To make fusion power, researchers must keep plasma (a gas heated until its particles carry electric charge) away from a reactor’s walls. Magnetic fields do the holding. But in a stellarator, those fields twist through a doughnut-shaped chamber, making their behavior difficult to describe exactly.
This month, two independent papers found three mathematical examples of something researchers had long struggled to establish: a smooth, three-dimensional plasma equilibrium. That means the plasma’s outward pressure and the magnetic forces balance throughout the chamber.
The discovery settles a long-running question about what the equations permit. It does not, by itself, produce a better reactor.
The question behind the discovery
Imagine a doughnut filled with smaller doughnut-shaped layers. In the idealized picture of plasma confinement, magnetic field lines run along these nested surfaces, helping keep the plasma away from the chamber wall.
In 1967, physicist Harold Grad questioned whether such a neatly balanced arrangement could truly exist when the shape lacked certain symmetries and pressure changed from one layer to the next. Later versions of his conjecture focused on families of solutions: if you adjusted a configuration slightly, could you find another valid one nearby?
This mattered most for stellarators. Their external coils create deliberately twisted, three-dimensional magnetic fields. Researchers have simulated these fields for decades, but a simulation of an approximate solution is different from an exact proof that the equations allow one.
What the papers found
Javier Gómez-Serrano, Lukas Liehr, and Mitchell A. Taylor posted Counterexamples to Grad’s conjecture on September 21. They prove that smooth equilibria exist in a doughnut-shaped region with nested pressure surfaces. Their solutions repeat around the ring, but lack the continuous and mirror symmetries at issue in the conjecture. They also form a family: adjusting a parameter produces genuinely different, nearby equilibria.
The team supplied a version of its main theorem in Lean 4, software that checks whether a formal mathematical proof follows from its stated definitions and assumptions.
The next day, University of Maryland physicist Matt Landreman posted a second paper with two more families of solutions. His approach makes the magnetic field and pressure explicit in formulas. In one family, field lines trace repeating paths and eventually return to where they started. In the other, the amount of twist changes between layers, so some lines close while others keep winding without exactly retracing their route.
That gives the papers different strengths. The first proves a carefully specified class of equilibria exists. Landreman’s formulas give researchers exact answers against which to test their simulation software.
The AI connection
Landreman says he found his two families using GPT-6 Astra Pro and used it to help draft parts of the paper. He also says he checked every equation manually. The other paper presents an independent mathematical construction and does not attribute its discovery to that model.
The distinction matters. The AI-assisted work produced formulas that experts can inspect and test. The separate Lean proof checks a formal version of the other team’s theorem. Neither step alone tells researchers how well a real reactor would perform. As The Neuron has reported on AI-assisted mathematics, finding a result and understanding everything it means are different parts of research.
What could change
The immediate gain is clarity. These papers show that smooth, three-dimensional equilibria with nested surfaces are mathematically possible in the cases they construct. Landreman’s exact formulas could also help researchers find errors or limits in the computer codes used to study stellarators.
A working fusion reactor faces more tests. Its plasma must hold on to heat, resist disturbances, and confine energetic particles. Engineers also have to build coils that produce the intended magnetic field. The US Department of Energy describes those as continuing challenges for stellarators. One of the new mathematical constructions has a magnetic field that falls to zero along its central axis, for example, so it should not be mistaken for a ready-made device design.
The broader implication reaches beyond fusion. AI helped a researcher find two exact answers to a difficult question, while another team used a different method to resolve it. The useful next step is to learn from the solutions: which features can improve our understanding of plasma, which can sharpen simulation tools, and whether any can eventually guide a practical design.
Related reading
- Inside OpenAI’s Navier–Stokes Claim examines another AI-assisted mathematics claim and why a formal proof still needs careful interpretation.
- OpenAI Claims 100+ Math Solutions. Can Researchers Keep Up? explores the growing gap between generating mathematical results and reviewing them.
