BubbleID-Flow Multimodal Flow-Boiling Dataset

DOI: 10.5281/zenodo.22235802

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Model: the fine-tuned checkpoint is also published on its own at UARK-NED3/BubbleID-Flow, and remains included here under ModelWeights/ so this mirror stays complete against manifest.csv.

Mirror: this repository mirrors the Zenodo deposit of 1 September 2026. The file set was verified identical at upload time: 4,456 files, 24 GB. Every data file is byte-identical to the deposit, so the SHA-256 values in manifest.csv verify as published. The one exception is this file: Hugging Face requires README.md to be the dataset card, so it carries added front matter and the notes on this page. Its content is otherwise the deposit's README. For a byte-exact README.md matching the manifest.csv checksum, use the Zenodo deposit.

Dataset Viewer: the viewer shows manifest.csv, the 4,455-row release index (path, case, file class, byte size, SHA-256). It is an index, not the measurement data.

Release controls: manifest.csv, data_dictionary.json, timing_reconciliation.csv, event_definition.md, calibration/, and processing/.

Overview

This dataset supports the manuscript BubbleID-Flow: Machine-Vision Quantification of Vapor Area Fraction in Subcooled Flow Boiling. It contains high-speed optical images, manually annotated images, thermal workbooks, acoustic-emission (AE) records, and the fine-tuned Mask R-CNN checkpoint used in the reported analysis.

The experiments used FC-72 in a rectangular flow passage that was 2.5 mm wide and 5.0 mm high. The bottom wall was electrically heated over 114.6 mm. High-speed images were acquired with a FASTCAM SA-Z at 1024 x 1024 pixels and 3000 frames/s. The manuscript reports a two-dimensional, near-wall vapor area fraction measured from the image-plane union of predicted vapor masks. It is not a volumetric vapor fraction or void fraction.

Contents

AnnotatedImages/
  130 paired BMP images and LabelMe JSON annotation files.

ModelWeights/
  model_final.pth                 Fine-tuned one-class Mask R-CNN checkpoint.

MultimodalData/
  Images/                         Raw high-speed images and camera metadata.
  FlowLoopThermalData/            Thermal workbooks and derived thermal plots.
  AcousticDatta/                  AE event, continuous, waveform, and plot files.

Videos/
  BubbleID-Flow_45V_raw_vs_segmented.mp4
                                  Derived 16-second presentation video.
  BubbleID-Flow_45V_raw_vs_segmented.json
  Video frame-selection and inference manifest.
  BubbleID-Flow_45V_raw_top_segmented_bottom.mp4
                                  Derived 16-second presentation video.
  BubbleID-Flow_45V_raw_top_segmented_bottom.json
                                  Video frame-selection and inference manifest.

AnnotatedImages contains 130 valid LabelMe JSON files paired with 130 BMP images. The annotation labels are bubble and bubble_cluster; the BubbleID-Flow model merges both labels into one vapor-region class for union-mask vapor-area-fraction analysis.

MultimodalData/Images contains 3,977 BMP images and 37 .cihx camera metadata files. MultimodalData/FlowLoopThermalData contains three Excel workbooks per case: <case>.xlsx, <case>_IN_with_units.xlsx, and <case>_u_with_units.xlsx. The PNG files under Plot Images and the AE Plot_Hit and Plot_Time folders are derived visualizations, not primary measurement inputs.

Videos/BubbleID-Flow_45V_raw_vs_segmented.mp4 is a derived presentation output. It shows 20 uniformly sampled frames from the 45 V state of each case at 5 frames/s (16 seconds total), with a raw near-wall view and the corresponding Mask R-CNN union-mask overlay. Its JSON sidecar records the selected filenames, checkpoint checksum, ROI, score threshold, and detection cap.

Case Map

Canonical case ID	Nominal mass flow	Nominal inlet subcooling	Optical image folder	Thermal folder	AE folder
5gs_22C	5 g/s	22 degC	MultimodalData/Images/5gs_22C	MultimodalData/FlowLoopThermalData/5gs_22C	MultimodalData/AcousticDatta/5gs_22C
10gs_22C	10 g/s	22 degC	MultimodalData/Images/10gs_22C	MultimodalData/FlowLoopThermalData/10gs_22C	MultimodalData/AcousticDatta/10gs_22C
15gs_20C	15 g/s	20 degC	MultimodalData/Images/15gs_20C	MultimodalData/FlowLoopThermalData/15gs_20C	MultimodalData/AcousticDatta/15gs_22C (archive alias)
25gs_20C	25 g/s	20 degC	MultimodalData/Images/25gs_20C	MultimodalData/FlowLoopThermalData/25gs_20C	MultimodalData/AcousticDatta/25gs_22C (archive alias)

Archive naming notes

• AcousticDatta is the current directory name in this archive.
• The AE directories 15gs_22C and 25gs_22C are aliases only. Their contained HIT, TIME, DTA, and WFS filenames identify the canonical 15gs_20C and 25gs_20C cases. Use the canonical IDs in analyses and citations.
• Optical folder labels such as 25_ONSET, 55 CHF, and 59CHF are source acquisition labels. They are not independently validated onset or critical-heat-flux labels.

Modality Details

Optical images

The image folders are organized as <case>/<electrical-input folder>. The study uses a fixed near-wall region of interest (ROI) beginning at pixel coordinate (0, 485) with size 1024 x 70 pixels. Spatial calibration and registration between this ROI, the heated surface, and the seven temperature locations were not available. Therefore, this archive does not support calibrated bubble-size measurements, local optical-thermal registration, or volumetric void-fraction calculation.

