Weak Constraints 4D-Var - PDF Free Download

Weak Constraints 4D-Var Yannick Trémolet ECMWF Training Course - Data Assimilation May 1, 2012 Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 1 / 30

Outline 1 Introduction 2 The Maximum Likelihood Formulation 3 4D Variational Data Assimilation Model Error Forcing Control Variable 4D State Control Variable 4 Covariance matrix 5 Results Constant Model Error Forcing Systematic Model Error Is it model error? 6 Towards a long assimilation window 7 Summary Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 2 / 30

4D Variational Data Assimilation 4D-Var comprises the minimisation of: J(x) = 1 2 [H(x) y]t R 1 [H(x) y] + 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 F(x)T C 1 F(x) x is the 4D state of the atmosphere over the assimilation window. H is a 4D observation operator, accounting for the time dimension. F represents the remaining theoretical knowledge after background information has been accounted for (balance, DFI...). Control variable reduces to x 0 using the hypothesis: x i = M i (x i 1 ). The solution is a trajectory of the model M even though it is not perfect... Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 3 / 30

Weak Constraint 4D-Var Typical assumptions in data assimilation are to ignore: Observation bias, Observation error correlation, Model error (bias and random). The perfect model assumption limits the length of the analysis window that can be used to roughly 12 hours. Model bias can affect assimilation of some observations (radiance data in the stratosphere). In weak constraint 4D-Var, we define the model error as and we allow it to be non-zero. η i = x i M i (x i 1 ) for i = 1,..., n Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 4 / 30

Weak Constraint 4D-Var We can derive the weak constraint cost function using Bayes rule: p(x 0 x n x b ; y 0 y n ) = p(x b; y 0 y n x 0 x n )p(x 0 x n ) p(x b ; y 0 y n ) The denominator is independent of x 0 x n. The term p(x b ; y 0 y n x 0 x n ) simplifies to: p(x b x 0 ) n p(y i x i ) i=0 Hence [ n ] p(x 0 x n x b ; y 0 y n ) p(x b x 0 ) p(y i x i ) p(x 0 x n ) i=0 Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 5 / 30

Weak Constraint 4D-Var [ n ] p(x 0 x n x b ; y 0 y n ) p(x b x 0 ) p(y i x i ) p(x 0 x n ) i=0 Taking minus the logarithm gives the cost function: J(x 0 x n ) = log p(x b x 0 ) n log p(y i x i ) log p(x 0 x n ) The terms involving x b and y i are the background and observation terms of the strong constraint cost function. i=0 The final term is new. It represents the a priori probability of the sequence of states x 0 x n. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 6 / 30

Weak Constraint 4D-Var Given the sequence of states x 0 x n, we can calculate the corresponding model errors: η i = x i M i (x i 1 ) for i = 1,..., n We can use our knowledge of the statistics of model error to define p(x 0 x n ) p(x 0 ; η 1 η n ) One possibility is to assume that model error is uncorrelated in time. In this case: p(x 0 x n ) p(x 0 )p(η 1 ) p(η n ) If we take p(x 0 ) = const. (all states equally likely), and p(η i ) as Gaussian with covariance matrix Q i, weak constraint 4D-Var adds the following term to the cost function: 1 2 n i=1 ηi T Q 1 i η i Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 7 / 30

Weak Constraint 4D-Var For Gaussian, temporally-uncorrelated model error, the weak constraint 4D-Var cost function is: J(x) = 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 + 1 2 n [H i (x i ) y i ] T R 1 i [H i (x i ) y i ] i=0 n [x i M i (x i 1 )] T Q 1 i [x i M i (x i 1 )] i=1 Do not reduce the control variable using the model and retain the 4D nature of the control variable. Account for the fact that the model contains some information but is not exact by adding a model error term to the cost function. The model M is not verified exactly: it is a weak constraint. If model error is correlated in time, the model error term contains additional cross-correlation blocks. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 8 / 30

4D-Var with Model Error Forcing J(x 0, η) = 1 2 n [H(x i ) y i ] T R 1 i [H(x i ) y i ] i=0 + 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 ηt Q 1 η with x i = M i (x i 1 ) + η i. η i has the dimension of a 3D state, η i represents the instantaneous model error, η i is propagated by the model. All results shown later are for constant forcing over the length of one assimilation window, i.e. for correlated model error. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 9 / 30

4D-Var with Model Error Forcing δx 0 η i time TL and AD models can be used with little modification, Information is propagated between obervations and IC control variable by TL and AD models. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 10 / 30

