Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics

arXiv:2604.08596 · nlin.CD, cs.NA, math.NA, math.OC · Submitted 2026-04-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Nonlinear Sciences. The day in nonlinear science: chaos, patterns, networks and adaptation.

Kai: We're looking at the paper: "Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics".

Mira: Here's the gist of the paper. In this paper, an Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics The results show that for 10% of noise at the IC, the results show a perfect fitting between assimilation curves and the Control This comparison is important because it examines the effectiveness of these two methods in tracking control under different levels of initial condition error.

Kai: Comparison of Assimilation Methods The research compares an Ensemble Kalman Filter (EnKF) to a 4DVAR Data Assimilation System using the Lorenz equations.

Mira: The study examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions.

Kai: With 10% of noise, the trajectories of both are almost perfect.

Title and authors: Kai: This paper, "Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics," is essentially setting up a head-to-head comparison between these two data assimilation techniques when dealing with chaotic dynamics using the Lorenz model <ref:2604.08596#pg1,a 4DVAR Data Assimilation System in Chaotic Dynamics>.

Mira: The authors are Harter and Corrêa, and they're looking at how each method handles errors at the initial condition, specifically testing ten percent, twenty percent, and forty percent noise levels.

Lev: They also look into a more realistic scenario where you only feed one observation—the X variable—at the one eighty time step into the Ensemble Kalman Filter to see how it performs compared to the 4DVAR system <ref:2604.08596#pg1>.

Kai: The title itself tells you that they are testing these methods in chaotic dynamics, and it points toward how reliable they are when the underlying system is inherently unpredictable.

Mira: It also points toward an important implication: that the choice between EnKF and 4DVAR really depends on the noise level and how much data you actually have available for assimilation <ref:2604.08596#pg1>.

Lev: The paper highlights that while 4DVAR is more accurate because it's model-dependent, it's also computationally much more expensive than simpler methods like the three deeVAR system, which is a key practical consideration for real hardware <ref:2604.08596#pg1>.

Kai: And they touch on the idea that even with these advanced tools, there are still limitations regarding what kind of observation coverage you need to get good results.

Mira: So what this means practically is that if you're dealing with chaotic systems, you need to understand exactly how much initial error your starting point has because the system can diverge quickly.

The paper's summary: Kai: To summarize, the paper examines how effective these two systems are at tracking the control when we introduce varying levels of noise in the initial conditions, using ten percent, twenty percent, and forty percent error as benchmarks.

Mira: The core finding is that at ten percent noise, both methods give almost perfect results for trajectories. But as you go to twenty percent noise, the Ensemble Kalman Filter starts showing larger deviations from the true trajectory later on due to the chaotic nature of the system.

Lev: So we see a trade-off: EnKF gets away with a bit of error initially but struggles with long-term tracking in chaotic regimes, whereas 4DVAR seems to maintain better accuracy for twenty percent noise <ref:2604.08596#pg1>.

Kai: They also investigated what happens when you have forty percent initial condition error, and they found that neither the Ensemble Kalman Filter nor the 4DVAR could track the control if you only fed them three observations <ref:2604.08596#pg1,neither the Ensemble Kalman Filter>.

Mira: That result suggests a clear demand: for forty percent initial condition error in this type of system, you need either a significantly larger number of observations for the EnKF assimilation or a much larger data assimilation window for the 4DVAR <ref:2604.08596#pg1>.

Lev: It reinforces that these systems aren't magic; they are heavily dependent on how much information you feed them and how unstable the underlying dynamics actually are.

The paper's improvements: Kai: The paper suggests a few ways to improve the methods themselves. First, they point out that the Ensemble Kalman Filter can be improved by replacing its reliance on linearization with methods that handle non-linearity more robustly, because EKF can drift away from the true solution if the linearization isn't good enough.

Mira: I agree with that. And they also suggest enhancing 4DVAR by incorporating a better way to calculate the cost function gradient, specifically by using an adjoint model integrated backwards in time to get an exact cost function gradient, which is computationally tough but would be more accurate <ref:2604.08596#pg1>.

Lev: From a hardware standpoint, the paper points out that if you want to run these systems operationally on real hardware, like for weather forecasting, you can't always afford the full complexity of 4DVAR because it's too expensive; therefore, methods like reduced rank Kalman Filters become more interesting because they let you estimate the error covariance <ref:2604.08596#pg1>.

