Prince Mario-Max Schaumburg-Lippe: AI Helps Crack a Plasma Puzzle From 1967

Harold Grad had a good reason for his guess. In 1967, the New York mathematician argued that smooth three-dimensional plasma equilibria, the kind with neatly nested flux surfaces and pressure that actually changes from place to place, should need symmetry to exist. It sounded tidy. It also sounded plausible for almost sixty years.

Then September happened. Twice.

Two separate papers, posted a day apart, gave counterexamples to that conjecture. Neither team claimed a machine did the work alone. Both said something more interesting, and more useful. The models helped them search. The humans checked everything.

A conjecture that held for 59 years

Start with what Grad was really asking. A plasma inside a fusion device does not sit still like water in a glass. It flows, it presses outward, and magnetic fields have to hold it in a shape that can survive. An equilibrium is a balance point, where magnetic forces and plasma pressure cancel cleanly.

The neat version of that balance has flux surfaces nested inside each other, like layers, with a well defined inside and outside. If the pressure is constant everywhere, the math gets easier. If the pressure varies, which is what a real device wants, the question sharpens.

Do you need symmetry to get that kind of smooth, varying-pressure equilibrium in three dimensions? Grad thought yes. For decades, nobody produced a clear no.

That matters beyond pure math because stellarators are built to be deliberately non-symmetric. Their coils twist in complex three-dimensional shapes, by design. If non-symmetric equilibria with the right properties simply did not exist, that would be an awkward fact to design around. Knowing they can exist does not build a reactor, but it removes a quiet worry from the floor.

The Lean proof, posted September 21

The first counterexample came from Javier Gomez-Serrano, Lukas Liehr and Mitchell Taylor, in a paper posted to arXiv on September 21. Their approach was a formal existence proof, written and checked in Lean 4, the proof assistant that forces every logical step to be made explicit.

That word, checked, needs stressing. Lean does not accept a hand wave. If a step is missing, the proof does not compile. Getting a difficult existence result into that form is slow, technical work, and the authors are open about how they did parts of it.

In their disclosure, they say large language models, including GPT and Claude models, helped fill in technical details and produce Lean code under their guidance. They also say they checked everything. It is not a story about pressing a button. It is a story about guidance, drafting, correction and verification, repeated until the formal checker said yes.

Readers who followed a related formal effort will recognise the pattern in the work on redesigning quantum circuits with AI assistance, where the value was not in trusting the model blindly but in testing its output against strict rules.

The explicit solutions, posted September 22

A day later, Matt Landreman at the University of Maryland posted something different in character. Not an existence proof in the abstract, but explicit analytic solutions, written out in elementary functions.

He gave two families. One has uniform rotational transform. The other has sheared rotational transform, meaning the twist varies across the plasma. Both are counterexamples in the sense Grad cared about, and both also work as steady flows for the Euler equations, the classic equations for fluid motion.

That second point matters. A solution you can write down can be fed into codes. Other researchers can differentiate it, test it, break it, see where a numerical solver drifts from the exact answer. It becomes a benchmark. Benchmarks are unglamorous and invaluable.

Landreman’s acknowledgement is unusually direct. He says the solutions were discovered using GPT-6 Astra Pro, that the model also helped draft parts of the manuscript, and that he confirmed all equations manually. Again, the same split. Discovery help from the machine. Responsibility with the person whose name is on the paper.

An October 5 follow-on then generalized the families, extending the result beyond the first examples. That is often how a closed question opens up. First someone finds one door. Then everyone sees the corridor behind it.

Priority, stated plainly

Precision matters here, so here it is. The Lean team published a day earlier, on September 21. Theirs was the first counterexample to appear. Landreman’s, on September 22, are the first explicit closed-form examples, written in elementary functions that anyone can inspect directly.

Those are different achievements. One proves existence with formal certainty. The other hands you the formulas. Science can use both, and in this case it got both within about 24 hours.

AI as a search partner, not an oracle

What did the models actually do well? Search, mostly. Exact mathematics is full of plausible forms that almost work. Try a term here, adjust a coefficient there, check a condition, fail, try again. A human can do that for weeks. A model can propose a lot of candidates quickly, including odd ones a tired researcher might not try at midnight.

That strength shows up in other physics work too. A recent collaboration described in the CERN and Mistral FLARE physics model project took a similar view, using AI to move faster through a space that is too large to wander by hand alone.

But proposing is not proving. In both September papers, the safeguard is human verification, and it is stated, not hidden. The Lean proof has a machine checker, overseen by the authors. Landreman checked the equations himself. If an equation was wrong, the paper would say so quickly, because explicit formulas invite explicit testing. Someone will test them. That is the point.

There is a cheapness to this kind of breakthrough that deserves a note. No new telescope. No new reactor. No nine-figure instrument. A conjecture from 1967 fell to careful thinking, modern proof software and models used as assistants. Cheap, checkable and genuinely new is a rare combination. Take it when it appears.

Why stellarator designers should care

Stellarators have had a strong few years in the public imagination, often as the twisty alternative to the more familiar tokamak donut. Their promise is steady operation without some of the disruptions that trouble other designs. Their cost is complexity. Every coil shape has to be calculated, then built to tight tolerances.

Design codes for these machines need test cases with known answers. Landreman’s families give them new ones, with properties that were once thought to require symmetry. A code that reproduces the exact solution can be trusted a bit more on the next, harder case where no exact answer exists. A code that misses will show its error plainly. Either outcome helps.

And the broader forecasting lesson applies too. Whether the task is plasma or weather, as seen in new systems that learn to forecast weather weeks ahead, models earn trust when their outputs can be compared against something solid.

Grad asked a sharp question in 1967. He got a clear answer in 2026, from two teams, in two different styles, with help they were happy to disclose. That honesty shows other researchers exactly how to check it. More of that, please.

Also published on: