A pitching wing that briefly exceeds its stall angle sheds a large vortex and loses lift abruptly, a phenomenon called dynamic stall. Simulations of the standard laboratory case all disagree with the experiment in the same way, whatever their fidelity. These include cheap time-averaged turbulence models, hybrid methods that model turbulence near the wall and resolve it in the wake (hybrid RANS-LES, short for Reynolds-Averaged Navier-Stokes / Large Eddy Simulation), and a 374-million-node simulation that resolves the wall layer directly. Each one under-predicts downstroke lift by about 0.25 in lift coefficient and puts peak lift about 2° too early. The authors of that comparison conclude that the error is not in the turbulence model. They attribute it to things the simulations leave out: tunnel walls, inflow turbulence, and surface roughness (Wall-resolved large eddy simulations of a pitching airfoil incurring in deep dynamic stall).
We propose to treat these facility conditions the way time-series forecasting treats weather or holidays. They become exogenous variables: external drivers that enter the model through dedicated, structured channels instead of being folded into the flow state. The key point is that one of these drivers is not a fixed number. Wall interference depends on the wing's own instantaneous lift, so it is a time-varying input that changes most during stall. The project builds three such physically motivated channels into a hybrid RANS-LES solver and an exogenous-conditioned surrogate of it. It then asks a falsifiable question: is the shared offset explained by a lift-dependent, time-varying facility input that transfers across observables, or by a constant correction that a two-parameter fit would recover anyway?
Core Claim
The claim has two parts. The first is that the experiment–simulation gap in deep dynamic stall can be reduced substantially by injecting a small number of physically structured exogenous inputs into a hybrid solver. These inputs are an effective-incidence history coupled to the instantaneous lift, the inflow turbulence intensity and length scale, and a roughness or receptivity parameter in the transition equation. No re-meshing of the tunnel is required. The second part matters more. These inputs should be identifiable: inferred from the lift loop alone, they should predict the moment and drag loops and the vortex-detachment angle, which were not used in fitting. If they cannot, then the "exogenous factors" explanation in the 374-million-node study cannot be separated from an ordinary curve fit, and that negative result is worth publishing.
The central exogenous channel follows classical image-vortex reasoning for a wing between tunnel walls. In that reasoning, the wall-induced upwash scales with the circulation and therefore with lift:
$$\alpha_{\text{eff}}(t) = \alpha_{\text{geo}}(t) + \Delta\alpha_0 + \lambda\, C_L(t),$$
Here $\alpha_{\text{geo}}$ is the prescribed pitch angle $10^\circ + 15^\circ\sin(\omega t)$, $\Delta\alpha_0$ is a constant incidence offset, $C_L(t)$ is the instantaneous lift coefficient computed by the solver, and $\lambda$ is a coupling gain that theory places on the order of $(c/h)^2$, with $c$ the chord and $h$ the tunnel height. Because $C_L$ collapses during stall, this term shifts the effective incidence by different amounts during upstroke and downstroke. A constant shift cannot do this. That difference is what the experiments are designed to detect.
Research Program
The physical substrate is the configuration of The effects of turbulence modeling on dynamic stall. That study ran 3D Stress-Blended Eddy Simulation (SBES), a hybrid that explicitly blends modeled near-wall stresses with resolved wake stresses, on meshes of up to 13 million cells. It found that the intermittency transition equation is required to capture the laminar separation bubble that triggers stall. We keep that transition equation, and following the span study we use a span of $L_z/c \ge 1.0$ to avoid artificially two-dimensional vortices. The three exogenous channels are implemented as follows. The lift-coupled incidence enters as a time-varying rotation of the far-field velocity, updated from the running lift each step. Inflow turbulence enters as synthetic fluctuations at the inlet, with intensity $Tu$ and integral length $L_t$. Roughness and receptivity enter as a parameter of the intermittency model. All three are boundary or closure inputs, so the near-wall shielding logic of the hybrid model is left untouched.
Hybrid runs are too expensive to search over these parameters directly. We therefore build a surrogate that predicts surface pressure, skin friction, and force coefficients over a cycle from the exogenous inputs. It extends the physics-informed convolutional reduced-order model (a fast learned approximation of the solver) of A deep neural network physics-based reduced order model for dynamic stall. That model reached about 1.1 s inference with a loss that penalizes both surface distributions and integrated forces, but it accepts only static cycle scalars and was trained on 2D URANS. We change its input side using two ideas from the exogenous-forecasting literature, which supply the method. From Aura: Universal Multi-dimensional Exogenous Integration for Aviation Time Series we take the separation of static attributes ($Tu$, $L_t$, roughness) from a continuous co-evolving series (the incidence history), each with its own encoder. From EgPDE-Net: Building Continuous Neural Networks for Time Series Prediction With Exogenous Variables we take the idea that a continuous exogenous trajectory can steer the time derivative of a latent state, so the surrogate responds to incidence histories it was not trained on rather than to cycle labels. We deliberately do not use spectral operators on uniform grids such as Domain Agnostic Fourier Neural Operators. That method reaches 0.59% error on steady transonic airfoils, but the hundreds of POD modes needed for dynamic-stall velocity fields reported in Pattern recognition of the flow around a pitching NACA 0012 airfoil in dynamic stall conditions make full-field spectral prediction a poor match. We predict surface quantities instead.
