Ask five manufacturers “does your technology suit my process?”: you will get five yeses. Each of them sells only one answer.
Product sheets give the instrument specifications: resolution, spectral range, ingress protection rating. None of those three decides the fate of a project.
Four questions complete them, and they are the ones that decide.
- The model: does one have to be built, and on which population of samples?
- The matrix: is it compatible with the measurement principle?
- The access: is there an installation point at the right place in the process?
- The pace: how often do you have to decide?
We integrate nine families of non-destructive measurement, each carried by at least one manufacturer partner. The choice between them is made on your question. A scoping phase can conclude that your quantity is best followed indirectly, by tracking the parameter that governs it.
In short
What separates the nine families is where the displayed value comes from. Either it comes from a model learned on data, to be built, validated and maintained: NIR, NIR-HPTLS, RAMAN, MIR depending on the matrix, TERAHERTZ as soon as a content is targeted, hyperspectral imaging. Or it comes from the physics of the measurement: electrical tomography, OCT, SR-DLS, machine vision. The parameter to measure then narrows the list, often to two candidates.
The nine families, side by side
The table reads by column, not by row. A technology is chosen on three points: the quantity followed, the matrix, and the calibration burden accepted.
| Technology | What it follows | Suitable matrices | Calibration burden | Installation | Known limit | Reference texts |
|---|---|---|---|---|---|---|
| NIR spectroscopy | Moisture content, content of a constituent, blend homogeneity, composition of a liquid | Powders, granules, tablets and divided solids in diffuse reflection; liquids, suspensions and emulsions in transmission or transflection | Full, PLS or PCR for a content. None where the criterion is a stabilisation | Probe in contact, window, or contact-free above the flow | Probes a few milligrams per measurement. Penetration depth varies with the material | In pharma, Ph. Eur. 2.2.40 and USP ‹856›; outside pharma, ISO 12099, ASTM E1655 and D6122 |
| NIR-HPTLS | Concentration of the components of a liquid | Bioproduction media only | None on a defined medium: decomposition onto pure-component spectra. In a living culture, calibration on samples | Fixed optical path | Calls for a composition that is known and exhaustively declared. Works from what it has been told, and discovers nothing that was not declared | Those of NIR |
| RAMAN, 785 nm excitation in handhelds | Identity, polymorphism, reaction progress | Liquids, suspensions, solids, polymers | None to identify: a library and a threshold. Full for the quantitative | Immersed probe, contact-free measurement, or in contact for a handheld | Fluorescence of the matrix. Dark materials heat, which calls for reduced power | In pharma, Ph. Eur. 2.2.48 and USP ‹858›; outside pharma, ASTM D6122 and E1840 |
| MIR spectroscopy | Composition in the liquid phase, reaction progress and endpoint | Liquids, emulsions, viscous and loaded media | Full as soon as bands overlap. Following one isolated band can be enough for a switching point | Short path length cell, in-line or on a bypass loop | The optical path stays short because water absorbs strongly. It does tolerate viscosity | In pharma, Ph. Eur. 2.2.24 and USP ‹854›; outside pharma, ASTM E1655 and D6122 |
| TERAHERTZ | Identity, integrity and content of a packed product | Packed solids, transparent, translucent or opaque packaging, metal excepted | A model to identify or class. None for a packaging inspection, which is a geometric criterion | Contact-free. Non-ionising radiation, so no radiation protection and no controlled area | Polar liquids are the demanding case. Strong spectral distortion through packaging. Metal reflects the radiation, and it does not pass through | In pharma, Ph. Eur. 5.24 names terahertz; elsewhere, no text: the method is validated against your reference method |
| Electrical tomography ECT and ERT | Phase distribution, end of homogenisation, passage of a product front | Vessels, pipes, opaque volumes | None for a distribution; a few calibration points for a content | Flush wall electrodes, or a fitted belt | Spatial resolution of a few per cent of the sensor diameter. Returns a distribution with no model, a content only through a calibrated soft sensor | None, in pharma or elsewhere: the method is validated against your reference method |
| Optical coherence tomography OCT | Layer thickness, dispersion of that thickness, surface roughness | Coatings, tablet coating | None | Contact-free, in a coating pan or in-line | Probe to sample distance held to within a few tens of microns. Heavily pigmented coatings are the demanding case | In pharma, Ph. Eur. 5.24 names OCT; for roughness, ISO 21920-2 and 21920-3 |
| SR-DLS | Particle size and size distribution, with no dilution | Emulsions, suspensions, nanosuspensions, concentrated media | None | In-line or on a bypass loop, with no sample preparation | Laminar flow required, bubbles to be excluded. It returns the size and the distribution; particle shape and surface charge belong to other methods | None for the spatially resolved variant; USP <430> in pharma and ISO 22412 elsewhere cover classical scattering, on a diluted sample |
| Imaging and machine vision | Number, size and shape of objects. Spatial distribution of a constituent | Surfaces, countable objects, tablets, films, product moving in a sheet | None for a geometry. Full for a classification or a content per pixel | Camera and dedicated lighting, contact-free. The lighting decides rather than the camera | Sees the surface. The absolute value per pixel is noisier than a point spectrum, and the hyperspectral cube raises a data volume question | In pharma, Ph. Eur. 5.24 and USP <1776>; outside pharma, ISO 13322-1 and -2 for particle size |
Of the six columns, two come from field experience: “Calibration burden” and “Known limit”. The first sets the cost of ownership and the start-up lead time. The second depends on your matrix and your measurement point, and it is established at scoping. What a control costs, and when an in-line measurement pays for itself.
