README.md

---
license: cc-by-sa-4.0
task_categories:
  - graph-ml
tags:
  - cfd
  - aerodynamics
  - point-cloud
  - surrogate-modeling
  - automotive
  - drivaer
  - drivaerml
size_categories:
  - n<1K
---

# DrivAerML Point Clouds

A preprocessed, point-cloud version of the [**DrivAerML**](https://huggingface.co/datasets/neashton/drivaerml) high-fidelity CFD dataset, ready for training point-based deep learning surrogates (PointNet, PCT, DGCNN, Graph Neural Operators, etc.) for automotive external aerodynamics.

The original DrivAerML release contains 500 scale-resolving CFD simulations of parametrically morphed DrivAer notchback geometries and ships as ~31 TB of raw STL / VTP / VTU / OpenFOAM data. This release distills the **surface boundary** of each run down to a single compressed `.npz` file (~10 MB) containing the STL point coordinates and the CFD surface fields interpolated onto those points — so the full usable set fits in **~5 GB** instead of multiple terabytes.

## What's in here

For each run:

- `point_cloud_{i}.npz` — STL surface points with interpolated CFD fields
- `force_mom_{i}.csv` — time-averaged force and moment coefficients (Cd, Cl, Clf, Clr, Cs)

Plus at the dataset root:

- `splits.json` — reproducible train/val/test assignment (seed 42, 80/10/10)
- `processing_log.json` — per-run processing status and nearest-neighbor distance diagnostics

### Directory layout

```
Drivaerml_point_clouds/
├── train/          # 387 runs
│   ├── run_1/
│   │   ├── point_cloud_1.npz
│   │   └── force_mom_1.csv
│   └── ...
├── val/            # 46 runs
│   └── ...
├── test/           # 52 runs
│   └── ...
├── splits.json
└── processing_log.json
```

Total: **485 runs** (of the 500 original designs, 15 runs are missing from the source dataset: 167, 211, 218, 221, 248, 282, 291, 295, 316, 325, 329, 364, 370, 376, 403, 473).

> **Note on the Hugging Face dataset viewer:** the viewer only previews the tabular `force_mom_*.csv` files (the Cd/Cl/Clf/Clr/Cs coefficients). The actual point-cloud payload lives in the `.npz` files, which the viewer does not render — clone the repo or stream it with `huggingface_hub` to access the geometry and fields.

## `.npz` contents

Each `point_cloud_{i}.npz` has three arrays:

| Key           | Shape        | Dtype      | Description                                           |
| ------------- | ------------ | ---------- | ----------------------------------------------------- |
| `points`      | `(N, 3)`     | `float32`  | XYZ coordinates of STL surface points                 |
| `fields`      | `(N, 5)`     | `float32`  | CFD surface fields interpolated to each STL point     |
| `field_names` | `(5,)`       | `str`      | Names of the columns in `fields`                      |

`N` varies per run (roughly 300k points, matching the STL resolution).

The five field columns are:

1. `CpMeanTrim` — time-averaged pressure coefficient
2. `wallShearStressMeanTrim_mag` — magnitude of time-averaged wall shear stress
3. `wallShearStressMeanTrim_x` — x-component
4. `wallShearStressMeanTrim_y` — y-component
5. `wallShearStressMeanTrim_z` — z-component

## How the preprocessing works

The original DrivAerML surface data is stored in **`boundary_{i}.vtp`**: a high-resolution (~8M point) mesh with CFD fields attached as **cell data**. The STL geometry **`drivaer_{i}.stl`** is a lower-resolution (~300k point) representation of the same surface but without any fields.

To produce a clean point cloud where every geometry point carries its own CFD values, the pipeline runs the following per run:

1. Load the `.vtp` boundary mesh and the `.stl` geometry.
2. Convert the VTP's cell-centered fields to point-centered via PyVista's `cell_data_to_point_data()` (averaging adjacent cells to the shared vertex).
3. Build a `scipy.spatial.cKDTree` over the VTP point coordinates.
4. For each STL point, take the **K=1 nearest neighbor** in the VTP point cloud and copy its field values.
5. Save `points`, `fields`, and `field_names` as a compressed `.npz`.
6. Delete the raw VTP and STL to keep disk usage bounded while processing.

The K=1 nearest-neighbor mapping was chosen deliberately for **benchmark comparability** with existing DrivAerML / DrivAerNet++ leaderboard models (TripNet, FIGConvNet, RegDGCNN), all of which operate on the STL vertices directly. The nearest-neighbor distances are logged in `processing_log.json` for every run so this mapping can be audited.

The 80/10/10 train/val/test split is computed with `numpy.random.default_rng(seed=42)` over a sorted list of successfully-processed run IDs, making it fully reproducible from the `splits.json` manifest.