4D State Control Variable Use x = {x i } i=0,...,n as the control variable. Nonlinear cost function: J(x) = 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 + 1 2 n [H(x i ) y i ] T R 1 i [H(x i ) y i ] i=0 n [M(x i 1 ) x i ] T Q 1 i [M(x i 1 ) x i ] i=1 In principle, the model is not needed to compute the J o term. In practice, the control variable will be defined at regular intervals in the assimilation window and the model used to fill the gaps. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 11 / 30

4D State Control Variable J q J b δx 0 x b δx 1 J q δx 2 J q δx 3 time Model integrations within each time-step (or sub-window) are independent: Information is not propagated across sub-windows by TL/AD models, Natural parallel implementation. Tangent linear and adjoint models: Can be used without modification, Propagate information between observations and control variable within each sub-window. Several 4D-Var cycles are coupled and optimised together. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 12 / 30

Model Error Covariance Matrix An easy choice is Q = αb. If Q and B are proportional, δx 0 and η are constrained in the same directions, may be with different relative amplitudes. They both predominantly retrieve the same information. B can be estimated from an ensemble of 4D-Var assimilations. Considering the forecasts run from the 4D-Var members: At a given step, each model state is supposed to represent the same true atmospheric state, The tendencies from each of these model states should represent possible evolutions of the atmosphere from that same true atmospheric state, The differences between these tendencies can be interpreted as possible uncertainties in the model or realisations of model error. Q can be estimated by applying the statistical model used for B to tendencies instead of analysis increments. Q has narrower correlations and smaller amplitudes than B. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 13 / 30

Model Error Covariance Matrix Currently, tendency differences between integrations of the members of an ensemble are used as a proxy for samples of model error. Use results from stochastic representation of uncertainties in EPS. Compare the covariances of η produced by the current system with the matrix Q being used. It is possible to derive an estimate of HQH T from cross-covariances between observation departures produced from pairs of analyses with different length windows (R. Todling). Is it possible to extract model error information using the relation P f = MP a M T + Q? Model error is correlated in time: Q should account for time correlations. Account for effect of model bias. Characterising the statistical properties of model error is one of the main current problems in data assimilation. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 14 / 30

Results: Fit to observations AMprofiler-windspeed Std Dev N.Amer Background Departure Analysis Departure 4 3.2 2.4 1.6 0.8 0 9 10 11 12 13 14 15 16 17 18 19 20 21 Model Error 4 3.2 2.4 1.6 0.8 0 9 10 11 12 13 14 15 16 17 18 19 20 21 Control Fit to observations is more uniform over the assimilation window. Background fit improved only at the start: error varies in time? Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 15 / 30

Mean Model Error Forcing Temperature Model level 11 ( 5hPa) July 2004 Mean M.E. Forcing M.E. Mean Increment Control Mean Increment Monday 5 July 2004 00UTC ECMWF Mean Increment (enrc) Temperature, Model Level 11 Min = -1.97, Max = 1.61, RMS Global=0.66, N.hem=0.54, S.hem=0.65, Tropics=0.77 2 1.8 1.6 1.4 1.2 1 0.8 0.6 - - -0.6-0.8-1 -1.2-1.4-1.6-1.8-2 Wednesday 30 June 2004 21UTC ECMWF Mean Model Error Forcing (eptg) Temperature, Model Level 11 Min = -5, Max = 0, RMS Global=2, N.hem=1, S.hem=3, Tropics=1 Monday 5 July 2004 00UTC ECMWF Mean Increment (eptg) Temperature, Model Level 11 Min = -1.60, Max = 1.15, RMS Global=0.55, N.hem=0.51, S.hem=1, Tropics=0.69 9 8 7 6 5 4 3 2 1-1 -2-3 -4-5 -6-7 -8-9 - 2 1.8 1.6 1.4 1.2 1 0.8 0.6 - - -0.6-0.8-1 -1.2-1.4-1.6-1.8-2 Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 16 / 30

Weak Constraints 4D-Var with Cycling Term Model error is not only random: there are biases. For random model error, the 4D-Var cost function is: J(x 0, η) = 1 2 n [H(x i ) y i ] T R 1 i [H(x i ) y i ] i=0 + 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 ηt Q 1 η For systematic model error, we might consider: J(x 0, η) = 1 2 n [H(x i ) y i ] T R 1 i [H(x i ) y i ] i=0 + 1 2 (x 0 x b ) T B 1 (x 0 x b ) + 1 2 (η η b) T Q 1 (η η b ) Test case: can we address the model bias in the stratosphere? Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 17 / 30

