Physically Admissible Generative Flow Matching for Sparse Flow Reconstruction
FSC keeps generated fluid fields physically valid throughout reconstruction, with reported error reductions of 50% and 46% against an unconditioned prior.
The framework constrains the entire generative trajectory to obey discrete conservation laws while preserving the kinetic-energy metric. Reconstruction from sparse observations uses a frozen prior, so every posterior sample inherits incompressibility and boundary conditions without per-sample projection loops. Tests on periodic Kolmogorov flow and a cylinder wake increased mode coverage from 0.000 under terminal projection to 0.790 with FSC, while eliminating high-wavenumber spectral distortion. The authors also report lower sensor residuals and calibrated uncertainty that varies across the flow field.
ArXiv · AI/CL/LG's note
The framework constrains the entire generative trajectory to obey discrete conservation laws while preserving the kinetic-energy metric. Reconstruction from sparse observations uses a frozen prior, so every posterior sample inherits incompressibility and boundary conditions without per-sample projection loops. Tests on periodic Kolmogorov flow and a cylinder wake increased mode coverage from 0.000 under terminal projection to 0.790 with FSC, while eliminating high-wavenumber spectral distortion. The authors also report lower sensor residuals and calibrated uncertainty that varies across the flow field.
ArXiv · AI/CL/LG's note
score 4