Jrhoss/DrivaerML-PCTR

---
license: cc-by-sa-4.0
task_categories:
- other
tags:
- CFD
- aerodynamics
- point-cloud
- pytorch
- fluid-dynamics
- regression
size_categories:
- 100K<n<1M
splits:
train:
num_examples: 387
validation:
num_examples: 48
test:
num_examples: 49
---

# DrivaerML-PCTR

Processed version of [neashton/drivaerml](https://huggingface.co/datasets/neashton/drivaerml)
for use with the PCT-R (Point Cloud Transformer for Regression) model.

## Processing

Each `boundary_i.vtp` file (~500MB, ~8.8M cells) was processed as follows:

1. **Random sampling:** 100,000 cells per run, fixed seed `seed = run_id + 42`.
2. **Cell centres:** Coordinates are CFD cell centres — exact locations where OpenFOAM computed field values.
3. **Coordinate normalisation:** Local per-run, zero mean and unit std per axis. Stored in each file for unnormalisation.
4. **Target fields:** Raw, no normalisation. All 500 runs share identical CFD boundary conditions.

## Dataset Splits

The dataset is split **80/10/10** (train/val/test) using a random shuffle with `seed=42`.

| Split | Runs |
|------------|------|
| Train | 387 |
| Validation | 48 |
| Test | 49 |
| **Total** | **484** |

> **Missing runs:** `167, 211, 218, 221, 248, 282, 291, 295, 316, 325, 329, 364, 370, 376, 403, 473`

## File Structure

```
train/
run_i/
boundary_i.pt
val/
run_i/
boundary_i.pt
test/
run_i/
boundary_i.pt
```

Each `.pt` file contains:

```python
{
'coords': torch.float32, # [100000, 3] locally normalised x, y, z
'targets': torch.float32, # [100000, 4] raw CFD field values
'sample_idx': torch.int64, # [100000] indices into original VTP
'coords_mean': torch.float32, # [3] per-axis mean
'coords_std': torch.float32, # [3] per-axis std
'run_id': int,
}
```

## Target Field Order

| Index | Field | Units | Description |
|-------|------------------------|-----------|----------------------------------------------------------|
| 0 | `CpMeanTrim` | [-] | Time-averaged static pressure coefficient |
| 1 | `pMeanTrim` | [m²/s²] | Time-averaged kinematic pressure |
| 2 | `pPrime2MeanTrim` | [m⁴/s⁴] | Time-averaged square of pressure fluctuations |
| 3 | `wallShearStressMeanTrim` | [m²/s²] | Magnitude of time-averaged wall shear stress vector |

## Loading the Dataset

```python
from huggingface_hub import snapshot_download
import torch, os

# Download the full dataset
local_dir = snapshot_download(repo_id="Jrhoss/DrivaerML-PCTR", repo_type="dataset")

# Load a single run from the training split
data = torch.load(os.path.join(local_dir, "train", "run_1", "boundary_1.pt"))
print(data['coords'].shape) # [100000, 3]
print(data['targets'].shape) # [100000, 4]
```

## Citation

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

## License

CC-BY-SA 4.0 (inherited from original dataset)