Datasets:

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 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:

Random sampling: 100,000 cells per run, fixed seed seed = run_id + 42.
Cell centres: Coordinates are CFD cell centres — exact locations where OpenFOAM computed field values.
Coordinate normalisation: Local per-run, zero mean and unit std per axis. Stored in each file for unnormalisation.
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:

{
    '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

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

@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)