Errors, Uncertainties in Data Assimilation

1
Errors, Uncertainties in Data Assimilation

• François-Xavier LE DIMET
• Université Joseph FourierINRIA
• Projet IDOPT, Grenoble, France

2
Acknowlegment

• Pierre Ngnepieba ( FSU)
• Youssuf Hussaini ( FSU)
• Arthur Vidard ( ECMWF)
• Victor Shutyaev ( Russ. Acad. Sci.)
• Junqing Yang ( LMC , IDOPT)

3
Prediction What information is necessary ?

• Model
• law of conservation mass, energy
• Laws of behaviour
• Parametrization of physical processes
• Observations in situ and/or remote
• Statistics
• Images

4
Forecast..

• Produced by the integration of the model from an
  initial condition
• Problem how to link together heterogeneous
  sources of information
• Heterogeneity in
• Nature
• Quality
• Density

5
Basic Problem

• U and V control variables, V being and error on
  the model
• J cost function
• U and V minimizes J

6
Optimality System

• P is the adjoint variable.
• Gradients are couputed by solving the adjoint
  model then an optimization method is performed.

7
Remark on statistical information

• Statistical information is included in the
  assimilation
• In the norm of the discrepancy between the
  solution of the model ( approximation of the
  inverse of the covariance matrix)
• In the background term ( error covariance matrix)

8
Remarks

• This method is used since May 2000 for
  operational prediction at ECMWF and MétéoFrance,
  Japanese Meteorological Agency ( 2005) with huge
  models ( 10 millions of variable.
• The Optimality System countains all the available
  information
• The O.S. should be considered as a  Generalized
  Model 
• Only the O.S. makes sense.

9
Errors

• On the model
• Physical approximation (e.g. parametrization of
  subgrid processes)
• Numerical discretization
• Numerical algorithms ( stopping criterions for
  iterative methods
• On the observations
• Physical measurement
• Sampling
• Some  pseudo-observations , from remote
  sensing, are obtained by solving an inverse
  problem.

10
Sensitivity of the initial condition with respect
to errors on the models and on the observations.

• The prediction is highly dependant on the initial
  condition.
• Models have errors
• Observations have errors.
• What is the sensitivity of the initial condition
  to these errors ?

11
Optimality System including errors on the model
and on the observation
12
Second order adjoint
13
(No Transcript)
14
(No Transcript)
15
Models and Data

• Is it necessary to improve a model if data are
  not changed ?
• For a given model what is the  best  set of
  data?
• What is the adequation between models and data?

16
A simple numerical experiment.

• Burgers equation with homegeneous B.C.s
• Exact solution is known
• Observations are without error
• Numerical solution with different discretization
• The assimilation is performed between T0 and T1
• Then the flow is predicted at t2.

17
(No Transcript)
18
(No Transcript)
19
Partial Conclusion

• The error in the model is introduced through the
  discretization
• The observations remain the same whatever be the
  discretization
• It shows that the forecast can be downgraded if
  the model is upgraded.
• Only the quality of the O.S. makes sense.

20
Remark 1

• How to improve the link between data and models?
• C is the operator mapping the space of the state
  variable into the space of observations
• We considered the liear case.

21
(No Transcript)
22
(No Transcript)
23
Remark 2 ensemble prediction

• To estimate the impact of uncertainies on the
  prediction several prediction are performed with
  perturbed initial conditions
• But the initial condition is an artefact there
  is no natural error on it . The error comes from
  the data throughthe data assimilation process
• If the error on the data are gaussian what
  about the initial condition?

24
Because D.A. is a non linear process then the
initial condition is no longer gaussian
25
Control of the error
26
(No Transcript)
27
Choice of the base
28
Remark .

• The model has several sources of errors
• Discretization errors may depends on the second
  derivative we can identify this error in a base
  of the first eigenvalues of the Laplacian
• The systematic error may depends be estimated
  using the eigenvalues of the correlation matrix

29
Numerical experiment

• With Burgers equation
• Laplacian and covariance matrix have considered
  separately then jointly
• The number of vectors considered in the correctin
  term varies

30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
(No Transcript)
37
An application in oceanographyin A. Vidards
Ph.D.

• Shallow water on a square domain with a flat
  bottom.
• An bias term is atted into the equation and
  controlled

38
(No Transcript)
39
(No Transcript)
40
(No Transcript)
41
(No Transcript)
42
(No Transcript)
43
RMS ot the sea surface height with or without
control of the bias
44
An application in hydrology(Yang Junqing )

• Retrieve the evolution of a river
• With transportsedimentation

45
Physical phenomena

• fluid and solid transport
• different time scales

46
N
2D sedimentation modeling
1. Shallow-water equations
2. Equation of constituent concentration
3. Equation of the riverbed evolution
47
Semi-empirical formulas

• Bed load function

• Suspended sediment transport rate

are empirical constants
48
An example of simulation
Initial river bed

• Domain

• Space step 2 km in two directions
• Time step 120 seconds

Simulated evolution of river bed (50 years)
49

• Model error estimation controlled system
• model

• cost function

• optimality conditions

• adjoint system(to calculate the gradient)

50
Reduction of the size of the controlled problem

• Change the space bases

Suppose is a base of the
phase space and is time-dependent
base function on 0, T, so that
then the controlled variables are changed to
with controlled space size
51
Optimality conditions for the estimation of
model errors after size reduction
If P is the solution of adjoint system, we
search for optimal values of
to minimize J
52

• Problem how to choose the spatial base
  ?
• Consider the fastest error propagation direction
• Amplification factor
• Choose as leading eigenvectors of
• Calculus of
• - Lanczos Algorithm

53
Numerical experiments with another base

• Choice of correct model
• - fine discretization domain with 41 times
  41 grid points
• To get the simulated observation
• - simulation results of correct model
• Choice of incorrect model
• - coarse discretization domain with 21
  times 21 grid points

54
The difference of potential field between two
models after 8 hours integration
55
Experiments without size reduction (108348)
the discrepancy of models at the end of
integration
before optimization
after optimization
56
Experiments with size reduction (38048)
the discrepancy of models at the end of
integration
before optimization
after optimization
57
Experiments with size reduction (3808)
the discrepancy of models at the end of
integration
before optimization
after optimization
58
Conclusion

• For Data assimilation, Controlling the model
  error is a significant improvement .
• In term of software development its cheap.
• In term of computational cost it could be
  expensive.
• It is a powerful tool for the analysis and
  identification of errors