Glossary of process measurement

The sensor measures in the stream itself: nothing is removed. In contact with the product, behind a window, or with no contact above the material — the criterion is not contact, it is that nothing leaves the process. The one position that allows an operation to be stopped at the right moment. What it costs is mechanical rather than optical: access, cleaning, temperature, window fouling.

A fraction of the material is diverted automatically from the stream to a measurement cell, then most often returned to the process. It suits an operation whose transit time stays short against its dynamics. A time that is quantified.

A sample removed from the process and measured beside the line, in seconds instead of tens of minutes. A real gain in development, and the pace of decision is what it turns on: the sampling adds to the analysis.

Sent to the laboratory. Presented as what you want to remove, it is what you never do without: it supplies the reference values that build the model, and then the ones that watch it.

These four modes are not trade vocabulary: they are defined by Ph. Eur. 5.25, which separates them by what is removed from the stream, where the equipment sits, and the delay before the result can be used.

A framework rather than an instrument: designing, analysing and controlling a process from measurements made in time to act. The decision loop is what makes a PAT approach, beyond the instrument.

The approach carried by the ICH Q8(R2) to Q12 guidelines: quality is built into the design of the product and the process rather than observed on the finished product. It assumes the variables that make it vary have been identified, and so measured.

It brings the vocabulary the rest of this site uses: the quality target product profile, the critical quality attributes and the critical process parameters that move them, the design space within which the process stays conforming, and the control strategy that keeps it there. In-line measurement is what makes that space observable during manufacture rather than after it.

The ICH guidelines, with their references.

A property of an incoming material whose variation moves a critical attribute of the product: excipient particle size, water content, lot origin. The concept is ICH’s — the annex to ICH Q8(R2), § 2.3 Risk Assessment: Linking Material Attributes and Process Parameters to Drug Product CQAs, and ICH Q11 § 3.1.5 for drug substance CQAs — the acronym is usage.

Mind the reference: ICH Q8(R2) comes in two parts, each with its own numbering. The § 2.3 of the guideline itself is Manufacturing Process Development; it is the § 2.3 of the annex that links material attributes to CQAs.

What it changes in a measurement project: it is the first cause of model drift. A CMA not identified at scoping becomes an unexplained deviation two years later, at the first change of supplier.

The method that decides what is critical. The operation is described as inputs, transformation and outputs, then the outputs go through a failure mode, effects and criticality analysis. ICH Q9(R1) names it in annex I.3 among quality risk management tools.

What it changes in a measurement project: it is what says where to measure. A measurement chosen without it is justified by what the instrument can do, not by what the process asks.

ICH Q8(R2): “an alternative approach to process validation in which manufacturing process performance is continuously monitored and evaluated”. It replaces validation on three batches.

⚠️ Not the FDA’s continued process verification, stage 3 of its validation lifecycle, which extends it. One word apart, two different commitments in a dossier.

“A condition in which the set of controls consistently provides assurance of continued process performance and product quality” (ICH Q10). The quality system must “establish and maintain” it.

What it changes in a measurement project: it is not a state you reach, it is a state you hold. Continuous measurement makes it observable rather than assumed.

The starting point of Quality by Design: what the product has to achieve in quality, safety and efficacy to do its job. Route of administration, strength, dosage form, release, shelf life, container. It is settled before the process, and the critical quality attributes are derived from it.

What it changes in a measurement project: a written QTPP is what lets you say why a given quantity is worth measuring in line. Without it the question becomes “what can this sensor do”, which has no useful answer.

A property of the product that stays within limits: content, residual moisture, particle size, dissolution. Rarely measurable in-line, which is why a related indicator is sought.

An operating parameter whose variation affects a quality attribute: temperature, duration, agitation, spray rate. The distinction with the CQA is the one between the lever and the result. The value sits in the link.

The documented set of controls that secures quality: in-process controls, specifications, process parameters, in-line measurements. A measurement carries regulatory weight once it is written into that document.

The range of process parameters and material attributes within which quality has been demonstrated. Moving inside that space is not a change in the regulatory sense; leaving it is. It is the central deliverable of the enhanced approach in ICH Q8(R2) and ICH Q11.

What it changes in a measurement project: a design space is only worth having if you can tell where you are inside it during manufacture. Established by design of experiments, it stays a map; the in-line measurement is what gives the position. What makes a measurement hold.