Weak Constraints 4D-Var with Cycling Term No Cycling Term With Cycling Term 90N/ 90S/ MetgraF -3.0-2.0-1.5-1.0 0-0.5-0.30-0 -5-0 -7-4 -2-1 1 2 4 7 0 5 0 0.3 0.50 0.7 1.0 1.5 2.0 - - - 1000mb 850mb 700mb 500mb 300mb 200mb 150mb 100mb 70mb 50mb 30mb 20mb 10mb 5mb 2mb 1.00mb 0.50mb 0mb 0mb 5mb 2mb 1mb 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 68 72 76 80 84 88 90N/ 90S/ MetgraF -3.0-2.0-1.5-1.0 0-0.5-0.30-0 -5-0 -7-4 -2-1 1 2 4 7 0 5 0 0.3 0.50 0.7 1.0 1.5 2.0-1.0-1.0-0.5-0.5-0.5-0.5-0.3-0.3 - - - - - - - - - - - - - - - - - - 0.3 0.3 0.3 0.3 0.3 0.3 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.7 0.7 0.7 0.7 0.7 0.7 1.0 1.0 1.0 1.0 1.5 1.5 2.0 2.0 1000mb 850mb 700mb 500mb 300mb 200mb 150mb 100mb 70mb 50mb 30mb 20mb 10mb 5mb 2mb 1.00mb 0.50mb 0mb 0mb 5mb 2mb 1mb 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 68 72 76 80 84 88 Monthly Mean Model Error (Temperature (K/12h), July 2008) Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 18 / 30

Weak Constraints 4D-Var with Cycling Term No Cycling Term With Cycling Term 90N/ 90S/ MetgraF -30-20 -15-10 -7-5 -3.0-2.0-1.5-1.0 - -0-0 0.7 1.0 1.5 2.0 3 5 7 10 15 20-1.5-1.5-1.5-1.0-1.0-1.0 - - - - - - - - - - - - - - - - - - - - - - - - - 0.7 0.7 0.7 0.7 0.7 1.0 1.0 1.0 1.5 1.5 1000mb 850mb 700mb 500mb 300mb 200mb 150mb 100mb 70mb 50mb 30mb 20mb 10mb 5mb 2mb 1.00mb 0.50mb 0mb 0mb 5mb 2mb 1mb 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 68 72 76 80 84 88 90N/ 90S/ MetgraF -30-20 -15-10 -7-5 -3.0-2.0-1.5-1.0 - -0-0 0.7 1.0 1.5 2.0 3 5 7 10 15 20-1.0-1.0 - - - - - - - - - - - - - - - - - - - - - - - - 0.7 0.7 1.0 1.0 1.5 1000mb 850mb 700mb 500mb 300mb 200mb 150mb 100mb 70mb 50mb 30mb 20mb 10mb 5mb 2mb 1.00mb 0.50mb 0mb 0mb 5mb 2mb 1mb 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 68 72 76 80 84 88 Monthly Mean Analysis Increment (Temperature (K), July 2008) Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 19 / 30

Weak Constraints 4D-Var with Cycling Term AMSU-A Background departures, Channels 13 and 14 RADIANCES FROM METOP / AMSU-A CHANNEL 13 MEAN FIRST GUESS DEPARTURE (OBS-FG) (USED) DATA PERIOD = 2008070100-2008073112 EXP = f57z Min: -0.883688 Max: 0.90642 Mean: -84109 RADIANCES FROM METOP / AMSU-A CHANNEL 13 MEAN FIRST GUESS DEPARTURE (OBS-FG) (USED) DATA PERIOD = 2008070100-2008073112 EXP = f8j2 Min: -0.592767 Max: 8862 Mean: -26685 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E 60 N 60 N 30 N 30 N 0 0 30 S 30 S 60 S 60 S 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E RADIANCES FROM METOP / AMSU-A CHANNEL 14 MEAN FIRST GUESS DEPARTURE (OBS-FG) (USED) DATA PERIOD = 2008070100-2008073112 EXP = f57z Min: -1.6020 Max: 1.7330 Mean: 16017 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E 1.001 0.875 0.75 60 N 60 N 0.625 0.5 0.375 30 N 30 N 5 25 0 0 0-25 -5 30 S 30 S -0.375-0.5-0.625 60 S 60 S 5-0.875-0.999 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E RADIANCES FROM METOP / AMSU-A CHANNEL 14 MEAN FIRST GUESS DEPARTURE (OBS-FG) (USED) DATA PERIOD = 2008070100-2008073112 EXP = f8j2 Min: -1.0986 Max: 0.973196 Mean: 99372 1.001 0.875 0.75 0.625 0.5 0.375 5 25 0-25 -5-0.375-0.5-0.625 5-0.875-0.999 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E 2.002 2.002 1.75 1.75 1.5 1.5 60 N 60 N 60 N 60 N 1.25 1.25 1 1 30 N 30 N 0.75 30 N 30 N 0.75 0.5 0.5 5 5 0 0 0 0 0 0-5 -5-0.5-0.5 30 S 30 S 5 30 S 30 S 5-1 -1-1.25-1.25 60 S 60 S 60 S 60 S -1.5-1.5-1.75-1.75-1.998-1.998 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E 150 W 120 W 90 W 60 W 30 W 0 30 E 60 E 90 E 120 E 150 E Control Model Error Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 20 / 30

