Gas choice, shape-shifting magnets, and teaching AI reactor safety

Choosing the right gas could push ARC fusion reactor past one gigawatt

Which gas you inject at the hot exhaust zone of a fusion plasma turns out to decide whether you hit one gigawatt — or fall hundreds of megawatts short.

Picture a wood-burning stove with a flue problem. The fire in the firebox needs to stay ferociously hot, but the pipe carrying exhaust to the outside can't be so hot that it melts the wall. Fusion reactors face an almost identical problem, except the 'firebox' is a plasma at 100 million degrees and the 'flue' is called the divertor. The exhaust heat flux is so intense that without deliberate cooling, the metal walls would erode in minutes. A team modeling the ARC reactor — a compact, high-field tokamak concept with roots at MIT and now central to Commonwealth Fusion Sciences' roadmap — ran integrated simulations to figure out which gas to seed into that exhaust zone. They tested two candidates: argon and neon. Argon won decisively. With argon, the divertor plasma cooled reliably below 2 electron volts — a threshold called 'detachment,' where the gas absorbs heat before it hits the wall — while the fusion core stayed hot enough to produce 750 to 1,000 megawatts. Neon underperformed at 600 to 850 megawatts, because neon atoms drifted too far inward, polluting the plasma core and degrading performance. The finding matters because the divertor is one of fusion's most stubborn engineering constraints. Demonstrating that a single gas choice, at the right density, can solve both problems simultaneously is a useful design anchor. The catch is large though: every number here comes from a chain of simulation codes — ASTRA, TGLF, FACIT, and others — not from an actual experiment. ARC has not been built. How real impurities behave against real wall materials could diverge from the model. This is a credible design step, not a confirmed result.

One set of magnets that can be a tokamak or a stellarator — your choice

What if one set of magnets could run as either of fusion's two main machine types, just by changing which coil carries how much current?

In fusion research, two main machine families have spent decades in a kind of scientific rivalry. Tokamaks — the more common type, including ITER — use strong external coils plus a current driven through the plasma itself to confine the fuel. Stellarators twist the entire magnetic cage into an intricate 3D shape, avoiding the plasma current entirely and sidestepping some instability risks that come with it. The tradeoff is that stellarators are brutally expensive to design and manufacture. A team at Columbia University asked a different question: could you build a single machine that morphs between the two? They designed a ring of 32 small dipole coils — picture them like individually adjustable valve handles arranged around a donut, each tilted at a programmable angle — and showed, through computation, that by varying the current in each coil you can dial in tokamak geometry, a basic stellarator configuration, or hybrid states in between. The coils are made from high-temperature superconductors, meaning they conduct electricity with essentially zero resistance when cooled, which is what lets them generate powerful magnetic fields in a compact package. Peak mechanical forces on the coils stayed within safe limits across all tested configurations. Why does this matter? Each machine type teaches physicists different things, and right now you have to build separate machines to learn them. A shape-shifting experiment could cut that cost. Here is the honest asterisk: this is a purely computational design study on a small test-scale geometry — one-metre radius, half-Tesla field. No plasma was involved. The stellarator twist they achieved is modest, and real manufacturing tolerances will be messier than the simulations.

An AI trained on one fusion reactor learned to predict crashes on another

ITER will be a brand-new machine with zero historical crash data — so how do you build a safety system for something that's never been turned on?

A disruption is fusion's version of a car skidding on black ice — a sudden, uncontrolled plasma collapse that dumps enormous energy into the reactor wall in milliseconds. For ITER, a single bad disruption could cause millions of euros of damage. So you need a warning system that fires early enough to intervene. The problem: AI disruption predictors are trained on data from specific machines, and they fail badly when moved to a different reactor with different sensors and different plasma behaviour. A research team tackled the cold-start version of this problem: can you build a working predictor for a machine it has never seen, using zero experimental data from that machine? Their target was J-TEXT, a smaller Chinese tokamak. Their source of training data was EAST, a larger Chinese machine. To bridge the gap, they ran physics simulations using the NIMROD code to generate synthetic sensor readings — fake-but-physics-based signals that mimic what J-TEXT's specific instruments would see during a disruption. They then applied a technique called Fourier Domain Adaptation, which you can think of as an audio equaliser that adjusts the 'sound' of data from one machine so it matches the sonic signature of another. The result across 1,596 J-TEXT plasma shots: the zero-shot early warning rate improved from 50% to 57%. The relevance to ITER is direct: any safety system has to work on day one, before years of operational data accumulate. The honest catch: 57% is a long way from the near-certain reliability you would want on a billion-dollar reactor. The authors are explicit that this is a feasibility demonstration — which is accurate and appropriately modest.

Look at what these three papers are actually doing together. The ARC modeling work is pushing on a specific number — one gigawatt — and asking what engineering choices get you there. The hybrid tokamak-stellarator study is asking whether the physical hardware can be made more flexible, so that the next generation of machines doesn't have to be locked into a single design philosophy before you've learned enough to know which philosophy is right. And the disruption prediction work is asking how safety systems survive the leap from a known machine to an unknown one. All three are, at root, about the transition from one-off experiments to a fleet. You can't run a power grid on a single proof-of-concept reactor. You need multiple machines, replicable designs, and safety systems that generalise. None of today's papers solves that transition — but all three are pointed squarely at it. That coherence is not accidental. It reflects where serious fusion engineering attention has shifted.

On the ARC front, Commonwealth Fusion Sciences is still on track to start SPARC operations in the late 2020s — SPARC being the actual experimental machine that will test many of the assumptions behind today's simulations. Watch for any updates on SPARC's first plasma timeline. On disruption prediction, the open question I'd want answered next is simple but hard: what does the early warning rate look like at 90% or 95%? The gap between 57% and 'good enough for ITER' is enormous, and no paper has yet shown a credible path across it.

• Wikipedia: ARC fusion reactor concept
• Wikipedia: Plasma disruption
• Wikipedia: Stellarator