Generative models can bypass expensive transient dynamics in physics simulations by directly modeling steady-state distributions—useful when autoregressive approaches accumulate errors and reduced-order models aren't available.
This paper introduces GyroFlow, a generative model that skips the slow transient phase of turbulent simulations and directly generates statistically steady-state turbulence. Instead of simulating time-evolution step-by-step, it learns to sample from the distribution of saturated states, achieving massive speedup while maintaining accuracy for gyrokinetic (plasma) turbulence.