---

name: data-and-download description: "Safely inspect PDEBench dataset metadata and local HDF5 data, visualize downloaded files, and convert 3D CFD velocity fields to vorticity with verified CPU NumPy/JAX paths." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Data and download

Use this sub-skill for metadata-first, local-data workflows: identify a PDEBench shard, inspect its expected HDF5 names and shapes, visualize a file that is already local, or convert a local 3D CFD velocity field to vorticity. Do not begin a multi-gigabyte download, upload, simulation, training run, or benchmark evaluation without an explicit user request and a storage/network/credential check.

Route quickly

1. Read data formats before opening an HDF5 file or constructing a filename.
2. Use the CLI reference for commands. Start with the metadata checker and --help; these have no network side effect.
3. Use the troubleshooting guide when imports, paths, schemas, coordinates, or spectral backends fail.
4. For a local 3D CFD file, use the bundled converter with an explicit output path and --overwrite only when replacement is intentional. Its default backend is NumPy/CPU; JAX is an opt-in acceleration path.
5. Run the deterministic vorticity smoke after an install change or API change.

Scope and boundaries

• Source-evidence modules: pdebench.data_download.download_direct, pdebench.data_download.visualize_pdes, and pdebench.data_gen.velocity2vorticity establish native names, dispatch, and schemas. Their source artifacts are not copied into this runtime tree. The native downloader must never be treated as a harmless default action.
• Bundled skill helpers: metadata checker reads a local CSV only; safe converter reads/writes local HDF5 only with explicit output/overwrite behavior; and vorticity smoke uses a deterministic tiny fixture.
• Full PDE simulation/Hydra generation routes to data-generation. Model training, evaluation, and metrics route to models-and-evaluation.
• The EasyDataverse downloader and uploader are reference-only. They require external service access and possibly credentials; neither is a default route.

Verified operating facts

Recorded inspection evidence verified Python 3.10, pdebench 0.1.0, NumPy 1.26.4, JAX 0.4.38 on CPU, and PyTorch 1.13.1 on CPU. NumPy and JAX spectral vorticity APIs and the CPU path were verified in that recorded environment. The current review shell used Python 3.13 and did not have JAX installed, so those facts are not current-shell verification. CUDA may be useful for acceleration, but is optional and unverified; do not claim GPU verification.

Bundled files

• HDF5 layouts, metadata names, sizes, and filename conventions
• Safe commands and accepted native flags
• Recovery and failure diagnosis
• Local CSV/PDE metadata validator
• Explicit-output velocity-to-vorticity converter
• Deterministic NumPy/JAX API smoke