A batch is a quantity produced in one operation, as opposed to continuous manufacturing where material enters and leaves without interruption. In GMP usage a lot is a defined portion of a batch, identified for release.

The distinction is more than vocabulary: it decides what an in-line measurement can do. On a batch it serves to decide a stopping point. In continuous manufacturing it serves to drive a setpoint.

The part of a pharmaceutical dossier describing how the product is made, controlled and released. That is where an in-line measurement is written in, or is not. An instrument can work perfectly and stay outside the dossier: two different demonstrations. What a procedure demonstrates.

The moment an operation has reached its target: end of drying, end of blending, end of reaction.

Detecting it by measurement, rather than running the operation to a fixed duration, removes the safety margin paid on every batch, including the ones that did not need it. The corresponding page.

A blend endpoint criterion. The standard deviation of successive spectra is computed over a window of a few acquisitions, and the window is slid forward: while the blend is still homogenising the value falls, and when it settles low the operation has converged.

Its appeal is that it asks for no reference values: it follows a convergence rather than an absolute content. Its limit is the same one: it says the blend has stopped moving, not that it conforms. The two criteria complete each other rather than replace each other.

The trajectory a measurement describes over a whole operation, rather than the single value reached at the end. It repeats from batch to batch when the process is in command, so a running batch compares with the reference and a departure shows during the operation.

That is the fundamental difference between a Quality by Design approach and a final quality control: the laboratory verifies a result, the signature watches a behaviour. And it is what makes continuous manufacturing steerable. On a continuous line there is no endpoint: the batch is defined by a run time or a quantity (ICH Q13), not by the end of an operation. The signature then becomes the main object of control.

It calls for less calibration than a content measurement: comparing trajectories does not require returning an absolute value.

Concluding on the conformity of a batch from process data, without waiting for the final control. A reachable outcome, never an automatic one: it takes a full method validation and an accepted control strategy.

The region of overtones and combination bands of vibrations, mainly of C-H, O-H and N-H bonds. Broad, overlapping bands: a signal that is easy to obtain and rarely assignable to a single constituent, which is why a model is almost always involved.

FT-NIR is not a synonym of NIR. The Fourier transform names the way the spectrum is acquired, and an instrument can work in the near infrared without being an FT instrument. The technology page.

Tunable laser near infrared. A laser sweeps the useful wavelengths one by one instead of illuminating the whole band at once, so the flux available at each wavelength is of another order, which allows the material to be traversed rather than probed at the surface.

The practical consequence is the calibration burden rather than the spectral fineness: on a liquid whose composition is known, finite and declarable, the measurement decomposes onto pure component spectra and deploys quickly. A measurement with no chemometrics is still chemometrics, with its entry price elsewhere. The technology page.

The light enters the product, scatters inside it and leaves on the source side. The common mode on powders. One decisive particularity: the explored depth is not a constant, and it can vary during one process.

The light crosses the sample from side to side, so the information bears on a volume rather than a surface and the sampling is more representative. In return the geometry stays constant and the traversable thickness becomes limiting early.

Inelastic scattering of light: narrow bands, often assignable to a bond or a polymorph. Two conditions dominate. Fluorescence of the matrix, and heating of dark materials, which calls for reduced power. The technology page.

The region of fundamental vibrations: intense bands, far more specific than in NIR. Water absorbs very strongly there, so the optical path stays short, and it is that short path which makes the measurement tolerant of viscosity. The technology page.

FTIR is not a synonym of MIR. The Fourier transform names the mode of acquisition, used by the great majority of instruments in the field. It is a technique of MIR rather than MIR itself.

The beam reflects inside a crystal in contact with the sample and explores only a few microns beyond the interface. It is a surface measurement: in a heterogeneous medium the sampling is unrepresentative, and the crystal fouls.

Radiation that passes through transparent, translucent or opaque packaging, metal excepted. Its argument against X-rays is harmlessness: at 1 THz a photon carries about 4 meV against 26 meV for thermal agitation at 300 K, so no radiation protection. It works on non-metals, and metal reflects the radiation, which makes an indirect detection conceivable. The technology page.

Wall or belt electrodes measure the capacitance or the resistivity of the medium, from which the distribution of the phases is reconstructed. No multivariate model for the image. A few calibration points where a content is drawn from it. Resolution of the order of a few per cent of the sensor diameter. The technology page.

