A Shortcut to Statistically Steady-State Turbulence with Flow Matching

TL;DR

GyroFlow is a latent generative model that directly produces statistically steady-state turbulence snapshots in gyrokinetic systems without simulating transient dynamics. It leverages ergodicity to bypass time integration, achieving significant speedups and outperforming autoregressive and reduced-order methods in accuracy and efficiency.

Why It Matters

Simulating turbulent plasma states is computationally expensive, especially in fusion energy research. GyroFlow enables rapid generation of high-fidelity steady-state configurations, accelerating scientific discovery and improving the efficiency of numerical solvers through warm-starting.

Key Findings

• GyroFlow generates steady-state 5D phase space turbulence snapshots directly from noise, avoiding time integration of transient dynamics.
• The model achieves superior performance and speedup compared to autoregressive, reduced-order, and other generative approaches, and can effectively warm-start existing simulation codes.