deinal/spacecast-data-small

Small Vlasiator Dataset for Machine Learning Studies

The data is stored in Zarr format.

It can be downloaded to a local data_small directory with:

from huggingface_hub import snapshot_download

snapshot_download(
repo_id="deinal/spacecast-data-small",
repo_type="dataset",
local_dir="data_small"
)

This will yield a local data_small folder that can be used with spacecast:

data_small/
├── graph/                - Directory containing graphs for training
├── run.zarr/             - Vlasiator run with ρ = 1.0 cm⁻³ solar wind
├── static.zarr/          - Static features x, z, r coordinates
├── vlasiator_run.zarr    - Preprocessed Vlasiator run
├── vlasiator_config.yaml - Configuration file for neural-lam
└── vlasiator_run.yaml    - Configuration file for the datastore, referred to from vlasiator_config.yaml

The run was preprocessed with mllam-data-prep:

mllam_data_prep data_small/vlasiator_run.yaml

This produces a training-ready Zarr store in the data_small directory.

Simple, multiscale, and hierarchical graphs are included already, but can be created using the following commands:

python -m neural_lam.create_graph --config_path data_small/vlasiator_config.yaml --name simple --levels 1 --coarsen-factor 5 --plot
python -m neural_lam.create_graph --config_path data_small/vlasiator_config.yaml --name multiscale --levels 3 --coarsen-factor 5 --plot
python -m neural_lam.create_graph --config_path data_small/vlasiator_config.yaml --name hierarchical --levels 3 --coarsen-factor 5 --hierarchical --plot