Abstract
State estimation in multi-layer turbulent flow fields with only a single layer of partial observation remains a challenging yet practically important task. Applications include inferring the state of the deep ocean by exploiting surface observations. Directly implementing an ensemble Kalman filter based on the full forecast model is usually expensive. One widely used method in practice projects the information of the observed layer to other layers via linear regression. However, when nonlinearity in the highly turbulent flow field becomes dominant, the regression solution will suffer from large uncertainty errors. In this paper, we develop a multi-step nonlinear data assimilation method. A sequence of nonlinear assimilation steps is applied from layer to layer recurrently. Fundamentally different from the traditional linear regression approaches, a conditional Gaussian nonlinear system is adopted as the approximate forecast model to characterize the nonlinear dependence between adjacent layers. The estimated posterior is a Gaussian mixture, which can be highly non-Gaussian. Therefore, the multi-step nonlinear data assimilation method can capture strongly turbulent features, especially intermittency and extreme events, and better quantify the inherent uncertainty. Another notable advantage of the multi-step data assimilation method is that the posterior distribution can be solved using closed-form formulae under the conditional Gaussian framework. Applications to the two-layer quasi-geostrophic system with Lagrangian data assimilation show that the multi-step method outperforms the one-step method with linear stochastic flow models, especially as the tracer number and ensemble size increase.
Introduction
In geophysical fluid dynamics, understanding the dynamics of multi-layer systems—such as the ocean's stratified layers—is crucial. While surface data is abundant thanks to satellite observations and floating drifters, the deep ocean remains largely unobserved. Data Assimilation (DA) is the standard approach to bridge this gap by combining physical models with sparse observations.
However, existing DA methods face significant challenges. Simple linear regression models are computationally efficient but fail to capture the intense, nonlinear turbulence characteristic of oceanic flows. Conversely, sophisticated Ensemble Kalman Filters (EnKF) that utilize full physics models are often too computationally expensive for high-resolution applications. This paper introduces a "multi-step" nonlinear data assimilation algorithm designed to be both mathematically rigorous and computationally efficient.
Methods: Multi-Step Conditional Gaussian Data Assimilation
The proposed method leverages a multi-step nonlinear data assimilation scheme that proceeds layer-by-layer rather than attempting a single global mapping.
The process is structured as follows:
- Step 1: Surface Recovery: Surface tracer data is used to estimate the flow of the top layer using a one-step CGDA with linear stochastic models (LSMs).
- Step 2: Recursive Assimilation: The estimated state of the upper layer serves as "pseudo-observations" for the layer immediately below.
- Step 3: Iteration: This process is repeated for subsequent layers.
The "secret sauce" of this approach is the Conditional Gaussian Nonlinear System (CGNS) framework. By approximating the forecast model as a CGNS, the system maintains its nonlinear characteristics while allowing the posterior distribution to be solved using closed-form analytic formulae. Unlike standard filters that assume a Gaussian (bell curve) distribution, this method yields a Gaussian Mixture posterior. This is critical for capturing "intermittency"—the sudden, extreme velocity spikes common in turbulent flows—which linear models typically smooth out.
Experimental Setup
The algorithm was evaluated using a standard benchmark in fluid dynamics:
- Two-layer Quasi-Geostrophic (QG) System: A model simulating large-scale atmospheric and oceanic flows.
- Lagrangian Tracers: Observations consisted of 256 simulated passive tracers moving with the surface flow.
- Turbulent Regimes: The model was tested in both "moderately turbulent" and "strongly turbulent" scenarios to assess its performance under varying degrees of chaos.
Key Results and Insights
The multi-step CGDA significantly outperformed traditional one-step linear approaches across several metrics:
- Improved Accuracy: In strongly turbulent regimes, the multi-step method achieved a Root-Mean-Square Error (RMSE) of 0.291 for the lower layer, a 27% improvement over the 0.400 RMSE of the linear method.
- Structural Fidelity: Visual analysis confirmed that the multi-step method successfully recovers fine-scale vortices and structures in the deep layer that linear models fail to detect.
- Scalability and Robustness: The advantages of the multi-step approach became more pronounced as the number of tracers and ensemble size increased. Furthermore, because the solution is analytic, it avoids the sampling errors common in Monte Carlo-based methods.
- Computational Efficiency: By employing a "constant covariance" approximation, the authors reduced computational costs by nearly 50% while maintaining high accuracy (RMSE of 0.303).
Conclusion
This research demonstrates that non-Gaussianity is a vital feature for accurate data assimilation in turbulent systems. By producing a Gaussian mixture posterior, the multi-step CGDA can represent complex, skewed distributions essential for identifying extreme events.
Future work will focus on extending this recurrent approach to systems with dozens of layers, common in real-world ocean models. Additionally, the authors aim to develop a forward-only filter version to make the method more suitable for real-time "online" forecasting, as the current version relies on a backward smoother for generating pseudo-observations.
Summary updated