Koopman-Type Latent Data Assimilation
Data and pretrained models for When Latent Dynamics Meet Data Assimilation: A Benchmark of Koopman-Type Representations.
Code and experiment instructions: KoopmanTypeLatentDataAssimilation.
Contents
main/: pretrained models, fitted operators, and supporting experiment assets.datasets/: prepared training, validation, and test data.
The benchmark covers NL-AE, WL-AE, KAE, Tensor-Var, DMD, DBF, and CGKN on Lorenz–96, Cylinder, Shallow Water, ERA5–Coarse, and ERA5–High.
Download
From the code repository, run:
bash download_data.sh
bash download_ckpt.sh
To download a single data set and its checkpoints:
bash download_data.sh --only lorenz96
bash download_ckpt.sh --only lorenz96
The download scripts place files in the expected local directories and verify their SHA-256 hashes.
Licenses and data sources
Licenses apply by component, as specified in LICENSE.md.
| Component | Source | License |
|---|---|---|
| Model checkpoints and fitted operators produced for this benchmark | Benchmark authors | MIT |
| Lorenz–96 simulation data | Generated using the benchmark's Lorenz–96 simulator | MIT |
| Cylinder data | CFDBench, cylinder boundary-condition subset | Apache-2.0 |
| Shallow Water data | PDEArena ShallowWater-2D | MIT |
| ERA5 data at all included resolutions | ERA5 via WeatherBench 2 | Copernicus Products licence supplied with the WeatherBench 2 ERA5 archive |
Preparation details, original attribution, and copies of the applicable terms are provided in DATA_LICENSES.md.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support