Physically Admissible Generative Flow Matching for Sparse Flow Reconstruction

Yuhao Wang, Yuanxi PengarXiv 2026 · arXiv

TL;DR

The authors introduce Feasible Subspace Consistency (FSC), a generative flow matching framework that ensures fluid flow reconstructions strictly obey physical laws (like incompressibility) throughout the entire generation process. By treating physical admissibility as an invariant of the generative dynamics rather than a post-processing step, FSC achieves state-of-the-art accuracy in sparse flow reconstruction while providing reliable, spatially aware uncertainty estimates.

What Problem Does This Paper Address?

Reconstructing a high-resolution fluid velocity field from a tiny handful of sensors (covering <1% of the domain) is a classic "ill-posed" inverse problem. There are infinitely many possible flow fields that could explain the same sparse measurements. To be useful for engineering, a reconstruction must satisfy two criteria:

  1. Physical Admissibility: It must strictly follow conservation laws (e.g., mass conservation/incompressibility) and boundary conditions (e.g., no-slip at walls).
  2. Uncertainty Quantification: It should provide a range of likely solutions, not just a single "best guess," especially in unobserved turbulent regions.

Existing deep learning models often fail here. "Soft" constraints (penalizing physics violations in the loss function) don't guarantee a valid solution. "Hard" constraints applied only at the final output (terminal projection) often lead to mode collapse—where the model loses the diversity of turbulent structures—or spectral distortion, where the energy balance across different scales is ruined.

Methods

The core innovation is Feasible Subspace Consistency (FSC). Instead of fixing the output, FSC ensures the entire "path" the model takes to generate a sample stays within the physically allowed space.

  • Stage 1: Path-Feasible Prior. The model learns a generative trajectory where every step is constrained to the tangent space of the physical manifold. It uses mass-orthonormal coordinates, which ensures that the "distance" the model moves in its internal latent space directly corresponds to physical kinetic energy. This prevents the spectral distortion seen in previous methods.
  • Stage 2: Amortized Posterior Transport. To handle sparse sensors, the authors "freeze" the physical prior and train a second network to map sensor data into the prior's latent space. Because the prior is already guaranteed to be physical, any sample drawn from this conditioned latent space is automatically physical "for free," without needing expensive optimization loops during inference.
Figure 1: Overview of FSC. Stage 1 learns a tangent-space flow and defines a frozen decoder whose image lies on Mh(c). Stage 2 maps sparse observations to the prior latent space; every decoded posterior sample inherits the physical constraints under an exact projector (and up to solver tolerance in numerical realization).

Datasets

The authors evaluated FSC on two distinct and challenging flow topologies:

  1. System 1: Periodic Kolmogorov Flow. A standard benchmark for 2D turbulence on a flat torus. This was used to test the model's ability to capture complex spectral energy distributions.
  2. System 2: Wall-Bounded Cylinder Wake. A more "real-world" scenario involving flow past a circular cylinder in a channel with rigid walls and internal obstacles. This tested the model's handling of complex boundary conditions and immersed solids.

Key Insights and Ideas

  • Physics as an Invariant: The most important "aha" moment is that physical laws should be an invariant of the generative dynamics. If you start with a physical state and only move in "physical" directions, you never have to worry about drifting off the manifold.
  • The Danger of Terminal Projection: The paper clearly demonstrates that simply "fixing" a non-physical result at the end is destructive. As shown in Figure 2, models without path-consistency drift orders of magnitude away from physical reality during generation, and forcing them back at the end collapses their diversity.
  • Latent Space Conditioning: By performing the "Bayesian" part of the problem (conditioning on data) entirely in the latent space of a frozen physical decoder, the authors decouple the learning of physics from the learning of sensor patterns. This makes the model much more robust to different sensor layouts.
Comparison of trajectory feasibility. FSC (blue) maintains low divergence throughout, while other methods drift significantly.

Key Results

  • Mode Coverage: FSC maintained a mode coverage of 0.790, whereas models using terminal projection collapsed to 0.000 (meaning they failed to represent the variety of the data).
  • Error Reduction: Compared to unconditioned priors, amortized FSC reduced global reconstruction error by 50% on the Kolmogorov flow and 46% on the cylinder wake.
  • Sensor Fidelity: Sensor residuals (how well the reconstruction matches the actual measurements) were reduced by 63% and 55% respectively.
  • Qualitative Success: As seen in Figure 6, the model accurately recovers the Kármán vortex street behind a cylinder and provides standard deviation (Std) maps that correctly identify the wake as the region of highest uncertainty.
Qualitative results showing ground truth (GT), mean reconstruction, and uncertainty (Std) for both systems.

Limitations and Future Work

  • Resolution Caps: On periodic domains, the use of a spectral cutoff limits the finest scales the model can resolve.
  • Iterative Solver Dependence: For complex geometries (System 2), the model relies on iterative projectors (like PCG). While highly accurate, these are "physical" only up to a specified numerical tolerance, not perfectly zero.
  • Observation Operators: While the model handles direct velocity sensors well, generalizing to entirely different sensor types (like pressure or temperature) currently requires a small amount of test-time refinement (APLR) to reach peak accuracy. Future work could focus on making this cross-modality transfer fully amortized.

Summary updated

Abstract

Reconstructing full fluid velocity fields from sparse, localized observations is a severely ill-posed inverse problem: most of the spatial domain remains unobserved, discrete conservation laws must hold strictly, and predictive uncertainty must be reliable. Existing generative models typically enforce physical constraints only at the terminal output or via soft penalties, leaving intermediate generative trajectories unconstrained. We introduce Feasible Subspace Consistency (FSC), a framework that preserves physical admissibility as an invariant of continuous generative dynamics: tangent-space matching confines the entire ODE trajectory to the discrete conservation manifold, while mass-orthonormal coordinates preserve the kinetic-energy metric. Sparse reconstruction then reduces to amortized posterior transport in the latent space of a frozen FSC prior, ensuring every posterior sample inherits incompressibility and boundary conditions without per-sample projection loops. We validate FSC on a periodic Kolmogorov flow, where exact spectral projection isolates path tangency and metric isometry, and a wall-bounded cylinder wake with immersed obstacles, realized via matrix-free iterative projection. FSC prevents the mode-coverage collapse induced by terminal projection (increasing mode coverage from $0.000$ to $0.790$) while eliminating high-wavenumber spectral distortion. In sparse reconstruction, amortized FSC reduces global relative error by $50\%$ and $46\%$, and sensor residuals by $63\%$ and $55\%$, relative to the unconditioned prior, while delivering calibrated, spatially heterogeneous uncertainty.
BibTeX
@misc{wang2026physically,
  author = {Yuhao Wang and Yuanxi Peng},
  title = {{Physically Admissible Generative Flow Matching for Sparse Flow Reconstruction}},
  year = {2026},
  eprint = {2611.22549},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22549},
}
Figure