Predicting Plasma Explosions 100 Milliseconds Before They Strike

An AI Trained to Sense When Plasma Is About to Burst

Your fusion reactor just started running — and 100 milliseconds from now, the plasma is going to explode outward. Can you see it coming?

Inside a tokamak — the donut-shaped magnetic bottle that most fusion programmes use — the plasma doesn't sit still. It periodically releases bursts of energy from its outer edge, called ELMs (Edge Localised Modes). Think of a pressure cooker that, every so often, vents a sharp puff of steam sideways. In a kitchen, that's fine. In a reactor, those sideways bursts hammer the surrounding walls with enough heat to cause serious wear over time. At ITER scale, unchecked ELMs are a known threat to the machine's lifetime. A team working with data from the DIII-D tokamak in San Diego — one of the world's most heavily instrumented fusion machines — trained a neural network to read the warning signs. Their input: 50 milliseconds of radar-like measurements called Doppler backscattering, or DBS, which tracks how turbulence in the plasma is evolving just before the edge. The network's job is to predict, in four time-windows, when the first ELM is likely to arrive. In early tests, it could reliably flag the event 100 milliseconds before it happened. A hundred milliseconds sounds tiny. But in a reactor with active control systems — magnetic coils that can fire faster than your blink reflex — 100ms is enough time to intervene. The catch is important here. This is a proof-of-concept paper. The authors do not publish accuracy scores, false alarm rates, or how many shots the model was trained on — that section of the paper is incomplete. What they have demonstrated is a working architecture and a plausible signal. Whether it holds up across a wider database, and whether it generalises to bigger machines, is entirely open. Treat this as promising early wiring, not a finished alarm system.

A Radio Wave That Stops Runaway Electrons Dead in Their Tracks

Imagine an electron inside a fusion reactor that keeps accelerating, and accelerating, and accelerating — until it hits the wall at nearly the speed of light.

During a plasma disruption — a sudden collapse of the magnetic confinement — electric fields can form inside the reactor that act like a long downhill slope for electrons. Some electrons catch that slope and accelerate relentlessly. We call them runaway electrons, and they are a genuine engineering headache: a beam of near-light-speed particles slamming into the reactor wall can melt through it. This paper, from a team using particle-in-cell simulations — think of it as a very detailed numerical crash-test, running millions of virtual particles through electromagnetic fields — proposes what they call a firewall effect. The idea: inject a specific type of radio wave (an R-wave, which rotates in sync with how electrons naturally spin in a magnetic field) into the plasma at the right moment. When a runaway electron races toward this wave and hits the frequency at which it resonates with the wave, something counterintuitive happens. The electron gets trapped in that resonance, and the parallel electric field that was accelerating it forward is suddenly converted into sideways acceleration instead. The electron's forward momentum is redirected — effectively bounced back — before it can reach damaging speeds. Think of it like a speed bump on a motorway, except the bump only activates for cars going exactly the wrong speed, and it redirects them sideways onto a slip road instead of into oncoming traffic. The honest limit: this is theory plus simulation, using tokamak-representative numbers but not a real tokamak. The key question nobody has answered yet is whether you can inject the right wave at the right moment during the chaotic few milliseconds of an actual disruption. That's a control engineering problem that remains wide open.

Mirror Fusion Simulations Just Got 30,000 Times Faster

A calculation that used to take weeks now takes minutes — and it opens a door for a type of fusion reactor most people have never heard of.

Most fusion coverage focuses on tokamaks and inertial confinement (the laser approach). But a quieter third path — the magnetic mirror — is attracting serious investment again, partly because high-temperature superconducting magnets now make it more practical. The idea: instead of a donut, you build a straight magnetic tube that squeezes tighter at both ends, like a hot-dog bun, trapping plasma in the middle. To design a good mirror machine, you need to simulate the plasma at a very fine level — tracking not just bulk flows but the individual energy distributions of particles bouncing back and forth (this is called gyrokinetic modelling, or tracking how particles spiral along magnetic field lines). The problem has always been that these simulations are brutally slow: the particles bounce thousands of times faster than they collide, so you either simulate at the fast timescale and wait forever, or you make approximations that lose accuracy. A team using the open-source Gkeyll code applied a technique called pseudo orbit-averaging — essentially, recognising that near the tight ends of the magnetic bottle, most of the fast motion is repetitive and unimportant, so you can mathematically time-skip through it. Combined with a second trick that relaxes the simulation's internal clock in those quiet zones, they report a 30,000-fold speedup for reaching stable equilibrium on the WHAM experiment geometry (a 17-Tesla machine being built at the University of Wisconsin). The honest catch: this result covers a single magnetic field line at a specific radius. The team hasn't yet demonstrated the method across the full three-dimensional volume of the machine, and no formal convergence study is shown in the paper. But 30,000x faster is a number that, if it holds, changes what is computationally feasible.

Look at what these three papers are circling around: the same core problem, approached from three different angles. Fusion plasmas are unstable in ways that are difficult to anticipate, difficult to stop, and expensive to simulate. The ELM prediction work is betting that machine learning can read the plasma's own 'body language' fast enough to act. The runaway electron firewall is betting that physics itself offers a natural trap — if you can tune into the right frequency. And the mirror simulation speedup is betting that smarter mathematics can finally make a long-neglected reactor concept tractable to design properly. None of these three papers are connected. They come from different machines, different teams, different traditions. But taken together they suggest something specific: the fusion field is moving from 'can we make this work in principle' to 'can we make this reliable and predictable in practice.' That second question is harder, less glamorous, and more important.

The DIII-D ELM prediction team will need to publish validation metrics — accuracy, false alarm rate, performance on unseen shots — before this becomes a real control candidate; watch for a follow-up paper with those numbers. On the mirror side, the WHAM experiment at Wisconsin is approaching its first high-field runs, which will give the Gkeyll team real data to test their equilibrium predictions against. The open question I'd most want answered: does the radio-wave firewall work when the disruption itself is generating broadband electromagnetic noise that could interfere with the injected R-wave?

• Wikipedia: Edge-localised modes explained
• Wikipedia: Magnetic mirror reactor concept
• Wikipedia: Runaway electrons in fusion plasmas