Flow Field Reconstruction with Sensor Placement Policy Learning

Technical Summary: Flow Field Reconstruction with Sensor Placement Policy Learning

Problem Statement
Reconstructing high-fidelity flow fields from sparse sensor measurements is a critical challenge in fluid dynamics, particularly for applications like aerodynamic optimization and active flow control. Existing deep learning approaches often rely on unrealistic assumptions that limit their applicability to real-world scenarios. Specifically, prior works frequently assume: (1) two-dimensional domains, whereas real applications are three-dimensional; (2) known governing physical equations (e.g., Navier-Stokes) that can be integrated as constraints, whereas empirical data often deviates from idealized PDEs; (3) synthetic datasets generated by simplified solvers (e.g., pseudo-spectral or RANS) that lack the complexity of real turbulence; and (4) unconstrained sensor placement, ignoring the physical reality that sensors are typically fixed at domain boundaries or advected with the flow. Furthermore, traditional sensor placement strategies (e.g., SVD, QR pivoting) often yield suboptimal configurations when paired with complex reconstruction models, failing to account for the model's performance disparity across different sensor layouts.

Methodology
The authors propose a unified framework addressing both the reconstruction model and the sensor placement strategy under realistic constraints.

1. Dataset Generation: To bridge the gap between simulation and reality, the authors generated four 3D turbulent flow datasets using Direct Numerical Simulation (DNS) in COMSOL. These datasets feature varying geometries (sphere, ellipsoid, cylinder, NACA airfoil) with time-dependent, stochastic inlet velocities and random initial fields. Crucially, sensors are restricted to the domain boundaries to reflect practical deployment limits.
2. Directional Transport-Aware GNN (DTA-GNN): The reconstruction model is an Encoder-Processor-Decoder Graph Neural Network designed to operate without explicit physical priors.
   • Encoding: Nodes and edges are embedded into a latent space using distinct MLPs.
   • Processing: A novel processor explicitly encodes flow directionality and information transport. It computes a directional score ($d$) as the inner product between a node's latent state and edge features, acting as a proxy for the advection operator ($\nu \cdot \nabla \phi$). This score weights the differences between adjacent node states, mimicking continuous advection in a discrete mesh setting.
   • Decoding: A final MLP maps the updated node embeddings to estimated velocity and pressure fields.
   • This design functions as a learnable interpolation algorithm that propagates information along physically motivated pathways.
3. Two-Step Constrained PPO for Sensor Placement: To optimize sensor locations, the authors employ Proximal Policy Optimization (PPO) with a novel two-stage training procedure to handle the constraint that the number of sensors must equal a fixed proportion ($k$) of the mesh nodes.
   • Stage 1 (Penalized Training): The policy is trained unconstrained but with a penalized reward function ($\hat{R} = R - \lambda(k|V| - ||a||_2^2)^2$) to softly enforce the sensor count. This allows the policy to learn the fluid field variability and model performance disparity without the initial restriction of hard constraints.
   • Stage 2 (Constrained Training): The policy is refined under hard constraints. Sampling is performed using a Gumbel Top-$k$ selection to ensure exactly $k|V|$ sensors are chosen. To compute the log-probability required for PPO updates, the authors employ a saddle-point approximation, which offers $O(|V|)$ complexity compared to the intractable exact computation.
   • The policy learns to place sensors in regions of high flow variability (e.g., high gradients) while simultaneously accounting for where the reconstruction model performs best.

Key Results

• Reconstruction Accuracy: Extensive experiments on the four 3D geometries demonstrate that the DTA-GNN consistently outperforms baselines, including MLP/CNN hybrids, physics-informed networks (PINNs), and other GNNs (e.g., MeshGraphNets, FCN). The model achieves the lowest Mean Squared Error (MSE) across various sensor densities (5%–30%) and placement strategies (uniform vs. random). In the challenging NACA airfoil scenario, improvements of approximately 10% were observed.
• Real-World Validation: The model was evaluated on the NOAA OISST V2 global sea surface temperature dataset. It achieved a 14.01% relative improvement in MSE over the strongest baseline (DiffusionPDE), successfully reconstructing both large-scale seasonal cycles and fine-scale anomalies like equatorial upwelling.
• Sensor Placement Efficacy: The learned sensor placement policy reduced the reconstruction MSE by approximately 15% compared to uniform sensor placement. The policy successfully identified high-variability regions and optimized for the specific performance characteristics of the reconstruction model.
• Efficiency: The proposed model is parameter-efficient compared to other deep learning baselines.

Significance and Claims
The paper claims to advance the state of the art by introducing a problem formulation that aligns closely with realistic fluid dynamics challenges: 3D domains, unknown governing equations, complex stochastic data, and boundary-constrained sensor placement. The authors argue that their directional transport-aware GNN provides a robust, geometry-agnostic solution that learns meaningful representations without relying on predefined PDEs. Furthermore, they demonstrate that conventional sensor placement algorithms are insufficient for modern deep learning reconstruction tasks and that a learned policy, trained via their Two-Step Constrained PPO, is necessary to jointly optimize sensor layout and model performance. The work establishes a new benchmark for flow field reconstruction under realistic constraints, offering a practical pathway for high-fidelity data recovery in aerospace and environmental monitoring.