SABER

SABER is the System Agnostic Background Error Representation.

It provides generic software utilities for computing and working with the background error covariance matrix, often referred to as the B matrix.

SABER Error Covariance Model

The B matrix is generally modeled as a series of linear operators, represented in SABER by “SABER blocks”. Such blocks, even if they come from different components of SABER, are often interoperable. The full series of blocks (linear operators) used in a model for B is referred to as a “block-chain”.

A block-chain is composed of a central block surrounded, symmetrically, by a ‘backward’ AKA ‘adjoint’ outer block chain on the left and a ‘forward’ AKA ‘tangent linear’ outer block chain on the right. The forward and backward outer block-chains are always mirrored images of each other as shown in Fig. 3. When reading a block-chain in a YAML configuration, saber blocks are listed from top-to-bottom in the ‘forward’ order, but are first applied to an incoming model increment in the backwards (bottom-to-top) order.

Fig. 3 An outline of a SABER block-chain.

The B matrix can be modeled in one of several ways, depending on the needs of the user. SABER has options for setting up parametric, ensemble, or hybrid background error covariances. A parametric B, sometimes called a “static” B in the literature, could be a model which does not evolve with time or a model that introduces some flow-dependence through dependence on the background state. An ensemble B uses an ensemble of forecasts to update/evolve the background error in time. A hybrid B combines a set of parametric and ensemble models using a weighted sum.

More details here:

• Introduction to SABER Error Covariance Model

SABER blocks

SABER blocks can encapsulate various components:

• BUMP: Background error on an Unstructured Mesh Package
  • General overview
  • BUMP SABER blocks
  • Workflows
• GSI: interface to the GSI covariance
• ID: identity operator
• SPECTRALB: spectral covariance/correlation
  • Spectral To Gauss
  • Spectral Analytical Filter
• StdDev: standard-deviation application
• DuplicateVariables: outer block to duplicate one variable into others
• Interpolation blocks
  • Rescaling Layer
  • Generic Interpolator
• UKMO-specfic saber blocks
  • UKMO_vertical_localization
  • Moisture Incrementing Operator

SABER applications

There are currently only two applications. The main application runs most of the functionality of saber and is called ErrorCovarianceToolbox. In addition we have an application that reads either an ensemble of states or perturbations. It then processes / filters the transformed increments, dumping them to file. More details are in

• The ProcessPerts Application
• The ErrorCovarianceToolbox Application

Calibration of SABER error covariance

A SABER error covariance can be calibrated from ensemble data:

• Calibration of a SABER error covariance model

SABER testing

SABER has its own pseudo-model for testing purposes, called QUENCH. Also, SABER has an automated testing process which will require a few more steps for adding new tests.

For more details here:

• QUENCH
• SABER block testing
• Adding a SABER test

A low-level description of the classes, functions, and subroutines is also available, produced by means of the Doxygen document generator.

Doxygen Documentation