[Paper] SurfPhase: 3D Interfacial Dynamics in Two-Phase Flows from Sparse Videos

Source: arXiv - 2602.11154v1

Overview

The paper SurfPhase tackles a long‑standing challenge in fluid dynamics: reconstructing the full 3D shape and motion of liquid‑vapor interfaces (e.g., boiling bubbles) from only a handful of video cameras. By marrying a dynamic surfel‑based geometry representation with a signed‑distance‑function (SDF) constraint and a video diffusion model for view synthesis, the authors can recover high‑fidelity 3‑D interfacial dynamics from as few as two sparse camera views. This opens the door to inexpensive, high‑speed measurement of two‑phase flows that were previously only accessible with costly, intrusive instrumentation.

Key Contributions

• SurfPhase pipeline – a unified framework that jointly optimizes a dynamic Gaussian‑surfel cloud, an SDF field, and a neural video diffusion model to reconstruct 3‑D interfacial geometry from sparse video inputs.
• Geometric consistency via SDF – the signed distance formulation enforces smoothness and prevents surfels from drifting apart, crucial for handling sharp, rapidly deforming liquid‑vapor boundaries.
• Diffusion‑based novel‑view synthesis – a pretrained video diffusion model generates high‑quality synthetic views, which are fed back into the reconstruction loop to close the gap caused by limited camera coverage.
• New high‑speed pool‑boiling dataset – over 1 k clips captured at >10 k fps with synchronized multi‑camera rigs, released for the community.
• Quantitative velocity estimation – the system can extract surface velocity fields directly from the reconstructed geometry, achieving errors <5 % compared to ground‑truth high‑speed laser profilometry.

Methodology

1. Dynamic Gaussian Surfel Representation

   • The liquid‑vapor interface is modeled as a cloud of Gaussian surfels (tiny oriented disks).
   • Each surfel carries position, normal, radius, and a learned feature vector that encodes appearance.

2. Signed Distance Function (SDF) Regularizer

   • An auxiliary SDF field is learned simultaneously.
   • The surfel positions are constrained to lie near the zero‑level set of the SDF, which enforces a coherent surface and prevents self‑intersection during rapid deformation.

3. Sparse Video Ingestion

   • Two (or more) calibrated cameras provide RGB frames at high frame‑rate.
   • For each frame, the current surfel cloud is rasterized into a differentiable rendering pipeline, producing synthetic views that can be compared to the real images.

4. Video Diffusion Model for View Synthesis

   • A pretrained diffusion model (e.g., Stable Diffusion‑Video) is conditioned on the observed views and the current surfel estimate to generate plausible novel‑view frames.
   • These synthetic frames act as “virtual cameras,” supplying extra photometric constraints that guide the surfel/SDF optimization.

5. Joint Optimization Loop

   • Losses: photometric reprojection error, SDF consistency, surfel smoothness, and diffusion‑based perceptual loss.
   • The system iterates over time, updating surfel positions and the SDF field frame‑by‑frame, yielding a temporally coherent 3‑D reconstruction.

Results & Findings

• High‑quality view synthesis: Even with only two cameras, the diffusion‑augmented pipeline produces realistic novel‑view videos that are visually indistinguishable from those generated with dense camera arrays.
• Accurate geometry: The surfel‑SDF combo captures sharp bubble edges and rapid topological changes (e.g., bubble coalescence) that classic neural radiance fields miss.
• Velocity extraction: By tracking surfel motion across frames, the authors recover surface velocity fields that match laser‑based measurements within a few percent.

Qualitative examples (available on the project site) show clean reconstructions of boiling bubbles, splashing droplets, and vapor pockets, all from just two synchronized high‑speed cameras.

Practical Implications

• Cost‑effective experimental setups – Labs can replace expensive multi‑camera rigs or intrusive laser scanners with a pair of off‑the‑shelf high‑speed cameras and run SurfPhase to obtain full 3‑D flow data.
• Real‑time monitoring – The pipeline runs at ~5 fps on a single RTX 4090, suggesting feasibility for near‑real‑time diagnostics in industrial boiling, spray cooling, or fuel‑injection systems.
• Data‑driven CFD validation – Engineers can feed the reconstructed surface meshes and velocity fields directly into CFD solvers for model validation or hybrid simulation‑data workflows.
• Cross‑domain applicability – The same surfel‑SDF + diffusion framework could be adapted to other sharp‑interface phenomena such as solid‑liquid melting, additive‑manufacturing melt pools, or even medical imaging of moving membranes.

Limitations & Future Work

• Dependence on high‑speed cameras – The current implementation assumes >10 k fps capture to resolve fast interface motion; slower cameras introduce motion blur that degrades reconstruction quality.
• Diffusion model generalization – The video diffusion model was fine‑tuned on boiling data; applying SurfPhase to drastically different fluids (e.g., opaque liquids) may require re‑training or domain adaptation.
• Scalability to larger volumes – The surfel cloud grows linearly with surface area; for very large reactors the memory footprint could become a bottleneck.
• Future directions highlighted by the authors include: integrating physics‑based priors (e.g., Navier‑Stokes constraints) into the optimization, extending the method to multi‑phase flows with more than two fluids, and exploring hardware‑accelerated implementations for true real‑time deployment.

Authors

• Yue Gao
• Hong‑Xing Yu
• Sanghyeon Chang
• Qianxi Fu
• Bo Zhu
• Yoonjin Won
• Juan Carlos Niebles
• Jiajun Wu

Paper Information

• arXiv ID: 2602.11154v1
• Categories: cs.CV
• Published: February 11, 2026
• PDF: Download PDF