Research map / Probabilistic and scientific modeling / Dynamical systems
Flow Field Reconstruction with Sensor Placement Policy Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a directional transport-aware GNN for 3D flow-field reconstruction from boundary sensors, plus a Two-Step Constrained PPO that learns sensor placements using a saddle-point log-probability for the exact-count constraint.
Problem
Reconstructing 3D turbulent flow fields from sparse boundary sensor measurements under realistic conditions (no known PDEs, irregular geometries, stochastic inlets), and choosing sensor locations that give accurate reconstructions when the reconstruction model's accuracy varies by placement.
Approach
A message-passing GNN whose messages are weighted by a directional score (node-edge inner product) applied to latent differences, mimicking the advection operator and a learnable interpolation; sensor placement is learned by PPO with an exact sensor-count constraint handled via penalized pretraining, Gumbel top-k sampling, and a saddle-point approximation of the constrained log-probability.
Key results
- On the Sphere dataset at 10% uniform sensors, DTA-GNN reaches MSE 3.092e-4 versus 3.537 for FCN and 3.755 for MeshGraphNets, and it has the lowest error in all 32 dataset/placement settings in the main table.
- On sea surface temperature, DTA-GNN gives a 14.01% relative MSE improvement over DiffusionPDE.
- The learned sensor placement policy gives about a 15% MSE reduction relative to uniformly placed sensors at 10% density.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DTA-GNN instead of the strongest listed baseline | Flow reconstruction · Sphere, Ellipsoid, Cylinder, and NACA 4-digit; 10% Uniform and Random | MSE (×10^-4): Sphere Uniform: 3.537 -> 3.092; Sphere Random: 4.541 -> 4.110; Ellipsoid Uniform: 18.105 -> 9.676; Ellipsoid Random: 25. | Helped |
| Use DTA-GNN instead of the strongest listed baseline | Flow reconstruction · Kolmogorov Flow and Taylor-Green Vortex, 5% Random | MSE: 10.180 -> 9.484; 8.430 -> 6.749 | Helped |
| Add directional information to message passing (ablation) | DTA-GNN · NACA 4-digit, 5% Uniform and 10% Uniform | MSE (×10^-4): 75.124 -> 59.308; 59.064 -> 43.647 | Helped |
| Use neighboring latent-state differences instead of concatenating neighboring states (ablation) | DTA-GNN · NACA 4-digit, 5% Uniform and 10% Uniform | MSE (×10^-4): 63.162 -> 59.308; 52.294 -> 43.647 | Helped |
| Add directional information and neighboring-state differences to message passing (ablation) | DTA-GNN · NACA 4-digit, 5% Uniform | MSE (×10^-4): 78.421 -> 59.308 | Helped |
| Use Two-Step Constrained PPO instead of starting with constrained PPO (ablation) | Sensor placement policy · Sphere, Ellipsoid, Cylinder, and Airfoil | MSE (×10^-4): 19.280 -> 3.270; 31.647 -> 11.002; 55.014 -> 9.490; 327.966 -> 46.431 | Helped |
| Add constrained second-stage training after penalized PPO (ablation) | Sensor placement policy · Sphere, Ellipsoid, Cylinder, and Airfoil | MSE (×10^-4): 3.270 -> 3.178; 11.002 -> 10.629; 9.490 -> 9.575; 46.431 -> 44.103 | Mixed |
| Use saddle point approximation for constrained log probability (ablation) | Two-Step Constrained PPO · Sphere, Ellipsoid, Cylinder, and Airfoil | MSE (×10^-4): 3.178 -> 3.270; 10.629 -> 11.002; 9.575 -> 9.490; 44.103 -> 46.431 | Mixed |
| Use learned sensor placements instead of uniform sensor placements | Sensor placement policy · Four flow datasets, 10% sensor density | MSE: +15% reduction in MSE | Helped |
| Use QR Pivoting or d-optimal sensor locations instead of uniform locations | Sensor placement · Sphere, 10% sensor density | MSE (×10^-4): 3.092 -> 6.988; 3.092 -> 6.309 | Hurt |
| Increase sensor density | Flow reconstruction · Flow datasets | MSE | Helped |
Surprising findings
- QR pivoting and D-optimal sensor placements gave higher reconstruction error than uniform or random placements (e.g., 6.988 and 6.309 vs 3.092 for uniform on Sphere).
- Constrained PPO trained from scratch performed far worse than the penalized variant (e.g., 19.280 vs 3.270 on Sphere), and without the saddle-point log-probability the learned policy failed to beat uniform placement.
Limitations
- Trained only on simulated datasets; real-world sensor performance is untested and may degrade.
- Sensor placement restricted to domain boundaries.
- Relies on an independence assumption for Bernoulli sensor variables, which the constraint breaks.
Experimental setup
Datasets: DNS-simulated 3D turbulent flows in COMSOL (Sphere, Ellipsoid, Cylinder, NACA 4-digit) NOAA OISST V2 sea surface temperature Kolmogorov Flow (appendix) Taylor-Green Vortex (appendix)
Benchmarks: COMSOL 3D turbulent flow datasets (Sphere, Ellipsoid, Cylinder, NACA 4-digit) NOAA OISST V2 sea surface temperature Kolmogorov Flow Taylor-Green Vortex
Metrics: MSE of normalized velocity and pressure
Baselines: Mean KNN OFormer GKO FCN MeshGraphNets DiffusionPDE QR pivoting D-optimality
Builds on: MeshGraphNets PPO Flow Completion Network DiffusionPDE OFormer
Closest papers by meaning
- PEINR: A Physics-enhanced Implicit Neural Representation for High-Fidelity Flow Field Reconstruction (ICML 2025 · similarity 0.86)
- PhySense: Sensor Placement Optimization for Accurate Physics Sensing (NeurIPS 2025 · similarity 0.86)
- Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions (NeurIPS 2026 · similarity 0.85)
- Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks (ICLR 2025 · similarity 0.84)
- Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates (ICML 2025 · similarity 0.84)
- Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.