From an OpenFOAM run to a dataset¶

FoamNordic owns solver execution and returns the generated case through result.case. StackNordic reads that case and exports stored fields. The hand-off uses files; it does not add Python to the solver loop or couple the two package APIs.

flowchart LR
    F[FoamNordic run] --> C[Completed OpenFOAM case]
    C --> S[StackNordic export]
    S --> H[HDF5 snapshots or time series]

Export a completed run¶

Start with a run returned by fno.Longship(...).launch() and wait for successful completion. The example uses the macOS OpenFOAM launcher; on HPC, use the same module command and shell selected for the original run.

from pathlib import Path
import stacknordic as sno

result = run.stop(timeout=900, progress=True)
if not result.success:
    raise RuntimeError(f"OpenFOAM run failed; inspect {result.logs}")

stored = sno.OpenFOAM.case(case_dir=result.case, of_cmd="openfoam", shell="zsh")
dataset = stored.export(
    output_dir=Path("datasets/cavity"),
    fields={"U": ("Ux", "Uy", "Uz"), "p": "p"},
    times=None, mode="direct", dtype="float32", n_jobs=1, progress=True,
)

times=None exports every complete stored time. Direct export prefers reconstructed fields and can also decode supported decomposed layouts. Keep source cases intact. See FoamNordic run control and StackNordic export for selection and runtime options.

Package and compile¶

Packaging requires a verified Cartesian dataset. Cell-indexed exports remain valid for stored-data analysis but are not eligible for this packaging path.

snapshots = Path("datasets/cavity-snapshots")
sno.Dataset.package(
    dataset_dir=Path("datasets/cavity"), output_dir=snapshots, name="cavity",
    fields={"U": "fields/U", "p": "fields/p"}, n_jobs=1, progress=True,
)
sno.Dataset.compile(
    dataset_dir=snapshots, output_file=Path("datasets/cavity.h5"),
    pattern="cavity_id*.h5", progress=True,
)

Use logical field names such as U when packaging; Ux, Uy and Uz are its exported components. The manifest preserves that association. See Stored-data layout and Storage API.

Reduce or deploy¶

Choose StackNordic filtering according to the physical quantity and available density. Align physical times, coordinates, units and component order before comparing independent simulations; matching array shapes alone is insufficient.

A trained model is a separate output. Declare its tensor contracts and export a FoamNordic FNOM artefact before coupling it to a solver. Dataset names do not implicitly define model tensor bindings.