Weak Constraints 4D-Var with Cycling Term 1.0 0.8 Weak constraints 4D-Var with cycling - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14 OBS-BG OBS-AN Obs Bias Increment Model Error 0.6 T (K) 0.6 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 Jun Jul Aug Sep Oct Nov Dec 1.0 0.8 Weak constraints - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14 OBS-BG OBS-AN Obs Bias Increment Model Error 0.6 T (K) 0.6 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 Jun Jul Aug Sep Oct Nov Dec The short term forecast is improved with the model error cycling. Weak constraints 4D-Var can correct for seasonal bias (partially). Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 21 / 30

Model Error or Observation Error? Friday 30 April 2004 21UTC ECMWF Mean Model Error (ej6a) Temperature, Model Level 60 Min = -0, Max = 5, RMS Global=0, N.hem=1, S.hem=0, Tropics=0 2 05 9 75 6 45 3 15 05-05 -15-3 -45-6 -75-9 -05-2 Friday 30 April 2004 21UTC ECMWF Mean Model Error (ej8k) Temperature, Model Level 60 Min = -7, Max = 6, RMS Global=0, N.hem=1, S.hem=0, Tropics=0 2 05 9 75 6 45 3 15 05-05 -15-3 -45-6 -75-9 -05-2 The only significant source of observations in the box is aircraft data (Denver airport). Removing aircraft data in the box eliminates the spurious forcing. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 22 / 30

Model Error or Observation Error? obs bg temperature bias (K) 0.5 0.3 0.3 0.5 0.6 0.7 0.8 Aircraft Temperature Bias USA ACARS and TEMP 00z temperature biases Monthly averages for asc, desc and cruise level Asc ACARS Desc ACARS Cruise ACARS TEMP 200 hpa Oct Nov Dec Jan Feb Mar Apr May June July Aug Sep 2000 2001 Observations are biased. Figure from Lars Isaksen Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 23 / 30

- - - Is it model error? 1mb 2mb 1mb 2mb Strong Constraint 5mb Weak Constraint 5mb 1 1.0 0mb 1 0mb 0mb 0mb 2 2-3 0.50mb 3 0.50mb 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 20 22 24 26 28 30 32 34 36 38 40 44 48 52 56 60-20mb 20 15 10 30mb 7 5 3 50mb 2.0 1.5 70mb 1.0 0.7 100mb 0 150mb - -0 200mb - -1.0 300mb -1.5-2.0-3.0 500mb -5-7 -10 700mb -15 850mb -20 1000mb -30 90N/ 90S/ MetgraF - -1.5 1.00mb 2mb 5mb 10mb 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 20 22 24 26 28 30 32 34 36 38 40 44 48 52 56 60-20mb 20 15 10 30mb 7 5 3 50mb 2.0 1.5 70mb 1.0 0.7 100mb 0 150mb - -0 200mb - -1.0 300mb -1.5-2.0-3.0 500mb -5-7 -10 700mb -15 850mb -20 1000mb -30 90N/ 90S/ MetgraF - -1.5 1.00mb 2mb 5mb 10mb The mean temperature increment is smaller with weak constraint 4D-Var (Stratosphere only, June 1993). Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 24 / 30

1.0 - - - Is it model error? 1mb 2mb 1mb 2mb ERA interim 5mb Weak Constraint 5mb 1 0mb 1 0mb -1.5 0mb 0mb 2-2 - 3 0.50mb 3 0.50mb 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 20 22 24 26 28 30 32 34 36 38 40 44 48 52 56 60 2.0-20mb 20 15 10 30mb 7 5 3 50mb 2.0 1.5 70mb 1.0 0.7 100mb 0 150mb - -0 200mb - -1.0 300mb -1.5-2.0-3.0 500mb -5-7 -10 700mb -15 850mb -20 1000mb -30 90N/ 90S/ MetgraF - - - -1.5 - -3.0-1.0 1.0-1.00mb 2mb 5mb 10mb 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 20 22 24 26 28 30 32 34 36 38 40 44 48 52 56 60-20mb 20 15 10 30mb 7 5 3 50mb 2.0 1.5 70mb 1.0 0.7 100mb 0 150mb - -0 200mb - -1.0 300mb -1.5-2.0-3.0 500mb -5-7 -10 700mb -15 850mb -20 1000mb -30 90N/ 90S/ MetgraF - -1.5 1.00mb 2mb 5mb 10mb The work on model error has helped identify other sources of error in the system (balance term). Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 24 / 30

