Magnets, Runaway Electrons, and AI: Fusion's Unglamorous Progress

A Thin Layer of Solder Fixes a Hidden Flaw in Fusion Magnets

Every time a fusion magnet is pressurized and relaxed, the electrical resistance between its tape layers can quietly multiply by ten thousand.

The most powerful compact fusion magnets being built today — including those at Commonwealth Fusion Systems for the SPARC reactor — are wound from a material called REBCO tape: thin ribbons of a ceramic superconductor that carry enormous currents without resistance when kept very cold. To build a coil, you stack thousands of turns of this tape together, separated by thin stainless-steel strips. The trick is that those interlayer contacts need a carefully controlled electrical resistance — not zero, not infinite, but a precise middle value — so the coil behaves predictably. Here is the problem the researchers found: every time the coil is mechanically loaded and unloaded — think of pressing down on a sponge and releasing it, over and over — the resistance between tape layers drops by up to 10,000 times from its starting value. After enough cycles, the coil's electrical properties bear no resemblance to what was designed. In a real reactor, where magnets will be cycled repeatedly over years, this is a serious engineering headache. The team, working with coil samples tested at liquid-helium temperature (-269°C), tried several fixes: oxidizing the steel strips, chemical blackening, exotic nanoscale coatings. Most helped a little. What actually worked was coating the REBCO tape itself with a layer of lead-tin solder about two to three micrometres thick — roughly forty times thinner than a human hair. This stabilized the resistance through 30,000 pressure cycles. They then validated the method in a real coil of nearly 1,500 turns tested in a strong magnetic field, and the results matched the small-sample tests. The catch: 30,000 cycles in a lab is not the same as decades in a reactor. And the solder process needs to be scaled up to tapes hundreds of metres long without degrading the superconducting layer. Real progress, but a long path still ahead.

Why Injecting Gas into a Dying Plasma Stops Its Most Dangerous Sparks

When a fusion plasma suddenly collapses, it can leave behind a focused beam of electrons moving close to the speed of light — and nobody wants that hitting the reactor wall.

Fusion plasmas don't always end gracefully. Sometimes the magnetic confinement fails and the plasma collapses in milliseconds — a disruption. In large tokamaks, the aftermath can include a beam of so-called runaway electrons: particles that get swept up and accelerated to near-light speed by the collapsing electric fields, carrying enough energy to punch holes in reactor components. For ITER, the big international tokamak under construction in France, this is one of the scariest engineering challenges on the list. One known method for dealing with runaway beams is injecting neutral gas into the plasma at just the right moment. Experiments on the DIII-D tokamak in San Diego and on TCV in Switzerland showed this can produce a 'benign termination' — the beam disperses safely rather than hitting the wall. But nobody was sure exactly why the gas injection worked. Was it because the gas reduced the number of free electrons, weakening the beam? Or something else? This paper, from a team using the JOREK simulation code combined with their own kinetic modeling, gives a clear answer: it is not about reducing the electron count. It is about resistivity — how much the plasma resists electrical current, like the difference between copper wire and damp string. When neutral gas recombines with free electrons, the plasma's resistivity jumps by roughly ten times. That spike preferentially amplifies a specific magnetic instability at the plasma edge, which tears the beam apart from the outside in, rather than from the inside out. The catch: the simulations are calibrated to medium-size tokamaks. Whether the same mechanism dominates in ITER's much larger plasma — and over what gas injection window — still needs experimental confirmation.

An AI Learns Plasma Physics from Equations Alone, No Lab Data Needed

What if you could teach a neural network to solve a hard physics equation by handing it the laws of physics directly, with no experimental data at all?

Running a tokamak is a constant act of balancing. One of the quantities engineers need to track in real time is something called neoclassical toroidal viscosity torque — NTV for short. It is essentially a rotational drag force that acts on the spinning plasma, and controlling it matters for keeping the plasma stable. The problem: calculating NTV from first principles requires solving a famously stiff equation (the drift kinetic equation, or DKE) that, even on fast computers, takes long enough to be impractical for real-time control. The team behind this paper, working with data from the EAST tokamak in China, trained a neural network to act as a fast stand-in for that solver. The twist is how they trained it. Most neural networks for physics learn by example: you show them thousands of input-output pairs from a traditional solver, and they learn to mimic it. This team took a different approach — imagine teaching someone to cook not by having them taste dishes, but by giving them the laws of chemistry and asking them to reason from there. The network was trained purely on the governing equations themselves, with no labeled solution data. The physics constraints were baked directly into the network's architecture and loss function. The result is a surrogate that computes NTV torque much faster than the original solver, while also being more physically consistent than conventional data-trained models — meaning it is less likely to produce answers that technically fit past data but violate physical laws in new situations. The honest caveat: the speedup factor, while described as significant, was not fully quantified in the available paper text. And this has only been tested on EAST's conditions. Generalizing to other tokamaks with different geometries is not guaranteed without retraining.

Look at what these three papers are actually doing: one is making fusion magnets more durable under mechanical stress, one is figuring out how to safely kill a runaway plasma beam, and one is making the calculations needed to control a plasma fast enough to be useful. None of these is 'fusion is solved.' All of them are necessary for fusion to work in practice. There is a pattern here worth naming: the frontier of fusion engineering right now is about closing the gap between 'it works in an experiment' and 'it works reliably over years.' The REBCO solder result is a good example — the physics was not mysterious, but nobody had systematically solved the manufacturing problem at coil scale. The runaway electron work is similar: the gas injection trick was already used in experiments, but the mechanism was debated. Knowing the mechanism means you can optimize the intervention. The AI surrogate work is about speed — knowing the physics is not enough if the calculation takes too long to be useful in a live reactor. These are engineering maturity problems. They are unglamorous, and they are exactly what has to happen next.

The immediate thing to watch is whether the REBCO solder result gets picked up by companies actively winding HTS coils — Commonwealth Fusion Systems in particular is racing toward their SPARC prototype and will face exactly this problem at scale. On the disruption side, ITER's disruption mitigation system is scheduled for design finalization in the next year or two, so papers that sharpen our understanding of the physics behind gas injection are landing at a useful moment. The open question I'd most want answered: does the resistivity-driven disruption mechanism hold in ITER-scale plasmas, where the runaway beam carries far more energy than in DIII-D or TCV?

• Wikipedia: Runaway electrons in tokamaks
• Wikipedia: High-temperature superconductivity and REBCO
• Wikipedia: Physics-informed neural networks