README.md

---
license: cc0-1.0
language:
- en
task_categories:
- image-segmentation
- object-detection
tags:
- pool-boiling
- two-phase-flow
- thermal-management
- bubble-morphology
- unsupervised-learning
- principal-component-analysis
- high-speed-imaging
- scientific-imaging
size_categories:
- n<1K
pretty_name: Pool Boiling High-Speed Video Datasets and Analysis Toolkit
viewer: false
---

# Pool Boiling High-Speed Video Datasets and Analysis Toolkit

Four high-speed video datasets of pool boiling across two working fluids and
four heater surfaces, the manually annotated bubble-contour subsets derived
from them, and the MATLAB toolkit that extracts both conventional and
PCA-derived physical descriptors.

**Dataset DOI:** [10.5061/dryad.kh18932mw](https://doi.org/10.5061/dryad.kh18932mw)
· **Primary article:** *International Journal of Heat and Mass Transfer* **255**, 127894,
[10.1016/j.ijheatmasstransfer.2025.127894](https://doi.org/10.1016/j.ijheatmasstransfer.2025.127894)
· **License:** CC0 1.0 (public domain dedication)

Mirror of the Dryad deposit published 30 October 2025.

## The measurement problem

Boiling is hard to quantify from images. Bubbles overlap, merge, and leave the
field of view, so the conventional descriptors — bubble count, mean bubble
area and radius, vapor area fraction — are recovered by hand-annotating
contours frame by frame, which does not scale to the thousands of frames a
high-speed camera produces in a second.

The associated study asks whether an unsupervised decomposition can stand in
for that labor. Principal component analysis over image sequences yields two
descriptors — **Dominant Amplitude (Dₐ)** and **Dominant Frequency (D𝑓)** —
taken from the FFT of the leading PC's time series. The paper reports strong
positive correlations between these and the manual measurements: Dₐ with
bubble size and vapor area fraction, D𝑓 with bubble count.

**There is no pretrained model in this deposit, and none is published
alongside it.** The PCA is re-fit at run time, per dataset and per heat load,
and the scripts write PC scores to CSV rather than persisting a basis. The
method lives in the toolkit below, not in a set of weights.

## Datasets

Two boiling regimes appear across the videos: steady-state nucleate boiling at
lower heat loads, and the transient excursion to film boiling at the highest.
Videos were captured with Phantom high-speed cameras and are distributed as
downsized `.mp4` converted from the camera's native format with Phantom Camera
Control (PCC).

| Dataset | Surface | Fluid | Videos | Heat loads | Annotated images | Size |
| --- | --- | --- | ---: | --- | ---: | ---: |
| `FCu-H2O` | Copper foam | Deionized water | 7 | 15, 30, 60, 100, 150, 200, 220 W | 91 | 11 GB |
| `PCu-H2O` | Plain copper | Deionized water | 7 | 10, 40, 60, 90, 110, 120, 130 W | 91 | 2.3 GB |
| `PSi-HFE` | Plain silicon | HFE-7100 | 9 | 6, 8, 10, 12, 14, 16, 17.6, 20, 22 W | 119 | 3.5 GB |
| `SSi-HFE` | Structured silicon | HFE-7100 | 8 | 6, 12, 18, 24, 30, 36, 42, 44 W | 56 | 1.5 GB |

Provenance differs by fluid, and the difference matters when interpreting
cross-dataset comparisons:

- **Cu-H₂O** (`PCu-H2O`, `FCu-H2O`) — Nano Energy and Data-Driven Discovery
  Laboratory, University of Arkansas (Prof. Han Hu).
- **Si-HFE** (`PSi-HFE`, `SSi-HFE`) — Cooling Technologies Research Center,
  Purdue University (Prof. Justin Weibel).

Fluid, surface, facility, and optical arrangement all vary together between
the water and HFE-7100 groups. Treat a water-versus-HFE difference as a
compound effect, not a fluid-only one.

## Layout

```text
DataProcessTools/
  Conventional_descriptors/   bubbles_annotator_v6.m, bubbleStat_calculator_v4.m
  Generalized_descriptors/    PC_calculator_v4.m, dominantDescriptor_calculator_v4.m
<Dataset>/                    downsized .mp4 videos, one per heat load
<Dataset>/annotatedBubbles/   megapixel .jpg frames + <Dataset>-BubbleContours.json
```

Images are named `<Dataset>_heatLoad_imageSequenceNum.jpg`. Each regime's
`BubbleContours.json` holds one object per image, carrying a `FileName` and a
`Bubbles` object with the coordinates of each labeled bubble (`b0001`,
`b0002`, …).

> **Path note.** The deposit README documents the PCA scripts under
> `./Generalizable_descriptors/`. The directory is actually named
> `Generalized_descriptors`. Use the latter.

## Toolkit

MATLAB R2021a or newer, with the Image Processing Toolbox and the Statistics
and Machine Learning Toolbox.

### Conventional descriptors

`bubbles_annotator_v6.m` — a GUI for drawing and labeling bubble contours.
Prompts for a source image folder and a destination, then opens each image:
left-click places polygon vertices, Backspace undoes the last point,
double-click finalizes a bubble. Writes annotated images and a single
`bubble_labeled_data.json`.

`bubbleStat_calculator_v4.m` — reads that JSON and computes bubble count
(N_b), mean bubble area (A_b), mean bubble radius (R_b), and vapor area
fraction (VAF), grouped by heat load. **Set the pixel resolution `l_px` for
your dataset before running** — the physical statistics scale directly with
it. Optionally writes contour overlays to a `BubbleContours` subfolder.

