FLUIDSPLAT: Flow Field Reconstruction via Gaussian Primitives
A new method called FLUIDSPLAT reconstructs continuous flow fields from sparse surface-mounted sensors. Inspired by 3D Gaussian Splatting, it predicts anisotropic Gaussian primitives forming a partition-of-unity scaffold, offering spatial interpretability. Theoretical analysis proves approximation rates and risk decomposition, guiding optimal primitive count scaling with observations. The approach targets aerodynamic design, flow control, and digital-twin instrumentation.
Key facts
- FLUIDSPLAT reconstructs flow fields from sparse sensors.
- Model predicts K anisotropic Gaussian primitives.
- Primitives form a partition-of-unity scaffold.
- Inspired by 3D Gaussian Splatting.
- Proves O(K^{-s/d}) approximation rate for Sobolev smoothness s.
- Squared-risk decomposition: bias O(K^{-2s/d}), variance O(σ^2 K/N).
- Optimal K* ~ (N/σ^2)^{d/(2s+d)}.
- Targets aerodynamic design, flow control, digital twins.
Entities
—