MSEG-VCUQ: Multimodal SEGmentation with Enhanced Vision Foundation Models, Convolutional Neural Networks, and Uncertainty Quantification for High-Speed Video Phase Detection Data

1. Problem Statement

High-speed video (HSV) phase detection (PD) is critical for monitoring vapor, liquid, and microlayer phases in industrial boiling processes (e.g., nuclear reactors, electronics cooling). However, analyzing these images presents significant challenges:

• Complexity of PD Data: Unlike standard shadowgraphy used in two-phase flow (TPF) studies, PD images feature complex bubble structures, diverse liquid-vapor interactions, and intricate contact lines that are difficult to segment.
• Limitations of Existing Models: Traditional Convolutional Neural Networks (CNNs), such as U-Net, struggle with generalization across varying experimental conditions and modalities. Conversely, Vision Foundation Models (VFMs) like the Segment Anything Model (SAM) have not yet been effectively adapted for scientific PD tasks, often failing to capture the specific nuances of boiling dynamics.
• Lack of Reliability Metrics: Existing uncertainty quantification (UQ) methods lack pixel-level reliability for critical metrics like contact line density and dry area fraction, which are essential for modeling heat transfer.
• Data Scarcity: There is a notable absence of large-scale, open-source, multimodal experimental datasets tailored specifically for PD segmentation, hindering the development of robust, generalizable models.

2. Methodology

The authors propose MSEG-VCUQ, a hybrid framework designed to enhance segmentation accuracy, generalization, and reliability.

A. Hybrid Model Architecture (VideoSAM)

The core of the framework is VideoSAM, a model that integrates the strengths of CNNs and Transformers:

1. Initial Segmentation (U-Net): A fine-tuned U-Net CNN generates preliminary segmentation masks. This stage captures primary liquid-vapor boundaries and fine-grained features using transfer learning from biological cell imaging domains.
2. Refinement (SAM Transformer): The preliminary masks are passed to a transformer-based module (derived from the Segment Anything Model, SAM). This stage uses attention mechanisms to refine the masks, improving boundary precision and handling complex, overlapping bubble structures.
3. Training Strategy: The model utilizes a combined loss function (Binary Cross-Entropy + Dice Loss) to optimize both pixel classification and boundary overlap. The image and prompt encoders from SAM are frozen to leverage pre-trained features, while the mask decoder is fine-tuned for HSV data.

B. Uncertainty Quantification (UQ)

To ensure the reliability of boiling metrics, the framework incorporates a rigorous UQ pipeline:

• Discretization Error Analysis: The authors quantify errors introduced by representing continuous bubble shapes on discrete pixel grids.
• Metrics: They calculate Percentage Relative Error (PRE) and Mean Error (ME) for perimeter and area measurements.
• Simulation & Validation: Iterative simulations varying grid resolution and bubble radius are compared against theoretical values. The framework also analyzes the impact of morphological operations (erosion and dilation) on boundary accuracy.

C. Data Curation

The team curated and released the first large-scale, open-source multimodal HSV PD dataset.

• Scope: Includes data from four distinct fluids: Argon, Nitrogen, FC-72, and Water.
• Conditions: Covers both Saturated Pool Boiling (SPB) and Flow Boiling (FB) under varying heat fluxes and pressures.
• Annotation: Features high-resolution frame-mask pairs with semi-automated and expert-refined annotations for liquid, vapor, and microlayer phases.

3. Key Contributions

1. MSEG-VCUQ Framework: A novel hybrid architecture combining U-Net and SAM (VideoSAM) that outperforms standalone CNNs and VFMs in complex boiling scenarios.
2. Open-Source Multimodal Dataset: The release of a comprehensive dataset covering diverse fluids and boiling regimes, addressing the scarcity of representative PD data.
3. Integrated Uncertainty Quantification: A systematic UQ pipeline that provides pixel-level error estimates for critical metrics (contact line density, dry area fraction), moving beyond general-purpose UQ methods.
4. Transfer Learning Adaptation: Demonstrating the successful adaptation of biological imaging transfer learning techniques to fluid dynamics, enabling models to capture fine distinctions between phases.
5. First Application of VFMs to PD: Bridging the gap by fine-tuning SAM for scientific phase detection, proving its viability for high-speed video analysis.

4. Key Results

• Performance Superiority:
  • VideoSAM consistently outperformed baseline U-Net and standard SAM in complex modalities (Nitrogen, FC-72, Argon). For example, in the Nitrogen dataset, VideoSAM achieved an IoU of 0.8317 compared to SAM's 0.6702.
  • In simpler modalities (Water), standard U-Net performed slightly better (IoU 0.5619) than VideoSAM (IoU 0.1894), suggesting VideoSAM is optimized for complex, high-density structures.
• Metric Accuracy:
  • U-Net/VideoSAM segmentation showed significantly higher sensitivity to small bubbles and fine vapor structures compared to adaptive thresholding.
  • The models achieved deviations of <2% for dry area fraction and <8% for contact line density when compared to ground truth annotations from five domain experts.
• Uncertainty Insights:
  • Perimeter measurements were found to be highly sensitive to grid resolution, transitioning from overestimation (fine grids) to underestimation (coarse grids).
  • Dilation operations were shown to reduce perimeter PRE from 13.8% to 6.6%, significantly improving boundary representation.
  • LN2 (Liquid Nitrogen) exhibited the highest perimeter uncertainty (-8.4% PRE) due to pronounced boundary irregularities.

5. Significance and Impact

• Scientific Advancement: This work provides a scalable, reproducible solution for analyzing boiling dynamics, enabling more accurate quantification of heat transfer mechanisms (nucleation, dry-out, microlayer evaporation).
• Industrial Application: The framework supports the development of autonomous experiments and real-time monitoring systems for critical infrastructure like nuclear reactors and electronics cooling.
• Methodological Shift: By successfully integrating Vision Foundation Models with CNNs and rigorous UQ, the paper sets a new standard for scientific computer vision, moving beyond simple object detection to reliable, uncertainty-aware phase segmentation in complex fluid dynamics.
• Community Resource: The open-source dataset and tools lower the barrier to entry for researchers in multiphase flow, fostering further innovation in thermal management and phase detection research.