Local Ensemble data assimilation in OOPS¶

The Local Ensemble DA application is a generic application for running data assimilation with Local Ensemble Kalman filters. It can be extended to use any Local EnKF that updates state gridpoints independently from each other by using observations within a localization distance from a gridpoint.

Configuration file (e.g. letkf.yaml) for running LocalEnsembleDA application has to contain the following sections:

geometry - geometry for the background and analysis files
background - ensemble background members (currently used both for computing H(x) and as backgrounds)
observations - describes observations, observation errors, observation operators used in the assimilation, and the horizontal localization
driver - describes optional modifications to the behavior of the LocalEnsembleDA driver
local ensemble DA - configuration of the local ensemble DA solver package

Supported modifications to the driver¶

Read HX from disk instead of computing it at run-time.

driver:
  read HX from disk: false #default value

Compute posterior observer and output test prints for the oma statistics. One might choose to set this flag to false in order to speed up completion of the localEnsembleDA solver.

driver:
  do posterior observer: true #default value

Run LocalEnsembleDA in observer mode to compute HX offline. This works hand-in-hand with read HX from disk. One might choose to separate this into two steps because it is possible to use more efficient round-robin distribution if run as observer only: true.

driver:
  run as observer only: false #default value

Save posterior mean. Requires “output” section in the yaml file.

driver:
  save posterior mean: false #default value

Save posterior ensemble. Requires “output” section in the yaml file.

driver:
  save posterior ensemble: true #default value

Save prior mean. Requires “output mean prior” section in the yaml file.

driver:
  save prior mean: false #default value

save posterior mean increment. Requires “output increment” section in the yaml file.

driver:
  save posterior mean increment: false #default value

save prior variance. Requires “output variance prior” section in the yaml file.

driver:
  save prior variance: false #default value

save posterior variance. Requires “output variance posterior” section in the yaml file.

driver:
  save posterior variance: false #default value

Default behavior is for the LocalEnsembleDa to update the obs config with the geometry info relevant to this PE. This is needed for Halo distribution to work properly. If not using Halo distribution or using models that do not implement grid decomposition (e.g. l95) one might choose to not update obs config by setting update obs config with geometry info : false.

driver:
  update obs config with geometry info: true #default value

Supported Local Ensemble Kalman filters¶

LETKF¶

Two Local Ensemble Transform Kalman Filter (Hunt et al 2007) implementations are supported:

C++ implementation using Eigen (double precision).

This implementation is used when LETKF keyword is used in solver section of configuration file:

local ensemble DA:
  solver: LETKF

GSI-LETKF Fortran implementation using LAPACK (single precision).

This implementation is used when GSI LETKF keywords are used in solver section of configuration file:

local ensemble DA:
  solver: GSI LETKF

LGETKF¶

Another available solver is Local GETKF (Gain form of the Ensemble Transform Kalman filter, Bishop et al 2017) using modulated ensembles to emulate model-space localization in vertical. The implementation calls GSI-GETKF Fortran implementation and follows Lei et al 2018.

To use LGETKF, specify GETKF in solver section. Using LGETKF also requires specifying parameters for the modulation product that emulates model-space localization in vertical:

fraction of retained variance - fraction of the variance retained after the eigenspectrum of the vertical localization function is truncated (1 – retain all eigenvectors, 0 – retain the first eigenvector)
lengthscale units - name of variable for vertical localization. FV3 implementation currently supports two types of units: logp – logarithm of pressure at mid level of the vertical column with surface pressure set to 1e5 at all points, and levels – indices of vertical levels.
lengthscale - localization distance in the above units, at which Gaspari-Cohn localization function is zero.

An example of using LGETKF solver in FV3:

local ensemble DA:
  solver: GETKF
  vertical localization:
    fraction of retained variance: .95
    lengthscale: 1.5
    lengthscale units: logP

Localization supported in the ensemble solvers¶

Observation-space \(R-localization\) is used in all local solvers following Frolov et al, 2024. The obs localizations syntax specifies a sequence of obs localizations for each obs space. Localization is initialized to all ones internally and is refined (multiplied) with each subsequent localization in the list. In other words, we assume that localizations are separable.

