AI steers fusion plasma in real time. Here's how.

A Real-Time AI Steers Fusion Heating Even When Equipment Breaks

Halfway through a fusion experiment, one of your heating devices fails — and the AI quietly redistributes the load without missing a beat.

The DIII-D tokamak, run by General Atomics in San Diego, uses beams of microwave energy — called ECH, or Electron Cyclotron Heating — to warm plasma from the outside. The challenge: you need to hit the plasma in exactly the right spot. Too far in or out, and you waste energy or destabilize the whole thing. The old system could only locate the peak of the heating, not its full shape across the plasma cross-section. And it was too slow for real-time adjustment anyway. A team at General Atomics built an algorithm called ECHO (ECH Optimization). The trick was replacing a slow physics simulation, called TORBEAM, with a neural network trained to mimic it. That network runs fast enough to search through thousands of antenna-angle combinations in real time — think of it like a DJ who has pre-loaded every possible remix and can switch tracks instantly, rather than composing music live. ECHO picks the combination that best matches a target heating shape, then sends those instructions to the actual machine. They tested it on DIII-D itself, not just in simulation. It worked. Better still: when one of the heating devices — a gyrotron, which is the microwave generator — failed mid-experiment, ECHO automatically redistributed the remaining devices to compensate, without human intervention. The catch: DIII-D is a research machine, not a power plant. Scaling this kind of real-time control to ITER or a commercial reactor — which will have far more heating systems and far tighter tolerances — is a separate challenge this paper doesn't address. A promising rehearsal. The performance isn't yet on the stage it will eventually need to play.

A Neural Network Reconstructs Plasma Shape Ten Times Faster Than Before

Imagine driving by looking at a rearview mirror that only updates once per second — that's roughly what plasma shape control looked like until this week.

Inside a tokamak, the plasma — a cloud of superheated gas — is held in place by magnetic fields. But the shape of that plasma is constantly shifting, and keeping it stable requires a feedback loop: measure the shape, adjust the magnets, measure again. The faster you measure, the smoother the control. The standard real-time tool at the TCV tokamak in Switzerland, run by EPFL, managed one shape reconstruction per second (technically 1 kHz). A team at EPFL trained a physics-informed neural network — a type of AI that learns from both real experimental data and the underlying physics equations simultaneously — on roughly 10,000 past plasma shots from TCV. The result: a model that reconstructs the plasma shape in under 100 microseconds on a standard single-core processor. That's ten times faster, enabling ten updates per second (10 kHz). They didn't just run this in a simulator. They deployed it on TCV's actual plasma control system — the computer that runs the real machine. The network was trained on a wide variety of plasma shapes, including the unusual snowflake and negative-triangularity configurations TCV likes to experiment with, so it handles the full range of what the machine can produce. Faster shape measurement means the control system can catch and correct small deviations before they snowball into disruptions — the fusion equivalent of a skid turning into a crash. The catch: this was built and validated specifically for TCV's geometry, coil layout, and historical data. Retraining it for a different machine — including ITER, which has a very different geometry — is a real effort that hasn't been demonstrated yet. Portability is an open question.

One Set of Magnets That Can Run Two Different Reactor Designs

What if you didn't have to choose between building a tokamak and building a stellarator — because the same magnets could do both?

There are two leading designs for a fusion reactor. The tokamak — think of a hollow doughnut — is easier to build but needs constant electrical current flowing through the plasma to maintain its shape. The stellarator uses a more complex twisted magnet arrangement to achieve the same thing passively, without that current, which makes it inherently more stable — but it's far harder to design and build. Researchers have traditionally had to commit to one or the other. A team at Columbia University ran a computational study asking whether a single array of ring-shaped superconducting magnets — called dipole coils — could be electronically reconfigured to produce either type of plasma in the same device. Think of it like a modular kitchen: same appliances, different arrangement, and you can switch between a home setup and a professional one by moving components around. In simulation, the answer is yes. By adjusting the electrical current through 32 coils made from HTS — high-temperature superconductor, a material that carries electricity with zero resistance at achievable cold temperatures — the team showed they could produce stellarator-like magnetic field configurations, hybrid arrangements, and standard tokamak shapes, all within the same physical envelope. The peak forces on the HTS coils stayed well below the material's known tolerance limits. The catch: this is a purely computational feasibility study, based on a small university-scale machine. No prototype has been built. The force calculations use a simplified analytic model, and real-world complications — how the superconductor behaves during a magnet quench, manufacturing tolerances, cryogenic plumbing — haven't been addressed. This is a 'worth trying' result, not a 'we built it' result.

Look at what these three papers have in common: none of them are about whether fusion works. They're all about whether we can control it precisely enough, quickly enough, and flexibly enough to be useful. The ECHO algorithm and the TCV equilibrium network are both tackling the same underlying problem — that plasma moves faster than traditional computers can track, and that even small delays in measurement or correction can let problems compound. Together they suggest that real-time AI control of plasma is crossing from 'promising simulation result' into 'deployed on real machines.' That's a meaningful shift. The hybrid tokamak-stellarator paper is a different kind of signal: it suggests that the binary choice between reactor designs may be more flexible than assumed, which matters enormously for the economics of research. You can't afford to build ten different machines to test ten different ideas. A machine that can explore multiple configurations is worth a serious look. These three papers together are telling you that the bottleneck in fusion is increasingly about software, control architecture, and smart engineering — not fundamental physics.

The most immediate thing to watch is whether the TCV equilibrium network gets tested on a second tokamak — that's the moment it becomes a general tool rather than a TCV-specific one. On the magnet side, Columbia's hybrid concept needs an actual hardware prototype to move beyond simulation; watch for any university-scale experiment announcements in the next year. And if you want a bigger event: ITER's first plasma operations remain the horizon that will test whether real-time control tools like these can scale to a machine ten times the size of the ones they were trained on.

• Wikipedia: How a tokamak works
• Wikipedia: Stellarators explained
• Wikipedia: Electron cyclotron resonance heating