Changes between Version 4 and Version 5 of EnsembleGeneration


Ignore:
Timestamp:
Dec 11, 2019, 1:08:42 PM (4 years ago)
Author:
lnerger
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • EnsembleGeneration

    v4 v5  
    2222For step 2, the routine [wiki:PDAF_eofcovar PDAF_eofcovar] helps to compute the singular value decomposition. For this task, one provides the routine with the matrix holding the model snapshots. One can also let the routine compute the temporal mean state, or one can compute and subtract the mean state before (this is particularly useful if one does not want to subtract the mean state over the full model trajectory). The routine computes the singular values decomposition and returns the singular vectors and values.
    2323
    24 Using `PDAF_eofcovar`, one has to read in the trajectory. Then one calls the routine, and finally one stores the singular vectors and values into a file. This procedure is demonstrated for the Lorenz-96 model in `testsuite/src/lorenz96/tools/generate_covar.F90`.
     24Using `PDAF_eofcovar`, one has to read in the trajectory. Then one calls the routine, and finally one stores the singular vectors and values into a file. This procedure is demonstrated for the Lorenz-96 model in `models/lorenz96/tools/generate_covar.F90`.
    2525
    2626For the documentation of the calling interface of `PDAF_eofcovar` see [wiki:PDAF_eofcovar the detail page about the routine].
     
    3333
    3434There are two variants to use `PDAF_sampleens`:
    35 1. One can use the routine in the data assimilation program in the call-back routine that initializes the ensemble array for PDAF (usually `init_ens_pdaf`). Thus one starts the data assimilation program. when `PDAF_init` is executed, it calls `init_ens_pdaf`. Here one can use `PDAF_sampleens` to generate the ensemble from the stored singular values and vectors and directly store it in the ensemble array provided by PDAF. Thus, one can directly start the data assimilation sequence. This variant is demonstrated for the Lorenz-96 model in `testsuite/src/lorenz96/init_ens_eof.F90`.
     351. One can use the routine in the data assimilation program in the call-back routine that initializes the ensemble array for PDAF (usually `init_ens_pdaf`). Thus one starts the data assimilation program. when `PDAF_init` is executed, it calls `init_ens_pdaf`. Here one can use `PDAF_sampleens` to generate the ensemble from the stored singular values and vectors and directly store it in the ensemble array provided by PDAF. Thus, one can directly start the data assimilation sequence. This variant is demonstrated for the Lorenz-96 model in `models/lorenz96/init_ens_eof.F90`.
    36361. One can use a separate program that reads the singular values and vectors, and a central ensemble state and then calls `PDAF_sampleens` to generate the ensemble. Then one can store the ensemble state in ensemble restart files. This variant can be useful for both the online- and offline modes of PDAF: [[BR]]In the online mode, one would implement `init_ens_pdaf` as a routine reading the ensemble states from files. This is useful, e.g., for the case of a large scale data assimilation when the execution time for the forecasts and the analysis steps is so long that one needs to checkpoint and restart the data assimilation. In this case the checkpointing could write ensemble restart files in the same format as the ensemble generation program does. Thus, one can use `init_ens_pdaf` to read ensemble restart files either at the very beginning of a data assimilation sequence or when the assimilation program is restarted. [[BR]] In the offline-mode the data assimilation program only contains the analysis step, while the forecasts are computed by the model program. Thus, the model program just reads model restart files, but does not execute `init_ens_pdaf`. In this case one would let the program that generates the ensemble used `PDAF_sampleens` write restart files for the model. The model will use these files to initialize the model fields so that the forecast for an ensemble state can be computed. At the end of a forecast phase, the model writes new restart files. The restart files for each ensemble state are then read by the data assimialtion program to compute the analysis step.
    3737