$

$ teds read --post boilingbench-multimodal-fiftyone-analysis

BoilingBench-Multimodal: Scrub Boiling Video Against Temperature and Sound

BoilingBench-Multimodal syncs high-speed video, surface temperature and three acoustic sensors on one clock across 7 boiling and cooling runs.

Boiling changes character before you can see it change. BoilingBench-Multimodal heats a copper surface past the onset of boiling while a high-speed camera, a hydrophone, a microphone and an acoustic-emission sensor record it. Every stream sits on one timeline, so you can scrub the video against the temperature curve and the sound.

## What was recorded

The data comes from the NED³ laboratory at the University of Arkansas (BoilingBench-Multimodal), converted by Voxel51 to native multimodal MCAP episodes. The set has seven episodes and 1.72 hours of recording, holding 389,483 camera frames, 3,392,459 thermal samples and 171,970 acoustic-emission hits.

Episode Surface Condition Duration Capture
ambient-subcooled-flat-cu flat copper ambient subcooled 16m29s 97 fps
subatmospheric-saturated-flat-cu flat copper subatmospheric saturated 25m57s 62 fps
ambient-saturated-cu-foam copper foam ambient saturated 9m16s 149 fps
ambient-saturated-flat-cu flat copper ambient saturated 6m26s 153 fps
ambient-saturated-hydrophone flat copper ambient saturated 4m25s acoustic only
immersion-water-100w flat device water, 100 W 13m43s 18 fps IR
immersion-hfe7100-50w flat device HFE-7100, 50 W 27m01s 25 fps IR

The four pool-boiling episodes each carry the side-view video, surface temperature and heat flux, four embedded thermocouples, band-integrated acoustic power, acoustic frequency features, per-hit acoustic-emission records, and event markers. The markers are departure from nucleate boiling, the surface temperature peak, the critical-heat-flux marker, the return to nucleate boiling, and the DC power start and shutoff.

How hot each of the four pool-boiling runs got:

Episode Peak wall temperature Peak heat flux
subatmospheric-saturated-flat-cu 250.1 °C 43.3 W/cm²
ambient-saturated-cu-foam 236.9 °C 165.4 W/cm²
ambient-saturated-flat-cu 173.1 °C 88.6 W/cm²
ambient-subcooled-flat-cu 172.9 °C 278.0 W/cm²

Scrub video against temperature

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub(
    "Voxel51/BoilingBench-Multimodal",
    name="BoilingBench-Multimodal",
    persistent=True,
)

view = dataset.sort_by("max_surface_temp_C", reverse=True)
fo.launch_app(dataset, view=view)

That puts the run that reached the highest wall temperature first. Open it and watch the video while the temperature, acoustic power and event markers move beside it.

How the timeline was built

Every sensor stream is placed on the temperature acquisition clock using the offset table the release publishes. The acoustic-emission sensor opens about eight seconds before the temperature record and is placed there.

The camera is the exception. It carries no offset record, and its container duration is the run stretched for playback at 30 fps, so frames are placed by scaling container time onto the run. That placement is checked against the logged DC power events: on ambient-subcooled-flat-cu the onset of boiling brackets the power start at 5.2 s, and the decay begins where shutoff is logged at 673.9 s. Each episode stores the video_time_scale and video_capture_fps it was placed with.

Plot channels are thinned to 10 Hz, because the thermal acquisition runs at 3 kHz on two of the experiments and the full rate is not readable on a timeline. The release also ships a human-annotated still-image set and the raw acquisition files, including a 12.6 GB acoustic-emission waveform per experiment. Neither is in this conversion.

The critical-heat-flux figures are the release’s own screening markers, not independently validated measurements. Treat them as markers for finding the right frames, not as ground truth.

The source release is CC BY 4.0, and the conversion uses the same license. The data is at Voxel51/BoilingBench-Multimodal.

For another dataset that pairs a contact image with a measured signal, see Robots Can See. They Still Can’t Feel. For more of my FiftyOne dataset notes, see my Voxel51 author page, and to inspect and curate your own multimodal data in the same tool, try Voxel51.

Frequently asked questions

What does BoilingBench-Multimodal record?

A copper surface driven past the onset of boiling, watched by a high-speed camera from the side while a hydrophone, a microphone and an acoustic-emission sensor listen. It has seven episodes and 1.72 hours of recording, with 389,483 camera frames and 171,970 acoustic-emission hits.

Why put sound and video on the same timeline?

Boiling changes character before it changes appearance. Every modality carries a recorded clock offset against the temperature acquisition, so the sound, the surface temperature and the frames can be read against each other.

Are the critical-heat-flux numbers independent measurements?

No. The critical-heat-flux figures are the release's own screening markers rather than independently validated measurements. Each episode carries the source's chf_event_status alongside them.

Does the camera have its own timing record?

No. The camera carries no offset record, and its container duration is the run stretched for playback at 30 fps. Frames are placed by scaling container time onto the run, and each episode stores the video_time_scale and video_capture_fps it was placed with.

Which runs are not pool boiling?

Two closed-loop immersion-cooling runs, recorded in infrared with HFE-7100 and water over a flat device. A hydrophone reference run is also included, and it ships no video.