Kai: And they also touch on designing a system that handles chaotic dynamics better by recognizing that for forty percent initial condition error, the chaos demands a relatively large number of observations in the EnKF assimilation and a large data assimilation window in 4DVAR <ref:2604.08596#pg1>.

Mira: They also suggest making sure observation coverage is comprehensive; the results showed that you get better results when your observation covers all parts of the model vector, which makes sense because you need information across all dimensions to handle that chaos.

Conclusion: Kai: So wrapping up this paper on "Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics," we see that both methods have their limits, especially when dealing with the inherent sensitivity of chaotic dynamics and high initial condition errors <ref:2604.08596#pg1,a 4DVAR Data Assimilation System in Chaotic Dynamics>.

Mira: The main implication for users is that you need to be honest about your initial state error because the performance of these systems drops off significantly as you increase it, and the required observation density changes depending on whether you use EnKF or 4DVAR <ref:2604.08596#pg1>.

Lev: For me, the final thing is that operational implementation means we can't always run the most accurate system like full 4DVAR; we have to settle for simplified methods like Ensemble Kalman Filter because of the computational cost <ref:2604.08596#pg1>.

Kai: So, this paper shows us that both EnKF and 4DVAR are still state-of-the-art techniques in data assimilation, but they demand different things from the observational setup <ref:2604.08596#pg1>.

Mira: Exactly. The finding that EnKF needs more frequent observations and 4DVAR demands observation in more than one variable of the Lorenz system really guides how we set up future experiments in these areas <ref:2604.08596#pg1>.

Lev: I'm just thinking about what this means for building real hardware; it tells us exactly where the computational overhead lands when you try to push these systems into a real operational environment.

Universidade Federal de Pelotas - Faculdade de Meteorologia - Departamento de Meteorologia · Departamento de Ci๺ e Tecnologia Aeroespacial - Instituto de Aeronía e Espaśco

nlin.CD, cs.NA, math.NA, math.OC

Submitted: 2026-04-07

Updated: 2026-04-07

Comments: 7 pages, 2 figures, J. Aerosp. Technol. Manag., 9(4), 469--475 (2017)

Journal ref: Journal of Aerospace Technology and Management, 9(4), 469--475 (2017)

DOI: 10.5028/jatm.v9i4.811

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 58/100

The gist: In this paper, an Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics The results show that for 10% of noise at the IC…

Key concepts

Ensemble Kalman Filter (EnKF)
A data assimilation technique that uses an ensemble of model runs to estimate the state of a system. It is compared here against 4DVAR to see how well it tracks a control signal when starting with imperfect initial conditions, especially in chaotic dynamics.
4DVAR Data Assimilation System
A sophisticated method that integrates a forecast model forward in time and uses an adjoint model to calculate the cost function gradient. It is compared against EnKF to assess its accuracy in tracking the control signal under various noise levels and initial condition errors.
Chaotic Dynamics (Lorenz Equations)
The Lorenz equations are a set of non-linear differential equations used as a simplified model for chaotic systems, similar to those found in weather forecasting. The study uses this model to test the tracking capabilities of the EnKF and 4DVAR under conditions where small initial errors lead to large differences over time.
Initial Condition (IC) Error
This refers to the inaccuracy in knowing the exact starting state of a model when running simulations. The research specifically tests how much error (10%, 20%, or 40%) at this starting point affects whether the EnKF or 4DVAR can successfully track a desired control trajectory.

Terminology

Summary

In this paper, an Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics The results show that for 10% of noise at the IC, the results show a perfect fitting between assimilation curves and the Control This comparison is important because it examines the effectiveness of these two methods in tracking control under different levels of initial condition error. p. 5

Where the paper says it

Considering 10% of noise at the IC, the results show a perfect fitting between assimilation curves and the Control. These results are still quite good for 20% of noise, but there is a disagreement between the “truth” and the estimation, especially at the end of integration period for EnKF, due the chaotic nature of the system.

