FLUIDSPLAT: Reconstructing Physical Fields from Sparse Sensors via Gaussian Primitives

Researchers introduce FLUIDSPLAT, a neural network model that reconstructs continuous flow fields from sparse sensor data using anisotropic Gaussian primitives. The approach provides theoretical guarantees on approximation rates and demonstrates 11-28% error improvements over existing methods across multiple aerodynamic benchmarks.

FLUIDSPLAT addresses a fundamental challenge in computational fluid dynamics and aerodynamic design: reconstructing complete flow fields from limited sensor measurements. Traditional neural approaches encode sensor data into opaque latent representations that lack spatial interpretability and provide no formal guidance on scaling computational resources with observation density. This new model inverts that approach by predicting explicit Gaussian primitives that form an interpretable spatial scaffold, bringing transparency to the learned representations.

The theoretical contribution establishes formal bounds on approximation performance, proving that primitive count scales as (N/σ²)^(d/(2s+d)) where N is observation count and σ² is noise level. This reveals a critical variance bottleneck inherent to sparse sensing regimes—additional primitives beyond this optimal point increase estimation variance without sufficient bias reduction. Rather than treating this as a limitation, the authors leverage this insight to design a complementary residual decoder that captures high-frequency variations the scaffold cannot represent.

The empirical validation spans diverse scenarios from 2D cylinder flows to 3D automotive aerodynamics, consistently outperforming strong baselines by substantial margins. This breadth of testing suggests the method generalizes across different flow regimes and dimensionalities. The work has direct applications in digital-twin systems for real-time monitoring, wind tunnel testing optimization, and flow control design where sensor placement is constrained by cost and physical limitations. The combination of theoretical rigor with practical performance improvements positions this as a meaningful advance in physics-informed machine learning for continuous field reconstruction.

• →FLUIDSPLAT uses interpretable Gaussian primitives as spatial scaffolds, improving transparency over black-box latent representations
• →Theoretical analysis proves optimal primitive count scales with (N/σ²)^(d/(2s+d)), revealing fundamental variance-bias tradeoffs in sparse sensing
• →11-28% error reduction across four benchmarks demonstrates consistent improvements over existing neural baselines
• →Residual decoder architecture directly addresses the variance bottleneck identified by theory, closing the gap between prediction and practice
• →Method applicable to real-world aerodynamic design, flow control, and digital-twin instrumentation with practical sensor constraints