Data assimilation in slow-fast systems using homogenized climate models

A deterministic multiscale toy model is studied in which a chaotic fast subsystem triggers rare transitions between slow regimes, akin to weather or climate regimes. Using homogenization techniques, a reduced stochastic parametrization model is derived for the slow dynamics. The reliability of this reduced climate model in reproducing the statistics of the slow dynamics of the full deterministic model for finite values of the time scale separation is numerically established. The statistics however is sensitive to uncertainties in the parameters of the stochastic model. It is investigated whether the stochastic climate model can be beneficial as a forecast model in an ensemble data assimilation setting, in particular in the realistic setting when observations are only available for the slow variables. The main result is that reduced stochastic models can indeed improve the analysis skill, when used as forecast models instead of the perfect full deterministic model. The stochastic climate model is far superior at detecting transitions between regimes. The observation intervals for which skill improvement can be obtained are related to the characteristic time scales involved. The reason why stochastic climate models are capable of producing superior skill in an ensemble setting is due to the finite ensemble size; ensembles obtained from the perfect deterministic forecast model lacks sufficient spread even for moderate ensemble sizes. Stochastic climate models provide a natural way to provide sufficient ensemble spread to detect transitions between regimes. This is corroborated with numerical simulations. The conclusion is that stochastic parametrizations are attractive for data assimilation despite their sensitivity to uncertainties in the parameters.

💡 Research Summary

The paper investigates a deterministic multiscale toy model in which a slow variable x evolves in a double‑well potential V(x)=x⁴/4−x²/2 and is continuously “kicked’’ by a fast chaotic Lorenz‑63 subsystem (variables y₁, y₂, y₃). The time‑scale separation is moderate (ε²=0.01), so the fast dynamics are not infinitely fast, yet they induce rare transitions of x between the two metastable states at x≈±1.

Using stochastic homogenization, the authors rigorously derive a reduced one‑dimensional stochastic differential equation (SDE) for the slow variable:
 dX = X(1−X²) dt + σ dW,
where the diffusion coefficient σ is given by the integral of the autocorrelation function of the fast variable y₂. The derivation proceeds via a perturbation expansion of the backward Kolmogorov (or Fokker‑Planck) operator L = ε⁻²L₀ + ε⁻¹L₁ + L₂, applying the Fredholm alternative and exploiting the ergodicity of the Lorenz fast dynamics. The resulting SDE captures the effective drift (the deterministic double‑well term) and the stochastic forcing arising from the fast chaos.

Numerical experiments show that, despite the finite separation, the SDE reproduces the stationary distribution of x, the exponential distribution of residence times, and the autocorrelation function of the full deterministic system with high fidelity. However, the statistics are sensitive to the precise values of the drift and diffusion parameters; small errors in σ can noticeably alter transition rates.

The core of the study is the impact of this reduced stochastic model on data assimilation. The authors employ an Ensemble Kalman Filter (EnKF) where only the slow variable x is observed at discrete times. Two forecast models are compared: (i) the full deterministic 4‑D system and (ii) the reduced stochastic SDE. With modest ensemble sizes (≈20), the deterministic forecast suffers from sampling error: the ensemble spread is too narrow, leading to filter divergence and a failure to capture regime switches. The stochastic model, by virtue of its intrinsic diffusion, naturally provides a larger spread, effectively acting like covariance inflation or an enlarged ensemble.

A systematic sweep over observation intervals Δt reveals a window where the stochastic model yields the greatest skill improvement. Δt must be larger than the typical waiting time for a regime transition (so that a transition can be observed) but smaller than the decay time of the slow variable’s autocorrelation (so that the filter still benefits from recent observations). Within this window, the root‑mean‑square error (RMSE) of the analysis is reduced by up to 30 % compared with the deterministic forecast.

Sensitivity tests on σ show that a slightly over‑estimated diffusion coefficient often yields the best performance: it supplies enough ensemble spread to detect transitions without overwhelming the signal with noise. Conversely, under‑estimating σ leads to insufficient spread and loss of skill. This highlights a trade‑off between accurate statistical representation of the slow dynamics and practical ensemble dispersion needed for robust assimilation.

The authors conclude that stochastic parametrizations derived via homogenization are attractive for data assimilation in multiscale settings. Even though the reduced model may not perfectly match the true statistics, its built‑in stochasticity supplies the ensemble diversity required to detect rare regime switches, especially when observations are sparse and only the slow variables are measured. Careful calibration of the diffusion term and appropriate choice of observation frequency are essential to reap the benefits.