Comparison of Assimilation Methods

The research compares an Ensemble Kalman Filter (EnKF) to a 4DVAR Data Assimilation System using the Lorenz equations. The study examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions. p. 0

Where the paper says it

The Lorenz model is chosen for its simplicity in structure and the dynamic similarities with primitive equations models, such as modern numerical weather forecasting. It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions.

With 10% of noise, the trajectories of both are almost perfect. p. 0
Where the paper says it

It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions. With 10% of noise, the trajectories of both are almost perfect.

p. 0
Where the paper says it

It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions. With 10% of noise, the trajectories of both are almost perfect.

p. 0
Where the paper says it

It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions. With 10% of noise, the trajectories of both are almost perfect.

With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. p. 0
Where the paper says it

With 10% of noise, the trajectories of both are almost perfect. With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR.

p. 0
Where the paper says it

With 10% of noise, the trajectories of both are almost perfect. With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR.

However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due to chaotic behavior system. p. 0
Where the paper says it

With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due the chaotic behavior system.

Handling Initial Condition Errors

The study investigated how different initial condition errors affect tracking capability. For the case with 40% error at the Initial Condition, neither the Ensemble Kalman Filter nor 4DVAR could track the Control with only 3 observations ingested. p. 0

Where the paper says it

However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due the chaotic behavior system. However, for the case with 40% error at the Initial Condition, neither the Ensemble Kalman Filter or 4DVAR could track the Control with only 3 observations ingested.

The literature suggests that given a 40% error at the IC, the chaotic nature of the model solution demands a relatively large number of observations in the EnKF assimilation and large data assimilation window in 4DVAR. p. 4
Where the paper says it

However, for the case with 40% error at the IC, neither EnKF or 4DVAR could track the Control with only 3 observations ingested. In fact, according to the literature, given a 40% error at the IC, the chaotic nature of the model solution demands a relatively large number of observations in the EnKF assimilation and large data assimilation window in 4DVAR.

Application to Realistic Scenarios

To evaluate a more realistic assimilation application, an experiment was created where the Ensemble Kalman Filter ingested single observation at the 180th time step in the X, Y, and Z Lorenz variables and only in the X variable. The results show a perfect fit of 4DVAR and the Control during a complete integrations period, but the Ensemble Kalman Filter has a disagreement after the 80th time step. p. 0

Where the paper says it

To evaluate a more realistic assimilation application, it was created an experiment in which the Ensemble Kalman Filter ingested single observation at the 180th time step in the X, Y, and Z Lorenz variables and only in the X variable. The results show a perfect fit of 4DVAR and the Control during a complete integrations period, but the Ensemble Kalman Filter has a disagreement after the 80th time step.

On the other hand, it was shown a considerable disagreement between the Ensemble Kalman Filter trajectories and the Control as well as a total fail of 4DVAR. p. 0
Where the paper says it

The results show a perfect fit of 4DVAR and the Control during a complete integrations period, but the Ensemble Kalman Filter has a disagreement after the 80th time step. On the other hand, it was shown a considerable disagreement between the Ensemble Kalman Filter trajectories and the Control as well as a total fail of 4DVAR.

Better results were obtained for the case in which observation covers all the components of the model vector. p. 0
Where the paper says it

On the other hand, it was shown a considerable disagreement between the Ensemble Kalman Filter trajectories and the Control as well as a total fail of 4DVAR. Better results were obtained for the case in which observation covers all the components of the model vector.

Conclusion on Methodology

The final comments state that EnKF and 4DVAR are state-of-the-art techniques in Data Assimilation implemented in numerical weather forecasting. Despite the limitations, EnKF and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the IC. p. 5

Where the paper says it

FINAL COMMENTS In this study, an EnKF and 4DVAR were available in a Data Assimilation context in chaotic regime. Both methodologies were examined and are effective to track the Control for 10, 20 and 40% of error at the IC.

The study concludes that EnKF needs more frequent observations, and the 4DVAR demands for observation in more than 1 variable of the Lorenz system. p. 5
Where the paper says it

Considering 40% of noise at the IC, both EnKF and 4DVAR fail. Regarding an experiment in which it is explored the assimilation over-determined by lack of observation, results show that EnKF needs more frequent observations, and the 4DVAR demands for observation in more than 1 variable of the Lorenz system.