Low coherence interferometry: layer thickness and roughness read directly, with no multivariate model. The factor that decides these projects is mechanical, the probe to sample distance, held to within a few tens of microns. The technology page.

Particle size is derived from Brownian motion. In the spatially resolved version, multiple scattering is excluded rather than corrected, which allows measurement with no dilution. Laminar flow required, bubbles governing. The technology page.

The statistical methods that link an instrumental signal to a chemical or physical quantity. The word worries people because it is taken as a synonym for black box. It covers two opposite levels of cost.

Two ways of obtaining a content from a spectrum, and they cost differently. By decomposition, the spectrum is projected onto those of the pure components: the composition is known, finite and declared, and what is undeclared is not seen. By regression, the signal is linked to laboratory values, which calls for a population of samples covering the future variability. Both are chemometrics. Where the entry price is paid.

The most widespread inverse method: it builds combinations of wavelengths oriented towards the quantity to predict and tolerates highly correlated variables, which is the case of a spectrum. It stays inside its calibration range.

A neighbour of PLS: the spectrum is first reduced to its principal components, then regressed. That reduction ignores the quantity to predict. Often more stable, sometimes less performant.

The spectrum of the mixture is fitted as a sum of pure spectra. The oldest of the multivariate calibration methods, and therefore chemometrics in full. Structural advantage: extrapolation stays legitimate as long as the linearity established on the dilution series holds.

A direction built in the space of the spectra, meant to carry a real variation. Their number is a choice rather than a result: too few and the model explains little, too many and it learns the noise.

Corrections applied before modelling, derivatives, scatter correction, normalisation, to remove physical effects such as particle size. It is part of the model: changing it after validation changes the model.

The typical gap between predicted value and reference value, on samples the model has not seen. It adds the model error, the laboratory error and the sampling error. Quoted with its reference method and its matrix, it is a figure you can use.

The distance between a new sample and the centre of the learnt domain. It answers whether this sample resembles the ones used to calibrate. Without it, a model answers with no basis for knowing.

The part of the spectrum the model leaves unexplained. A high Q signals an unknown phenomenon: a contaminant, a new band, an instrument drift. T squared detects the unusual, Q the unexpected.

Where the texts name them: ICH Q14 cites both families as examples of diagnostics — « Examination of residuals […] (e.g., x-residuals or F-probability) » and « Outlier diagnostics […] (e.g., Hotelling’s T-squared or Mahalanobis distance) » — without requiring them. USP ⟨1039⟩ names them in the vocabulary used here, « Hotelling’s T2 and Q residuals (DmodX) », and Ph. Eur. 5.21 cites Hotelling’s T². ICH Q2(R2) names neither: it asks for the residuals to be examined, without designating a metric.

The region where the model is entitled to answer. Leaving it produces no visible error: the model returns a value silently. Out-of-domain detection is therefore part of the design. Keeping a model alive.

Running on a second instrument a model built on the first. The difficulty sits in the response of the instruments, and sometimes in a convention: in granulometry two instruments differ by about 11 per cent on an intensity basis and converge on a volume basis. Two instruments, two results.

The slow gap between what the model predicts and what the reference measures, as the process, the raw materials or the instrument evolve. The least covered subject in the market: monitoring tools exist everywhere, recalibration triggers nowhere.

A model that computes a quantity no sensor measures directly, from quantities that are measured. Electrical tomography returns a phase distribution with no multivariate model; getting a dry matter content out of it takes a soft sensor, which ties the measured conductivity to the content you are after.

That is the nuance that decides the cost of ownership. “Measurement without calibration” stays true for the physical quantity, and stops being true as soon as a concentration is wanted: a soft sensor is calibrated, monitored and maintained like any other model. What each approach costs.

The portion of material a measurement really takes into account. A reflection acquisition analyses a few milligrams: an increment rather than a batch. The mass analysed by the sensor.

The aggregation of several increments taken at different places or instants: the one way to approach the composition of a batch. Measuring the same point ten times reduces the instrument noise.

The gap between the portion measured and the batch it is meant to represent. No instrument performance reduces it. It is the one attributed to the sensor when a model comes up short. Sampling and representativeness.

The property of a sample in which every element of the batch had an equal chance of taking part. It is designed before measuring: a sensor placed in a dead zone will faithfully measure the dead zone.