Thermal data

The thermal workbooks contain time, mass flow, voltage, electrical power, inlet subcooling, inlet/outlet temperature, pressure, and seven test-section surface-temperature streams. The manuscript derives heat flux from electrical power and the documented heated area and derives mass flux from the measured flow area. Reported standard deviations characterize temporal variability within selected windows; they are not instrument uncertainty estimates.

For each canonical case, the BubbleID-Flow manuscript pipeline uses both:

• <case>.xlsx for thermal variables; and
• <case>_IN_with_units.xlsx to obtain voltage and define/match state windows.

Acoustic-emission data

Each AE case directory contains:

• HIT_<case>.TXT: event-level hit data;
• TIME_<case>.TXT: continuous time-domain summary data;
• Raw Data/<case>.wfs: raw waveform file;
• Raw Data/<case>.DTA: acquisition/export file;
• notes.txt: available operator notes; and
• Plot_Hit and Plot_Time: derived diagnostic figures.

The optical, thermal, and AE systems do not have a verified common hardware trigger. The manuscript's supplementary 10 g/s AE screen uses a 25.395 s archive-start timestamp offset for descriptive state comparison only. These data do not establish optical-acoustic synchronization, causal coupling, source localization, flow-regime classification, or calibrated AE energy.

Annotations and model checkpoint

The model was fine-tuned on 130 manually annotated images. The published reconstruction evaluation uses a deterministic same-sequence split: every fifth sorted image is assigned to holdout, yielding 104 training images and 26 holdout images. This is not an experiment-held-out generalization test; the source-sequence proximity between training and holdout images must be considered when interpreting performance.

The included checkpoint is:

ModelWeights/model_final.pth
SHA-256: 6f60969ce876f57a78b53fc61895c99b92cb2e1b6f0e0b30f01fa301236042cd

The manuscript operating settings are a 0.30 score threshold and a 300-detection cap per image. Standard COCO AP evaluation uses a 0.05 score threshold. The training/inference code and dependency guidance are maintained in the BubbleID-Flow repository: https://github.com/UARK-NED3/BubbleID-Flow.

Reuse and Reproduction

The analysis code and dependency guidance are maintained in the BubbleID-Flow repository: https://github.com/UARK-NED3/BubbleID-Flow. The principal scripts are:

• scripts/convert_labelme_to_coco.py for LabelMe-to-COCO conversion and ROI cropping;
• scripts/train_detectron2.py for fine-tuning;
• scripts/evaluate_detectron2.py and scripts/analyze_segmentation_robustness.py for segmentation evaluation;
• scripts/audit_thermal_reduction.py for the thermal state reduction; and
• scripts/analyze_optical_thermal_association.py for the state-level optical-thermal screening analysis.

This archive does not include a frozen file-level split manifest, the original training trace, random seed, augmentation record, or checkpoint-selection history. Consequently, it supports use of the released data and checkpoint, but does not independently establish an exact reconstruction of the reported 104-image training and 26-image holdout evaluation.

Corrections to the deposit metadata

Two author records in the Zenodo deposit and CITATION.cff are being corrected. The authoritative forms are:

• Kharangate, Chirag R (Case Western Reserve University). The Zenodo creator list spells this "Karangate, Chirag R", and CITATION.cff omits the middle initial. Cite the name as given here.
• Curl, Daniel is at the University of Arkansas, Fayetteville only. The Zenodo record's affiliation for him is incorrect.

The archived files are mirrored unmodified so that manifest.csv checksums continue to verify, so CITATION.cff in this repository still carries the uncorrected spelling. It is scheduled for correction in a future release.

Dataset Citation

Fahim, Md Abrar Hoq, Farshad Barghi Golezani, Daniel Curl, Mohammad Ishraq Hossain, Stephen Pierson, Sruthi S. Pydimarry, Chirag R. Kharangate, and Han Hu. 2026. BubbleID-Flow Multimodal Flow-Boiling Dataset. Zenodo. https://doi.org/10.5281/zenodo.22235802.

The machine-readable citation is available in CITATION.cff.

License

This dataset is distributed under the Creative Commons Attribution 4.0 International License. Users must provide appropriate attribution, link to the license, and indicate whether changes were made.

Funding

This study was supported by the U.S. National Science Foundation (NSF) under Grant No. CBET-2323022.

Release Records

• manifest.csv records the release payload path, canonical case identifier, file class, byte size, SHA-256 checksum, and release version. It excludes itself because a file cannot contain a stable checksum of its own final bytes.
• data_dictionary.json defines the released file classes and fields, units, dtypes, shapes, and missing-value rules. Vendor-defined AE fields and binary waveform formats are explicitly identified where no public unit/schema record is available.
• timing_reconciliation.csv records modality-specific time bases and the absence of verified common-trigger synchronization.
• event_definition.md distinguishes retained source folder labels from validated ONB, CHF, maximum-temperature, and NBR events.
• calibration/README.md records the authorized calibration and uncertainty evidence included in this release and the missing records.
• processing/ contains the release processing snapshot and environment specification.

Limitations and Appropriate Use

• The image-derived metric is two-dimensional vapor area fraction in a fixed camera-plane ROI, not volumetric void fraction.
• The manual holdout is same-sequence, not independent by experiment, facility, or operating path.
• Thermal calibration records, heat-loss characterization, sensor placement corrections, and full uncertainty information were not available.
• The archive contains raw AE waveforms and event records, but no verified cross-modality synchronization or AE sensor calibration.
• The folder labels with onset/CHF terminology are retained for traceability only and must not be used as validated physical transition labels.

Contact

Han Hu, Associate Professor of Mechanical Engineering, University of Arkansas

Email: hanhu@uark.edu