Experimental Setup

The numerical experiments explored forward from here, where the Lorenz equations are solved by finite differences with a nondimensional time increment of 0.01 for 200 time steps integration length. p. 3

Where the paper says it

RESULTS NUMERICAL EXPERIMENTS In the results explored forward from here, the Lorenz equations are solved by finite differences with a nondimensional time increment of 0. 01 for 200 time steps integration length.

The sensitivity of the model to the IC was evaluated in Fig. p. 3
Where the paper says it

In this paper, only the variable X is depicted, just because Y and Z take to the same conclusions. In this study, the sensitivity of the model (Classic Experiment) to the IC is evaluated.

1a, depicting resulting model state trajectories for σ = 10, b = 8/3 and r = 10, assuming the IC to be X0 = 1. p. 3
Where the paper says it

In this study, the sensitivity of the model (Classic Experiment) to the IC is evaluated. 1a, it is depicted the resulting model state trajectories for σ = 10, b = 8/3 and r = 10, assuming the IC to be X0 = 1.

00, Y0= 3.00, Z0 = 5.00 (Case 1) and assuming IC to be X0 = 1. p. 3
Where the paper says it

00, Z0 = 5. 00 (Case 1) and assuming IC to be X0 = 1.

10, Y0 = 3.30, Z0 = 5.50 (Case 2). p. 0
Where the paper says it

However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due the chaotic behavior system. However, for the case with 40% error at the Initial Condition, neither the Ensemble Kalman Filter or 4DVAR could track the Control with only 3 observations ingested.

In the chaotic regime (r = 32), comparing the trajectories of Case 1 and Case 2, it is clearly seen that the bifurcation in the model solution where L T K is obtained. p. 3
Where the paper says it

By other side, for chaotic regime (r = 32), comparing the trajectories of Case 1 and Case 2, it is clearly seen that the bifurcation in the model solution where: Substituting Eq. 15, it is obtained where L T K is the transpose of LK.

The experiment also explored a realistic application where EnKF ingests a single observation at the 180th time step in the X, Y, and Z Lorenz variables. p. 4
Where the paper says it

Thus, the amount of observed variables applied hitherto is somewhat overestimated. By this reason, Experiment 2 represents a more realist assimilation application, where the EnKF ingests a single observation at the 180th time step in the X, Y, and Z Lorenz variables Figure 2.

Key Model Formulation Details

The Methodology section describes a brief theoretical formulation of Lorenz equations. The Kalman Filter method for non-linear models is called Extended Kalman Filter (EKF) and is given by the definition involving forecast and analysis procedures. p. 1

Where the paper says it

ENSEMBLE KALMAN FILTER Kalman Filter is the best linear unbiased estimator for a linear model under Gaussian assumption for the measurements and model random errors. The Kalman Filter method for non-linear models is called Extended Kalman Filter (EKF) and is given by the following definition, considering forecast and analysis procedures: where τ ≡ π 2 H–2 (1 + a2 ) κ t is the non-dimensional time, being H, a, κ and t layer height, thermal conductivity, wave number Forecast Analysis (1) (2) (3) (4) (5) (6) (7) (8) wn+1 a = wn+1 f + Kn+1 n+1 o y − H( n+1 f w ) " # $ % (7) where Fn is our mathematical model; μn is the stochastic forcing (modeling noise error); subscript n denotes discrete time-step; superscript f represents the forecasting value.

The state vector is defined as wn+1 = [Xn+1, Yn+1, Zn+1] and estimated through the recursion w a n+1 = w f n+1 + Knn[y o n+1 –H(w f n+1)]. p. 1
Where the paper says it

The typical gaussianity, 0-mean and ortogonality hypotheses for the noises are adopted. Th e state vector is defi ned as wn+1 = [Xn+1, Yn+1, Zn+1] and estimated through the recursion w a n+1 = w f n+1 + Kn+1[y o n+1 –H(w f n+1)], where w a n+1 is the analysis value, Kn is the Kalman gain, computed from the minimization of the estimation error variance J n+1 (Lorenz 1963), being J n+1 =E{(w a n+1 – w f n+1) T (w a n+1 – w f n+1)},being E{.

