BubbleID-Flow

Fine-tuned one-class Mask R-CNN checkpoint for segmenting near-wall vapor regions in subcooled flow-boiling high-speed imagery.

Companion to the dataset at UARK-NED3/BubbleID-Flow (DOI 10.5281/zenodo.22235802) and the code at github.com/UARK-NED3/BubbleID-Flow.

What it measures, and what it does not

In subcooled flow boiling, vapor generated at the heated wall governs how much heat the surface can remove before the wall dries out. This checkpoint segments vapor regions in the image plane so that a two-dimensional near-wall vapor area fraction can be computed from the union of predicted masks.

That quantity is not volumetric void fraction, and the checkpoint does not produce one. Spatial calibration and registration between the imaging region of interest, the heated surface, and the temperature sensors were not available for this release, so the outputs do not support calibrated bubble-size measurement, local optical-thermal registration, or void-fraction calculation.

Checkpoint

File	model_final.pth
Architecture	Mask R-CNN, Detectron2 0.6
Classes	One foreground class (vapor region)
SHA-256	6f60969ce876f57a78b53fc61895c99b92cb2e1b6f0e0b30f01fa301236042cd

The LabelMe source annotations carry two labels, bubble and bubble_cluster. Both are merged into a single vapor-region class for union-mask analysis.

Operating settings

The reported analysis uses a 0.30 score threshold and a 300-detection cap per image, over a fixed near-wall region of interest beginning at pixel (0, 485) with size 1024 x 70. Standard COCO AP evaluation instead uses a 0.05 score threshold — the two settings answer different questions, so state which one a reported number came from.

Training data

Fine-tuned on 130 manually annotated images from a FC-72 subcooled flow-boiling facility. The flow passage was 2.5 mm wide and 5.0 mm high with the bottom wall electrically heated over 114.6 mm. Images were acquired with a FASTCAM SA-Z at 1024 x 1024 pixels and 3000 frames/s.

The published evaluation uses a deterministic same-sequence split: every fifth sorted image goes to holdout, giving 104 training and 26 holdout images.

Limitations

• The holdout is not an independent generalization test. Training and holdout frames come from the same sequences, so they are adjacent in time and visually similar. Performance on it indicates reconstruction quality within these sequences, not transfer to a new facility, fluid, or operating path.
• Exact retraining is not reproducible from this release. No frozen file-level split manifest, training seed, augmentation record, or checkpoint-selection history is included. The release supports reanalysis and new split construction; it does not reconstruct the reported 104/26 evaluation exactly.
• Out-of-domain behavior is documented as weak. In an independent off-the-shelf cross-domain screening on HFE-7100 pool-boiling frames (BoilingBench-CV), this checkpoint scored pixel IoU 0.0585 with recall 0.0590 against pixel precision 0.8631 — conservative, with substantial undercounting. It was applied there outside the domain it was fitted for, and no adaptation was performed. Expect to fine-tune before using it on pool boiling or on another facility.

Environment

Python 3.10.19, PyTorch 2.9.1 (CPU), Detectron2 0.6, OpenCV 4.11.0, NumPy 1.23.0, pandas 2.3.3. The dataset's processing/environment.yml is a dependency specification, not a platform-independent lockfile.

Usage

from huggingface_hub import hf_hub_download
from detectron2.config import get_cfg
from detectron2.engine import DefaultPredictor
from detectron2 import model_zoo

weights = hf_hub_download("UARK-NED3/BubbleID-Flow", "model_final.pth")

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file(
    "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1
cfg.MODEL.WEIGHTS = weights
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.30   # manuscript setting
cfg.TEST.DETECTIONS_PER_IMAGE = 300
cfg.MODEL.DEVICE = "cpu"                        # set "cuda" if available

predictor = DefaultPredictor(cfg)
# crop to the near-wall ROI before inference: image[485:555, 0:1024]

Training and evaluation scripts live in the dataset repository under processing/scripts/ and in the GitHub repository.

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 Zenodo creator list spells the seventh author "Karangate, Chirag R" and the deposit's CITATION.cff omits his middle initial. Kharangate, Chirag R (Case Western Reserve University) is the correct form.

The checkpoint accompanies the manuscript BubbleID-Flow: Machine-Vision Quantification of Vapor Area Fraction in Subcooled Flow Boiling.

Funding

U.S. National Science Foundation, Grant No. CBET-2323022.

Contact

Han Hu, Associate Professor of Mechanical Engineering, University of Arkansas — hanhu@uark.edu