The horizontal localization sequence is specified as following for each obs space:

observations:
  observers:
  - obs space:
      name: radiosonde
    ...
    obs localizations:
    - localization method: Horizontal Gaspari-Cohn    # inflate errors with Gaspari-Cohn function, based on the
                                                      #   horizontal distance from the updated grid point
      lengthscale: 1000e3                             # horizontal localization distance in meters

Other options for obs-space localization are available outside of OOPS. Specifically, UFO supports Gaspari-Cohn, SOAR, and Box Car localizations with kd-tree distance search (e.g., localization method: Horizontal SOAR). Additional localizations are supported in SOCA (Rossby radius based) and FV3-JEDI (soil-specific localization).

Similarly, the vertical localization sequence is specified as:

observations:
  observers:
  - obs space:
      name: radiosonde
    ...
    obs localizations:
    - localization method: Vertical localization      # As above but for vertical localization
      localization function: Gaspari Cohn             # Function for vertical localization
      ioda vertical coordinate group: MetaData        # Group containing the below vertical coordinate
      ioda vertical coordinate: height                # Name of UFO variable storing the vertical coordinate
                                                      #   of the observation locations
      vertical lengthscale: 6e3                       # vertical localization distance in units of given coord

The Gaspari-Cohn, SOAR, and Box Car methods are also supported for vertical localization (e.g., localization function: SOAR). If using vertical localization for LETKF (or GSI LETKF), the 3D Geometry Iterator must be enabled to carry model height information into the vertical localization routines following:

geometry:
  iterator dimension: 3

Finally, the LGETKF implementation uses ensemble modulation to approximate model-space vertical localization (see above section for details).

Inflation supported in the ensemble solvers¶

Several covariance inflation methods are supported:

multiplicative prior inflation:

\[{P^{b}}'=\alpha P^{b}\]

Parameter of multiplicative inflation is controlled by inflation.mult configuration value, for example:

local ensemble DA:
  inflation:
    mult: 1.1

RTPP (relaxation to prior perturbation), Zhang et al, 2004

\[{X_{a}^{i}}' = \alpha X_{b}^{i} + (1-\alpha) X_{a}^{i}\]

Parameter of RTPP inflation is controlled by inflation.rtpp configuration value, for example:

local ensemble DA:
  inflation:
    rtpp: 0.5

RTPS (relaxation to prior spread), Whitaker and Hamill, 2012

\[{X_{a}^{i}}' = X_{a}^{i} (\alpha \frac{\sigma_{b}-\sigma_{a}}{\sigma_{a}}+1)\]

Parameter of RTPS inflation is controlled by inflation.rtps configuration value, for example:

local ensemble DA:
  inflation:
    rtps: 0.6

Table 1 Inflation options available in different solvers¶
Solver	Inflation options
LETKF	Multiplicative inflation, RTPP, RTPS
GSI LETKF	RTPP, RTPS
GETKF	RTPP, RTPS

NOTE about obs distributions¶

Currently Local Ensemble DA supports InefficientDistribution and Halo obs distribution. When InefficientDistribution distribution is used, all observations and H(x) are replicated on all PEs. When Halo distribution is used, only observations needed on this PE are stored on each PE. Halo distribution allows for more efficient memory management compared to distribution.name: InefficientDistribution, however at the expense of potentially poor load management compared to distribution.name: RoundRobin. For optimal combination of memory and load balancing, we developed an option to run Local Ensemble DA in the observer-only mode with distribution.name: RoundRobin. Then one can read ensemble of H(x) from disk using driver.read HX from disk == true, distribution.name: Halo obs distribution, and driver.do posterior observer == false.

The type of the obs. distribution is specified for each ObsSpace. For example:

observations:
  observers:
  - obs space:
      distribution:
        name: Halo
        halo size: 5000e3