Grasping extreme aerodynamics on a low-dimensional manifold

The study addresses how small air vehicles can maintain stable flight in extreme gusty environments such as urban canyons and mountainous terrains, where strong vortical disturbances pose severe challenges. Classical aerodynamics largely treats steady or quasi-steady conditions and small perturbations, leaving a gap in theory for extreme, highly unsteady vortex-wing interactions. The parameter space governing gust encounters (vortex strength, size, position, orientation, and geometry) is vast, making exhaustive experimentation or simulation infeasible. The authors focus on the gust ratio G = Ugust/Uc and target extreme cases with G > 1, considering 0 ≤ G ≤ 10 in general. The research question is whether the rich, nonlinear dynamics of extreme gust-airfoil interactions possess a unifying, low-dimensional structure that can be discovered and exploited for modeling, estimation, and control. The purpose is to identify a compact set of variables that capture the essential physics, enabling real-time reconstruction and control despite the high-dimensional nature of the flow. The authors hypothesize that a physically informed nonlinear compression (autoencoder augmented with lift information) can discover a low-dimensional manifold governing extreme aerodynamic responses.

The paper situates its contribution within several areas: classical and low-speed aerodynamics and rotorcraft theory that emphasize steady or weakly unsteady regimes; gust load literature and recent analyses of discrete gust encounters; and challenges for urban air mobility where gusty, vortex-dominated environments are prevalent. For data-driven modeling of flows, principal component analysis/proper orthogonal decomposition (PCA/POD) is a standard linear reduction method but can fail for strongly nonlinear dynamics. Nonlinear autoencoders and convolutional neural networks have been proposed for fluid data compression and nonlinear mode decomposition. Concepts from inertial manifolds suggest low-dimensional structures can underlie complex dynamics. Prior sparse-sensor reconstruction and decoder-type neural networks indicate feasibility of real-time state estimation from limited measurements. The paper builds on these by integrating a physical observable (lift) into a nonlinear autoencoder to uncover a universal manifold for extreme gust-airfoil interactions.

Flow simulations: Two-dimensional incompressible direct numerical simulations of a NACA 0012 airfoil at chord-based Reynolds number Re = uc/ν = 100. Angles of attack α ∈ {20°, 30°, 40°, 50°, 60°}; undisturbed flow is steady at 20° and exhibits periodic shedding at α ≥ 30°. Computational domain (x, y)/c ∈ [-15, 30] × [-20, 20], leading edge at the origin. Gust model: a strong upstream vortex with angular velocity profile Ug = Ue,max (r/R) exp[1/2 − r²/(2R²)], introduced at x0/c = −2 and y0/c ∈ [−0.5, 0.5]. Parameters: gust ratio G = Ue,max/U ∈ [−10, 10], size L = 2R/c ∈ [0.5, 2], and vertical position y0/c. Dataset: For each α, 40 randomly sampled disturbed cases (from G ∈ [−4, 4], L ∈ [0.5, 2], y0/c ∈ [−0.5, 0.5]) plus undisturbed baselines. For each case, 1200 vorticity snapshots over 10.2 convective time units are collected. Analysis subdomain (x, y)/c ∈ [−1.4, 4] × [−1.2, 1.2] with grid 240 × 120. Total training data: 100 disturbed cases + 5 baselines, yielding 1.26 × 10^5 frames; split 20 train and 20 test cases per α. Baseline reduction: PCA/POD applied to vorticity fields to assess linear compression performance. Autoencoder architectures: A convolutional autoencoder (encoder: CNN to capture global vortical features, flatten/reshape, MLP to latent; decoder is symmetric CNN/MLP) with hyperbolic tangent activation. Latent dimension m = 3. Training objective for regular autoencoder minimizes L2 reconstruction error of vorticity. Lift-augmented autoencoder: adds a side MLP head to predict lift coefficient Cl from the latent vector, encouraging latent variables to retain lift-relevant information. Combined loss: ||q − q̂||2 + β ||Cl − Ĉl||2 with β = 0.05 (chosen via L-curve analysis). Optimization via Adam. Evaluation: Reconstruction accuracy assessed via structural similarity index (SSIM) between decoded and reference vorticity fields; assessment of latent-space trajectories across parameters, and generalization to interpolation (|G| ≤ 4), extrapolation (|G| > 4), noisy inputs (30% Gaussian noise), and multi-vortex disturbances (two vortices arranged vertically/horizontally; five random vortices).

