ibm-research/hemolab-bench

Hemolab Bench

A dataset of parameterized 3-D CFD surface meshes (UNSTRUCTURED_GRID, triangulated) with per-vertex pressure and wall shear stress (WSS) fields.

Each sample corresponds to a distinct geometry generated from a 17-parameter family (curvature, twist, taper, Reynolds number, Taylor number). The dataset contains 49,660 samples, all sharing the same triangulation topology (9,600 triangles, 25,600 vertices).

Train / Validation / Test split

The dataset ships with pre-built splits so experiments are reproducible without re-implementing the shuffle logic.

Split	Fraction	Approx. samples
train	80 %	~40 000
validation	10 %	~5 000
test	10 %	~5 000

Splits are assigned by a deterministic shuffle of all CSV ids:

import torch

all_ids = sorted(full_csv_ids)          # all ids, sorted ascending
g = torch.Generator().manual_seed(42)
perm = torch.randperm(len(all_ids), generator=g).tolist()
shuffled = [all_ids[i] for i in perm]

n = len(shuffled)
train_ids      = shuffled[:int(n * 0.8)]
validation_ids = shuffled[int(n * 0.8):int(n * 0.9)]
test_ids       = shuffled[int(n * 0.9):]

Test-set band labels

Two categorical columns (alpha_split, reynold_split) are populated only for test-split rows (they are null in train and validation). They indicate which distribution region each test sample belongs to, enabling per-regime evaluation.

alpha_split

Bands are symmetric around zero (checked on |alpha|):

| Label | |alpha| range | |-------|--------------| | IR2 | [0.000, 0.013] | | IR1 | (0.013, 0.040) | | ID | [0.040, 0.120) | | OD1 | [0.120, 0.160) | | OD2 | [0.160, 0.200] |

reynold_split

Label	Reynolds range
OD	[100, 150) ∪ (450, 500]
ID	[150, 250) ∪ [350, 450]
IR	[250, 350)

Usage

from datasets import load_dataset

# Full dataset
ds = load_dataset("ibm-research/hemolab-bench")
ds_train = ds["train"]
ds_val   = ds["validation"]
ds_test  = ds["test"]

print(f"{len(ds_train)} train  |  {len(ds_val)} val  |  {len(ds_test)} test")

# Single split
ds_train = load_dataset("ibm-research/hemolab-bench", split="train")

Access a sample:

sample = ds_train[0]

import numpy as np

points   = np.array(sample["points"]).reshape(25600, 3)   # (N, 3) xyz
pressure = np.array(sample["pressure"])                    # (N,)
wss      = np.array(sample["wss"]).reshape(25600, 3)       # (N, 3) vector

print(sample["reynolds"], sample["alpha"])

Access band labels (test split only):

sample = ds_test[0]
print(sample["alpha_split"], sample["reynold_split"])  # e.g. "ID", "IR"

The shared mesh topology (cell connectivity) is stored once in topology.parquet at the dataset root:

import pandas as pd

topology = pd.read_parquet(
    "hf://datasets/ibm-research/hemolab-bench/topology.parquet"
)
cells = topology["cells"].to_numpy().reshape(-1, 4)  # (9600, 4) quad indices

Dataset schema

Column	Type	Shape	Description
id	int32	—	Sample identifier (0-based)
points	float32	76800 (= 25600×3, flattened)	Vertex coordinates
pressure	float32	25600	Per-vertex pressure
wss	float32	76800 (= 25600×3, flattened)	Per-vertex wall shear stress vector
alpha	float64	—	Geometry parameter
gammaY0..gammaY3	float64	—	Y-curvature parameters
gammaZ0..gammaZ3	float64	—	Z-curvature parameters
beta0..beta3	float64	—	Taper parameters
noise	float64	—	Geometry noise level
iseed	int64	—	Random seed used for geometry generation
reynolds	float64	—	Reynolds number
taylor	float64	—	Taylor number
alpha_split	string	—	Alpha band label (test rows only; null elsewhere)
reynold_split	string	—	Reynolds band label (test rows only; null elsewhere)

Citation

TBD — paper/preprint forthcoming.

License

CDLA Permissive 2.0