Early results from our sampling wheel research with UCC
A shared research update from our ongoing collaboration with University College Cork.
For the past year Expo has been building a relationship with Dr. Kevin Cronin and Dr. Vivek Verma’s process engineering group at University College Cork. It’s already produced one direct result: Giovana Teixeira, now our Application Scientist, came out of their group. This year the collaboration moved into a more technical phase, with UCC students starting academic work directly on the rotating wheel sampling geometry used in our µSolis interface.
The wheel is central to how µSolis works. It’s the mechanism that presents powder to the NIR probe, one pocket at a time, and it’s what allows µSolis to take a measurement of moving powder rather than a fixed point in a pipe. We’ve written before about the representative-sampling problem this is meant to solve, and about optimising the physical wheel geometry itself; the question UCC is asking here, whether a well-filled pocket actually contains a representative sample of the bulk, goes directly to the core assumption both of those rest on.
The latest piece of that work is from Ke Lin, an MEngSc student supervised by Dr. Cronin and Dr. Verma. UCC has given us permission to share it here, with full credit to them.
The question
A rotating sampling wheel looks stable once it’s running. The pockets fill, the video feed is consistent, nothing looks wrong. Ke Lin’s project asked a more precise question: does a stable-looking fill actually mean the powder captured in each pocket matches the true composition of the bulk stream?
What was built and what it showed
Ke Lin built a bench-scale replica of the wheel geometry, first with glass beads (a controlled, non-pharmaceutical stand-in), then with two single powders, then with a real 1:1 blend of recrystallized paracetamol and microcrystalline cellulose (MCC).
The glass bead test is the clearest illustration. Two bead sizes were mixed at a nominal 1:1 ratio. The combined, fully recovered material stayed close to that ratio (1.04:1). But the local sample collected from the wheel’s pocket was skewed to 1.58:1, enriched in the smaller beads.
The pharmaceutical blend showed the same pattern. Across every crystallization condition, particle size, and wheel speed tested, the local sample was consistently depleted in paracetamol relative to cellulose: ratios of 0.64 to 0.83 in the local sample, against 0.86 to 0.89 in the bulk collector, versus a nominal 1.00.
The fill looked stable throughout. The composition captured did not match the bulk.
The open question this raises
Worth being clear about what this study is and isn’t. It’s a first attempt, deliberately simple: a bench-scale replica, glass beads before powders, a single blend ratio, a short list of wheel speeds. Both UCC and we see it that way. It’s the right kind of study to start with, and it’s exactly what it looks like: a first pass, not a finished answer.
The obvious reading of the result is “the wheel is biased.” We don’t think that conclusion is earned yet, on a study this early, and we asked Ke Lin about it directly.
There are at least two distinct mechanisms that could produce exactly this result. One is that the wheel’s pocket doesn’t capture a fully representative cross-section of the falling powder stream, in which case any pre-existing differences in the stream get sampled unevenly. The other is that the powder is already segregating by particle size or density before it reaches the wheel at all, on the chute or in the feed, and the wheel is simply revealing a bias that already existed upstream.
These aren’t equivalent findings. The first points to sampling wheel design and commissioning. The second points to the powder handling upstream of it. Distinguishing them matters, because the fix looks completely different depending on which one is at play, and because it’s not obvious from a stable-looking fill signal which one you’re dealing with.
Ke Lin’s own data already contains a clue: a separate test using the inclined chute with no wheel at all still showed the falling stream itself varying in composition depending on where across its width you collected, which is direct evidence that some degree of upstream segregation is real and independent of the wheel. What hasn’t yet been measured is how much of the wheel’s capture area the stream actually fills, or how much segregation has occurred by the time the powder reaches the wheel with the pharmaceutical blend specifically. Both are open, and both are needed to separate the two effects properly.
A Theory of Sampling reading
The distinction above isn’t just semantics. It maps directly onto the error taxonomy in DS 3077:2024, the Theory of Sampling (TOS) horizontal standard, and naming it that way is what turns “the wheel looks biased” into a testable engineering question.
TOS separates sampling errors into two categories. Correct Sampling Errors are inherent to heterogeneous material and are managed, not eliminated: Grouping and Segregation Error (GSE), material segregating locally before it’s sampled, is one of these, and it’s reduced by mixing upstream, not by redesigning the sampler. Incorrect Sampling Errors are a property of the sampling device itself and must be eliminated outright; Increment Delimitation Error (IDE) is the relevant one here, meaning the device’s capture geometry doesn’t give every part of the stream cross-section an equal chance of being collected.
The standard’s Fundamental Sampling Principle is unambiguous on what this implies: a correctly delimited increment, one that captures the entire cross-section of the stream at each instant, is supposed to be robust to whatever segregation (GSE) already exists in the material. It doesn’t need the stream to already be homogeneous. So the practical test is exactly the one raised above: if the wheel’s pocket spans the full stream cross-section and the compositional bias still shows up, that’s GSE alone, a process/material characteristic no sampler design can fix. If the bias disappears once full-cross-section capture is confirmed, what was being observed was IDE, a sampling design and commissioning issue, fixable at the wheel.
This is also the direct link back to how µSolis itself is meant to handle exactly this problem. The working principle isn’t that any one pocket is an imperfect sample that later gets averaged into a good one; it’s that the wheel’s capture geometry is designed so every particle in the stream has an equal probability of being presented to the probe at each pocket, which is the Fundamental Sampling Principle applied at the level of a single increment (the geometry behind our own representative sampling application note). Get that right and IDE is designed out from the start, rather than something averaging is asked to paper over afterwards.
What’s next
This is the first round of a longer collaboration, not a closed study. UCC will continue this line of work through further undergraduate and Masters final year projects, with the next step being exactly the measurements above: how much of the wheel’s capture area is actually filled, and how much segregation occurs before the powder reaches it. From there the natural progression is testing against an actual µSolis unit, where the pocket geometry is designed for equal-probability capture in the first place and the wiper mechanism keeps the probe window from fouling, rather than relying on averaging to fix a bias at the source.
We’ll share the next round of results as they come in.
Credits
This work was carried out by Ke Lin at University College Cork, under the supervision of Dr. Kevin Cronin and Dr. Vivek Verma. Shared here with their permission.