Batch control by sampling and analysis, usually in the laboratory and after the operation. It verifies a result on a few grabs; between two samples the process is not observed. In-process measurement completes it rather than replacing it: QC stays the reference the models are built on. In French, CQ.

A result outside the acceptance criteria. It triggers a formal investigation, first in the laboratory and then in manufacturing, and a batch decision. It is the only one of the three notions covered by a dedicated FDA guidance.

What it changes in a measurement project: an OOS found in the laboratory arrives once the batch is already made. In-process measurement aims to make it rarer, not to replace it.

A compliant result within a series that is drifting from its history. The failure is not there yet, but it is approaching. The FDA OOS guidance mentions it only in a footnote, saying its principles may be useful for examining such results.

What it changes in a measurement project: a trajectory followed continuously shows the drift during the operation, not only batch after batch. Not to be confused with model drift, which concerns the measurement itself.

A compliant result that departs from what process knowledge predicted. It is a term of usage with no regulatory definition. The FDA OOS guidance nonetheless recommends that a low assay, even within specification, should raise a question.

What it changes in a measurement project: to call a result unexpected, you must know what was expected. That is exactly what a process signature provides.

A good practice guide for computerised systems: it classes software into categories and proportions the validation effort. The category follows the real use, and a configurable software always used in a frozen configuration stays in a low category.

The attributes expected of a datum: attributable, legible, contemporaneous, original, accurate. Then complete, consistent, enduring, available. In-line, the two that call for attention are attributable and original, since the raw datum travels and is transformed before storage.

The unalterable record of who did what, when and why. Its existence is one half: an auditor asks how it is reviewed, and whether the data passed to a third-party system carries the same information. Compliant is separate from validated.

Verifies that the system is installed in line with the specification: references, versions, connections, siting. On a probe, this is where the geometric parameters the whole measurement will depend on are checked.

Verifies that the system works as intended: wavelength accuracy, photometric linearity, drift. OQ bears on the instrument; the relevance of the model belongs to method validation. Telling them apart early sets the right scope.

Verifies that the whole holds in real conditions, on the real product: the one step that reveals fouling, vibration or a thermal gradient. Some chains call for it at every power-up.

The tests run at the supplier before shipment. A question to ask systematically: does the protocol qualify the measurement in flow, or the instrument at rest? Often the qualification bears on the static and the promise on the running.

The guideline on validation of analytical procedures, revised to cover multivariate methods: its § 2.5 is devoted to them. It expects the demonstration of range, accuracy, precision and robustness, and for a multivariate model a validation on a set of independent samples.

The handling of the out-of-domain case is not here: § 2.5 defers explicitly to ICH Q14 for the detail, and it is Q14’s diagnostics and maintenance plan that carry it.

The guideline on development of analytical procedures, complementary to Q2(R2). It installs a lifecycle logic: a procedure is designed, watched, and evolves. It is the official frame for the upkeep of a model.

The chapter on uniformity of dosage units. At the first stage, ten units are examined with an acceptance value of at most 15.0. Otherwise the test continues on twenty further units, thirty in all. Deriving content uniformity from it assumes cumulative conditions. Content uniformity.

The software layer between the instruments and the control system. It calls the method at the right moment, checks that every instrument is online, synchronises the start on the process event, applies the models and sends back the prediction. On a single measurement it is not needed; as soon as there are several measurement points, several linked steps or a release to prepare, it is what holds the whole together. What is asked of it, whichever the publisher: an audit trail, model version control, a separation between sandbox and GMP mode. Where the measurement stops.

The supervisory system. It receives the values coming from sensors and controllers, displays them to the operator, archives them and carries the alarms. It does not produce the measurement, it decides what is done with it: a PAT measurement usually reaches it over OPC UA. What the sharing of roles settles.

The same function as a SCADA, in an architecture where control is distributed across the plant rather than centralised. It is found mostly on continuous processes and large installations; SCADA is more common on batch processes and standalone equipment. For a PAT measurement the choice changes neither the interface nor the demonstration to be produced.

The part that acts: it opens a valve, stops a motor, changes a setpoint. It — not the chemometric model — carries the fallback behaviour on the day a measurement becomes unavailable. That behaviour is written into the requirement specification, before the wiring.

In open loop the measurement informs: it is displayed, it alarms, and an operator decides. In closed loop it commands: the value triggers the action with no human in between. Moving from one to the other is not a setting but a design decision — the measurement enters the control strategy, and a fault in it becomes a process fault.