Signal-Aware Conditional Diffusion Surrogates for Transonic Wing Pressure Prediction

Carlos Sanmiguel Vila; Rodrigo Castellanos; V\'ictor Franc\'es-Belda

arxiv: 2604.11263 · v1 · submitted 2026-04-13 · ⚛️ physics.flu-dyn · cs.LG

Signal-Aware Conditional Diffusion Surrogates for Transonic Wing Pressure Prediction

V\'ictor Franc\'es-Belda , Carlos Sanmiguel Vila , Rodrigo Castellanos This is my paper

Pith reviewed 2026-05-10 15:43 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn cs.LG

keywords diffusion modelsaerodynamic surrogatestransonic flowwing pressure predictionconditional generationprincipal component analysisreliability indicators

0 comments

The pith

A conditional diffusion model with signal-aware loss predicts transonic wing pressures more accurately than deterministic baselines.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces a conditional denoising diffusion probabilistic model to predict surface pressure fields on a transonic aircraft wing across ranges of Mach number, angle of attack, and control surface settings. Deterministic regressors often smooth sharp nonlinear features such as suction peaks and shocks when trained with pointwise losses. The proposed approach first reparameterizes the unstructured pressure data via a non-truncated principal component representation, then trains a fully connected network by propagating a reconstruction loss through the diffusion process to create timestep-dependent weights that prioritize high-gradient regions. Repeated conditional sampling produces output variability whose spread tracks reconstruction error, which the authors interpret as a qualitative reliability indicator. Accurate fast surrogates of this kind would let designers evaluate many more configurations without repeated expensive computational fluid dynamics runs.

Core claim

The central claim is that a conditional denoising diffusion probabilistic model operating on PCA-reparameterized unstructured surface pressure data, trained with a signal-aware objective obtained by propagating reconstruction loss through the diffusion chain, reduces mean absolute error relative to deterministic baselines and improves capture of suction peaks, shock structures, and control surface discontinuities, while the spread across repeated samples correlates strongly with pointwise error and functions as a qualitative reliability measure rather than calibrated uncertainty.

What carries the argument

Conditional denoising diffusion model whose training objective is reweighted by propagating a reconstruction loss through the noising schedule, applied after a reversible non-truncated PCA reparameterization of the pressure field.

If this is right

The formulation reduces mean absolute error relative to the considered deterministic baselines.
Reconstruction of suction peaks, shock structures, and control surface discontinuities is improved.
Sampling-induced spread exhibits strong correspondence with surrogate error and supports use as a qualitative reliability indicator.
The model handles prediction under varying Mach number, angle of attack, and four control surface deflections on the NASA Common Research Model wing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same signal-aware diffusion construction could be tested on other fluid fields that contain sharp discontinuities, such as density or vorticity distributions.
The Local and Global Reliability Indices might be calibrated against held-out data to produce quantitative confidence intervals.
Embedding the surrogate inside an optimization loop would let designers trade off accuracy against the cost of additional CFD evaluations in a controlled way.

Load-bearing premise

That propagating a reconstruction loss through the diffusion process produces a timestep-dependent weighting that genuinely improves fidelity in strong-gradient regions, and that the non-truncated PCA representation preserves all information needed to reconstruct the original unstructured pressure field without introducing artifacts.

What would settle it

A test case containing a strong shock or suction peak where the diffusion surrogate shows no reduction in local error compared with a standard neural-network regressor, or where the PCA-reconstructed pressure field exhibits visible smoothing or spurious oscillations absent from the original data.

Figures

Figures reproduced from arXiv: 2604.11263 by Carlos Sanmiguel Vila, Rodrigo Castellanos, V\'ictor Franc\'es-Belda.

**Figure 2.** Figure 2: (a) Training process of the surrogate. (b) Conditional sample generation from Gaussian noise. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: MAE distribution between actual and predicted [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Chordwise pressure distributions for actual and surrogate [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Error distribution over the upper surface of the wing for [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 7.** Figure 7: Convergence of the model’s statistics as a function [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

**Figure 6.** Figure 6: Visualization test cases together with the three [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 8.** Figure 8: Spatial deviation of the model. The rows display the actual fields ( [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

**Figure 9.** Figure 9: a presents the 𝜔-LRI profiles for the 𝑙 testing conditions. To facilitate the comparison, the LRI is normalized to the unit interval [0, 1] for each flight condition, denoted herein as LRI d. In general, a consistent monotonic increase is observed across the dataset, confirming that regions of higher standard deviation also tend to exhibit higher high reconstruction error. In addition, the 𝑙 curves are so… view at source ↗

read the original abstract

Accurate and efficient surrogate models for aerodynamic surface pressure fields are essential for accelerating aircraft design and analysis, yet deterministic regressors trained with pointwise losses often smooth sharp nonlinear features. This work presents a conditional denoising diffusion probabilistic model for predicting surface pressure distributions on the NASA Common Research Model wing under varying conditions of Mach number, angle of attack, and four control surface deflections. The framework operates on unstructured surface data through a principal component representation used as a non-truncated, reversible linear reparameterization of the pressure field, enabling a fully connected architecture. A signal-aware training objective is derived by propagating a reconstruction loss through the diffusion process, yielding a timestep-dependent weighting that improves fidelity in regions with strong pressure gradients. The stochastic sampling process is analyzed through repeated conditional generations, and two diagnostic metrics are introduced, the Local Reliability Index and Global Reliability Index, to relate sampling-induced spread to reconstruction error. Relative to the considered deterministic baselines, the proposed formulation reduces mean absolute error and improves the reconstruction of suction peaks, shock structures, and control surface discontinuities. The sampling-induced spread exhibits strong correspondence with surrogate error, supporting its interpretation as a qualitative reliability indicator rather than calibrated uncertainty quantification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Signal-aware diffusion surrogate for wing pressures shows promise for sharp features but requires verification on PCA reversibility and quantitative results.