## Loading

### With `datasets` (CSV force/moment preview only)

The Hugging Face dataset loader will pick up the `force_mom_*.csv` files automatically:

```python
from datasets import load_dataset
ds = load_dataset("Jrhoss/Drivaerml_point_clouds")
# ds["train"][0] -> {"Cd": 0.30, "Cl": 0.07, "Clf": -0.04, "Clr": 0.10, "Cs": 0.05}
```

### Point clouds (recommended)

Clone the repo or snapshot-download it, then load `.npz` files directly:

```python
from huggingface_hub import snapshot_download
import numpy as np
from pathlib import Path

root = Path(snapshot_download("Jrhoss/Drivaerml_point_clouds", repo_type="dataset"))

run = np.load(root / "train" / "run_1" / "point_cloud_1.npz", allow_pickle=True)
points = run["points"]            # (N, 3)
fields = run["fields"]            # (N, 5)
names  = run["field_names"]       # ['CpMeanTrim', 'wallShearStressMeanTrim_mag', ...]
```

### Minimal PyTorch `Dataset`

```python
import numpy as np
import pandas as pd
import torch
from pathlib import Path
from torch.utils.data import Dataset

class DrivAerMLPointClouds(Dataset):
    def __init__(self, root, split="train", n_points=16384):
        self.run_dirs = sorted((Path(root) / split).glob("run_*"),
                               key=lambda p: int(p.name.split("_")[1]))
        self.n_points = n_points

    def __len__(self):
        return len(self.run_dirs)

    def __getitem__(self, idx):
        run_dir = self.run_dirs[idx]
        run_id = int(run_dir.name.split("_")[1])

        npz = np.load(run_dir / f"point_cloud_{run_id}.npz", allow_pickle=True)
        points, fields = npz["points"], npz["fields"]

        # Random subsample for batching
        sel = np.random.choice(len(points), self.n_points, replace=False)
        points, fields = points[sel], fields[sel]

        # Integrated force/moment coefficients
        fm = pd.read_csv(run_dir / f"force_mom_{run_id}.csv").iloc[0]
        coeffs = torch.tensor([fm["Cd"], fm["Cl"], fm["Clf"], fm["Clr"], fm["Cs"]],
                              dtype=torch.float32)

        return {
            "points":  torch.from_numpy(points),   # (n_points, 3)
            "fields":  torch.from_numpy(fields),   # (n_points, 5)
            "coeffs":  coeffs,                     # (5,)
            "run_id":  run_id,
        }
```

## Suggested tasks

- **Per-point surface field regression** — predict `CpMeanTrim` and wall shear stress vectors from geometry alone. Comparable to the TripNet / FIGConvNet / RegDGCNN benchmarks (which report ~20% relative L2 error).
- **Integrated coefficient regression** — predict `Cd`, `Cl`, etc. from the point cloud (global pooling over the surface).
- **Coupled prediction** — joint learning of per-point fields and integrated coefficients, using the integrated values as a physics-informed auxiliary loss.

## Known limitations

- **K=1 interpolation is not exact.** It preserves the CFD field values faithfully where VTP and STL points are co-located, but introduces small errors at points where the STL has higher local resolution than the VTP. `processing_log.json` reports `nn_dist_mean`, `nn_dist_max`, and `nn_dist_p99` per run so you can filter out any pathological cases.
- **Fields are time-averaged only.** Transient information (e.g. unsteady vortex shedding in the wake) is not preserved; the source dataset contains scale-resolving data but only the trim-averaged fields are interpolated here.
- **Surface only.** The volumetric flow field (the 50 GB-per-run `volume_{i}.vtu`) is not included — go to the source dataset for volumetric surrogate modeling.

## Reproducing this dataset

The full preprocessing script is shipped alongside this repo (`process_drivaerml.py`). To regenerate from scratch:

```bash
pip install pyvista numpy scipy tqdm huggingface_hub
python process_drivaerml.py --output_dir ./drivaerml_processed --start 1 --end 500
```

The script streams one run at a time — downloading the VTP + STL + force/moment CSV from `neashton/drivaerml`, producing the `.npz`, then deleting the raw files — so it runs comfortably in a few tens of GB of free disk regardless of total dataset size.

## License

**CC-BY-SA 4.0**, inherited from the source dataset. If you use this data you must give appropriate credit, indicate any changes, and distribute any derivative works under the same license.

## Citation

Please cite the original DrivAerML paper:

```bibtex
@article{ashton2024drivaer,
  title   = {DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics},
  author  = {Ashton, N. and Mockett, C. and Fuchs, M. and Fliessbach, L. and Hetmann, H.
             and Knacke, T. and Schonwald, N. and Skaperdas, V. and Fotiadis, G.
             and Walle, A. and Hupertz, B. and Maddix, D.},
  journal = {arXiv preprint arXiv:2408.11969},
  year    = {2024},
  url     = {https://arxiv.org/abs/2408.11969}
}
```

## Acknowledgments

All credit for the underlying CFD data goes to the DrivAerML team (Neil Ashton et al., AWS / UpstreamCFD / BETA-CAE / Siemens Energy / Ford). This repository only redistributes a preprocessed surface-point-cloud view of that work. For the full multi-terabyte dataset including volumetric fields, OpenFOAM meshes, slice images, and residual plots, see [`neashton/drivaerml`](https://huggingface.co/datasets/neashton/drivaerml).