nansencenter/DAPPER

DAPPER is a set of templates for benchmarking the performance of data assimilation (DA) methods. The tests provide experimental support and guidance for new developments in DA. The typical set-up is a synthetic (twin) experiment, where you specify a dynamic model and an observational model, and use these to generate a synthetic truth (multivariate time series), and then estimate that truth given the models and noisy observations.

Getting started

Install, then read, run and try to understand examples/basic_{1,2,3}.py. Some of the examples can also be opened in Jupyter, and thereby run in the cloud (i.e. without installation, but requiring Google login): . This screencast provides an introduction. The documentation includes general guidelines and the API, but for any serious use you will want to read and adapt the code yourself. If you use it in a publication, please cite, e.g., The experiments used (inspiration from) DAPPER [ref], version 1.2.1, where [ref] points to . Lastly, for an introduction to DA theory also using Python, see these tutorials.

Highlights

DAPPER enables the numerical investigation of DA methods through a variety of typical test cases and statistics. It (a) reproduces numerical benchmarks results reported in the literature, and (b) facilitates comparative studies, thus promoting the (a) reliability and (b) relevance of the results. For example, this figure is generated by examples/basic_3.py, making use of built-in tools for experiment and result management, reproduces figure 5.7 of these lecture notes.

DAPPER is (c) open source, written in Python, and (d) focuses on readability; this promotes the (c) reproduction and (d) dissemination of the underlying science, and makes it easy to adapt and extend.

It also illustrates how to parallelise ensemble forecasts (e.g. the QG model), local analyses (e.g. the LETKF), and independent experiments (e.g. examples/basic_3.py). It comes with a battery of diagnostics and statistics. These all get averaged over subdomains (e..g "ocean" and "land") and then in time. Confidence intervals are computed, including correction for auto-correlations, and used for uncertainty quantification, and significant digits printing. Several diagnostics are included in the on-line "liveplotting" illustrated below, which may be paused for further interactive inspection.

In summary, DAPPER is well suited for teaching and fundamental DA research. Also see its drawbacks.

Installation

Successfully tested on Linux/Mac/Windows.

Prerequisite: Python>=3.9

If you're an expert, setup a python environment however you like. Otherwise: Install Anaconda, then open the Anaconda terminal and run the following commands:

conda create --yes --name dapper-env python=3.9
conda activate dapper-env
python --version

Ensure the printed version is 3.9 or more.
Keep using the same terminal for the commands below.

Install

Either: Install for development (recommended)

Do you want the DAPPER code available to play around with? Then

• Download and unzip (or git clone) DAPPER.
• Move the resulting folder wherever you like,
  and cd into it (ensure you're in the folder with a setup.py file).
• pip install -e '.[dev]'
  You can omit [dev] if you don't need to do serious development.

Or: Install as library

Do you just want to run a script that requires DAPPER? Then

• If the script comes with a requirements.txt file, then do
  pip install -r path/to/requirements.txt.
• If not, hopefully you know the version of DAPPER needed. Run
  pip install dapper==1.5.1 to get version 1.5.1 (as an example).

Finally: Test the installation

You should now be able to do run your script with python path/to/script.py.
For example, if you are in the DAPPER dir,

python examples/basic_1.py

PS: If you closed the terminal (or shut down your computer), you'll first need to run conda activate dapper-env

DA methods

¹: Stochastic, DEnKF (i.e. half-update), ETKF (i.e. sym. sqrt.). Serial forms are also available.
Tuned with inflation and "random, orthogonal rotations".
²: Also supports the bundle version, and "EnKF-N"-type inflation.
³: Resampling: multinomial (including systematic/universal and residual).
The particle filter is tuned with "effective-N monitoring", "regularization/jittering" strength, and more.

For a list of ready-made experiments with suitable, tuned settings for a given method (e.g. the iEnKS), use:

grep -r "xp.*iEnKS" dapper/mods

Test cases (models)

• *: Flexible; set as necessary
• **: Tangent Linear Model included?

The models are found as subdirectories within dapper/mods. A model should be defined in a file named __init__.py, and illustrated by a file named demo.py. Most other files within a model subdirectory are usually named authorYEAR.py and define a HMM object, which holds the settings of a specific twin experiment, using that model, as detailed in the corresponding author/year's paper. A list of these files can be obtained using

find dapper/mods -iname '[a-z]*[0-9]*.py'

Some files contain settings used by several papers. Moreover, at the bottom of each such file should be (in comments) a list of suitable, tuned settings for various DA methods, along with their expected, average rmse.a score for that experiment. As mentioned above, DAPPER reproduces literature results. You will also find results that were not reproduced by DAPPER.

Similar projects

DAPPER is aimed at research and teaching (see discussion up top). Example of limitations:

• It is not suited for very big models (>60k unknowns).
• Time-dependent length of state vector (but this can be emulated).
• Non-uniform time sequences (TODO).

The scope of DAPPER is restricted because

Moreover, even straying beyond basic configurability appears unrewarding when already building on a high-level language such as Python. Indeed, you may freely fork and modify the code of DAPPER, which should be seen as a set of templates, and not a framework.

Also, DAPPER comes with no guarantees/support. Therefore, if you have an operational or real-world application, such as WRF, you should look into one of the alternatives, sorted by approximate project size.

Below is a list of projects with a purpose more similar to DAPPER's (research in DA, and not so much using DA):

The EnKF-Matlab and IEnKS codes have been inspirational in the development of DAPPER.

*: AWI/Liege/CNRS/NERSC/Reading/Delft

Contributors

Patrick N. Raanes, Yumeng Chen, Colin Grudzien, Maxime Tondeur, Remy Dubois

DAPPER is developed and maintained at NORCE (Norwegian Research Institute) and the Nansen Environmental and Remote Sensing Center (NERSC), in collaboration with the University of Reading, the UK National Centre for Earth Observation (NCEO), and the Center for Western Weather and Water Extremes (CW3E).

Publication list

• Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the Lorenz 96 model
• Adaptive covariance inflation in the ensemble Kalman filter by Gaussian scale mixtures
• Revising the stochastic iterative ensemble smoother
• p-Kernel Stein Variational Gradient Descent for Data Assimilation and History Matching
• Springer book chapter: Data Assimilation for Chaotic Dynamics