DingDong1921/mgn-cylinderflow-stride8-75frames

MeshGraphNets CylinderFlow: phase-0 stride-8, 75 frames

This is a temporal derivative of the CylinderFlow data released with MeshGraphNets. It is not an independently simulated flow family.

Each of the 1,100 original 600-frame trajectories is sampled once with the phase-zero rule raw_uvp[0:600:8]. The HDF5 retains all 75 selected frames with raw indices 0, 8, ..., 592 and physical spacing 0.08 when the original solver output spacing is 0.01. Trajectory-level membership preserves the upstream split: 1,000 Train trajectories and 100 Validation trajectories. Test fields were not accessed and no Test data is included.

The feature/cylinderflow-stride8-1plus64 Text2PDE workflow reads exactly one sample per trajectory: the first 65 stored frames, corresponding to raw indices 0, 8, ..., 512. Frame 0 is the clean condition and frames 1 through 64 are the prediction targets. It does not create sliding windows or alternate temporal phases. The remaining 10 stored frames stay in the release for provenance and reuse.

Files

• data/cylinderflow_stride8_75frames.h5: complete 75-frame HDF5, 1,772,387,753 bytes, SHA-256 d416be274e03a5d77f1cf2dffc4be8abcfc63ff9d6bf5a9cca19a44b43b36533.
• metadata/cylinderflow_stride8_75frames_manifest.json: trajectory identities, split membership, raw-frame mapping, and Train-only statistics.
• metadata/dataset_audit.json: exact source-to-derived checks and data identity.
• metadata/train_normal_stats.pkl: Text2PDE-compatible [u_mean,u_std,v_mean,v_std,p_mean,p_std] pickle.
• LICENSE, NOTICE, and TEST_NOT_ACCESSED.txt: license, attribution, transformation, and access-boundary records.

The HDF5 root format is dgn4cfd.mgn_cylinderflow_temporal_stride.v1. Each trajectory_XXXX group contains:

• uvp: [75, N, 3] float32 with channels (u, v, p);
• mesh_pos: [N, 2] float32;
• cells: [C, 3] int32 triangle connectivity;
• node_type: [N] int32 MeshGraphNets boundary labels.

Node and cell counts vary by trajectory. Load groups lazily and use a batch size of one unless a mesh-aware collator is supplied. The file uses HDF5 2.0 object layouts and requires an h5py build linked against HDF5 2.0 or newer (the accompanying Text2PDE environment pins h5py==3.16.0).

Normalization

The released normalizer uses every one of the 75 stored frames from the 1,000 Train trajectories only:

field	mean	standard deviation
u	0.5661443793145688	0.57014025360096
v	0.0006225839965382098	0.1436773068500826
p	0.04184413450776907	0.34627331301859265

Validation fields do not contribute to these statistics.

Provenance and references

MeshGraphNets' upstream README states that its release contains the full datasets. This derivative follows the upstream Apache License 2.0 and preserves attribution in NOTICE.

• Tobias Pfaff et al., Learning Mesh-Based Simulation with Graph Networks, ICLR 2021.
• MeshGraphNets dataset and code release.
• Text2PDE: Latent Diffusion Models for Accessible Physics Simulation.

Please cite the upstream MeshGraphNets work when using this derivative and cite Text2PDE when using the accompanying Text2PDE training/evaluation workflow.