• Extreme gusts induce violent, rapid force transients: in a representative case (α = 20°, G = 3.8), lift increased by 714% and then dropped by 656% within 1.8 convective time units, highlighting control challenges and structural risks.
• PCA/POD fails to produce a meaningful, discriminative low-dimensional representation: trajectories in the first three PCA components appear incoherent and overlap across different angles of attack, leading to inaccurate reconstructions.
• A nonlinear convolutional autoencoder accurately compresses and reconstructs vorticity fields using only three latent variables, confirming shared flow features across cases; however, without physical guidance, latent trajectories do not collapse onto a simple structure.
• Incorporating lift into the autoencoder (lift-augmented autoencoder) reveals a universal, low-dimensional manifold: trajectories for all extreme gust-airfoil interaction cases collapse onto or near a cone-shaped, hourglass-like inertial manifold in 3D latent space (ξ1, ξ2, ξ3), with undisturbed periodic shedding states forming a cone-like structure.
• The latent coordinate ξ3 correlates with effective angle of attack and lift response; disturbed trajectories move toward neighboring undisturbed orbits consistent with transient changes in effective incidence.
• Reconstruction and estimation performance is high across conditions: decoded fields achieve SSIM values commonly in the 0.78–0.997 range (e.g., examples report 99.7%, 98.3%, 94.3%, 90.8%, 95.9%, 88.1%, etc.), while lift is accurately predicted from the latent variables.
• Generalization beyond training: the learned manifold and decoder handle extrapolative gust strengths (|G| ≥ 4), with trajectories exhibiting larger radial excursions in ξ1–ξ2 due to stronger disturbances, yet maintaining accurate flow and lift reconstructions.
• Robustness to noise and complex inputs: with 30% Gaussian noise added to inputs, the model reconstructs flow and lift well (e.g., SSIM ≈ 90.8–95.9%). The model, trained on single-vortex cases, also reconstructs and estimates lift for two-vortex configurations (vertical/horizontal) and for a severe five-vortex scenario, capturing multi-dip lift signatures and corresponding latent trajectory features.
• Overall, extreme aerodynamic flows can be nearly losslessly compressed to three variables that encode physically meaningful, universal dynamics suitable for real-time estimation and control.

The results show that highly nonlinear, extreme gust-airfoil interactions possess an unexpectedly low-dimensional structure. By embedding lift into the compression process, the autoencoder uncovers a cone-shaped inertial manifold onto which diverse gust responses collapse, providing a unified description across parameters. The physical interpretability of the latent space (e.g., ξ3 reflecting effective angle of attack) links latent dynamics to aerodynamic forces relevant for flight stability. The findings imply that only a small number of sensors may suffice to estimate latent states and reconstruct high-dimensional flows in real time via decoder-type networks. This paves the way for reduced-order models evolving on the manifold, potentially leveraging phase-amplitude or phase-reduction analyses for timing-sensitive control. Robust performance under noise and multi-vortex inputs suggests applicability to realistic turbulent environments. The manifold framework thus offers a path to situational awareness, dynamic modeling, and active control strategies to mitigate extreme gust effects on small air vehicles.

The study demonstrates that extreme, vortex-dominated gust-airfoil interactions can be compressed to a three-variable latent representation when lift is incorporated into a nonlinear autoencoder. This lift-augmented compression reveals a universal, cone-shaped inertial manifold capturing the essential dynamics across a wide parameter space, enabling accurate reconstruction of flow fields and lift, including extrapolative gust strengths, noisy conditions, and multi-vortex disturbances. These findings provide a foundation for real-time flow estimation, reduced-order modeling on the manifold, and control strategies to stabilize flight in environments traditionally deemed unflyable. Future work should extend to three-dimensional flows and higher Reynolds numbers, broaden gust models to more realistic turbulence structures, and integrate sparse sensing and control design operating directly on the discovered manifold.

• The study is restricted to two-dimensional incompressible flows at a fixed low Reynolds number (Re = 100), whereas real atmospheric gust interactions are three-dimensional and often at much higher Reynolds numbers.
• The primary training data involve single-vortex disturbances with simplified profiles; although the model generalizes to multiple vortices and noise, broader turbulence spectra and more complex gust structures were not used for training.
• The manifold and reconstructions are demonstrated for a canonical NACA 0012 airfoil and a limited set of angles of attack; generalization across airfoil geometries and operating conditions remains to be verified.
• Real-time deployment aspects (sensor placement/number, onboard computation constraints) are discussed conceptually but not experimentally validated in this work.