Lorenz96 model

A tutorial that validates the NEDAS ensemble Kalman filter implementation against the classic Lorenz 1996 model.

The model is a one-dimensional, cyclic 40-variable system representing mid-latitude zonal atmospheric circulation. With a forcing parameter of F = 8 the system is chaotic, making it a standard benchmark for data assimilation algorithms. One model time step (Δt = 0.05) corresponds to approximately 6 hours.

Results are cross-checked against the deterministic EnKF derivation in Bocquet’s lecture notes (section 5.5).

Topics covered:

• Configuring ensemble size, observation error variance, localization radius, and observation network

• Running the main cycling scheme (with Numba JIT compilation on first cycle)

• Collecting in-memory diagnostics over multiple assimilation cycles

• Animating ensemble spread and truth trajectory

• Plotting error time series to verify filter convergence

The notebook can be run in several environments:

• Google Colab

• Docker (see below)

• Native Python — refer to the environment setup guide

docker pull myying/nedas-tutorials
docker run -it --rm -p 8888:8888 myying/nedas-tutorials

Then open the URL printed in the terminal and navigate to 2.validation_with_lorenz96.ipynb.

The full notebook is available on GitHub.