anderson C. A., Defining the Range of Calibration Burden: From Full Calibration to Calibration-Free, Journal of Pharmaceutical Innovation 19(3), article 39, 2024, doi:10.1007/s12247-024-09839-5. This is the text that sets out the axis, and it defines calibration burden as the summation of the time, material and financial demands across the whole calibration process.
- Rish et al., Lean chemometrics in spectroscopic process analytical technology, Journal of Chemometrics, 2026, doi:10.1002/cem.70105, which defines lean chemometrics as the set of techniques fitted to the question that minimise the calibration burden.
- Besseling et al., Journal of Pharmaceutical and Biomedical Analysis 114, 2015. The moving window F test for blend homogeneity monitoring.
- Muteki et al., Industrial and Engineering Chemistry Research 52(35), 2013. Composition prediction by iterative optimisation, a calibration-free or minimally calibrated approach.
- Zacour et al., Journal of Pharmaceutical Innovation 6(1), 2011. Reduced calibration methods for blend monitoring.
Once the level is known, the question becomes how many
Where your case falls to a full calibration, the next question follows immediately: how many samples, taken how, and over what period. How many samples a model takes.
Where your case falls to the other two levels, the arbitration moves to the representativeness of what the sensor actually sees. How much material a sensor analyses.
Buy, rent, or build
The number of samples is not the only variable: the model may already exist. On agricultural and food materials, the calibration sets of the CRA-W and ZEISS cover whole families of products, bought or rented.
The saving is not free, which is what makes it countable. It is paid in transfer onto your instrument, in a ring test to establish the gap on your batches, and in annual updating: three bounded items, against a sampling campaign whose cost depends on the variability of your material.
Tell us what you measure. You will know what you would have to supply.
Forty-five minutes is enough: placing your process on this axis, saying what each level would cost at your site, and knowing whether a sample campaign is avoidable. Or unavoidable.