Presentation is loading. Please wait.

Presentation is loading. Please wait.

Assimilation Algorithms: Tangent Linear and Adjoint models Yannick Trémolet ECMWF Data Assimilation Training Course March 2006.

Similar presentations


Presentation on theme: "Assimilation Algorithms: Tangent Linear and Adjoint models Yannick Trémolet ECMWF Data Assimilation Training Course March 2006."— Presentation transcript:

1 Assimilation Algorithms: Tangent Linear and Adjoint models Yannick Trémolet ECMWF Data Assimilation Training Course March 2006

2 Tangent Linear and Adjoint Models 2 Introduction o 4D-Var is based on minimization algorithms:  Conjugate Gradient,  Newton methods (quasi-Newton, truncated…). o The cost function and its gradient are needed. o Overview:  Writing tangent linear and adjoint models,  Diagnostics.

3 Tangent Linear and Adjoint Models 3 4D Variational Assimilation o Model: o Observations: o Background: o Observation operator: o Cost function to minimise: o At the minimum

4 Tangent Linear and Adjoint Models 4 4D Variational Assimilation o Discretisation: o Observations: o Model: o Observation operator: o The cost function becomes:

5 Tangent Linear and Adjoint Models 5 Linearisation In the first order approximation, a perturbation of the control variable (initial condition) evolves acccording to the tangent linear model: The perturbation of the cost function is: Where is the linearised version of about, and are the departures from observations.

6 Tangent Linear and Adjoint Models 6 Linearisation And the gradient is given by:

7 Tangent Linear and Adjoint Models 7 Adjoint Equation We choose solution of the adjoint system: Then, we can show that: (Multiply by and sum over ). The gradient of the cost function is obtained by a backward integration of the adjoint model.

8 Tangent Linear and Adjoint Models 8 4D-Var Algorithm 1.Integrate forward model gives. 2.Integrate adjoint model backwards gives. 3.If STOP. 4.Compute descent direction (Newton, CG, …). 5.Compute step size : 6.Update initial condition:

9 Tangent Linear and Adjoint Models 9 Example: the Lorenz model Where is the time, the Prandtl number, the Rayleigh number, the aspect ratio, the intensity of convection, the maximum temperature difference and the stratification change due to convection (see references).

10 Tangent Linear and Adjoint Models 10 Lorenz model code SUBROUTINE model(x,dt,nstep) DO i=1,nstep CALL lorenz(x,dxdt) CALL step(x,dxdt,dt) ENDDO END SUBROUTINE model ! --------------------------- SUBROUTINE lorenz(x,dxdt) dxdt(1)=-p*x(1)+p*x(2) dxdt(2)=x(1)*(r-x(3))-x(2) dxdt(3)=x(1)*x(2)-b*x(3) END SUBROUTINE lorenz ! --------------------------- SUBROUTINE step(x,dxdt,dt) DO i=1,3 x(i)=x(i)+dt*dxdt(i) ENDDO END SUBROUTINE step

11 Tangent Linear and Adjoint Models 11 The linear code Linearise each line of the code one by one: dxdt(1) =-p*x(1) +p*x(2) -> dxdt_tl(1)=-p*x_tl(1)+p*x_tl(2) dxdt(2) = x(1)*(r-x(3)) -x(2) -> dxdt_tl(2)=x_tl(1)*(r-x(3)) -x(1)*x_tl(3) -x_tl(2) These can also be written as: dxdt(1)=-p*x(1)+p*x(2) dxdt(2)=x(1)*(r-xtraj(3))-xtraj(1)*x(3)-x(2) This writing convention can save a lot of typing... and errors !

12 Tangent Linear and Adjoint Models 12 Trajectory The trajectory has to be available. It can be: saved which costs memory, recomputed which costs CPU time. Intermediate options exist using checkpointing methods.