Observation Error or Model Error? 2.5 2.0 1.5 OB-FG OB-AN Obs Bias Increment Model Error Weak constraints 4D-Var with cycling - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14 T (K) 1.0 0.5 0.5 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 Oct Nov Dec Jan Feb 2.5 2.0 1.5 OB-FG OB-AN Obs Bias Increment CY35R2 - Metop-A AMSU-A Tb 13 N.Hemis - Model level 14 T (K) 1.0 0.5 0.5 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 25 01 05 10 15 20 Oct Nov Dec Jan Feb Observation error bias correction can compensate for model error. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 25 / 30

Weak Constraint 4D-Var Configurations 6-hour sub-windows: Better than 6-hour 4D-Var: two cycles are coupled through Jq, Better than 12-hour 4D-Var: more information (imperfect model), more control. Single time-step sub-windows: Each assimilation problem is instantaneous = 3D-Var, Equivalent to a string of 3D-Var problems coupled together and solved as a single minimisation problem, Approximation can be extended to non instantaneous sub-windows. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 26 / 30

Weak Constraint 4D-Var: Sliding Window (1) Weak constraint 4D-Var Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 27 / 30

Weak Constraint 4D-Var: Sliding Window (1) Weak constraint 4D-Var (2) Extended window Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 27 / 30

Weak Constraint 4D-Var: Sliding Window (1) Weak constraint 4D-Var (2) Extended window (3) Initial term has converged Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 27 / 30

Weak Constraint 4D-Var: Sliding Window (1) Weak constraint 4D-Var (2) Extended window (3) Initial term has converged (4) Assimilation window is moved forward Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 27 / 30

Weak Constraint 4D-Var: Sliding Window (1) Weak constraint 4D-Var (2) Extended window (3) Initial term has converged (4) Assimilation window is moved forward This implementation is an approximation of weak contraint 4D-Var with an assimilation window that extends indefinitely in the past......which is equivalent to a Kalman smoother that has been running indefinitely. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 27 / 30

4D State Control Variable: Questions Condition number: The maximum eigenvalue of the minimisation problem is approximately the same as the strong constraint 4D-Var problems for the sub-windows. The smallest eigenvalue is roughly in 1/n 2. The condition number is larger than for strong constraint 4D-Var, Increases with the number of sub-windows (it takes n iterations to propagate information). Simplified Hessian of the cost function is close to a Laplacian operator: small eigenvalues are obtained for constant perturbations which might be well observed and project onto eigenvectors of J o associated with large eigenvalues. Using the square root of this tri-diagonal matrix to precondition the minimisation is equivalent to using the initial state and forcing formulation. Can we combine the benefits of treating sub-windows in parallel with efficient minimization? Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 28 / 30

Future Developments in Weak Constraints 4D-Var In the current formulation of weak constraints 4D-Var (model error forcing): Background term to address systematic error, Interactions with variational observation bias correction, 24h assimilation window, Extend model error to the troposphere and to other variables (humidity). Weak constraint 4D-Var with a 4D state control variable: Four dimensional problem with a coupling term between sub-windows and can be interpreted as a smoother over assimilation cycles. Can we extend the incremental formulation? The two weak constraint 4D-Var approaches are mathematically equivalent (for linear problems) but lead to very different minimization problems. Can we combine the benefits of treating sub-windows in parallel with efficient preconditioning? 4D-Var scales well up to 1,000s of processors, it has to scale to 10,000s of processors in the future. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 29 / 30

Weak Constraints 4D-Var: Open Questions Weak Constraints 4D-Var allows the perfect model assumption to be removed and the use of longer assimilation windows. How much benefit can we expect from long window 4D-Var? Weak Constraints 4D-Var requires knowledge of the statistical properties of model error (covariance matrix). The forecast model is such an important component of the data assimilation system. It is surprising how little we know about its error characteristics. How can we access realistic samples of model error? How can observations be used? 4D-Var can handle time-correlated model error. What type of correlation model should be used? Can we distinguish model error from observation bias or other errors? Is there a need to anchor the system? The statistical description of model error is one of the main current challenges in data assimilation. Yannick Trémolet Weak Constraints 4D-Var May 1, 2012 30 / 30