¯Ö¤ü−ÖÖ´Ö-ÃÖÆüÖµµÖ ú †¬µÖÖ¯Ö ú ´ÖÖ¬µÖ´Ö …15×“Ö Ö»Öß “ÖÖÓ¬Ö‡Ô ¯ÖôûÃÖ Ö›ü ¤üÖî. ÁÖß´ÖŸÖß ÃÖ‘ÓÖ×´Ö¡ÖÖ
Andrew Poje (1) , Anne Molcard (2,3), Tamay Ö zg Ö kmen (4)
description
Transcript of Andrew Poje (1) , Anne Molcard (2,3), Tamay Ö zg Ö kmen (4)
Andrew Poje (1), Anne Molcard (2,3), Tamay ÖzgÖkmen (4)
1 Dept of Mathematics, CSI-CUNY,USA2 LSEET - Universite de Toulon et du Var, France3 ISAC-CNR Torino, Italy4 RSMAS/MPO, Univ. of MIAMI, USA
Direct drifter launch strategies for Lagrangian data assimilation using hyperbolic trajectories
Problem
• Given a model assimilating Lagrangian data, determine initial launch locations which accelerate convergence of model to ‘truth’.
• Well defined optimization problem?
Xo -> Nonlinear ODE’s -> +Assimilation + Nonlinear PDE -> Model Output
• High dimensionality• Unique solutions?• ‘Simple Case’ - perfect model
•How long?• What domain?
QuickTime™ and aMicrosoft Video 1 decompressorare needed to see this picture.
Properties of “Good” Lagrangian Data
• Span velocity space• Span physical space• Independent measurements
• Provide continuous velocity corrections• Capture phase and position of energetic structures in field
Approach
• Molcard et al. Lagrangian Assimilation + Lagrangian C.S. Launch
Strategy
•Target hyperbolic trajectories -
• maximize relative dispersion
• track coherent feature boundaries
The assimilation scheme (OI)
Ua=Ub+K (Yo-H(Ub))
K=BGT(GBGT+E)-1
Uaij=Ub
ij+-1 ijm(Vom-Vb
m)
=1 + o2/b
2
Oth order reduction:assimilation frequency which resolve the gradients;Observed and simulated variables non correlated in space and time; Gaussian distribution of the correction;Only 1 t or 2 successive positions.
Ua analysis variable Ub model calculated variableYo observed variableH(Ub) observation matrix that transforms the forecast data in the observed variable at the observed pointG=dH(Ub)/dUb sensitivity matrixB, E model and observation covariance matrices
A1 :Observation (t0)A2: Observation (t0+t)C1: First guess (t0+t)C2: Corrected forecast (t0+t)
A1
C1
C2
A2
Assimilation results for double-gyre, 1.5 layer MICOM
inflow
outflow
A direct launch strategy, based on tracking the Lagrangian manifolds emanating from strong hyperbolic regions in the flow
field is developed by computing the eigen-structure of the Eulerian fied at ti and tf and use it to initialize the manifold
calculation.
Optimal deployment strategy
Manifolds
Ws inflowing: attracted toward the hyperbolic trajectoryin forward time
Wu outflowing: repelled away from hyperbolic point
FSLE Finite Scale Lyapunov Exponents(measure of time required for a pair of particle trajectories to separate)
Deployments: Each 3 lines of 4 drifters:
•1 ‘Optimal’ (Directed) on Wu
•50 Randomly Choice
Analysis: 50 random deployments vs 1 optimal in terms of:overall coverage (or spreading), integrated Kinetic energy,
integrated distance between drifter observations and simulated trajectories.
Spreading KE
YO-YB Correction
The product of the overall coverage and the integrated distance difference provides a good proxy for both terms in the correction term in the assimilation scheme
ijm(Vom-Vb
m)
E=100% E=60% E=40%
Control
Assimilation
Free
DAY 0 DAY 30 DAY 0 DAY 30
Dependance on the length of the segment and on the “exact point” error
Dependance on the length of the segment and on the “exact point” error
Conclusion
Alignment of initial drifter positions along the out-flowing branch of identifiable Lagrangian boundaries is shown to
optimize both the relative dispersion of the drifters and the sampling of high kinetic energy features in the flow.
The convergence of the assimilation scheme is consistent and considerably improved by this optimal launch strategy.
•‘Real’ ocean? • Identifiable structures• Appropriate time/space scales? (Dart05?)
•Something slightly deeper here?• Fast convergence to attractor (Wu)• Knowing location of attractor => more information
Questions
Preprint available: Ocean Modeling, 2006