13 Tangent Linear and Adjoint Models 13 The tangent linear code SUBROUTINE model_tl(x,dt,nstep) DO i=1,nstep CALL get_trajectory(xtraj,i) CALL lorenz_tl(x,dxdt,xtraj) CALL step_tl(x,dxdt,dt) ENDDO END SUBROUTINE model_tl ! --------------------------- SUBROUTINE lorenz_tl(x,dxdt,xtraj) dxdt(1)=-p*x(1)+p*x(2) dxdt(2)=x(1)*(r-xtraj(3))-xtraj(1)*x(3)-x(2) dxdt(3)=x(1)*xtraj(2)+xtraj(1)*x(2)-b*x(3) END SUBROUTINE lorenz_tl ! --------------------------- SUBROUTINE step_tl(x,dxdt,dt) DO i=1,3 x(i)=x(i)+dt*dxdt(i) ENDDO END SUBROUTINE step_tl

14 Tangent Linear and Adjoint Models 14 Adjoint of one instruction From the tangent linear code: dxdt(1)=-p*x(1)+p*x(2) In matrix form, it can be written as: which can easily be transposed: The corresponding code is: x(1)=x(1)-p*dxdt(1) x(2)=x(2)+p*dxdt(1) dxdt(1)=0.

15 Tangent Linear and Adjoint Models 15 Adjoint of one instruction From the tangent linear code: dxdt(2)=x(1)*(r-xtraj(3))-xtraj(1)*x(3)-x(2) The adjoint is: x(1)=x(1)+(r-xtraj(3))*dxdt(2) x(2)=x(2)-dxdt(2) x(3)=x(3)-xtraj(1)*dxdt(2) dxdt(2)=0. Here again, the trajectory is needed.

16 Tangent Linear and Adjoint Models 16 The Adjoint Code Property of adjoints (transposition): Application: where represents the line of the tangent linear model. The adjoint code is made of the transpose of each line of the tangent linear code in reverse order.

17 Tangent Linear and Adjoint Models 17 Adjoint of loops In the TL code we had: DO i=1,3 x(i)=x(i)+dt*dxdt(i) ENDDO which is equivalent to x(1)=x(1)+dt*dxdt(1) x(2)=x(2)+dt*dxdt(2) x(3)=x(3)+dt*dxdt(3) We can reverse the order and transpose the lines: dxdt(3)=dxdt(3)+dt*x(3) dxdt(2)=dxdt(2)+dt*x(2) dxdt(1)=dxdt(1)+dt*x(1) which is equivalent to DO i=3,1,-1 dxdt(i)=dxdt(i)+dt*x(i) ENDDO

18 Tangent Linear and Adjoint Models 18 Conditional statements o What we want is the adjoint of the statements which were actually executed in the direct model. o We need to know which branch was executed o The result of the conditional has to be stored: it is part of the trajectory !!!

19 Tangent Linear and Adjoint Models 19 Remarks about adjoints o The adjoint always exists and it is unique, assuming spaces of finite dimension. Hence, coding the adjoint does not raise questions about its existence, only questions of technical implementation. o In the meteorological literature, the term adjoint is often improperly used to denote the adjoint of the tangent linear of a non-linear operator. One must be aware that discussions about the existence of the adjoint usually address the existence of the tangent linear model. o Without recomputation, the cost of the TL is usually about 1.5 times that of the non-linear code, the cost of the adjoint between 2 and 3 times that. o The tangent linear model is not necessary to run 4D-Var. It is needed as an intermediate step to write the adjoint.

20 Tangent Linear and Adjoint Models 20 Automatic differentiation o Writing the adjoint of a full meteorological model is a very long and very tedious process. o Existing tools: TAF (TAMC), TAPENADE (Odyssée),...  Reverse the order of instructions,  Transpose instructions instantly without typos !!! o There are still unresolved issues:  It is NOT a black box tool,  Cannot handle non-differentiable instructions (TL is wrong),  Can create huge arrays to store the trajectory,  The code needs to be cleaned-up and optimised.

21 Tangent Linear and Adjoint Models 21 Use of Tangent Linear and Adjoint Models In Data Assimilation At ECMWF

22 Tangent Linear and Adjoint Models 22 Incremental 4D-VAR where : o o computed at tangent linear The cost function can be computed by integrating the tangent linear model.

