README.md

---
license: cc-by-4.0
pretty_name: Blade3DNO Structured CFD WebDataset
size_categories:
- 1K<n<10K
tags:
- webdataset
- cfd
- scientific-machine-learning
- neural-operator
- turbomachinery
configs:
- config_name: default
  data_files:
  - split: train
    path: train.tar
  - split: validation
    path: val.tar
  - split: test
    path: test.tar
---

# Blade3DNO Structured CFD WebDataset

This repository provides an ML-ready WebDataset release associated with
**Blade3DNO**, a geometry-consistent spectral operator learning framework for
three-dimensional transonic compressor flows.

The data contain reconstructed structured tensors derived from steady RANS
solutions of multi-circular-arc compressor blade configurations under different
operating conditions. Each sample combines geometry, operating-condition
features, grid metrics, and three-dimensional flow fields on a logical grid of
`256 × 64 × 32`.

The paper associated with this release is:

> Liangrui Wei, Yuxin Zhao, Zhou Du, Quanyong Xu, and Feng Zhou.
> [Blade3DNO: Geometry-consistent spectral operator learning for transonic 3D
> compressor flows](https://doi.org/10.1016/j.ast.2026.112334).
> *Aerospace Science and Technology*, 177:112334, 2026.

## Dataset summary

The public release contains 1,828 samples:

| Split | Archive | Samples | Design-ID manifest |
|---|---:|---:|---|
| Train | `train.tar` | 1,279 | `train_design_ids.txt` |
| Validation | `val.tar` | 274 | `val_design_ids.txt` |
| Test | `test.tar` | 275 | `test_design_ids.txt` |
| **Total** |  | **1,828** |  |

The split manifests define membership in each split. Archive member names such
as `000795.npz` are packaging keys rather than design identifiers. Always use
the `design_id` stored inside each NPZ sample, or join against
`coefficients.csv`.

> **Release note:** the article reports 1,827 samples in its experiment
> snapshot, whereas this public package contains 1,828 samples. The archives
> and manifests above are authoritative for this release.

## Repository contents

```text
.
├── train.tar
├── val.tar
├── test.tar
├── coefficients.csv
├── train_design_ids.txt
├── val_design_ids.txt
└── test_design_ids.txt
```

- The three TAR files are WebDataset archives.
- Every TAR member is a compressed NumPy archive with an `.npz` extension.
- `coefficients.csv` contains one `design_id` column and 30 raw scalar
  operating/boundary-condition columns.
- The three text files contain one `design_id` per line.

## Sample schema

Unless noted otherwise, arrays are stored as `float32`.

| Field | Shape / type | Description |
|---|---|---|
| `design_id` | scalar string | Canonical identifier used by the manifests and `coefficients.csv`. |
| `features` | scalar Python object containing a dictionary | Min-max-normalized operating and boundary-condition features. |
| `coordinates` | `(256, 64, 32, 3)` | Physical-grid coordinates, with Cartesian coordinates in the last dimension. |
| `coordinates_centered` | `(256, 64, 32, 3)` | Centered coordinate representation used during preprocessing. |
| `density` | `(256, 64, 32)` | Density field. |
| `velocity` | `(256, 64, 32)` | Velocity-magnitude field. |
| `mach` | `(256, 64, 32)` | Mach-number field. |
| `temperature` | `(256, 64, 32)` | Temperature field. |
| `pressure` | `(256, 64, 32)` | Pressure field. |
| `density_std` | `(256, 64, 32)` | Precomputed normalized density field; the `_std` suffix is retained for compatibility. |
| `mach_std` | `(256, 64, 32)` | Precomputed normalized Mach-number field; the `_std` suffix is retained for compatibility. |
| `temperature_std` | `(256, 64, 32)` | Precomputed normalized temperature field; the `_std` suffix is retained for compatibility. |
| `pressure_std` | `(256, 64, 32)` | Precomputed normalized pressure field; the `_std` suffix is retained for compatibility. |
| `sdf` | `(256, 64, 32)` | Signed-distance-function geometry representation. |
| `normals_x`, `normals_y`, `normals_z` | `(256, 64, 32)` each | Cartesian components of the stored geometry-normal representation. |
| `wall_mask` | `(256, 64, 32)` | Binary wall-region mask stored as `float32`. |
| `metrics_dxi_dx`, `metrics_dxi_dy`, `metrics_dxi_dz` | `(256, 64, 32)` each | Stored derivatives of the logical coordinate `xi`. |
| `metrics_deta_dx`, `metrics_deta_dy`, `metrics_deta_dz` | `(256, 64, 32)` each | Stored derivatives of the logical coordinate `eta`. |
| `metrics_dzeta_dx`, `metrics_dzeta_dy`, `metrics_dzeta_dz` | `(256, 64, 32)` each | Stored derivatives of the logical coordinate `zeta`. |

The directory name used during preprocessing contains `256_32_32`, but the
serialized arrays have the verified shape `256 × 64 × 32`. Downstream code
should use the array shapes stored in the NPZ files.

