Synchronisation in two-dimensional damped-driven Navier-Stokes turbulence: insights from data assimilation and Lyapunov analysis

In Navier–Stokes (NS) turbulence, large-scale turbulent flows inevitably determine small-scale flows. Previous studies using data assimilation with the three-dimensional NS equations indicate that employing observational data resolved down to a specific length scale, $\ell^{3D}_{\ast}$, enables the successful reconstruction of small-scale flows. Such a length scale of essential resolution of observation' for reconstruction $\ell^{3D}_{\ast}$ is close to the dissipation scale in three-dimensional NS turbulence. % Here we study the equivalent length scale in {\it two}-dimensional NS turbulence, $\ell^{2D}_{\ast}$, and compare with the three-dimensional case. Our numerical studies using data assimilation and conditional Lyapunov exponents reveal that, for Kolmogorov flows with Ekman drag, the length scale $\ell^{2D}_{\ast}$ is actually close to the forcing scale, substantially larger than the dissipation scale. Furthermore, we discuss the origin of the significant relative difference between the length scales, $\ell^{2D}_{\ast}$ and $\ell^{3D}_{\ast}$, based on inter-scale interactions, cascades’ and orbital instabilities in turbulence dynamics.

💡 Research Summary

The paper investigates how much spatial resolution is required in observational data to reconstruct the full state of two‑dimensional (2‑D) Navier‑Stokes turbulence using continuous data assimilation (CDA). In three‑dimensional (3‑D) turbulence, earlier work showed that the “essential observation resolution” ℓ*₃ᴰ is close to the Kolmogorov dissipation scale η, meaning that one must resolve down to the smallest dynamically active eddies to synchronize the unobserved small‑scale motions. The authors ask whether the same holds for 2‑D turbulence, which is known to exhibit an inverse energy cascade and a forward enstrophy cascade.

The governing equations are the incompressible Navier‑Stokes system on a periodic torus, augmented with a linear Ekman drag term (αu) and a single‑mode Kolmogorov forcing f = sin(k_f y) e_x. The velocity field u is split in Fourier space into a low‑pass component p = P_{k_a}u (the observable) and a high‑pass component q = Q_{k_a}u (the unobservable). In CDA the observable p(t) is inserted directly (direct insertion), while the evolution of the unobserved component is integrated using the same nonlinear operator G(p,q) that governs q in the full system. Successful assimilation means that the approximated small‑scale field \tilde q(t) converges to the true q(t) as t→∞, i.e., the system synchronizes.

To quantify synchronization, the authors introduce the conditional Lyapunov exponent (CLE) λ_c(k_a) = lim_{T→∞} (1/T) ln‖δq(T)‖/‖δq(0)‖, where δq is a perturbation confined to the high‑pass subspace. λ_c(k_a) < 0 indicates exponential decay of the error and thus successful synchronization; λ_c(k_a) > 0 signals divergence. When k_a = 1 (no observation), λ_c reduces to the maximal Lyapunov exponent λ_1 > 0, confirming the definition.

Numerical experiments are performed with a pseudo‑spectral method (128×128 grid, 3/2 de‑aliasing) and a fourth‑order Runge‑Kutta time integrator. The reference case uses ν = 1.0×10⁻³, α = 1.0×10⁻¹, and forcing wavenumber k_f = 4. The authors vary the observation cutoff k_a from 3 to 7, compute λ_c(k_a) and monitor the enstrophy error ΔΩ(t) = ½‖ω(t) – \tilde ω(t)‖². They find a sharp transition: for k_a ≥ 4, λ_c becomes negative and ΔΩ rapidly decays to zero; for k_a = 3, λ_c remains positive and the error persists. The critical observation wavenumber k_a* therefore coincides with the forcing wavenumber k_f, implying that the essential observation length scale ℓ*₂ᴰ ≈ L_f (the forcing scale), not the viscous dissipation scale.

Parameter sweeps in viscosity ν and drag coefficient α show that k_a* is robust: even when ν or α are varied by an order of magnitude, the critical cutoff stays near k_f. This robustness indicates that the synchronization threshold is dictated by the large‑scale energy injection mechanism rather than by small‑scale dissipation.

The authors interpret these findings through the lens of 2‑D turbulence phenomenology. In 2‑D flows, energy cascades upscale (inverse cascade) while enstrophy cascades downscale. Because the large‑scale flow receives energy directly from the forcing, the low‑pass observable contains the dominant dynamical information that controls the evolution of the high‑pass modes. Consequently, once the observable resolves the forcing scale, the unobserved small‑scale motions are slaved to it and synchronize automatically. In contrast, 3‑D turbulence transfers energy downscale; the small scales are dynamically active and must be directly observed to suppress their chaotic growth, which explains why ℓ*₃ᴰ aligns with the Kolmogorov scale.

The sign change of λ_c(k_a) is identified as a “blow‑out bifurcation,” a well‑known transition in chaos synchronization theory where the transverse Lyapunov exponent crosses zero. The paper thus bridges data‑assimilation practice with rigorous dynamical‑systems concepts.

In conclusion, the study demonstrates that for 2‑D Navier‑Stokes turbulence with Ekman drag, the minimal observational resolution required for successful data assimilation is set by the forcing scale, not by the dissipation scale. This result has practical implications for atmospheric and oceanic data assimilation, where observations are often sparse and coarse, and for machine‑learning turbulence models that rely on partial field information. Future work suggested includes exploring non‑periodic domains, time‑varying forcing, and applying the methodology to real geophysical datasets.