Turbulent fluid flow is deterministic chaos: governed by the Navier-Stokes equations, sensitive to initial conditions, and apparently random despite being fully determined by its starting state. The dynamical systems approach to turbulence proposes that behind the apparent randomness lies a skeleton of exact solutions — unstable periodic orbits, traveling waves, equilibria — that the turbulent trajectory visits transiently as it wanders through phase space. Find enough of these exact solutions and you can describe turbulence as a choreography of near-misses with unstable coherent structures. The problem is finding them. Periodic orbits in the Navier-Stokes equations are unstable, isolated, and embedded in a phase space of extremely high dimension. Standard numerical methods require good initial guesses, and good guesses require already knowing roughly where the solution is.
Parker and Schneider (arXiv 2602.23181, February 2026) use a generative diffusion model to produce the guesses.
The approach trains a diffusion model on time series data from a direct numerical simulation of two-dimensional turbulence. The model learns the statistical structure of turbulent trajectories — not the equations, not the physics, but the distribution of flow fields that turbulence visits. Once trained, the model generates synthetic trajectories that resemble plausible turbulent evolution. These synthetic trajectories are not periodic — the training data contains no periodic orbits, since turbulent flow never exactly repeats.
The key modification: the temporal structure of the generative model is altered to enforce time-periodicity, and the spatial symmetries of the Navier-Stokes equations are imposed as hard constraints. The model, which learned what turbulence looks like, is asked to generate trajectories that look like turbulence but also repeat exactly in time and respect the symmetries. These synthetic periodic trajectories are not solutions — they are physically plausible guesses that satisfy the periodicity and symmetry constraints approximately.
The guesses are then refined into true solutions using a Newton-type iterative solver. The refinement converges because the guesses are close enough to actual periodic orbits that the solver's basin of attraction captures them. Out of the synthetic trajectories produced by the modified diffusion model, 111 converge to genuine periodic orbits of the Navier-Stokes equations — mathematically exact solutions, verified to numerical precision.
These 111 orbits have very short periods and were previously unknown. Their discovery reveals structure in the periodic orbit skeleton of two-dimensional turbulence that existing methods had missed, not because the methods were wrong but because they lacked initial guesses in the right regions of phase space. The diffusion model accesses these regions because it has learned the global distribution of turbulent states, not just the local neighborhoods of known solutions.
The division of labor is precise. The generative model produces candidates. The physics-based solver verifies them. The AI contributes intuition — pattern recognition across the full distribution of turbulent states — while the mathematics contributes certainty. The result is not an approximation or a statistical prediction but 111 exact solutions of a nonlinear partial differential equation, each verifiable independently of the method used to find it. The guess was artificial. The solution is not.