Blueprints, AI tutors, and a turbulence shortcut for fusion

Scientists draw up detailed blueprints for a 1-gigawatt fusion power plant

Before a single brick is laid, someone has to finish the drawings — and a team just finished very detailed drawings for a fusion power plant the size of a small city block.

Think of this as the architectural blueprints for a building that hasn't been built yet — drawn with impressive precision, but still blueprints. The GIGA project team used computer modeling to design a stellarator — a type of fusion reactor that looks like a twisted magnetic pretzel — capable of producing 3 gigawatts of fusion heat and delivering 1 gigawatt of electricity to the grid. That is roughly what a medium-sized nuclear fission plant outputs today. The key challenge with stellarators is that their twisted magnetic cages are devilishly hard to design. You want them to trap the hot plasma tightly — specifically, you need the fast-moving helium nuclei produced by the fusion reaction (called alpha particles) to stay inside long enough to heat the rest of the plasma. Think of alpha particles as little heating pellets: if they escape before depositing their warmth, you lose free energy the reactor desperately needs. The team achieved alpha confinement above 85%, meaning fewer than 15 in 100 of those heating pellets escape uselessly. They also reduced something called neoclassical effective ripple below 0.01 — think of smoothing a bumpy road. The bumpier the magnetic field structure, the more particles leak out sideways. Below 0.01 is considered a meaningful threshold for a viable power plant. Here is the honest catch: none of this has been built or tested. This is a pure computer design study using simulation codes called VMEC and STELLOPT, with no experimental validation. The team also fixed the plasma boundary in place during the calculation — a simplification that real-world conditions will eventually challenge. A one-gigawatt stellarator power plant remains a dream on paper. An increasingly detailed, credible dream, but a dream nonetheless.

A new AI-powered simulator lets scientists tune a fusion reactor like a GPS

Tuning a fusion reactor normally means turning knobs blindfolded and waiting days to hear what happened — TokaGrad just gave scientists a map.

Optimizing a fusion reactor is normally like trying to tune a complicated instrument in a pitch-dark room: you adjust one setting, run an expensive simulation for hours or days to hear what changed, then adjust again. TokaGrad, built by a team using JAX — a programming framework developed at Google — changes this by making the entire simulation what engineers call differentiable. Here is what that means in plain terms. A normal simulator is a black box: you put settings in, you get results out, but the box cannot tell you which direction to turn the knob next. A differentiable simulator can calculate, at every step, exactly how sensitive the output is to each input — like a GPS that tells you not just your current location but precisely which direction is uphill. That lets you use gradient-based optimization — the same mathematics behind training AI models — to home in on better settings far faster, without guessing. What makes TokaGrad specifically interesting is scope. It is the first simulator to handle a full reactor discharge in one connected calculation: from start-up, through the critical L-to-H transition (the moment when plasma suddenly shifts from a leaky, turbulent state to a well-confined one), through pedestal formation. Earlier tools handled these stages separately, so gradients could not flow across the whole sequence. Now they can. The honest catch: the detailed numbers comparing TokaGrad's accuracy against established tools like ASTRA and TRANSP are not fully visible in the published work yet, so it is genuinely unclear how precise the gradient-guided answers are in practice. The concept is compelling. The validation still needs scrutiny.

An AI model skips weeks of fusion turbulence simulation and jumps to the answer

Normally, predicting fusion turbulence means watching a slow storm build from a single gust — GyroFlow skips straight to the weather forecast.

Running a standard turbulence simulation inside a fusion plasma — what physicists call a gyrokinetic simulation — is like growing a plant from seed to study what a fully grown garden looks like. You have to simulate every early stage of chaotic growth, for days or weeks on a supercomputer, before arriving at the mature, settled state you actually care about. And you typically repeat this dozens of times as you vary conditions. GyroFlow takes a different approach. Instead of growing each plant from scratch, it learns what a mature garden looks like across thousands of past simulations, then uses that knowledge to generate a plausible fully grown example directly — skipping the growth phase entirely. The technical machinery it uses is called flow matching, a type of generative model related to the same technology behind AI image generators. You give it a plasma condition; it produces a sample of what settled turbulence probably looks like in that condition, without simulating the transient storm. The team tested this in what is called five-dimensional phase space — tracking where particles are and how fast they are moving simultaneously — and GyroFlow outperformed other machine-learning approaches including autoregressive models and reduced-order methods. Why does speed matter? If each turbulence estimate takes a week, you can test seven reactor configurations per year. If GyroFlow cuts that to minutes, you can test thousands. The catch to keep in mind: GyroFlow relies on an assumption called ergodicity — the idea that averaging many snapshots gives the same answer as watching one long simulation. This usually holds, but not always, and the method needs stress-testing in conditions it has not been trained on. I simplified here — the real validity boundary is an open question.

Put these three papers next to each other and a pattern emerges: the bottleneck in fusion right now is not just plasma physics — it is the cost of exploring the design space. Designing a stellarator on paper with VMEC (story one), optimizing a full tokamak discharge with gradient mathematics (story two), and skipping directly to settled turbulence with a generative model (story three) are all answers to the same underlying question: how do you search a vast, complicated space of possible reactor configurations without running out of time and compute budget? The honest read is that none of these papers puts fusion electricity on the grid. What they do — collectively — is shrink the gap between a promising idea and a testable design. The field is not moving fast; it is moving smarter. And smarter, in this case, looks a lot like the tools AI researchers spent the last decade building for entirely different purposes, now being quietly repurposed for the inside of a magnetic bottle.

The stellarator design from story one is explicitly a starting point for a power plant concept called GIGA — worth tracking whether the team publishes the next step, which would be a free-boundary equilibrium study allowing the plasma shape to respond to the surrounding coils. On the AI simulation side, the open question I would want answered is whether GyroFlow holds up on reactor-relevant plasma conditions it was not trained on — that is the test that will tell us if it is a genuine shortcut or a sophisticated interpolator. No specific conference or announcement is on the calendar yet, but the overlap between these computational approaches and NIF, ITER, and private fusion announcements will be worth watching this autumn.

• Wikipedia: Stellarator — background on how the twisted magnetic cage works
• Wikipedia: Tokamak — the donut-shaped reactor at the center of most fusion efforts
• Wikipedia: Gyrokinetics — the physics framework behind turbulence simulations in fusion