“When do I cut my fraction? Has my emulsification reached its plateau? Is my reaction finished, or only twelve minutes old as the recipe says?” Three questions whose answer is settled during the operation — and, on a liquid, in the most favourable situation in process measurement: you know what you put in.
A distillation column steered on head temperature, with a sample sent to chromatography every hour. Chromatography gives the exact composition of the sample; between two samples, the moment the fraction turns remains to be estimated, and the cut is made wide.
An emulsification run by the clock because the recipe says twelve minutes. Was it stabilised at eight? Is it really stabilised at twelve? The recipe does not say. Viscosity and turbidity barely move during the phase that matters.
A polymer functionalisation whose degree of substitution is checked afterwards, by titration or by magnetic resonance. By the time the result comes back, the reaction has carried on. When it has gone too far, there is no going back.
In all three cases, the reference value is right, and it remains the reference. What is missing is the same information while the operation is running.
In short
MIR follows a concentration from 500 ppm to 100 % in organic liquids, without calibration if the composition is declared. NIR-HPTLS targets defined aqueous bioprocess media, by decomposition on pure components.
Depending on your sector: distillation in chemicals, emulsification in cosmetics, bioproduction in biopharma.
What is actually measured
A spectroscopic measurement records the absorption of the medium over a range of wavelengths, and the concentrations come back from a calculation. What separates short projects from long ones holds in three questions, in this order.
Do you know exactly what is in it?
Where the list of constituents is known, finite and fully declarable, a calibration route opens that exists nowhere else. Where the medium is natural or partly characterised, the route is regression on reference values.
Are you after a concentration or a property?
A concentration has a spectrum. A viscosity, a stability, a rheological behaviour are correlated with the spectrum through the matrix rather than by a direct physical law. That boundary alone decides the calibration burden.
Is the optical path fixed?
The condition most easily overlooked. Decomposition onto pure components assumes a constant and controlled path length. A probe geometry that varies, progressive fouling, a passing bubble: each returns a concentration error rather than an error message.
Where the composition is known, finite and declarable, a pure-component calibration becomes possible, and it changes the economics of the project entirely. The measured spectrum is decomposed onto a library of pure component spectra acquired in the real solvent. You solve a physical system rather than regressing on a population.
And the exact reverse of that advantage: this approach works from what it has been told. An undeclared component is distributed over the others, which is why the choice between routes is made early. Pure-component model or model calibrated on samples.
Which technology for which case
| Technology | What it follows | Suitable matrices | Calibration burden | Installation | Known condition |
|---|---|---|---|---|---|
| MIR spectroscopy | Concentrations from 500 ppm to 100 per cent, reaction progress, cut point, supersaturation in crystallisation | Organic liquids, viscous media up to about 10 000 cP, up to 60 °C | None where the composition is declared, inverse otherwise | Very short path length cell, on a recirculation loop. No moving part and no optical fibre | Water absorbs strongly. The bounds of the calibration domain are hard bounds |
| NIR-HPTLS | Concentrations in bioproduction media, buffers, nutrients | Defined aqueous media before inoculation, buffers, formulations | None on a defined medium: decomposition onto pure components. In a living culture, calibration on samples | Fixed optical path, sterilisable probes | Works from an exhaustively declared composition |
| RAMAN spectroscopy | Reaction progress, polymorphism, constituents in an aqueous medium | Aqueous liquids, suspensions, culture media | Full, or by following one characteristic band | Immersed probe or measurement through a transparent wall | Fluorescence of the matrix. Local heating possible |
| Classical NIR | Content of a constituent, water, overall formulation monitoring | Liquids, emulsions, syrups, natural matrices | Full, PLS or PCR | Immersion probe, transflection, or flow cell | Broad overlapping bands, model to be kept alive |
A word on supersaturation, since it is among the quantities followed. A spectroscopic measurement returns a concentration in the liquid phase; supersaturation is the gap between that concentration and the solubility at the temperature of the moment. It therefore asks for two things: the in-line measurement, and a solubility curve established for your system. Without the second, what is followed is a concentration — often enough to set a seeding point, but it does not carry the name supersaturation.
The calibration burden column separates a light project from a heavy one. It is a column a manufacturer fills in from the technology they make, and so from the calibration burden that comes with it.
Where the measurement sits in the process
On a liquid the installation is simpler than on a divided solid. The product circulates, it is homogeneous at the scale of the cell, and a bypass is often enough.
| Mode | What it means for a composition measurement |
|---|---|
| In-line | Probe immersed in the reactor, or a cell inserted in the main line. No latency. Cleaning and the chemical resistance of the wetted parts are part of the design |
| On-line | A dedicated recirculation loop: pump, in-line filter, three-way valve, return to process. A frequent configuration, and the most controllable, where the transit time stays short against the dynamics followed |
| At-line | A sample measured beside the line, in seconds. Valuable in development and for building the pure spectra library, and the pace of decision is what it turns on |
| Off-line | Stays the reference that builds and watches the model. Two of the documented applications below sit in this mode, with the path to in-line laid out |
One integration point carries weight in a regulated environment. The loop can carry cleaning in place, drying with nitrogen or dry compressed air, and a system verification with calibration products, with the process left running. That is the answer to the first point raised in an audit.
What the pure-component route changes, and what it calls for
Fast deployment
Nothing of the future population to cover, so deployment is fast, with no sample campaign.
A transferable model
From one instrument to another without redoing the calibration campaign, because it is physical rather than statistical. Provided both instruments share the range, the resolution and a verified wavelength axis: the transfer is qualified, not presumed.
Legitimate extrapolation
Within the range covered by the dilution series, where a classical regression stays inside its calibration set.
