Why Our Microbead R&D and Manufacturing Process Suits AI and Machine Learning
Investors, partners, and scientists often ask us this question. The honest answer starts somewhere less obvious than algorithms: with how microbeads have been made for more than a century, and how we make ours. In this post, we compare our continuous, flow-based droplet process with conventional batch suspension polymerization. We explain why AI models trained on lab-scale batch data often fail when the process moves to large manufacturing reactors, and how we designed our process to avoid that problem. In short, we run the step that sets each bead's size under the same conditions at every scale, so what our models learn in development stays relevant in production.
How beads are usually made
Most polymer beads, including many used in chromatography, are made by suspension polymerization. You take a liquid mixture, break it into tiny droplets in a second liquid it does not mix with (usually water), and heat it until each droplet hardens into a solid bead. Think of making tiny pearls in a pot of soup. The method has been used for over a century for everyday plastics like PVC and polystyrene, as well as many resins used in water treatment and purification.
In the conventional process, all of this happens in a stirred tank. Stirring breaks the liquid into droplets, and the droplets harden while the tank keeps mixing. In the lab, that tank is a hand-sized flask. In a factory, it can hold tens of thousands of liters.
The giant cake problem
Imagine you have perfected a cupcake recipe. You have even built a model that predicts how the cupcake turns out from mixing speed, batter thickness, and oven temperature.
Now someone asks you to bake one cupcake the size of a delivery truck with the same settings. The outside burns long before the heat reaches the middle, and the center stays raw.
The recipe did not fail. The geometry did. When something gets bigger, its volume grows much faster than its surface: double the size and the volume grows eightfold, but the surface only fourfold. A giant cake has far less surface area per unit of batter for heat to enter. Your cupcake model was right about cupcakes. It just assumed, without saying so, that the whole cupcake was at the same temperature, and at truck size that is no longer true.
Engineers call these effects scaling artifacts: changes in outcome that come from equipment size, not the recipe.
What changes in a big tank
The same thing happens when bead manufacturing moves from a flask to a factory tank. Two changes matter most.
Stirring is no longer even. In a small flask, the whole liquid gets roughly the same treatment. In a large tank, the stirring blade churns violently right next to it, while liquid near the walls barely moves. Droplets that pass close to the blade get torn into very small pieces. Droplets that drift in the quiet zones bump into each other and merge into larger, irregular ones. The bead sizes you end up with depend on the tank's shape and size, not just on the recipe.
Heat is no longer even. A large tank is usually heated or cooled through its walls. As with the giant cake, there is much less wall per liter of liquid, so heat moves in and out slowly. Liquid near the walls and liquid in the middle can sit at different temperatures, and the droplets harden at different rates.
Even in a single batch, the conditions keep shifting. Early on, the droplets are runny and break apart easily. Later, as they thicken and get sticky, they tend to merge. So bead size drifts over the run and with tank size.
Biologics process development teams know this problem: a cell culture that behaves one way in a bench bioreactor often behaves differently in a production vessel for many of the same reasons.
Why this breaks an AI model
When you train a model on flask experiments, it feels like you are teaching it chemistry. In practice, you are also teaching it the habits of that flask: how it stirs, how fast it heats, how evenly the beads harden.
Take an illustrative example. A model trained in the lab learns which stirring speed and temperature give a 40-micron bead. In the factory tank, the same stirring speed creates very different conditions near the blade and near the walls, and the same heating setting leaves the batch's outside hotter than its middle. The model reproduces the settings exactly, but the beads don't experience the conditions the model learned from. The prediction fails exactly where it matters most.
The usual workaround is to collect new data at each larger scale and re-tune by trial and error. That is slow and expensive, and every new material has to go through it again. It also means lab data, the cheapest and fastest to collect, is worth less than it should be.
Why we did not just use off-the-shelf microfluidic chips
A well-known way to make very regular droplets is with standard microfluidic chips. These are small, sealed devices with hair-thin channels that pinch off droplets one at a time, at a fixed point, with very even sizes. We started there.
They didn't work for our materials. Our bead precursors are thick, carbon-based slurries, roughly 200 times as viscous as water at the upper end, and they clogged the tiny channels. The chips are also extremely slow. At the rates reported in the published literature, one chip would need somewhere between about nine months and more than fifty years of nonstop operation to make a single 10 kg lot.
What we built instead
We kept the idea that makes chips work, forming each droplet at a single, well-defined point in a controlled flow, and rebuilt it at a larger scale with open channels that handle thick, particle-filled slurries and can be cleaned.
This is still a microfluidic approach, and we describe it that way in our published work. Each droplet forms the same way it does in a chip: two liquids meet at a single point, and the flow pinches off one droplet at a time. We changed everything that made off-the-shelf chips impractical for us. We sized the channels for thick carbon slurries; they can be opened and cleaned, and they are designed to run at manufacturing rates.
The important part is how we make more beads. We do not build a bigger tank. We run the same droplet-forming setup for longer. A droplet formed in a small development run and one formed in a production run are made in the same place, under the same flow conditions, with the same geometry.
Chromatography scientists will recognize the logic. We scale up a column by widening it while keeping the bed height and flow velocity the same, so the molecules experience the same conditions. We apply that idea to making the beads themselves.
Why this matters for our data and AI
Because the step that sets each bead's size behaves the same way at every scale, what we learn in a quick development run describes the same droplet-forming conditions we use in production. We don't have to throw away lab data or relearn it when we scale up.
That turns our manufacturing into a learning loop:
Design a bead with a target composition, size, and structure.
Make it and record every controllable setting.
Test it in our lab and in customer programs.
Record how it was made alongside how it performed.
Predict better designs and settings from everything recorded so far.
Scale with fewer surprises because we collected the data under the conditions we manufacture in.
Each batch adds to the dataset, and because the data stays relevant at scale, each batch improves the next prediction. A dataset built this way takes years of manufacturing history to accumulate, and you can't recreate it after the fact.
To be clear: the dataset is real and growing, and it already links how we make our materials to how they perform. We're heading toward fully predictive design and AI-guided manufacturing, not claiming we've finished them. We believe a manufacturing process whose conditions don't shift with scale is a prerequisite for getting there, and we've chosen that path from the start.
What it means if you use chromatography media
The same property that keeps our data useful also matters to anyone packing a column. Bead size and how tightly sizes are grouped affect column pressure, separation quality, and how reproducibly a column packs. A process that forms droplets the same way at every scale delivers lot-to-lot consistency. NanoPak-C, our all-carbon reversed-phase microbeads for peptide purification (we describe how we make these carbon microbeads in ACS Omega), and NanoPak-B, our composite bead for large-biologic purification, are both made with this process.
If you are thinking about AI in materials development, bead consistency, or scale-up challenges, we would be glad to talk. Contact Millennial Scientific at inquiry@millennialscientific.com or use our contact form.
