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.
Figures & tables
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).
Configuration
Mode cov. ↑
SWD ↓
δdiv↓
Vanilla
0.790
0.362
3.9×10−1
+ soft penalty
0.799
0.369
9.8×10−3
+ endpoint projection
0.000
0.116
6.9×10−6
+ step retraction
0.317
0.438
7.6×10−6
FSC full-chain
0.790
0.073
7.2×10−6
Table 1: Constraint placement on System 1. Vanilla is an unconstrained velocity-space flow-matching model; the + rows add, respectively, endpoint projection of the terminal output, retraction onto the band-limited subspace ( ∣k∣≤13 ) after each ODE step, and a soft divergence penalty λ∥∇⋅v∥2/∥v∥2 . FSC full-chain enforces feasibility consistently across parameterization, training path, vector field, ODE state, and decoder.
Configuration
Mode cov. ↑
k13 ( →1.0 )
OOB ↓
δdiv↓
Unbounded
0.083
0.887
0.0351
7.0×10−6
Band-limited ψ
0.735
13.719
0.0000
7.3×10−6
Band-limited mass
0.790
1.301
0.0000
7.2×10−6
Table 2: Subspace geometry and metric alignment on System 1. Unbounded retains full-chain feasibility without a hard spectral cutoff; band-limited ψ and band-limited mass both enforce full-chain feasibility on ∣k∣≤13 , differing solely in the coordinate metric.
Configuration
E-cov ↑
SWD ↓
δdiv↓
Vanilla FM
0.738
0.1356
1.26×10−1
+ soft penalty
0.738
0.1464
2.75×10−2
+ endpoint projection
0.740
0.1354
1.55×10−6
+ step retraction
0.739
0.1360
7.05×10−7
FSC full-chain
0.746
0.1190
6.85×10−7
Table 3: Constraint placement on System 2 (wall-bounded cylinder wake), with rows aligned to Table 1 .
Figure 2: Inference-time trajectory ablations on System 2 ( 768 samples, 20 -step ODE solver, flux-corrected prior; FSC curves share the same checkpoint). (a) Fluid-cell relative divergence δdiv along integration steps. (b) Relative projection correction ∥Phδ(u)−u∥2/∥u∥2 required upon early exit at steps 0,5,10,15 , and 20 .
System 1 (Periodic Kolmogorov Flow)
System 2 (Wall-Bounded Cylinder Wake)
Method
e2↓
ey↓
δdiv↓
e2↓
ey↓
δdiv↓
Unconditioned prior
0.9871
1.0001
3.66×10−6
0.3165
0.2252
8.33×10−7
Supervised regression
0.2901
0.1127
5.06×10−1
0.0887
0.0323
1.53×10−1
EnKF+proj
0.4601
0.2071
5.14×10−6
0.1638
0.0411
1.18×10−6
FSC posterior (amortized)
0.4968
0.3730
4.30×10−6
0.1702
0.1020
1.72×10−6
FSC + APLR (refinement)
0.2880
0.0850
6.10×10−6
0.1574
0.0137
3.21×10−6
Table 4: Sparse posterior reconstruction on the test-in split across both physical benchmarks ( S=64 uniform sensors, K=32 ensemble members, σϵ=0.01 ). Unconditioned prior denotes the frozen Stage 1 generator without sensor conditioning; FSC + APLR adds 20 latent gradient refinement steps. Boldface marks the best e2 and ey among physically feasible methods; divergence entries are not ranked, as every feasible method sits at the solver floor.
Figure 3: Qualitative sparse reconstruction with FSC + APLR on held-out test cases ( S=64 sensors, K=32 posterior samples, σϵ=0.01 ). Left: System 1 (Kolmogorov flow) at Re=220 and 320 ( 1282 grid). Right: System 2 (cylinder wake) at Re=100 and 120 ( 256×128 grid). Rows: ground truth, posterior ensemble mean, and posterior standard deviation s(x) ; columns: velocity components u and v , with sensor locations overlaid on the GT and mean panels. Within each system, all velocity panels share one symmetric color scale and all std panels share one scale; GT and mean use identical vmin/vmax .
System 1
System 2
Method
cov90↑
CRPS ↓
cov90↑
CRPS ↓
Unconditioned prior
0.827
0.105
0.776
0.131
Supervised regression
–
–
–
–
EnKF+proj
0.900
0.046
0.838
0.065
FSC posterior
0.875
0.051
0.921
0.073
FSC + APLR (refinement)
0.871
0.028
0.953
0.068
Table 5: Uncertainty calibration on the test-in split. Supervised regression is deterministic, so cov90 and CRPS are undefined.
Figure 4: Uncertainty diagnostics. (a) Ratio of posterior to prior ensemble standard deviation on System 1 for two test-in cases ( Re=130 and 150 ), split into observed and sensor-blind regions. (b) Empirical 90% coverage ( cov90 ) on the test-in split for both systems; the dashed line marks the nominal 0.900 target.
Operator
Method
e2↓
ey↓
cov90↑
CRPS ↓
∥u∥2 (nonlinear)
Unconditioned prior
0.899
0.624
0.895
0.093
Supervised regression
0.771
0.635
–
–
EnKF+proj
0.872
0.662
0.966
0.093
FSC + APLR
0.837
0.126
0.966
0.084
ω (differential)
Unconditioned prior
0.989
4.136
0.858
0.102
Supervised regression
0.591
0.918
–
–
Table 6: Indirect and nonlinear observation operators on System 1 (interpolation split). For velocity magnitude ∥u∥2 , e2 , cov90 , and CRPS are evaluated after resolving global sign ambiguity. Boldface marks the best value per column, including the deterministic regression.
Department of Physics University of Cambridge · The Center for Computational Science University College London · Department of Computer Science University College London