Reactor walls, disruption warnings, and magnets that matter

AI learns to track radioactive hydrogen hiding inside reactor walls

Every fusion reactor wall slowly swallows its own fuel — and until now, watching that happen in a computer took longer than a coffee break per atom.

Here is the problem. Fusion reactors like ITER will use tritium — a radioactive form of hydrogen — as fuel. The inner wall is made of tungsten, a metal chosen because it can survive extreme heat. But tritium doesn't just bounce off tungsten. It wriggles in, gets trapped along the tiny internal cracks between metal grains, and stays there. That's bad for two reasons: you're losing expensive, radioactive fuel, and over time you're building up a radioactive inventory in the wall itself. To design around this, engineers run simulations of exactly how those hydrogen atoms move — which paths they take, where they get stuck. The standard approach, called kinetic Monte Carlo — think of it as a very patient bookkeeper tracking billions of coin-flip decisions one at a time — requires solving a separate energy calculation every time an atom wants to move. That's the slow part. It's like recalculating your entire grocery bill from scratch every time you add one item to the basket. A team working with polycrystalline tungsten models replaced those on-the-fly energy calculations with a chain of three neural networks: one predicts the local energy landscape, one finds trapping sites, and one evaluates how easy it is for an atom to hop between them. The result is that each local update goes from taking several minutes down to a fraction of a second — and the simulations correctly reproduce the known behaviour, with tritium piling up preferentially along grain boundaries. The catch: the paper is truncated, so we can't see the full validation numbers or know how the model performs outside the specific tungsten geometry tested. This is a simulation tool, not a measurement of a real reactor wall. There's a long road between a faster simulation and a wall that actually behaves as predicted under 15 million degrees of plasma.

A new graph-based method never missed a single plasma disruption in MAST data

What if you could predict a plasma catastrophe not by watching one signal, but by watching how all signals stop talking to each other?

A plasma disruption is fusion's version of an engine seizing — in milliseconds, the superheated plasma collapses, dumps enormous energy into the reactor wall, and can cause serious structural damage. The challenge isn't that disruptions are invisible: every tokamak has dozens of sensors watching the plasma. The challenge is reading all those signals together, fast enough to do something about it. Researchers working with data from MAST — the UK's spherical tokamak, now superseded by MAST-U — tried a different framing. Instead of watching individual signals, they built a graph: each diagnostic sensor is a node, and the strength of correlation between two sensors is the edge connecting them. Then they computed a single number called the algebraic connectivity — essentially how tightly the whole network of sensors is woven together at any moment. Think of it like monitoring how well the musicians in an orchestra are staying in sync. A healthy plasma has rich, coordinated patterns across its diagnostics. As a disruption approaches, that coordination starts to fray. The result: on the MAST dataset tested, the method caught every disruption with no misses (recall of 1.0), and it raised the alarm when roughly half the stable plasma phase still remained — giving meaningful warning time. Here is where I have to be honest with you. The full paper weighs just 70 kilobytes. We don't know the sample size, the number of false alarms, or whether this approach works on ITER-scale machines where the physics is different. A perfect recall score with no false-alarm rate reported is an incomplete story. The idea is genuinely interesting. The evidence base, as published, is thin.

Engineers simulate a 10,000-turn fusion magnet for the first time

The magnets that hold a fusion plasma in place are wound from thousands of turns of wire — and until now, nobody could simulate all of them at once.

The superconducting magnets in a fusion reactor — the kind being built for ITER, and the compact high-field variants in private-sector machines — are wound coils of special wire that carries electricity with zero resistance when kept very cold. They can be wound with hundreds or thousands of turns. When something goes wrong — a tiny defect, a sudden heat spike — the failure can cascade through the whole coil in milliseconds. Engineers need computer simulations that predict exactly how that happens. The problem is scale. Simulating each individual turn of wire in a realistic coil is like modelling every single thread in a sweater instead of treating it as fabric. For a coil with a hundred turns it's painful. For ten thousand turns it was, until now, impossible. A research team developed a method they call EXTRA — short for Explicit Turn Resolution with Anisotropic Homogenisation — which is a careful hybrid. Most of the coil is treated as a composite material with averaged-out electrical and thermal properties. Only the turns where interesting physics actually happens (the innermost, outermost, and any turns near a known defect) are simulated in full detail. The rest behaves like a bulk material, but one whose properties were tuned to match reality. The payoff: a 13-times speedup on a 150-turn test coil, and a successful 3D simulation of a 10,000-turn magnet — about ten times larger than anything done before. The method reproduces heat build-up and energy losses accurately against a detailed reference model. The catch: this was validated against a simulated reference, not against a physical magnet being tested to destruction. Real magnets have manufacturing quirks that no model fully captures yet.

Look at what these three stories share: they're all about the gap between what we can measure or build and what we can actually understand in detail. The tungsten paper says we can now track radioactive hydrogen migration faster, but we still can't watch it happen in a real reactor. The disruption-prediction paper says a clever graph metric can see trouble coming, but the evidence is too thin to deploy with confidence on a machine that costs billions. The magnet paper says we can now simulate coils at real-world scale, but only against a model, not a physical test. The pattern is: computation is catching up to the hardware faster than the hardware can be tested. That's useful — it means design cycles get cheaper. But it also means a lot of the fusion field is currently validating tools against other tools, not against experiments. The next five years of progress probably hinges on getting those two things — simulation and physical test — much closer together.

MAST-U at Culham continues to generate disruption data that could be used to properly validate the graph-connectivity approach described above — a multi-machine test across MAST-U and an international partner would be the paper this work needs to become. On the magnet side, Commonwealth Fusion Systems is due to report further high-field REBCO magnet test results through 2026, which would give methods like EXTRA their first real-world stress test. The open question I'd want answered: does the tritium trapping pattern in tungsten look the same in a real linear plasma device as in these accelerated simulations?

• Wikipedia: Tokamak — a solid primer on how magnetic fusion confinement works
• Wikipedia: ITER — background on the international fusion experiment
• Wikipedia: Plasma disruption — what actually happens when a plasma collapses