Getting Started

Create an environment with pyhton installed, e.g., with conda create -n dlwp-hpx python and activate it with conda activate dlwp-hpx.

Install dlwp-hpx package with pip install ., favorably in a dedicated environment (as shown above). Then cd into src/dlwp-hpx

Verify installation with python remap/healpix.py, which projects the exemplary data/era5_z500.nc file from the equirectangular LatLon convention of shape $[..., H, W]$ (with $H$ the latitudes and $W$ the longitudes) into the HEALPix mesh of shape $[..., F, H, W]$, where $F=12$ the number of faces and $H=W$ the side length of the squares.

Data

Download ERA5

Download data from the ERA5 dataset, i.e.,

1. constants
   • land-sea mask
   • orography
2. surface variables
   • two meter temperature
   • total column water vapor
3. pressure fields
   • geopotential 250 hPa
   • geopotential 300 hPa
   • geopotential 500 hPa
   • geopotential 700 hPa
   • geopotential 1000 hPa
   • air temperature 850 hPa

Subsequently, project the data to the HEALPix mesh, using the remapping functions in src/remapping/healpix.py. More details about data preprocessing are provided in data_pipeline.md.

Download preprocessed data for 2017 and 2018

TODO: Add link to data

Model Checkpoints

Pretrained weights for various model configurations (according to the iterative advancements) are provided for download at ...

TODO: Add link to checkpoint files

hpx32_unet_6var_up-inter_clrelu_300/ hpx32_unet_6var_dil_300/ hpx32_unet_6var_dil_cnxt_300/ hpx32_unet_6var_dil_cnxt_inv_300/ hpx32_unet_136-68-34_cnxt_skip_dil_gru_6h_300/ hpx64_unet_136-68-34_cnxt_skip_dil_gru_6h_300/ hpx64_unet_136-68-34_cnxt_skip_dil_gru_3h_300/

Train

Multiple exemplary commands to train a DLWP-HPX model are provided in src/dlwp-hpx/scripts/run_training.sh. To start a training of the HEALPix model, prepare an according RUN_CMND in run_training.sh and run sh scripts/train.py (from src/dlwp-hpx/).

Forecast

To generate a forecast with a trained model, consider the src/dlwp-hpx/scripts/forecast.py script, which requires the model directory as -m argument.

To convert the forecast file generated by a DLWP-HPX model to the equirectangular LatLon mesh, consider the src/dlwp-hpx/remap/healpix.py file.