Inspiration
Hundreds of billions of pounds are being poured into AI data centres, the liquid cooling market alone is heading past $17bn, and every single one of those GPUs is kept alive by water that is quietly being colonised by bacteria.
Bacteria stick to the pipe walls and build a slimy layer called a biofilm. It works like a blanket on the metal: heat stops moving, pumps work harder and the chips throttle.
The builds are getting stranger too. Data centres are going underwater, where nobody can open a pipe, and space is next. Cooling is critical infrastructure, so anything that touches it is an attack surface. Operators are a ransomware target, and a maintenance system that reaches into a live loop has to be safe by design.
Today operators pour biocide into the water. It kills the bacteria on the surface and leaves the glue they built, so the film grows back. Enzymes dissolve that glue properly, but ordinary enzymes fall apart in warm water within minutes.
Our solution
Sonazyme is a doctor for cooling pipes. We find the blockage with sound, then design an enzyme to dissolve it.
Listening. You cannot cut open a running pipe, so we ping it. The echo at a boundary depends on the two materials either side:
$$R = \frac{Z_2 - Z_1}{Z_2 + Z_1}$$
Biofilm is mostly water, so $Z_2$ sits close to $Z_1$ and the echo barely moves. That is the hard part, and our model shows it plainly: hard scale is easy to see and a soft film comes back with a wide error bar.
Designing. We take an enzyme that cuts the right bond, predict its shape with Boltz-2, then redesign it for heat with Tamarind Bio while holding the cutting site frozen. Every design is folded again and checked.
Deciding. We built a world model of the loop and used reinforcement learning to work out when and where to dose, so the pipes stay clean and the chips never run hot.
How we built it
Boltz-2 predicts every structure. Tamarind Bio runs HyperMPNN for heat-stable redesigns and MutRank for ranking mutations. Adaption fine-tunes a scorer on 5,000 enzyme data points. Synthesia turns each run into a short briefing.
Security shaped the build. The service is read-only, loopback only, serves an allowlist of files and runs with the network off. Every step records which model produced it and which file it wrote, so nothing on screen is unaccounted for.
Challenges we ran into
A ranking model's fourth-best mutation changed a catalytic residue, the part that does the cutting. A stable but dead enzyme at the top of a list is the worst failure, because it looks like a success. We now exclude the active site in the request itself. Forty-eight proposals were rejected on that rule.
Our first policy also learned on clean pipes and then met dirty ones. Random starting fouling fixed it.
Sound struggles to see biofilm at all. We wanted to claim clean detection and the physics would not allow it, so we kept the error bars on screen rather than hiding them.
What we learned
The biology was easy to read about and hard to be honest about. Most of the work became provenance: labelling what is a prediction, what comes from a paper and what is only a drawing. That discipline is what caught the broken enzyme above, so it paid for itself.
What's next
Anywhere water meets metal: shipping, water networks and spacecraft. The platform stays the same. Only the target file changes.
Built With
- ai
- biology
- boltz2
- data
- enzymes
- python
- rl
- synthesia
- tamarindbio
- worldmodels
