The emulsion that cleared every gate and broke in production

Scaling a recipe is not a single jump from pilot to production. It moves through gates. A formula that works at pilot goes to a factory trial — usually a single manufacturing batch on the production line first, then a longer run, often a half-day of around four hours' continuous production. Only once those pass is the recipe approved for first production, and many teams run several trials before they sign it off. Every one of those gates exists to catch a problem before it reaches the market. This dressing cleared all of them.

A creamy dressing — a stabilised oil-in-water emulsion, nothing exotic — clears eight pilot batches on a FrymaKoruma MaxxD Lab, the bench vacuum unit most teams develop sauces and dressings on, at around ten litres a batch. Production runs on a MaxxD 1800: the same family of machine, the same rotor-stator homogeniser under the vessel, the same scraper agitator — but 1,500 litres of usable volume, roughly a hundred and fifty times the batch. The factory trial goes clean: a full manufacturing batch fills tight and glossy, then a four-hour run behind it. The recipe is approved and first production starts.

The instability never shows at the line. It shows in the retains. A week after the trial, the held samples are oiling off at the neck — and by then first production is already running, separating the same way. Three production runs scrapped on retain, the launch slipping. The trial product that approved the recipe was separating in the warehouse while the line kept filling.

The investigation starts where it always starts: the formula and the lot. It was neither.

The scale-up they did by the book

The obvious suspects clear fast. Same emulsifier, same oil, same stabiliser across pilot, trial and production — a batch re-run against a quarantined pilot-grade lot separated the same way. Temperature checks out. Nothing in the recipe explains why a formula stable at ten litres won't hold at 1,500.

That leaves the machine — and the machine is the same design at both scales, which is exactly why the team trusted it. FrymaKoruma sells the MaxxD on reliable scale-up, and it delivers it, on one condition: that you hold the parameters that actually set droplet size. The team didn't scale those parameters. They ran the big unit the way they ran the lab.

In a rotor-stator system, droplet size is set by two numbers: how hard each pass shears, and how many passes the batch gets. Scale the unit up and both change — in opposite directions.

The homogeniser is a toothed rotor-stator mounted under the vessel; it pulls product through the mill and returns it through a recirculation line. The first number is rotor tip speed — the shear rate in the milling gap. The production rotor is larger, and at the same class of speed a larger rotor runs a higher tip speed, so every pass through the gap hits the droplets harder. The second number is the count of passes — how many times each volume element is drawn through the mill before the batch is done. A ten-litre batch turns over through the mill many times in its cycle. Size the loop for throughput rather than turnover, and a 1,500-litre batch gets a fraction of those passes in the same window.

So the production batch sees a harder hit and fewer of them: over-sheared at the gap on the passes it gets, under-worked in the bulk between them. The droplet-size distribution comes off the line far wider than pilot — and an emulsion fails on the spread of its droplet sizes, not the average.

The over-sheared droplets don't even survive as small ones: fresh interface is created faster than the emulsifier can cover it, so the bare droplets re-coalesce on the next pass. A wide distribution creams and separates — slowly, over days, which is why it cleared every gate at the filler and only declared itself in the retains. The recipe never moved; the droplet history did.

You cannot copy a result by copying the machine

The rule the failure exposes generalises well beyond dressings, and beyond this machine. Every piece of process equipment delivers its result through a specific mechanism, and the variables that govern that mechanism are what you hold constant on scale-up — not the size of the vessel, not the installed power, not how it was run last time. For a rotor-stator homogeniser, those variables are the shear rate and the shear dose.

Hold tip speed but let the loop run too few passes, and the batch is under-processed. Hold the cycle time but let the bigger rotor lift tip speed, and the batch is over-sheared — which is what happened here. The two move independently, and only one combination reproduces the lab's droplet history. "The same way" is not that combination; it isn't even a defined point in the space.

For a shear-sensitive emulsion you fix the target droplet distribution first, then set rotor speed and the number of recirculation passes on the production unit to hit it — and you confirm it with a particle-size measurement on the trial batch, not appearance at the filler. The team had the lab data to specify both numbers before the first production batch. They reproduced the machine instead of the mechanism.

And it is never only the homogeniser. The same scraper agitator that mixes the batch also drives heat transfer, and surface area per unit volume falls as the vessel grows, so the bigger MaxxD heats and cools more slowly — shifting the hold and cool profile the emulsion and its preservative system were validated against. Residence and hold times spread. Same trap, different parameter: the equipment scaled cleanly, and the physics underneath it did not.

What the trials actually proved

The pilot did not prove the process was ready, and neither did the factory trials. They proved the dressing could be made on a MaxxD 1800 and looked right off the filler — a narrower claim than approval implies, and not the same as proving it stable. Every gate judged the product at the line; none checked whether the shear rate and the number of passes matched the lab, and none waited for a failure that only shows on the shelf. A successful trial rarely shows a process is ready; more often it shows the process is ready for the next question. Here that question was specific and never asked: which variables govern this product on this machine, and did we actually hold them when the batch grew? The team had what it needed to answer it before the first batch ran.

Further reading