The 3D Euler equation describes how an ideal, inviscid fluid evolves. Whether its solutions can develop singularities — points where velocity becomes infinite in finite time — is one of the central open problems in mathematical physics, closely related to the Millennium Prize Problem for the Navier-Stokes equations. Stable singularities, those that persist under small perturbations to the initial conditions, have been found numerically before. But unstable singularities — those that exist only when initial conditions are tuned with effectively infinite precision — had never been found.
Wang, Bennani, Martens, and Lai (arXiv:2509.14185) found them, using physics-informed neural networks trained to near double-float machine precision. They discovered new families of unstable self-similar blow-up profiles for the 3D Euler equation with boundary, the incompressible porous media equation, and the Boussinesq equation. The solutions are real. They satisfy the governing equations to a precision constrained only by GPU hardware round-off — sufficient for computer-assisted mathematical proof.
The structural interest is in why these objects were invisible for so long. An unstable singularity, by definition, is destroyed by perturbation. Any infinitesimal deviation in initial conditions diverts the solution away from blow-up. The trajectory to singularity is a knife-edge: approach it exactly and the solution explodes; miss by any amount and it doesn't. Standard numerical methods introduce perturbation at every step — discretization error, floating-point arithmetic, iterative approximation. These are not flaws in the method. They are the method. And they are precisely what the unstable singularity cannot tolerate.
The neural network approach works not because it is smarter but because it is more precise. The physics-informed architecture enforces the governing equations as constraints rather than solving them step by step. The Gauss-Newton optimizer refines solutions to accuracies orders of magnitude beyond standard deep learning. The result is a search tool gentle enough to find objects that any rougher touch would destroy.
The through-claim is about compatibility between search method and search target. Conventional numerical methods are perturbative: they explore by varying. Unstable singularities are anti-perturbative: they exist only in the absence of variation. The method and the object have opposite relationships to perturbation, which is why decades of numerical work found every stable singularity but none of the unstable ones. The discovery required a tool whose mode of operation did not contradict the defining property of what it sought.
This is not specific to fluid dynamics. Any search for fragile structures — configurations that exist only under exact conditions — will fail if the search method introduces the inexactness that defines their absence. The unstable singularities were not hidden by complexity. They were hidden by the incompatibility between what they are and how we looked.