In return, five conditions come together, all five of them:
- a composition that is known, finite and declarable
- pure component spectra available in the real solvent, with traceable standards
- the quantity sought is a concentration rather than a correlated property
- a fixed and controlled optical path
- real spectral selectivity between the constituents
The fifth is the one usually established last, and it is the one that decides. Which route for your process, and what each one costs.
What is worth preparing on your side
- The complete list of constituents, written and owned. It is the entry price of the pure-component model. Complete means that nothing else is present in a detectable amount, traces of a previous cleaning or an antifoam added by habit included.
- Traceable standards and pure spectra in the real solvent. A pure component spectrum acquired in another solvent describes another situation. That point is settled in the laboratory, upstream.
- The real range of concentrations, bounds included, and the values met only in an incident. A calibration domain that covers the range is what makes the measurement hold on the day it counts.
- The reference method, frozen: chromatography, titration, with their exact conditions. A model is worth what its reference is worth, and changing it mid-project changes the quantity measured.
- Hydraulic access and cleaning. Nozzle, bypass loop, chemical resistance of the wetted materials, cleaning between products. On a liquid, most of the integration cost sits there.
Conditions for success and limits
A declared composition is what the model resolves
A pure-component calibration solves a system over the components it was given. What sits outside the list is redistributed over the others and the result stays plausible. Where your need includes detecting the unexpected, an impurity, a degradation, a cross contamination, the route is a model calibrated on samples with formalised anomaly detection, or a per-measurement quality indicator asked for explicitly.
Calibration domains have hard bounds
The available application notes announce their calibration domains and state explicitly that beyond those bounds the model saturates. A measurement outside the domain returns a value, which is why out-of-domain detection belongs to the method validation rather than to the options list.
Water absorbs very strongly in MIR
The major physical constraint of this spectral range. It is worked around by a very short optical path, a few tens of microns, which makes water tolerable. That short path has a second and counter-intuitive effect: it allows viscous media to be measured in transmission, up to about 10 000 cP, where a surface measurement would sample the surface.
A composition says nothing about the physical state
Two emulsions of rigorously identical composition can have entirely different droplet size distributions and stabilities. Where your critical attribute is a physical state, a second measurement supplies it, in-line particle size or rheology. It is a typical case of two technologies completing one another.
What it changes, in practice
Four applications are documented for this family of measurement. Two in line, two off line, with the path to in-line laid out but not yet taken.
In line, fractional distillation. Continuous monitoring allows real-time detection of the depletion of a constituent, and therefore a cut point made objective instead of a temperature set point. Calibration was carried out on pure components and on samples, validated against chromatography, with announced ranges and the statement that beyond them the model saturates.
In line, emulsification. In-situ monitoring of surfactant addition up to stabilisation brings out a fact that running by the clock hides. Emulsification is a delayed process, not an instantaneous one. The bands of the oil phase only appear after a measurable delay. Hence a spectral endpoint criterion triggered by the chemistry and not by time. And the revelation of stable states that no viscosity, turbidity or particle size measurement was signalling.
Off line. Assaying an oxidant in solution. The assay of hydrogen peroxide in acidified ethanol goes down to levels that other spectral ranges do not reach. With a model calibrated on samples with four latent variables. The transposition to at-line is described, not demonstrated.
Off line. Polymer functionalisation. Measurement of viscous samples, without dilution, follows a degree of substitution that titration and magnetic resonance establish accurately once the batch is finished. The reference value is right, and it stays right: it is what calibrates the model. What is missing is the same information while the reaction progresses. Moving to real time remains a project, not an achievement.
Frequently asked questions
How long before we have a usable model?
It follows the calibration burden. On the pure-component route, once the library of pure spectra is acquired in the real solvent, deployment is fast, since there is no representative design of experiments to build. On the sample-calibrated route, a sample campaign covering the future variability follows the production calendar, and the model is then watched. Establishing which of the two applies to you is the first thing we do.
What happens if an unexpected constituent appears?
With a pure-component model the returned value stays plausible. That is the exact counterpart of not having had to build a representative population. The answer is to ask for a fit quality indicator per measurement, a spectral residual, and to follow it like a control chart. It exists on some instruments, and it is worth asking for and checking that it is exported.
At what rate is the measurement available?
On the compact instruments of this family — the OrionIR from Paeonia Innovations for the one we deploy — the documentation announces one spectrum every 0.3 second in the configuration described. That figure holds with its configuration. And a maximum rate is not a working rate: pushed to its ceiling, it degrades the signal-to-noise ratio. On that instrument we work at 0.7 second, which gives the best signal-to-noise. The useful time includes the transit in the loop, the processing and the return to supervision. Set the rate your process calls for first, then verify it is reachable — the right setting is the slowest your process tolerates, not the fastest the instrument announces.
Does the instrument drift over time?
That is the right question to put to any supplier, and it calls for a figure rather than an adjective. Some architectures with no moving part, no cooling and no optical fibre document spectral stability over more than a hundred days. It still calls for a periodic verification with calibration products, since stability is one thing and its documented demonstration is another.
Can we release on this measurement?
An in-line composition measurement can contribute to a documented control strategy and, in time, to real time release testing. That rests on a full analytical validation file: domain of validity, anomaly detection, model lifecycle management. It is a project in itself. In pharma and biotech that file is the one ICH Q2(R2) and Q14 describe. In chemicals, food and anaerobic digestion there is no release in the regulatory sense: the underlying requirement is the same, but it is carried by your own quality system rather than by a compendial text. Validating an analytical procedure.
Tell us what is in your liquid. That answer decides everything else.
Forty-five minutes is enough: whether your medium falls under a pure-component or an inverse calibration, what each would cost, which spectral range suits your matrix, and whether a trial makes sense. Some cases call for another route — a composition that cannot be declared, a quantity that is not a concentration, a matrix that is too aqueous: you will then leave with the one that fits.