Zarr

A reader and writer for Zarr stores in NVIDIA PhysicsNeMo's own mesh layout — the one physicsnemo.mesh.io.from_zarr reads and its MeshReader accepts wherever it accepts a .pmsh. Like .pmsh a store is a directory, but a chunked and compressed one, so a large mesh can be read in pieces.

Format name	zarr
Extensions	.zarr
Read / Write	✓ / ✓
Extra dependencies	zarr (writing needs 3.x)

Reading & writing

import meshioplusplus

meshioplusplus.write("case.zarr", mesh)                  # a directory named case.zarr
mesh = meshioplusplus.read("case.zarr")

meshioplusplus.zarr.write(
    "case.zarr", mesh, manifold_dim="auto", float32=True, chunk_rows=200_000, zstd_level=3
)

manifold_dim/float32 are pmsh's. chunk_rows and zstd_level are upstream's own chunk and compression policy — matching them keeps a store written here indistinguishable from one written by to_zarr — and zstd_level=0 writes uncompressed.

Writing requires zarr 3.x, and the error says so by name: zarr-python 2.x cannot produce a v3 store at all, and a v2 store is not what upstream reads. Reading uses only open_group and is not version-gated in the same way.

File structure

case.zarr/
├── zarr.json          group metadata; root attributes carry the type tag
├── points/            (n_points, dim) float32/float64
├── cells/             (n_cells, k) int64
├── point_data/<name>
├── cell_data/<name>
└── global_data/<name> scalars keep shape ()

Root attributes: physicsnemo_mesh_type: "Mesh" (what upstream dispatches on), __tensordict__ ({"batch_size": [...], "version": 1}, also on each data group, so the store opens with tensordict.from_zarr too) and meshioplusplus:provenance. Metadata is consolidated, as upstream does.

Cell types

Identical to pmsh: one simplex kind per store, with the same tessellate-and-warn reduction.

Data mapping

point_data/cell_data map onto the groups of the same name and field_data onto global_data, with scalars kept genuinely 0-d rather than promoted to length-1 arrays. Regions and non-numeric arrays are dropped with a warning.

Quirks & limitations

• Not write_dataset(format="zarr"). That writes a training dataset — one subgroup per mesh, tabular columns, a manifest in the root attributes — which is a different thing entirely (see ML data handling). Such a store is refused here by name rather than half-read.
• A DomainMesh store is refused by name.
• A store with no type attribute but a points array is accepted, matching from_zarr's own rule.
• Nested groups inside a data group are dropped with a note: meshio++ data dictionaries are flat.
• Provenance rides meshioplusplus:provenance in the root attributes, which is plain JSON — so read_metadata recovers it without importing zarr at all.