The surrogate is trained on many cheap 2D URANS-with-intermittency runs and a smaller set of 3D SBES runs. These are combined as a learned low-to-high fidelity correction, so that the parameters inferred later are not distorted by the missing 3D vortex breakdown in 2D URANS. Inference fits the posterior over $(\Delta\alpha_0, \lambda, Tu, L_t, \text{roughness})$ to the measured lift loop of Lee and Gerontakos. The posterior mean is then verified with fresh SBES runs that use no surrogate.
Baselines
The naive reference is SBES as published, with no exogenous inputs. For this configuration the file reports a downstroke lift offset of about 0.25 in $C_L$ and peak lift at about 22° versus 24° experimentally, shared by URANS, SBES, and wall-resolved LES (Wall-resolved large eddy simulations of a pitching airfoil incurring in deep dynamic stall). We rerun it on our mesh so that every comparison uses the same discretization.
The strongest competitor is a classical constant correction: a fixed incidence shift plus a lift scaling, fit to the experimental lift loop. It has two free parameters and requires one or two SBES runs. It is strong because it is what experimentalists already apply, and it may absorb most of a gap that is roughly a shift. It is given the same fitting target (the lift loop) and the same number of verification SBES runs as our method. To count as beating it, our method must reduce held-out moment- and drag-loop error and the vortex-detachment-angle error, not only lift error. No reported result for such a corrected SBES on this case was found on Lacuna.
The second competitor is an output-discrepancy model: a Gaussian process or small network fitted to the experiment-minus-SBES lift residual as a function of phase. It is given identical data. It will match lift closely by construction, so it is judged only on held-out observables. It is also judged on whether a residual fitted on lift says anything about moment, which it has no mechanism to do.
The key ablation removes the lift coupling ($\lambda = 0$) while keeping the constant offset and the inflow and roughness channels. This isolates whether the time-varying exogenous input is what matters. A second ablation replaces the structured surrogate inputs with plain concatenation of all exogenous inputs into the baseline network. This checks whether the separate-stream design adds anything at our data scale, a question Select, then Balance: A Plug-and-Play Framework for Exogenous-Aware Spatio-Temporal Forecasting answered positively, with relative-error gains often above 20%, but only for urban forecasting.
Why not just…?
Why not just apply a constant incidence and lift correction? This is the real threat, and the proposal is built around it. If the constant fit closes the gap on lift and predicts moment, drag, and detachment angle as well as the lift-coupled model, then the facility effect is effectively static. In that case we report that the time-varying exogenous channel is unnecessary for this case. The physical reason to expect otherwise is that the offset is largest on the downstroke, after lift has collapsed. A lift-proportional correction changes sign in its effect there, and a constant one does not. The $\lambda = 0$ ablation settles this directly.
Why not just calibrate inflow turbulence in cheap 2D URANS? Bayesian calibration on URANS alone would be fast. However, The effects of turbulence modeling on dynamic stall shows that 2D URANS over-predicts vortex coherence, and Pattern recognition of the flow around a pitching NACA 0012 airfoil in dynamic stall conditions shows that it mistimes vortex formation. Calibration would therefore load model error onto the facility parameters. We test this by running the URANS-only calibration and checking whether its parameters, transplanted into SBES, still reduce the gap. If they do, the multi-fidelity surrogate is unnecessary and we say so.
Why not just fit an output residual model? It will reproduce lift, but it attributes no cause. It therefore cannot transfer to quantities it did not see or to configurations with different tunnel geometry. The held-out observable test measures exactly this.
Proposed Experiments
All experiments use the Lee and Gerontakos deep-stall case: NACA 0012, $Re = 1.35\times10^5$, $k = 0.1$, $\alpha = 10^\circ + 15^\circ\sin\omega t$. Metrics are phase-resolved RMSE of the $C_L$, $C_m$, and $C_D$ loops against experiment, the error in peak-lift angle, and the error in the angle at which the dynamic-stall vortex detaches. That detachment angle is the quantity SBES got wrong in the turbulence-modeling study. We also check that the half-pitching-frequency cycle-to-cycle peak reported there survives the added inputs.
The decisive check uses one fit, to the lift loop only, followed by fresh SBES verification. Four outcomes are informative. If the lift-coupled model beats the constant correction on held-out $C_m$, $C_D$, and detachment angle, the time-varying facility channel is supported. If the two tie, the facility effect is static and classical corrections suffice. If no physical channel closes the gap without implausible parameters (for example, $\lambda$ far from the $(c/h)^2$ scale, or $Tu$ outside typical tunnel ranges), the external-factor explanation is itself in doubt. If the URANS-only calibration transfers to SBES, the surrogate machinery is unnecessary. A secondary check measures how well the surrogate predicts held-out SBES runs at unseen exogenous settings, relative to the static-scalar model of Baldan and Guardone retrained on the same data. This tells us whether exogenous-trajectory conditioning earns its complexity in aerodynamics as it did in forecasting.
References
- Wall-resolved large eddy simulations of a pitching airfoil incurring in deep dynamic stall
- The effects of turbulence modeling on dynamic stall
- A deep neural network physics-based reduced order model for dynamic stall
- Aura: Universal Multi-dimensional Exogenous Integration for Aviation Time Series
- EgPDE-Net: Building Continuous Neural Networks for Time Series Prediction with Exogenous Variables
- Domain Agnostic Fourier Neural Operators
- Pattern recognition of the flow around a pitching NACA 0012 airfoil in dynamic stall conditions
- Select, then Balance: A Plug-and-Play Framework for Exogenous-Aware Spatio-Temporal Forecasting