Welcome to the Data Assimilation Research Testbed

The Data Assimilation Research Testbed (DART) is an open-source, freely available community facility for ensemble data assimilation (DA). [1] DART is developed and maintained by the Data Assimilation Research Section (DAReS) at the NSF National Center for Atmospheric Research (NSF NCAR).

spaghetti_square

assim_anim

Ensemble Data Assimilation

Ensemble DA is a technique for combining observations with numerical models to estimate the state of a physical system.

It enables modelers, observational scientists, and geophysicists to:

  • Generate initial conditions for forecasts.

  • Create a retrospective estimate of the state of a system, a practice known as producing a reanalysis.

  • Assess the relative value of specific observations on forecast skill, a practice known as conducting an observing system experiment (OSE).

  • Estimate the value of hypothetical observations in order to inform the design of an observing system, a practice known as conducting an observing system simulation experiment (OSSE).

  • Determine a model’s systematic bias in estimating the state of a system, a practice known as diagnosing model error.

The DART software environment makes it easy to explore a variety of data assimilation methods and observations with different numerical models. It provides powerful, flexible DA tools that are easy to use and customize to support efficient and reliable DA applications. While DART is primarily oriented for DA research, it has also been used in operational settings.

DART includes:

  • A comprehensive tutorial introducing the concepts of ensemble DA.

  • Extensive documentation of its source code.

  • Interfaces to a variety of models and observation sets that can be used to introduce new users or graduate students to ensemble DA.

  • Nonlinear and Non-Gaussian DA Capabilities

DART is also designed to facilitate the combination of assimilation algorithms, models, and real or synthetic observations to allow increased understanding of all three. It provides a framework for developing, testing, and distributing advances in ensemble DA to a broad community of users by removing the implementation-specific peculiarities of one-off DA systems.

These tools are intended for use by the full range of geosciencies community: beginners and experts; students and teachers; national centers and university research labs.

Nonlinear and Non-Gaussian Data Assimilation Capabilities in DART

One of the historical drawbacks of ensemble data assimilation techniques is the assumption that the quantities being assimilated obey a normal distribution. While this is often a safe assumption – distributions of temperature and pressure can be approximated using a normal distribution – many quantities such as precipitation, snow depth and tracer concentration, as well as many model parameters aren’t normally distributed.

Applying traditional ensemble data assimilation techniques in situations where assumptions of gaussianity are invalid can lead to poor forecast skill and inconclusive results.

To overcome these problems, DART now implements a novel data assimilation technique that no longer requires quantities to be normally distributed. The Quantile-Conserving Ensemble Filtering Framework (QCEFF) provides a general method of computing increments for the prior ensemble of an observed quantity by allowing the use of arbitrary distributions for the prior and the observation error. For a detailed description of the QCEFF, see Anderson (2022). [2]

While the QCEFF for computing observation increments can lead to significant improvements in analysis estimates for observed variables, those improvements can be lost when using standard linear regression of observation increments to update state variables. The QCEFF also implements a capability to do regression in a probit probability integral transformed space. Doing the regression of observation quantile increments in the transformed space guarantees that the posterior ensembles for state variables also retain the advantages of the observation space quantile conserving posteriors. For example, if state variables are bounded, then posterior ensembles will respect the bounds. The posterior ensembles also respect other aspects of the continuous prior distributions. For a detailed description of this process, see Anderson (2023) [3] and Anderson et al. (2023). [4]

Inflation and localization, methods that improve the quality of ensemble DA, can also negate the advantages of the QCEFF methods. For this reason, both localization and inflation can be done in the probit-transformed quantile space as well. Combining these new methods can significantly improve data assimilation for non-Gaussian quantities in Earth system models by extending the capabilities of ensemble DA to general non-Gaussian and nonlinear distributions. Transformative improvements in DA can result for many applications. The largest improvements are found for bounded variables like tracer concentrations, snow and ice depth, soil moisture, and similar quantities in other parts of the Earth system. Model parameters can also be estimated with DA and large improvements can occur for bounded parameters. Variables that have distinctly non-Gaussian prior distributions can also see large improvements. Examples can include atmospheric quantities like moisture and cloud amount in the presence of convection, and many land surface variables.

