GPT-6 Astra produced two exact plasma equilibrium families that overturned a 59-year fusion conjecture

A long-standing conjecture in plasma physics was challenged in late September by two papers posted one day apart, with one of them crediting GPT-6 Astra Pro for discovering two explicit families of exact three-dimensional, non-symmetric plasma equilibria. The work centers on a claim associated with Harold Grad, who argued in 1967 that smooth 3D plasma equilibria without symmetry were unlikely to exist, and later said in 1985 that smooth parameterized families of such solutions did not exist outside symmetric exceptions. According to the article, University of Maryland plasma physicist Matt Landreman prompted GPT-6 Astra Pro on Sept. 10 to design a non-axisymmetric magnetic configuration capable of confining plasma while satisfying nested magnetic surfaces, divergence-free fields, rotational transform, and MHD force balance. Astra first returned one exact family that met the hard constraints but had integer rotational transform, then later produced a second family in which the rotational transform varies across magnetic surfaces and is irrational on nearly every surface. Landreman said he verified the results with DESC, SymPy, and Mathematica, and his Sept. 22 arXiv paper states that the solutions were discovered by GPT-6 Astra Pro and that all equations were manually checked. A separate paper uploaded on Sept. 21 by Javier Gómez-Serrano, Mitchell Taylor, and Lukas Liehr also presented counterexamples to the Grad conjecture, using Nash-Moser iteration and a Lean 4 formal proof. Together, the two papers intensified discussion over AI-assisted scientific discovery.