The reading axis that is missing elsewhere
What separates these nine families is the origin of the displayed quantity. Wavelength and spectral fineness change nothing there. Either the quantity comes out of a model learnt on data, which will have to be built, validated, monitored and maintained. Or it comes out of the physics of the measurement itself.
| Call for a model to build and maintain | The quantity comes out of the physics |
|---|---|
| NIR spectroscopy, inverse model | Electrical tomography ECT and ERT. The distribution is reconstructed from electrical measurements |
| NIR-HPTLS, direct model | OCT. The thickness is an optical distance, and it reads |
| RAMAN, in quantification as in identification | SR-DLS. The size is derived from Brownian motion |
| MIR, with or without a model to calibrate depending on the matrix | Machine vision. A number, a length, an area are measured in the image |
| TERAHERTZ, as soon as the question is a content or a classification | none |
| Hyperspectral imaging, inverse model, applied pixel by pixel | none |
A measurement “without chemometrics” does not exist. A calibration on pure components is chemometrics, and it is even the oldest of the multivariate calibration methods. The entry price is paid in both cases, in different places.
- Model calibrated on samples: a design of experiments representative of the future population.
- Pure-component model: an exhaustively declared composition and a qualified spectral library.
The right-hand column of the table above is the only case without a multivariate model: the quantity there is geometric or physical. Four entries, not one more. Pure-component model or model calibrated on samples: what each costs, and when each is legitimate.
Start with the parameter, not with the technology
Coming in through the technology leads to comparing instruments before having written down what you are trying to know. The parameter to measure drives everything else: the accessible physical quantity, the candidate technologies, the calibration effort, and therefore the cost of the project.
- A moisture content to follow during drying: measuring moisture. The reference method decides before the technology: titration and loss on drying do not measure the same water.
- A blend where you need to know when it is homogeneous: content uniformity, and endpoint detection if the criterion is a convergence rather than a value.
- A raw material to identify on receipt: raw material identity. The value lies in the spectral library, and in extending it to the families the standard one does not cover.
- A particle size in suspension or emulsion: particle size. A single point settles it: measuring without dilution, or not.
- A coating thickness deposited on a tablet: coating thickness. TERAHERTZ does not answer it as things stand. OCT tomography does.
- A cut point in distillation: liquid composition. MIR follows it continuously in an organic liquid.
Once the parameter is set, the list of candidates shrinks by itself, often to two. When it falls to zero, another route remains open: following a quantity that governs the one you are after, the way viscosity is followed through what determines it. That is what the scoping phase settles, before any purchase.
Three warnings before choosing
A technology is chosen on a trial
Published specifications describe the instrument in the laboratory, on a clean matrix. They say nothing about what decides projects.
- Fouling of an optical window.
- The temperature gradient between the measurement point and the sampling point.
- Mechanical access, sometimes non-existent at the right place.
- Raw material variability that the model has never seen.
Admissible service conditions and behaviour in flow belong to a requirements specification. A trial on your product settles it.
The question of pace comes before that of precision
A very accurate measurement that arrives after the decision documents instead of steering. Before comparing accuracies, write down how often you have to decide something: stop an operation, divert a flow, alert an operator. It is that frequency that eliminates entire families.
It drives a second requirement: acquisition time and probed volume matched to the linear speed of the process. Without that, the measurement describes a state that has already passed.
Rates quoted out of context cannot be compared with each other. Public documentation gives very different values depending on the test configuration. Ask for the configuration, and the figures become comparable again.
A comparison published by a manufacturer is useful, but structurally partial
Some of these comparisons are technically careful. A single-technology challenger readily compares, since it is attacking a market that is already occupied. An established player never compares. In both cases the grid is built on the axes where the author is strong, and those where it is weak do not appear. Read them, and start with the list of missing criteria.
Frequently asked questions
Is a technology with no model always preferable?
It is preferable where the quantity it gives is the one you need. OCT returns a thickness rather than a composition, a size measurement returns neither shape nor surface charge. Electrical tomography returns a content through a soft sensor with a few calibration points, which takes it out of the no-calibration column. What is gained in calibration is traded against the reach of the answer. One question settles the choice: which quantity really drives your decision?
How do we verify a claim of regulatory compliance?
By asking for the document. A signed declaration of conformity carries a document number, a software version and a date. It also states, almost always, that overall compliance depends on the operator’s procedural controls. That reservation is what counts: compliant and validated system are two statements. The compliance matrix is requested with the declaration. What a declaration covers.
And the vocabulary? PLS, ATR, Q residuals, increment…
The terms used on this page are defined one by one, for a process engineer rather than for a spectroscopist. The glossary of process measurement.
Tell us what you are trying to know. You will know which families stay in the running.
Forty-five minutes is enough to place the families that hold on your matrix, the ones that deserve a trial, and the calibration burden of each. Where none of the nine follows your quantity directly, the indirect route stays open.