Many formulas, short runs, a sensory requirement that cannot be measured directly. No reference stays stable for long.
The same vessel makes an emulsion on Monday and a gel on Tuesday, and the range is reformulated constantly. One model per formula cannot keep pace: calibration is done by product family.
Butters, oils and natural extracts vary with the harvest and the supplier. A material can be the right one, compliant with its certificate, and still behave badly in production. So there are three separate questions: is it the right material, is it compliant, will it behave in your process?
A viscosity or a pH can be measured. The feel of a cream cannot, and yet it is what decides whether the product is accepted.
An instrument that no longer follows your formulas? An instrument bought for one formula, which stopped following once the range was reformulated. Before buying anything new, the diagnosis: often the hardware is sound, and it is everything else that is missing. What to do with an instrument that has fallen out of use.
Where to start, depending on your role
- You run the plant: start from the changeover that costs you most. Scoping.
- You formulate or develop: identity, compliance, fitness for use, what a proof of concept has to settle. Feasibility.
- You own quality: ISO 22716; the acceptance criterion for a measurement is yours. The texts.
- You decide the investment: the cost of a control, calculated on your runs. Cost of a control.
- You want examples: eight real situations, presented without names. Experience.
The three questions that come up most often
Is this the material I ordered?
Identity checking on receipt, across dozens of references, many of them of natural origin. It is quick to put in place, and exposed to the ageing of the library. Checking raw material identity.
Will my emulsion hold?
Drift in the droplet size distribution signals destabilisation: creaming, coalescence, Ostwald ripening. Often well before a viscosity or a turbidity moves. Measuring particle size.
How much water is left?
Water content of a powder, a stick, an extract, a product part-way through drying. A simple quantity to obtain and a difficult one to establish: everything depends on which laboratory method is taken as the reference. Measuring moisture.
Operation by operation, what is measured and what it changes
| Unit operation | What is measured | What it changes |
|---|---|---|
| Dispensing and identity checking | Nature of the raw material, discrimination between suppliers of the same reference | The check happens in the warehouse, without tying up the laboratory. Two chemically close materials stay hard to separate: that is verified before committing. |
| Emulsification | Surfactant incorporation, size of the dispersed phase, phase distribution and stabilisation | Emulsification is a delayed process rather than an instantaneous one. An end criterion based on the signal settling replaces a recipe time fixed once and for all. |
| Vessel mixing | Homogeneity, complete dissolution of a thickener, absence of lumps | The end of mixing is observed instead of assumed. Dissolution that the sight glass cannot show is read directly, before the sample goes to the laboratory. |
| Heating and cooling | How the medium changes on cooling, setting | The thermal profile is set on what the product actually does. This is also where temperature shifts the spectral bands: it belongs in the design of experiments rather than being discovered in production. |
| Stability testing | Drift of the size distribution, on fine emulsions and nanoemulsions | Following the dispersed state and viscosity from manufacturing onwards lets a change found at three years of stability be traced back to a process variable. |
| Filling | Presence and integrity of the product in the container | Units outside specification are diverted in-line rather than at final testing. The gain shows on packaging yield. |
Many of the properties that matter in cosmetics have no spectral signature at all. There is a spectrum of water, a spectrum of a surfactant, a spectrum of a polymer. There is no pure spectrum of a texture, nor of a hold, nor of a sensory quality. They can only be approached by correlation. Tying a spectrum to a panel score or to a rheological measurement means building a model calibrated on samples.
A model calibrated on samples has a range of validity, a maintenance cost and a shelf life. That is project information, and it belongs in the plan before the sample campaign. What makes a measurement hold is almost always knowable in advance.
Cleaning between two products can stay the one you already run. On a vessel that makes an emulsion on Monday and a gel on Tuesday, the probe goes through your existing CIP (cleaning in place) and SIP (sterilisation in place) cycles, with no additional procedure. What is checked at scoping: how the window material withstands those cycles.
Calibrate by family, not by formula
Thinking reference by reference gives thirty formulas, thirty sample campaigns and thirty models to maintain. Calibrating by family brings the same coverage back to three or four models, and that is what makes the project reach the thirtieth product.
The strategy that works rests on three choices. First, calibrate by family of formulas: the oil-in-water emulsions built on one base, the aqueous gels built on one thickening system. Building the sampling plan on the variability of the family rather than of a single product. Second, handle the outliers by residual monitoring. A rare formula does not need its own quantitative model, it needs someone able to say that it is departing from its usual behaviour. Third, and this is the most underrated lever, choose a measurement of state rather than a measurement of value wherever the question allows it. That a mix is homogeneous, that emulsification is finished, that the size distribution has stopped moving. Those read without predicting a figure, so with no calibration to build and none to maintain. What the question of sample numbers really covers.
Two questions that decide the project
How many products share the same base
Calibrating formula by formula is the default path, and the cost sits after the first calibration rather than in it: maintaining thirty models at once when an oil supplier changes. A model per formula is a recurring commitment rather than a one-off investment. So the opening question is how many products share the same base, and which one deserves a model. That is a portfolio decision, taken during scoping, not after the proof of concept.
What a spectral measurement adds alongside your panel
A sensor measures physical and chemical quantities that correlate with perception over a given range, and it holds while the formula stays inside that range. Two consequences follow, and both are useful. The model works alongside the panel and lets you convene it less often. And the way you detect a formula leaving the range is planned from the start, through spectral residuals and distance to the training population. A project scoped that way delivers exactly what the physics gives.
The texts that frame a measurement in cosmetics
There is no dosage unit and no compendial test: no text sets what an in-line measurement must demonstrate. What is framed is the product and the way it is made.
- Regulation (EC) No 1223/2009 — the cosmetic product frame: product information file, responsible person, safety assessment.
- ISO 22716 — cosmetic good manufacturing practice, covering incoming material control, production and in-process checks.
The practical consequence: the acceptance criterion for an in-line measurement is yours, and your own quality system carries it. The texts in full.
A new control is presented to the brand before it is put in place. When you manufacture for a brand, any change in how its products are controlled goes through its supplier quality team. The file that convinces them fits in a few pages: what the measurement controls, its reference method, what stays unchanged in your control plan, and what happens when the measurement leaves its domain. The existing control stays in place until the brand has accepted the new method.
Tell us how many formulas run on the same line. That is where the answer starts.
Forty-five minutes is enough to see which product families can share a model, which operations can be steered on a state rather than on a calibrated value, and what a first trial would take in samples and in line time. Your constraints belong in that conversation too: the containers you receive, the changeover rhythm, the panel you already run.