One gigawatt, smarter beams, and a plasma that predicts its own crashes

Argon gas keeps ARC's exhaust cool while fusion power soars

Picture running a car engine at ten times its rated power without the exhaust pipe melting — that is roughly the problem ARC's designers are trying to solve.

ARC is a compact fusion reactor design that aims to produce close to a gigawatt of fusion power — roughly what a mid-sized nuclear fission plant delivers. The problem is that all that energy has to go somewhere, and the edge of a fusion plasma is brutally hot. The region where the plasma touches the reactor wall, called the divertor, has to absorb enormous heat fluxes without vaporising. Think of it like the exhaust pipe on a blowtorch: the flame is useful, but the pipe has to survive it. The team behind this study ran sophisticated computer simulations linking the plasma core to the plasma edge, modelling how different gases mixed into the plasma change the heat load on the wall. They tested two options: argon and neon. The finding is that argon wins, clearly. Seeding the plasma with a small amount of argon keeps the divertor temperature below 2 electronvolts — a threshold below which the wall survives — while still letting the plasma reach 750 to 1,000 megawatts of fusion power. Neon, by contrast, drifts too far into the plasma core, cools the reaction prematurely, and often kills the high-performance mode entirely. The catch: this is a simulation, not an experiment. ARC does not yet exist as a physical machine. The model chains together several specialised codes, each with its own assumptions, and the team explored sensitivity by varying parameters rather than running formal statistical uncertainty tests. The core message — argon seeding can simultaneously protect the wall and sustain high fusion performance — is credible, but it will need experimental confirmation when hardware exists.

An AI trained on one reactor learned to predict crashes in another it had never seen

What if you could train a crash-prevention system on one reactor and deploy it, on day one, in a completely different reactor — without a single real experiment?

Disruptions are fusion's version of a blown fuse: the plasma suddenly loses control, dumps its energy into the wall in milliseconds, and can damage the machine. Predicting them early enough to intervene is critical, but current AI systems have to be trained on data from the specific reactor they will protect. A new reactor has no historical data. You are stuck in a circle. A research team tackled this by asking: can we fill the gap with synthetic data — basically, simulated sensor readings generated by physics codes rather than real experiments? They trained a model called TokaFormer on years of data from EAST, a tokamak in China, then tried to transfer it to J-TEXT, a different Chinese tokamak with a different shape, different sensors, and no shared training history. Without any adaptation, the zero-shot early warning rate was 50% — barely better than a coin flip. After adding synthetic J-TEXT sensor data generated by a magnetohydrodynamics simulator, plus two layers of signal alignment borrowed from image-processing research, the rate climbed to 57% across nearly 1,600 real J-TEXT plasma shots. That is a 7-percentage-point gain, and it sounds modest. It is modest. Think of it like training a lifeguard in one swimming pool, then sending them to a pool they have never visited but giving them an accurate video walkthrough first. They are better prepared — but not as good as someone who trained there for years. The honest take: 57% is not good enough for a power plant. But the direction matters enormously, because ITER and future machines will need exactly this kind of transfer learning before they run a single shot.

An AI steered fusion's heating beams in real time — and survived a mid-shot equipment failure

One of the hardest things about heating a plasma is that you need to hit an invisible, moving target inside a magnetic bottle — in real time, with no do-overs.

Fusion plasmas are heated partly by injecting microwave energy — beams fired into the plasma by devices called gyrotrons. Where those beams deposit their energy matters enormously: hit the right spot and you stabilise the plasma; miss and you can actually make things worse. The problem is that the 'right spot' moves as the plasma evolves, and calculating where to aim takes too long with traditional software. A team at the DIII-D tokamak in San Diego built a system called ECHO that replaces the slow physics code with a neural network trained to mimic it. The neural network runs fast enough — under 75 milliseconds — to steer all available gyrotrons collectively toward a desired heating profile in real time, adjusting mirror angles and power levels on the fly. Imagine a concert lighting rig where every spotlight is independently motorised and an algorithm keeps repositioning all of them together to hit a moving target on stage, even as the target changes shape. The experiment worked. And then, mid-shot, one of the gyrotrons failed. The system detected it, redistributed the load across the remaining gyrotrons, and kept the heating profile close to the target without human intervention. That robustness to hardware failure is the most practical result here — real machines break, and a control system that collapses when a component fails is not usable in a power plant. The catch: ECHO was trained specifically for DIII-D's geometry and plasma conditions. Retraining for a different machine, like ITER, is not trivial. The method is proven; the portability still needs work.

Look at these three stories side by side and a pattern becomes clear: fusion's engineering problems are being addressed from both ends simultaneously. The ARC study works on the physical hardware question — can we manage heat at gigawatt scale? The DIII-D paper works on real-time machine control — can software keep a plasma on target even when things break? And the disruption-prediction paper works on institutional knowledge transfer — can we carry what we learn from today's machines into tomorrow's without starting from scratch? None of these are solved. ARC is still a simulation. The disruption model barely clears 57%. The ECH controller needs retraining for every new device. But what strikes me is that the approaches — integrated physics modelling, neural network surrogates, transfer learning — are starting to speak to each other. The next generation of fusion control systems will probably combine all three. That is not imminent. It is, however, taking shape.

The ARC reactor is being developed by Commonwealth Fusion Systems, which has said it aims to demonstrate high-temperature superconducting magnet performance at scale before the end of the decade — watch for any updates on their SPARC device timeline, which will be the first real test of whether the integrated modelling in today's paper holds up in hardware. On the AI side, the disruption-prediction result will matter a lot more when someone tries it on ITER-relevant plasma parameters; that is the experiment worth asking about.

• Wikipedia: ARC fusion reactor — background on the design this paper models
• Wikipedia: DIII-D tokamak — the machine where the ECH control experiment ran
• Wikipedia: Plasma disruption — what disruptions are and why they matter