Ship reactors, liquid walls, and AI learning from old logs

Designing a fusion reactor small enough for a ship

What if a fusion reactor didn't have to power a city — just a large vessel at sea?

Most fusion reactor designs chase enormous scale: the logic being that bigger machines confine plasma more efficiently, so you go big or you go home. The Yinsen design flips that question. The team behind it — using a suite of computer modeling tools including FUSE, ASTRA, and OpenMC — asked instead: what is the *smallest* useful fusion reactor, if you anchor the design to what the materials can actually survive? Their key limit is something called DPA — Displacement Per Atom — which counts how many times neutron bombardment knocks atoms out of their places in the metal walls. Think of it like the tread on a car tyre: there is a ceiling, and once you hit it, you swap the part. Yinsen's inner vessel, made of a vanadium alloy called V-4Cr-4Ti, can absorb 35 DPA before it needs replacing — which works out to roughly once every 20 years at their design power level. Conveniently, the expensive superconducting magnets last about sixteen times longer than the vessel, so you are not replacing the priciest components on every cycle. The result is a reactor producing just over 25 megawatts of net electricity — enough to power around 20,000 homes — in a physically compact footprint suited to a large ship, a remote island, or an off-grid industrial site. The catch: every number here comes from simulations. No Yinsen reactor exists. No material has been tested in this specific configuration. Marine environments add corrosion, vibration, and space constraints that models do not capture. This is a credible design study, not a vessel you can charter. But the fact that engineers are asking 'what size is actually useful?' rather than 'how big can we go?' is a real shift in thinking.

Liquid metal walls that fix themselves by floating the right atoms up

The inside of a fusion reactor is so brutal that solid metal walls erode — so what if the wall was liquid instead?

One of fusion's nastier engineering problems is the wall. The plasma inside a tokamak runs hotter than the sun's surface, and it constantly flings particles outward. Solid metal walls erode. Worse, if heavy metal atoms sputter off the wall and get into the plasma, they contaminate it and cool it down — like dropping cold water into a hot pan. One idea: replace the solid wall with a liquid metal surface that flows, repairs itself, and can't be permanently damaged. But liquid metals bring their own question: which atoms end up at the surface where they touch the plasma? A team using a computational method called ReaxFF molecular dynamics — which models how individual atoms bond and move — looked at two candidate alloys: tin mixed with aluminium (Sn-Al) and tin mixed with lithium (Sn-Li). Think of it like a salad dressing. Oil always rises to the top. In these alloys, the lighter atoms — lithium or aluminium — act like the oil: they float to the surface. That is exactly what you want, because lithium and aluminium are 'low-Z' materials, meaning they contaminate the plasma far less than heavy tin would. The surprising finding: surface tension alone cannot explain how thoroughly the light atoms rise. When tiny amounts of oxygen or hydrogen are present — unavoidable in a real reactor — the chemical bonding energy they create is roughly ten times stronger than surface tension, and that force is what locks the light atoms at the surface. The catch: these are simulations of a slab of 2,000 atoms under idealised conditions. Real liquid metal walls operate under flowing magnetic fields, convection currents, and constant plasma bombardment. Whether this segregation holds up in those conditions is still an open question.

An AI trained on old reactor data was plugged into a real tokamak

A team trained an AI using only archived logs from a fusion reactor — then connected it to the actual machine.

Even if you build a fusion reactor that works in principle, you still have to control it in real time. The hot plasma inside a tokamak is constantly trying to go wrong — kinking, rotating unevenly, losing energy. Human operators cannot react fast enough; physics simulators are often too inaccurate to be trusted; and running live experiments on an expensive machine to train an AI is impractical. A team at General Atomics — the lab that runs the DIII-D tokamak in San Diego — tried a different approach. They trained an AI using nothing but historical data: the logs of what actuators were set to in thousands of past experiments, and what the plasma did in response. No live interaction, no physics simulator. Think of it like learning to drive by watching thousands of hours of dashcam footage and then being handed the car keys. The control task they targeted is called the plasma rotation profile — how fast the plasma spins at different depths inside the tokamak. Getting this right matters for keeping the plasma stable. It is genuinely hard because four actuators need to be tuned simultaneously (neutral beam injectors, radiofrequency heating, gas valves) and their effects interact in non-obvious ways. The AI — trained using a method called offline reinforcement learning — was then connected to the real DIII-D plasma control system and run on a live plasma discharge. The team describes the results as 'promising.' The catch: that word is doing a lot of work. The paper does not publish precise quantitative performance metrics, which makes it difficult to judge how much better the AI performed than existing controllers. One promising experiment is a long way from a robust system you would trust in a power plant. But getting the AI to run at all — trained only on records — is the meaningful step here.

These three papers are not random. They map three different layers of the same engineering problem. Yinsen asks: what shape and scale of reactor is actually useful in the real world, not just in a physics paper? The liquid metal wall research asks: what is that reactor made of on the inside, and how do we stop the walls from destroying the plasma or being destroyed by it? The DIII-D AI experiment asks: even if you get the design and materials right, how do you control the thing in real time when the physics is too complicated to model perfectly? What's notable is that all three are being worked on simultaneously, and none of them depends on the others being solved first. That's roughly what it looks like when a technology shifts from 'possible in principle' to 'possible in practice.' Each layer gets attacked in parallel. Progress is uneven and often quiet. This week was a quiet week — but a real one.

The DIII-D team's offline RL result will be worth following when they publish more detailed performance metrics — the 'promising' label needs numbers to be meaningful. More broadly, SPARC at Commonwealth Fusion Systems is targeting its first plasma in the late 2020s; any update to their construction timeline or magnet testing results would put the Yinsen-style compact reactor conversation on much firmer ground. The open question I'd most want answered: do those liquid metal surfaces actually maintain their light-atom coating under the kind of plasma exposure a real tokamak delivers?

• Wikipedia: Tokamak — a solid background on how these machines work
• Quanta Magazine: The quest to turn fusion power into a real technology
• Wikipedia: Plasma-facing material — the wall problem explained