### Generalizable descriptors

`PC_calculator_v4.m` — runs PCA over image frames in subsets. Configure
`MainData_DIR`, the `heatLoads` array, `totalImages`, `numSubsets`, and
`numPCs`; frames must be named sequentially (`00000.jpg`). Writes a
`./PC_Results` directory of CSV PC scores.

`dominantDescriptor_calculator_v4.m` — FFTs the time series of one principal
component (typically PC1) to extract Dₐ and D𝑓. Requires `PC_calculator_v4.m`
to have run first. Set `MainData_DIR`, `heatLoads`, `frameRate`, and
`pcToAnalyze`. Writes `./Dominant_Descriptors`, one CSV per heat load.

### Workflow

1. Extract frames from the videos into per-heat-load subfolders
   (e.g. `./FCu-H2O/15W/`).
2. Annotate a representative subset with `bubbles_annotator_v6.m`.
3. Run `bubbleStat_calculator_v4.m` for the conventional statistics.
4. Run `PC_calculator_v4.m`, then `dominantDescriptor_calculator_v4.m`.

Note that step 1 is a prerequisite the deposit does not ship: the videos are
distributed as `.mp4`, and the scripts consume extracted frame sequences.

## Relationship to BoilingBench-CV

The annotated subsets here also appear in
[UARK-NED3/BoilingBench-CV](https://huggingface.co/datasets/UARK-NED3/BoilingBench-CV),
which lists this deposit as one of its upstream sources. The overlap is
substantial and deliberate: all 357 annotated images, all 4 contour JSON
files, and 8 of the 9 `PSi-HFE` videos are common to both.

Which to use:

- **This repository** for the high-speed videos, the original contour JSON
  format, and the MATLAB descriptor toolkit.
- **BoilingBench-CV** for canonical COCO-style annotations, grouped benchmark
  splits with a documented leakage rule, and segmentation evaluation records.

The two carry different licenses — this deposit is CC0 1.0, BoilingBench-CV is
CC BY 4.0 — so check the terms for the copy you actually use.

## Loading

The Dataset Viewer is disabled: the payload is `.mp4` video and megapixel
`.jpg` frames with contours in a bespoke JSON schema, none of which the viewer
renders.

```python
from huggingface_hub import snapshot_download

# annotations and toolkit only, skipping ~18 GB of video
path = snapshot_download(
    "UARK-NED3/PoolBoiling-HighSpeedVideo", repo_type="dataset",
    allow_patterns=["*/annotatedBubbles/*", "DataProcessTools/*"],
)
```

```python
# one regime's videos
path = snapshot_download(
    "UARK-NED3/PoolBoiling-HighSpeedVideo", repo_type="dataset",
    allow_patterns=["SSi-HFE/*.mp4"],
)
```

## Limitations

- Videos are **downsized** conversions of the Phantom native recordings, not
  the original acquisitions. Frame rate and scale are not recorded in the file
  metadata; recover them from the paper or the acquisition records.
- `bubbleStat_calculator_v4.m` needs a per-dataset pixel resolution that the
  deposit does not tabulate. Conventional descriptors are not reproducible
  without it.
- Folder labels are acquisition heat loads, not validated onset or
  critical-heat-flux markers.
- The annotations are 2-D projected contours in image-pixel coordinates, not
  three-dimensional vapor measurements.

## Citation

Zhang, L., Soori, T., Bongarala, M., Li, C., Hu, H., Weibel, J. A., and
Sun, Y. 2026. Generalizable physical descriptors of pool boiling heat transfer
from unsupervised learning of images. *International Journal of Heat and Mass
Transfer* **255**, 127894.
https://doi.org/10.1016/j.ijheatmasstransfer.2025.127894

Data: Zhang, Lige; Soori, Tejaswi; Bongarala, Manohar; Li, Changgen; Hu, Han;
Weibel, Justin; Sun, Ying. *Data from: Generalizable physical descriptors of
pool boiling heat transfer from unsupervised learning of images.* Dryad.
https://doi.org/10.5061/dryad.kh18932mw

Affiliations: Drexel University (Zhang, Soori, Sun); Purdue University West
Lafayette (Bongarala, Weibel); University of Arkansas (Li, Hu); University of
Cincinnati (Sun).

## Funding

- U.S. National Science Foundation, CBET-2300317 and CBET-2323023 (Division of
  Chemical, Bioengineering, Environmental, and Transport Systems)
- Office of Naval Research, N000141613109
- U.S. National Science Foundation, 22-EPS4-0028, Arkansas EPSCoR DART
- U.S. National Science Foundation, OIA-1946391 (Office of Integrative
  Activities)

## Contact

Han Hu, Associate Professor of Mechanical Engineering, University of Arkansas
— hanhu@uark.edu
## Start here
**What this resource supports.** This release contains four high-speed pool-boiling video datasets, annotated two-dimensional bubble contours, and a MATLAB toolkit for conventional and PCA-derived image descriptors. It is intended for video-based bubble-morphology analysis under the released conditions.
**First five minutes.** Download annotations and DataProcessTools without the large video payload, then set the per-dataset pixel resolution before using bubbleStat_calculator_v4.m. To compute PCA-derived descriptors, first extract sequential video frames, run PC_calculator_v4.m, and then run dominantDescriptor_calculator_v4.m.
**Use with care.** The videos are downsized conversions rather than native camera recordings; frame rate and physical scale are not recorded in file metadata. The annotations are two-dimensional image-pixel contours, not three-dimensional vapor measurements. Comparisons between water and HFE-7100 combine fluid, surface, facility, and optical differences, so they must not be treated as fluid-only causal effects.
**Continue.** For a segmentation benchmark with canonical COCO annotations and grouped splits, use https://huggingface.co/datasets/UARK-NED3/BoilingBench-CV .
Dataset DOI: https://doi.org/10.5061/dryad.kh18932mw