train_idx list | val_idx list | test_idx list |
|---|---|---|
[
33,
2,
38,
99,
63,
104,
4,
3,
84,
94,
0,
50,
83,
24,
85,
76,
125,
25,
20,
117,
124,
93,
51,
103,
121,
86,
45,
18,
111,
28,
71,
37,
68,
127,
75,
27,
88,
115,
21,
79,
56,
73,
46,
87,
100,
7,
81,
5,
26,
78,
110,
89,
... | [
54,
49,
6,
102,
114,
11,
116,
41,
22,
19,
35,
64,
47,
113,
109,
12,
53,
107,
14,
13,
66,
36,
97,
65,
77,
8
] | [
900,
901,
902,
903,
904,
905,
906,
907,
908,
909,
910,
911,
912,
913,
914,
915,
916,
917,
918,
919,
920,
921,
922,
923,
924
] |
Rising Bubble Benchmark Dataset for Physics-Informed Machine Learning
Description
This dataset provides simulation data for the 2D rising bubble problem, a canonical benchmark in multi-phase computational fluid dynamics. It is intended for training and evaluating physics-informed machine learning (PIML) surrogate models on incompressible, isothermal, two-phase flow under gravity.
Associated Publication
Under review.
Problem Setup
Each trajectory simulates a single circular bubble (initial radius R = 0.25 m) rising in a denser surrounding fluid within a fixed 2D domain of [-0.5, 0.5] m x [0, 2] m, under constant gravity (0, -0.98 m/s²) and a fixed surface tension constant (1.96 N/m). The two fluids have fixed densities (1000 kg/m³ and 1 kg/m³) but their kinematic viscosities are varied across trajectories, spanning the ranges [0.001, 0.05] m²/s (phase 1) and [0.0005, 1] m²/s (phase 2). This viscosity variation is the primary axis of variability in the dataset and allows models to be tested across a wide range of flow regimes, from viscous to near-inviscid.
Data Generation
Ground truth trajectories were generated with the volume-of-fluid (VoF) solver interFoam in OpenFOAM, solving the continuity and Navier-Stokes equations for the coupled velocity, pressure, and phase-fraction fields, with surface tension modeled as a continuum surface force. Each simulation runs for 3 s of physical time. Volume fraction, dynamic pressure, and velocity fields were exported at 0.05 s intervals and spatially interpolated onto a uniform 128 x 256 mesh.
Dataset Composition
The dataset contains 100 training, 28 validation, and 25 test trajectories. Training and validation viscosities were sampled log-uniformly at random; test viscosities are spaced log-uniformly across the same range to form a systematic sweep, enabling controlled evaluation of model performance as a function of viscosity contrast.
Data Format
Each trajectory is stored as an individual HDF5 file, named according to the convention bubble_%03i.hdf5, where the three-digit index is the simulation ID used to define the train/validation/test split. Each HDF5 file contains the following keys:
| Key | Description |
|---|---|
snapshots |
Field data tensor (volume fraction, pressure, and velocity components) at each exported timestep |
time |
Array of physical time values (in seconds) corresponding to each snapshot |
xy |
Spatial coordinates of the interpolated 128 x 256 mesh |
channel_names |
Labels identifying each of the 4 field channels |
channel_units |
Physical units corresponding to each channel |
channel_max |
Per-channel maximum value for this trajectory |
channel_min |
Per-channel minimum value for this trajectory |
logvis |
Log-scaled kinematic viscosity values (ν1, ν2) for the trajectory, the varying simulation parameter |
mask |
Domain validity mask. As there is no immersed obstacle in the domain, this field is always 1 throughout this problem. |
Per-channel normalization constants (min/max used for scaling) are provided separately in normalization.json, and the train/validation/test split is provided in train_val_test.json.
- Downloads last month
- 90