23 Tangent Linear and Adjoint Models 23 Incremental 4D-VAR o Quadratic cost function allows for more efficient minimisation. o Because of the computational cost, the tangent linear and adjoint models are run at lower resolution. o The trajectory comes from a low resolution non-linear forecast. o Departures from observations are updated at high resolution twice (outer loop).

24 Tangent Linear and Adjoint Models 24 Incremental 4D-VAR

25 Tangent Linear and Adjoint Models 25 How accurate are our Tangent Linear and Adjoint Models ?

26 Tangent Linear and Adjoint Models 26 Diagnostics o Testing the linear and adjoint models. o Is the linear approximation valid in our 4D-VAR ? o Can we extend the 4D-VAR window ? o Can we increase the inner-loop resolution ? o Can we assimilate new weather parameters (rain, clouds…) ? o Can we evaluate algorithm changes ?

27 Tangent Linear and Adjoint Models 27 TL testing o Test of linear model based on Taylor series: o Valid for any perturbation (in practice a set of random vectors). o TL and NL run with the same setup:  Resolution  Physics  Time step  Simpler dynamics  Configuration (IFS)

28 Tangent Linear and Adjoint Models 28 The real world o In 4D-VAR the perturbation is not any vector, it is an analysis increment. It is not random and it is the result of a algorithm which involves the linear model. o The linear and non-linear models are used at different resolutions (T511/T159). o The non-linear model uses more physics. o Humidity is represented in spectral space in the linear model, in grid point space in the non-linear model.

29 Tangent Linear and Adjoint Models 29 Incremental test o Compare TL output with finite difference in 4D-VAR setup (resolution, physics, …). o All the necessary information is present during the minimisation: High resolution non-linear update Low resolution non-linear trajectory Low resolution TL (cost function) Minimisation (TL) Low resolution TL (diagnostic) High resolution non-linear update o All the components are used exactly in their 4D-VAR configuration.

30 Tangent Linear and Adjoint Models 30 Incremental test We can compute the relative error from 4D-Var output: Where is the pseudo-inverse of the simplification operator used in 4D-Var.

31 Tangent Linear and Adjoint Models 31 IFS: First minimisation Relative error TL/finite diff. Operational configuration, T511/T159, No TL physics.

32 Tangent Linear and Adjoint Models 32 IFS: Second minimisation Operational configuration, T511/T159, With linear physics.

33 Tangent Linear and Adjoint Models 33 Time evolution of the error T511/T159, With physics.

34 Tangent Linear and Adjoint Models 34 Time evolution of the error (adiab.) T511/T159, No physics.

35 Tangent Linear and Adjoint Models 35 Impact of TL resolution (adiabatic) T511 non linear, No physics.

36 Tangent Linear and Adjoint Models 36 Impact of TL resolution T511 non linear, TL physics.

37 Tangent Linear and Adjoint Models 37 Impact of the TL physics T511/T159, 12h, Adiabatic case represents the “upper bound”.

38 Tangent Linear and Adjoint Models 38 Longer 4D-VAR window T511/T159, 24h forecast. It is not much worse than 12h!

39 Tangent Linear and Adjoint Models 39 Concluding remarks… o Diagnostics which cannot be carried out with the traditional test are possible:  Vertical interpolation,  Impact of resolution. o Does not replace the usual validation test. o 4D-VAR is tested, not just the TL model.

40 Tangent Linear and Adjoint Models 40 Concluding remarks… o Errors seem very large but 4D-VAR works !!! o The resolution of the inner loop has reached a limit with the current IFS linear model. o Better TL physics is needed (humidity, new data). o Adiabatic tests at T255 show that there is room for improvement.

41 Tangent Linear and Adjoint Models 41 Is there something wrong ? 41 2 nd minimisation should be better (physics).

42 Tangent Linear and Adjoint Models 42 References o X. Y. Huang and X. Yang. Variational data assimilation with the Lorenz model. Technical Report 26, HIRLAM, April 1996. o E. Lorenz. Deterministic nonperiodic flow. J. Atmos. Sci., 20:130-141, 1963.


Download ppt "Assimilation Algorithms: Tangent Linear and Adjoint models Yannick Trémolet ECMWF Data Assimilation Training Course March 2006."

Similar presentations


Ads by Google