For instructions on how to use these tools, see Quantile-Conserving Ensemble Filter Framework.

For step-by-step examples of the QCEFF tools, you can work through examples with the Lorenz 96 tracer model

Organization of the documentation

Because of DART’s extensive scope, this documentation is detailed and carefully organized, enabling you to easily find the information you need. If you have any questions or suggestions for improvements, please contact DAReS staff by emailing dart@ucar.edu.

The documentation is partitioned into three parts:

  • a user guide that explains how to install DART and perform data assimilation

  • source code documentation that provides a detailed description of the programs and modules in the repository

  • a comprehensive description of data assimilation theory

Manhattan Release

The current version of DART is the Manhattan release. Email dart@ucar.edu for advice if you are interested in a model which has not been converted from the previous Lanai release.

Quick-start

DART is available through GitHub. To download the latest version of DART, use:

git clone https://github.com/NCAR/DART.git

Go into the build_templates directory and copy over the closest mkmf.template._compiler.system_ file into mkmf.template.

Edit it to set the NETCDF directory location if not in /usr/local or comment it out and set $NETCDF in your environment. This NetCDF library must have been compiled with the same compiler that you use to compile DART and must include the F90 interfaces.

Go into models/lorenz_63/work and run quickbuild.sh nompi.

$ cd models/lorenz_63/work
$ ./quickbuild.sh nompi

If it compiles, run this series of commands to do a very basic test:

$ ./perfect_model_obs
$ ./filter

If that runs and you have Matlab installed on your system add DART/diagnostics/matlab to your matlab search path and run the plot_total_err diagnostic script while in the models/lorenz_63/work directory. If the output plots and looks reasonable (error level stays around 2 and doesn’t grow unbounded) you have successfully installed DART and completed your first assimilation with it.

If you are planning to run one of the larger models and want to use the Lorenz 63 model as a test, run ./quickbuild.sh. It will build filter and any other MPI-capable executables with MPI.

./quickbuild.sh help will print out the quickbuild.sh usage.

Important

The mpif90 command you use must have been built with the same version of the compiler as you are using.

If any of these steps fail or you don’t know how to do them, go to the DART project web page listed above for very detailed instructions that should get you over any bumps in the process.

Quick-start for developers

To create a fork of DART for your own development you will need a GitHub account.

  1. fork the NCAR/DART repo on GitHub

  2. clone your (new) fork to your machine - this will set up a remote named ‘origin’.

git clone https://github.com/USERNAME/DART.git

where USERNAME is your GitHub username.

  1. create a remote to point back to the NCAR/DART repo. Convention dictates that this remote should be called ‘upstream’

git remote add upstream https://github.com/NCAR/DART.git

Use ‘upstream’ to keep your fork up to date with NCAR/DART. GitHub has documentation on working with forks.

  1. Download one of the tar files (listed below) of ‘large’ files so you can test your DART installation.

  2. If you want to contribute your work back to the DART community, create a feature branch with your work, then issue a pull request to propose changes to NCAR/DART.

There are several large files that are needed to run some of the tests and examples but are not included in order to keep the repository as small as possible. If you are interested in running bgrid_solo, cam-fv, or testing the NCEP/prep_bufr observation converter, you will need these files. These files are available at:

Release

Size

Filename

“Manhattan”

189M

Manhattan_large_files.tar.gz

“wrf-chem.r13172”

141M

wrf-chem.r13172_large_files.tar.gz

“Lanai”

158M

Lanai_large_files.tar.gz

“Kodiak”

158M

Kodiak_large_files.tar.gz

“Jamaica”

32M

Jamaica_large_files.tar.gz

“Hawaii”

32M

Hawaii_large_files.tar.gz

Download the appropriate tar file and untar it into your DART repository. Ignore any warnings about tar: Ignoring unknown extended header keyword.

Citing DART

Cite DART using the following text:

The Data Assimilation Research Testbed (Version X.Y.Z) [Software]. (2024). Boulder, Colorado: UCAR/NSF NCAR/CISL/DAReS. http://doi.org/10.5065/D6WQ0202

Update the DART version and year as appropriate.

References

Theory