The 4DVAR is summarized by procedures described below, including integration of forecast model forward in time and storage of the trajectories. p. 3
Where the paper says it

By definition, the adjoint model is: (17) (18) (19) (20) (21) (a) (b) Thus, 4DVAR can be summarized by the procedures described below: • Integration of forecast model (Eqs. 1 – 3) forward in time and storage of the trajectories.

The tangent linear operator L(w k f) is defined as the tangent linear equation of the forward model. p. 3
Where the paper says it

on the basic state w k f and time step k. Equation 16 is the tangent linear equation of the forward model (Eq.

The adjoint model is defined by (17) and (18). p. 2
Where the paper says it

However, for operational primitive equation models, where the number of degrees of freedom is of the order of 107 , this approach has some limitations. One of the limitations is overcoming by the adjoint model integrated backwards in time, which achieves an exact cost function gradient.

References

The references section lists numerous sources supporting the methodologies discussed in the paper Anderson JL (2001) An ensemble adjustment filter for data assimilation Bishop C, Etherton B, Majumdar S (2001) Adaptive sampling with the ensemble transform Kalman filter part I: The theoretical aspects Burgers G, Leeuwen PJV, Evensen G (1998) Analysis scheme in the EnKF Courtier P, Andersson E, Heckley W, Vasiljevic D, Hamrud M, Hollingsworth A, Rabier F, Fisher M, Pailleux J (1998) The European Centre for Medium-Range Weather Forecasting (ECMWF) implementation of three dimensional variational assimilation (3D-Var). I: Formulation Courtier P, Talagrand O (1998) Variational assimilation of meteorological observations with the adjoint vorticity equation. p. 5

Where the paper says it

Quart J Roy Met Soc 124(550):1783-1807. 49712455002 Courtier P, Talagrand O (1998) Variational assimilation of meteorological observations with the adjoint vorticity equation.

II: Numerical results Courtier P, Thépaut JN, Hollingsworth A (1994) A strategy for operational implementation of 4D-VAR, using an incremental approach Evensen G (1994) Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errors statistics Kalnay E (2004) Atmospheric modeling, data assimilation and predictability Lorenz EN (1963) Deterministic nonperiodic flow Miller R, Guil M, Gauthiez F (1994) Advanced data assimilation in strongly nonlinear dynamical systems Ott E, Hunt BR, Szunyogh I, Corazza M, Kalnay E, Patil DJ, Yorke JA, Zimin AV Ott E, Hunt BR, Szunyogh I Ott E, Hunt BR Parrish DF, Derber JC (1992) The National Meteorological Center’s spectral statistical-interpolation analysis system Rabier F, Courtier P (1992) Four-dimensional assimilation in the presence of baroclinic instability Rabier F, Jarvinen H, Klinker E, Mahfouf JF, Simmons A (2000) The ECMWF operational implementation of four-dimensional variational physics Saltzman B (1962) Finite amplitude free convection as an initial value problem Sasaki Y (1970) Some basic formalisms in numerical variational analysis Tippet MK, Anderson JL, Bishop CH, Hamill TM, Whitaker JS (2003) Ensemble square-root filters Whitaker JS, Hamill TM (2002) Ensemble data assimilation without perturbed observations.

--- Page 1 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol. p. 0

Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

p. 0
Where the paper says it

, São José dos Campos, Vol. 9, No 4, pp.

9, No 4, pp.469-475, Oct. p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

p. 0
Where the paper says it

9, No 4, pp. 469-475, Oct.

-Dec., 2017

ABSTRACT: In this paper, the Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics The Lorenz model is chosen for its simplicity in structure and the dynamic similarities with primitive equations models, such as modern numerical weather forecasting. It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions. p. 0

Where the paper says it

The Lorenz model is chosen for its simplicity in structure and the dynamic similarities with primitive equations models, such as modern numerical weather forecasting. It was examined if the Ensemble Kalman Filter and 4DVAR are effective to track the Control for 10, 20 and 40% of error at the Initial Conditions.

With 10% of noise, the trajectories of both are almost perfect. With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due to the chaotic behavior system. p. 0
Where the paper says it

With 20% of noise, the differences between the simulated trajectories and the observations as well as “true trajectories” are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due the chaotic behavior system.

