This repository contains an implementation of the local ensemble transform Kalman filter (LETKF) for performing data assimilation with the SPEEDY intermediate complexity atmospheric general circulation model. It is basically a copy of Takemasa Miyoshi's LETKF repository with some modifications for reducing precision in the forecast model.

The repository is a standard observing system simulation experiment (OSSE). The model is run once to generate a "nature run", from which synthetic observations are extracted. Then, the LETKF is used to perform data assimilation on these observations every 6 hours.

Go to the model/run repository and run bash run_first.sh to spin-up the nature model. Then run bash run_cycle.sh to run the actual nature run. This will output every 6 hours into the DATA/nature directory.

Go to the obs directory and run bash obsmake.sh to make the observations. These will be stored in DATA/obs.

Go to the letkf/run directory and first run bash init.sh. This builds the LETKF executable and sets up the output folders. Then run bash run_cycle.sh double to perform data assimilation with the double precision model or bash run_cycle.sh reduced to perform it with the reduced precision model. The output of the assimilation will be stored in DATA/ensemble.

The file config.sh controls the behaviour of the assimilation system.

The start and end date for the nature run/assimilation cycle can be set with the date parameters at the top of the config file (default: 1st January 1982 - 1st March 1983).

The number of ensemble members can be changed with the n_ens parameter (default: 20).

The number of MPI processes to use (and the "batch size" for the ensemble forecast) is set with the n_procs parameter (default: 4).

The resolution of the nature run can be set with the nat_res parameter, which currently only accepts t30 or t39. If t39 is selected, the nature run must be interpolated to T30 before generating the observations. This can be done by running bash interpolate.sh in the interpolate directory, which interpolates the nature run stored in DATA/nature_t39 to T30 and stores it in DATA/nature.

Two choices of observing network can be selected: real, for which observations are located mostly over the Northern Hemisphere continents like for the actual radiosonde network, and uniform, for which observations are uniformly distributed around the globe. The network can be chosen through the obs_network parameter in config.sh