Real-Time Aerodynamic Flow Reconstruction from Ultra-Sparse Sensors via Deep Permutation Networks (DPN) and Idempotent Physical Projections

Abstract

Real-time estimation of full unsteady aerodynamic flow fields $\mathbf{u}(\mathbf{x}, t) = (u, v)$ from ultra-sparse discrete sensor observations (e.g., $M \le 10$ point velocity probes across thousands of grid cells) is a foundational challenge in active flow control, wind tunnel telemetry, and unmanned aerial vehicle (UAV) gust alleviation. Physics-Informed Neural Networks (PINNs) have emerged as a dominant scientific machine learning paradigm; however, for online cold-start inverse reconstruction—where an unexpected, non-stationary flow snapshot must be inverted on-the-fly without prior model fine-tuning—PINNs require thousands of Adam/L-BFGS backpropagation iterations ($8.33\text{ s}$ per state), violating 1~kHz real-time deadlines by four orders of magnitude. While offline pre-trained neural surrogates evaluate rapidly, they demand extensive parameter sweeps and inherently suffer from non-physical divergence leakage and boundary penetration due to soft penalty loss formulations. Similarly, classical Gappy-POD provides an instantaneous linear estimate ($<0.05\text{ ms}$), yet modal truncation invariably leaks divergence ($\|\nabla \cdot \mathbf{u}\|_\infty \approx 1.84 \times 10^{-2}\text{ s}^{-1}$) and violates wall adherence. In this paper, we propose a two-tier discrete-continuous architecture unifying Deep Permutation Networks (DPN) with algebraic idempotent metric projections ($\boldsymbol{\Pi}^2 = \boldsymbol{\Pi}$) and sequential Projection Onto Convex Sets (POCS). In Tier 1, a discrete APM limit-cycle oscillator tracks the non-linear vortex shedding phase and filters modal coefficients along the invariant limit-cycle attractor in $\mathcal{O}(1)$ time with zero backpropagation. In Tier 2, four idempotent projectors cyclically interleave: An orthonormal proper orthogonal decomposition (POD) subspace projector $\mathbf{P}_{\text{POD}} = \boldsymbol{\Phi}\boldsymbol{\Phi}^\dagger$; A discrete Hodge-Helmholtz divergence-free projector $\boldsymbol{\Pi}_{\text{div}}$ enforcing $\nabla \cdot \mathbf{u} = 0$; A no-slip wall projector $\boldsymbol{\Pi}_{\text{wall}}$; and A Dirichlet sensor clamping projector $\boldsymbol{\Pi}_{\text{sensor}}$. In rigorous head-to-head empirical testing against PyTorch PINN and Gappy-POD on a Navier-Stokes Direct Numerical Simulation (DNS) wake ($Re=100$, $St=0.160$, 2048 cells, 10 sensors representing $0.49\%$ spatial coverage), the proposed DPN-IdemSolver achieves: An online execution latency of $0.275\text{ ms}$ ($275\ \mu\text{s}$), representing a $30,295\times$ speedup over online PINN backpropagation ($8,329.6\text{ ms}$); A reconstruction accuracy of $R^2 = 99.99\%$ (compared to $96.22\%$ for Gappy-POD and $69.88\%$ for PINN); A relative velocity $L_2$ error of $0.44\%$ ($11\times$ lower than Gappy-POD and $47.5\times$ lower than PINN); Machine-precision solenoidal compliance ($\|\nabla \cdot \mathbf{u}\|_\infty = 1.26 \times 10^{-11}\text{ s}^{-1}$); $3\times$ noise suppression under realistic probe turbulence; and Strictly zero GPU and zero neural network backpropagation training, operating within $5\text{ MB}$ of RAM on standard edge computing nodes.