However, for the case with 40% error at the Initial Condition, neither the Ensemble Kalman Filter or 4DVAR could track the Control with only 3 observations ingested. p. 0
Where the paper says it

However, the differences are increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due the chaotic behavior system. However, for the case with 40% error at the Initial Condition, neither the Ensemble Kalman Filter or 4DVAR could track the Control with only 3 observations ingested.

With 10% of noise, the trajectories of both are almost perfect.

--- Page 2 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics Fabrício Pereira Harter1, Cleber Souza Corrêa2

--- Page 3 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

Harter FP, Corrêa CS

--- Page 4 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

Th e 4DVAR is model-dependent and computationally much more expensive than 3DVAR, but more accurate In turn, only reduced rank Kalman Filters (KF) can be applied to operational numerical weather forecasting owing the computational cost. KF are especially interesting due the capacity to estimate the error covariance of the analysis. p. 1

Where the paper says it

In turn, only reduced rank Kalman Filters (KF) can be applied to operational numerical weather forecasting owing the computational cost. KF are especially interesting due the capacity to estimate the error covariance of the analysis.

--- Page 5 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

The large dimension of the models can prohibit the implementation of KF in operational numerical weather prediction.

--- Page 6 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

The purpose of this research was to compare an EnKF to a 4DVAR Data Assimilation system through the Lorenz equations and the sensitivity of the model to Initial Condition (IC) for non-linear and chaotic regimes. The main objective was to compare an implementation of the EnKF to a 4DVAR method with different noise levels at the IC and explore the assimilation systems over-determined by lack of observation. p. 1

Where the paper says it

The purpose of this research was to compare an EnKF to a 4DVAR Data Assimilation system through the Lorenz equations and the sensitivity of the model to Initial Condition (IC) for non-linear and chaotic regimes. The main objective was to compare an implementation of the EnKF to a 4DVAR method with different noise levels at the IC and explore the assimilation systems over-determined by lack of observation.

p. 1
Where the paper says it

The purpose of this research was to compare an EnKF to a 4DVAR Data Assimilation system through the Lorenz equations and the sensitivity of the model to Initial Condition (IC) for non-linear and chaotic regimes. The main objective was to compare an implementation of the EnKF to a 4DVAR method with different noise levels at the IC and explore the assimilation systems over-determined by lack of observation.

--- Page 7 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

The organization of the paper is as follows: the “Methodology” section describes a brief theoretical formulation of Lorenz equations and presents a basic introduction to EnKF and 4DVAR.

--- Page 8 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations. p. 1

Where the paper says it

METHODOLOGY LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations: (diameter of the Rayleigh-Bénard cell) and time, respectively; σ ≡ κ–1v is the Prandtl number, being v the kinematic viscosity; b ≡ 4(1 + a2 ) –1 is the relation between the height and the width of the rectangle (orbit travelled; Saltzman 1962); the parameter r = R/Rc ∞ ∆T is the Rayleigh number, being T the temperature and Rc the critical Rayleigh number.

p. 1
Where the paper says it

METHODOLOGY LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations: (diameter of the Rayleigh-Bénard cell) and time, respectively; σ ≡ κ–1v is the Prandtl number, being v the kinematic viscosity; b ≡ 4(1 + a2 ) –1 is the relation between the height and the width of the rectangle (orbit travelled; Saltzman 1962); the parameter r = R/Rc ∞ ∆T is the Rayleigh number, being T the temperature and Rc the critical Rayleigh number.

p. 1
Where the paper says it

METHODOLOGY LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations: (diameter of the Rayleigh-Bénard cell) and time, respectively; σ ≡ κ–1v is the Prandtl number, being v the kinematic viscosity; b ≡ 4(1 + a2 ) –1 is the relation between the height and the width of the rectangle (orbit travelled; Saltzman 1962); the parameter r = R/Rc ∞ ∆T is the Rayleigh number, being T the temperature and Rc the critical Rayleigh number.

p. 1
Where the paper says it

METHODOLOGY LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations: (diameter of the Rayleigh-Bénard cell) and time, respectively; σ ≡ κ–1v is the Prandtl number, being v the kinematic viscosity; b ≡ 4(1 + a2 ) –1 is the relation between the height and the width of the rectangle (orbit travelled; Saltzman 1962); the parameter r = R/Rc ∞ ∆T is the Rayleigh number, being T the temperature and Rc the critical Rayleigh number.

