# Implementation Guide

#### Implementation Guide

This page contains the implementation guide for the online mode of PDAF in which the numerical model is coupled with PDAF into a single program. The implementation guide for the offline mode, in which separate programs for model integrations and assimilation step are executed, is provided on a separate page. The differences between the online and offline modes is described on the page on the Implementation concept of PDAF.

For applied concrete examples on the implementation of a data assimilation system with PDAF, we recommend to first see our implementation tutorials. The implementation guide here will then give a more complete, but also more abstract, guide for all options.

## The implementation of PDAF with an existing model

For the online-mode of data assimilation, the numerical model has to be combined with PDAF into a single program to generate the assimilation system. The implementation of PDAF with an existing model can be performed in a few steps. A possible sequence is the following:

- Adaptation of the parallelization
- Initialization of PDAF and the ensemble by PDAF_init
- Modification of the model code for the ensemble integration
- Implementation of the analysis step
- Adding memory and timing information

In Steps 1 to 3, PDAF is attached to the numerical model to allow it to perform ensemble integrations. Step 4 completes the data assimilation program with the analysis step. Finally, step 5 is optional to obtain information on the computing time and required memory.
The implementation steps are described on the pages linked above. An example of the implementation is given in the test suite provided with PDAF with the routines in `testsuite/main/` and `testsuite/dummymodel_1D/`. Another example is provided in the tutorial implementations in the directory `tutorial/`.

Several subroutines that contain model and observation specific operations are called as call-back routines by PDAF. For this, PDAF defines interfaces for all routines. These routines need to be implemented by the user. The operations in all these routines are rather elementary like the initialization of the model's physical fields from a state vector provided by PDAF and the initialization of the vector of observations. The examples mentioned above also include example implementations for these routines. In addition, the directory `templates/` contains template implementations of the routines that can be used as a basis for new implementations.

An overview of the available options for each filter that can be specified when PDAF is initialized is given on the overview page on options.

## Generating initial ensembles

To perform ensemble data assimilation one has to create an initial ensemble which is then used in the data assimilation process. There are various different ways to generate an ensemble (like random picking of model states, breeding, short-term integrations of perturbed model fields, second-order exact sampling from EOFs). PDAF provides routines to use snapshots from a model integration to first compute EOFs (empirical orthogonal functions) and then to use the EOFs to perform a constrained random transformation to obtain ensemble perturbations to which a central state (i.e. ensemble mean) is added. This second-order exact sampling (Pham, 2001) showed good performance in our own data assimilation applications.

The ensemble generation is described on the separate page: