Sparse flow reconstruction stops relying on endpoint projection: FSC spreads conservation constraints across the whole generative trajectory, restoring 100% mode coverage
Synopsis
Recovering a full flow field from a few localized sensors is a severely ill-posed inverse problem in which discrete conservation laws must hold strictly, yet existing generative models enforce physics only at the terminal output or through soft penalties, leaving intermediate trajectories unconstrained. Feasible Subspace Consistency (FSC) writes tangent-space matching into the continuous generative dynamics so the entire ODE trajectory stays on the discrete conservation manifold, while mass-orthonormal coordinates preserve the kinetic-energy metric. On periodic Kolmogorov flow, full-chain FSC keeps 100% mode coverage while reaching the lowest sliced Wasserstein distance, avoiding the mode-coverage collapse caused by endpoint projection.
Interpretation
Physical constraints in sparse flow reconstruction move from the terminal output onto the entire generative trajectory, so intermediate states also satisfy incompressibility and boundary conditions. Earlier approaches either used soft penalties to lower equation residuals or projected only at the endpoint, leaving intermediate generative trajectories unconstrained. On periodic Kolmogorov flow, endpoint projection and step retraction both push relative divergence very low but collapse mode coverage to a lower level; full-chain FSC keeps 100% mode coverage and attains the lowest sliced Wasserstein distance.
The method keeps the generative velocity field inside the constraint tangent space and preserves the kinetic-energy metric with mass-orthonormal coordinates. Tangent-space matching means each step's generative velocity is projected onto the null space of the constraint operator, so the trajectory cannot leave the feasible manifold. On a wall-bounded cylinder wake with an immersed cylinder, a matrix-free iterative projection replaces the analytic Fourier basis, and full-chain FSC still attains the lowest sliced Wasserstein distance and the highest energy coverage.
Sparse reconstruction is recast as amortized posterior transport in the latent space of a frozen prior, so every posterior sample inherits incompressibility and boundary conditions by construction. Per-sample test-time projection loops are no longer needed, and the conditional model does not have to be retrained for each sensor layout. Relative to the unconditioned prior, amortized FSC lowers global relative error and sensor residuals on both the periodic and the wall-bounded system, while delivering calibrated, spatially heterogeneous uncertainty that retains spread in sensor blind spots.
Perspective
The demonstrated scope is limited to two benchmarks: two-dimensional periodic Kolmogorov flow and a wall-bounded cylinder wake with an immersed cylinder. Reconstruction uses uniformly distributed sparse sensors measuring both velocity components with additive Gaussian noise. On periodic domains the feasible subspace is band-limited by a spectral cutoff, while wall-bounded domains rely on iterative projection satisfying constraints to solver tolerance.
The spectral cutoff on periodic domains caps the finest resolvable scales, and wall-bounded geometries satisfy constraints only to iterative solver tolerance rather than by exact projection. Generalizing across unseen observation operators such as velocity magnitude or vorticity still requires test-time refinement, since the amortized network is trained only for direct velocity sensors.