--- Page 9 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

ENSEMBLE KALMAN FILTER Kalnay (2004), “Even if a system starts with a poor initial guess of the state of the atmosphere, the EKF may go through an initial transient period of a week or so, after which it should provide the best unbiased estimate of the state of the atmosphere and its error covariance”.

--- Page 10 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

The purpose of this research was to compare an EnKF to a 4DVAR Data Assimilation system through the Lorenz equations and the sensitivity of the model to Initial Condition (IC) for non-linear and chaotic regimes. The main objective was to compare an implementation of the EnKF to a 4DVAR method with different noise levels at the IC and explore the assimilation systems over-determined by lack of observation.

--- Page 11 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

The organization of the paper is as follows: the “Methodology” section describes a brief theoretical formulation of Lorenz equations and presents a basic introduction to EnKF and 4DVAR.

--- Page 12 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations.

--- Page 13 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

ENSEMBLE KALMAN FILTER Kalnay (2004), “Even if a system starts with a poor initial guess of the state of the atmosphere, the EKF may go through an initial transient period of a week or so, after which it should provide the best unbiased estimate of the state of the atmosphere and its error covariance”.

--- Page 14 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations.

--- Page 15 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

ENSEMBLE KALMAN FILTER Kalnay (2004), “Even if a system starts with a poor initial guess of the state of the atmosphere, the EKF may go through an initial transient period of a week or so, after which it should provide the best unbiased estimate of the state of the atmosphere and its error covariance”.

--- Page 16 ---

J. Aerosp. Technol. Manag., São José dos Campos, Vol.9, No 4, pp.469-475, Oct.-Dec., 2017

LORENZ EQUATIONS Lorenz was looking for the periodic solutions of the Saltzman’s model (1962), considering a spectral Fourier decomposition and only low-order terms. Then, he obtained the following system of non-linear coupled ordinary differential equations.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Ensemble Kalman Filter (EnKF) can be improved by replacing its reliance on linearization with methods that account for system non-linearity more robustly, as Miller et al. (1994) noted that EKF may drift away from the true solution if linearization becomes inaccurate.

  2. 4DVAR systems can be enhanced by incorporating a more sophisticated cost function gradient calculation, potentially through the use of an adjoint model integrated backwards in time to achieve an exact cost function gradient, as described in the Tangent Linear Operator and Tangent Linear Model section.

  3. The improved system can be designed to handle chaotic dynamics with increased observation requirements, specifically by recognizing that given a 40% error at the IC, the chaotic nature of the model solution demands a relatively large number of observations in the EnKF assimilation and large data assimilation window in 4DVAR.

  4. A more realistic application can be developed where single observations are ingested strategically, such as ingesting a single observation at the 180th time step in the X, Y, and Z Lorenz variables for an EnKF experiment to test its performance against 4DVAR.

  5. The system's performance can be optimized by ensuring that observation coverage is comprehensive; the results showed better results are obtained for the case in which observation covers all the components of the model vector.

Abstract

In this paper, the Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics. The Lorenz model is chosen for its simplicity in structure and its dynamical similarities with primitive equation models, such as modern numerical weather forecasting. It was examined whether the Ensemble Kalman Filter and 4DVAR are effective in tracking the control for 10%, 20%, and 40% of error in the initial conditions. With 10% of noise, the trajectories of both methods are almost perfect. With 20% of noise, the differences between the simulated trajectories and the observations, as well as the true trajectories, are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. However, the differences become increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due to the chaotic behavior of the system. For the case with 40% error in the initial conditions, neither the Ensemble Kalman Filter nor 4DVAR could track the control with only three observations ingested. To evaluate a more realistic assimilation application, an experiment was created in which the Ensemble Kalman Filter ingested a single observation at the 180th time step in the X, Y, and Z Lorenz variables, and only in the X variable. The results show a perfect fit of 4DVAR and the control during a complete integration period, but the Ensemble Kalman Filter shows disagreement after the 80th time step. On the other hand, a considerable disagreement between the Ensemble Kalman Filter trajectories and the control is observed, as well as a total failure of 4DVAR. Better results were obtained for the case in which observations cover all the components of the model vector.

Related papers