Surface Quasi-Geostrophic (SQG) Conditional Forecasting

This dataset contains conditional ensembles of trajectories from the surface quasi-geostrophic (SQG) equations. Each ensemble shares the same observed coarse initial state but differs in its hidden fine-scale initial conditions, producing multiple possible future trajectories.

The dataset accompanies an anonymous research submission.

Dataset

The observed state consists of 196 real-valued Fourier coefficients of surface buoyancy, recorded at 41 time points between (t=0) and (t=1).

The main file, data/trajectories_fine.npz, contains:

Array	Shape	Description
vecs_train	(80, 200, 41, 196)	Training ensembles
vecs_eval	(20, 200, 41, 196)	Validation and test ensembles

The main experiments used the 80 initial conditions in vecs_train (200 realizations each).

The file also contains group identifiers and time coordinates. All trajectory arrays are stored in float64.

Train, validation, and test conditions are specified in data/split.json and are disjoint. The dataset contains 80 training, 10 validation, and 10 test conditions.

Test reference ensembles

truth_bank/ contains 10,000 reference trajectories for each of the 10 test conditions. These are used to evaluate the predicted conditional distributions.

Files and usage

• data/trajectories_fine.npz — coarse trajectories
• data/split.json — group-level data splits
• data/normalization.npz — training-derived normalization parameters
• truth_bank/group_<gid>.npz — test reference ensembles
• SHA256SUMS — file checksums

Download using:

hf download coarse-grained-forecasting-benchmark/sqg --repo-type dataset --local-dir sqg

For the accompanying evaluation code, set SQG_DATA to sqg/data and SQG_TRUTH_BANK to sqg/truth_bank.

License

Creative Commons Attribution 4.0 International (CC BY 4.0).