### Scalar coefficients

The `features` dictionary contains normalized forms of the following 30 columns
from `coefficients.csv`:

```text
inlet_static_pressure
inlet_static_temperature
inlet_velocity_x
inlet_velocity_y
inlet_velocity_z
inlet_velocity_magnitude
inlet_mach_number
inlet_total_pressure
inlet_total_temperature
inlet_dynamic_pressure
inlet_density
inlet_mass_flow_rate
outlet_static_pressure
outlet_static_temperature
outlet_velocity_x
outlet_velocity_y
outlet_velocity_z
outlet_velocity_magnitude
outlet_mach_number
outlet_total_pressure
outlet_total_temperature
outlet_dynamic_pressure
outlet_density
outlet_mass_flow_rate
inlet_total_pressure_p01
inlet_static_pressure_p1
inlet_temperature_t1
velocity_y
velocity_z
outlet_static_pressure_p2
```

Use `coefficients.csv` when the original, non-normalized scalar values are
required.

## Loading the WebDataset

Install the two lightweight reader dependencies:

```bash
pip install numpy webdataset
```

The public archives can then be streamed directly:

```python
import io

import numpy as np
import webdataset as wds

url = (
    "https://huggingface.co/datasets/lrwei/bladenet/"
    "resolve/main/train.tar"
)

dataset = wds.WebDataset(url, shardshuffle=False)
sample = next(iter(dataset))

with np.load(io.BytesIO(sample["npz"]), allow_pickle=True) as data:
    design_id = str(data["design_id"].item())
    coordinates = data["coordinates"]       # (256, 64, 32, 3)
    sdf = data["sdf"]                       # (256, 64, 32)
    pressure = data["pressure"]             # (256, 64, 32)
    temperature = data["temperature"]       # (256, 64, 32)
    density = data["density"]               # (256, 64, 32)
    mach = data["mach"]                     # (256, 64, 32)
    features = data["features"].item()      # dict[str, np.float32]

print(design_id)
print(features)
```

`allow_pickle=True` is required because the scalar `features` field stores a
Python dictionary. Enable it only for dataset artifacts obtained from a trusted
source.

To read another split, replace `train.tar` with `val.tar` or `test.tar`.
Because the archives are large, streaming or copying them to fast local storage
is recommended.

## Data generation and processing

The Blade3DNO study constructs a parametric multi-circular-arc compressor-blade
design space and samples geometry and operating conditions before running
steady RANS simulations. Solver-native multi-block flow fields are then mapped
to globally continuous structured tensors suitable for convolutional models
and neural operators.

The packaged representation includes:

- blade geometry encoded by coordinates, SDF values, and normal components;
- operating and boundary conditions encoded by scalar features;
- raw and normalized flow variables;
- wall masks and logical-to-physical grid metric components.

Refer to the paper for the CFD setup, geometric parameterization, filtering
criteria, reconstruction method, and model experiments.

## Intended uses

This release is intended for research on:

- three-dimensional compressor-flow surrogate modeling;
- neural operators and structured-grid learning;
- geometry-aware scientific machine learning;
- full-field prediction of pressure, temperature, density, and Mach number;
- comparisons with 3D CNN and point-cloud baselines.

The data are numerical RANS results rather than experimental measurements.
Models trained on this release should be validated independently before use in
safety-critical or production engineering workflows.

## Limitations

- This is a processed, grid-aligned ML representation rather than the original
  solver-native multi-block archive.
- Interpolation and reconstruction may smooth or alter local flow features.
- The release contains one large TAR archive per split, so random access is less
  efficient than with many smaller shards.
- Exact reproduction of paper results also depends on the preprocessing,
  training code, random seeds, and experiment configuration used in the study.

## Citation

If you use this dataset, please cite the Blade3DNO paper:

```bibtex
@article{wei2026blade3dno,
  title   = {Blade3DNO: Geometry-consistent spectral operator learning for
             transonic 3D compressor flows},
  author  = {Wei, Liangrui and Zhao, Yuxin and Du, Zhou and
             Xu, Quanyong and Zhou, Feng},
  journal = {Aerospace Science and Technology},
  volume  = {177},
  pages   = {112334},
  year    = {2026},
  doi     = {10.1016/j.ast.2026.112334},
  url     = {https://doi.org/10.1016/j.ast.2026.112334}
}
```

When referring specifically to this packaged release, also include the
repository URL:
`https://huggingface.co/datasets/lrwei/bladenet`.

## License

This dataset is released under the
[Creative Commons Attribution 4.0 International license
(CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).
Redistribution and adaptation are permitted with appropriate attribution.