read the letter

The colleague should know upfront that this paper presents a conditional diffusion model for predicting surface pressure on the NASA Common Research Model wing, with a signal-aware objective and PCA-based handling of unstructured data, along with reliability indices from sampling spread. The central claim is better mean absolute error and feature recovery compared to deterministic baselines. What is actually new is the derivation of the timestep-dependent weighting in the loss by propagating reconstruction error through the diffusion chain. This focuses training on difficult regions with strong gradients. They also introduce the non-truncated PCA as a reversible linear reparameterization to feed the unstructured pressure fields into a fully connected architecture. Finally, the Local Reliability Index and Global Reliability Index provide diagnostics that link the variance across multiple conditional samples to the reconstruction error, offering a qualitative reliability signal. The paper does well in tackling a practical issue in aerodynamic surrogate modeling. Deterministic regressors with pointwise losses tend to smooth out suction peaks, shock structures, and control surface discontinuities, which are critical in transonic regimes. By using diffusion and the signal-aware loss, the approach claims to preserve these features better. The reliability indices are a thoughtful addition, turning the inherent stochasticity of the sampler into a useful indicator rather than a drawback. Soft spots are present in the supporting evidence. The abstract provides no numerical values for the error reductions, no details on the baselines used, validation splits, or error bars, so the magnitude of the improvement is not verifiable from the summary. The PCA step is another area of concern. The claim that it is non-truncated and preserves all information without artifacts assumes the training data spans the full space of possible pressure fields. However, in CFD applications the number of training cases is usually much smaller than the number of surface mesh points, so the PCA basis has rank limited by the sample count. Components orthogonal to this subspace get projected out and cannot be recovered in the inverse transform. This could mean that some of the reported gains in sharp feature reconstruction come from the preprocessing rather than the diffusion model or the new loss. The paper needs to report the actual dimensions and show PCA reconstruction errors separately to address this. This work is for aerospace engineers and computational fluid dynamics researchers who need fast surrogates for design optimization, as well as for machine learning practitioners applying generative models to physical field data. A reader focused on conditional generation or uncertainty-aware surrogates would extract useful technical ideas. It deserves a serious referee because the methods are clearly motivated by an application gap and introduce testable innovations. I recommend sending it for peer review, with the expectation that reviewers will request expanded quantitative results and explicit checks on the PCA fidelity.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a conditional denoising diffusion probabilistic model for surrogate prediction of surface pressure distributions on the NASA Common Research Model wing, conditioned on Mach number, angle of attack, and four control-surface deflections. It reparameterizes the unstructured pressure field via a non-truncated PCA representation to enable a fully connected network, derives a signal-aware training objective by propagating a reconstruction loss through the diffusion process to obtain timestep-dependent weighting, and introduces Local and Global Reliability Indices that relate sampling-induced variance to reconstruction error. The central claim is that the formulation reduces mean absolute error relative to deterministic baselines while improving recovery of suction peaks, shock structures, and control-surface discontinuities, with the sampling spread serving as a qualitative reliability indicator.

Significance. If the performance gains and reliability correspondence are confirmed, the work would provide a useful probabilistic surrogate framework for transonic aerodynamics that mitigates smoothing of sharp nonlinear features and supplies a built-in diagnostic for prediction reliability, potentially aiding uncertainty-aware design optimization.

major comments (3)

[PCA reparameterization] PCA reparameterization section: The assertion that the non-truncated PCA constitutes a reversible linear reparameterization that 'preserves all information needed to reconstruct the original unstructured pressure field without introducing artifacts' is not supported when the number of training samples N is smaller than the number of surface nodes D (typical for CFD meshes). The learned basis has rank at most N, so any test-field component orthogonal to the training subspace is projected away prior to diffusion; the inverse transform therefore cannot recover the true field. This directly undermines the claim of improved reconstruction of strong-gradient features (suction peaks, shocks, control-surface jumps), which may partly reflect the projection rather than the diffusion architecture or signal-aware loss.
[Signal-aware training objective] Signal-aware objective derivation: Propagating the reconstruction loss through the diffusion process to produce a timestep-dependent weighting is presented as improving fidelity in strong-gradient regions, but the manuscript does not demonstrate that this weighting is materially different from a standard diffusion objective or that it specifically targets gradient-rich areas. A quantitative ablation (signal-aware vs. standard loss on identical PCA inputs) is needed to establish that the claimed fidelity gains are attributable to this mechanism.
[Results] Results section: The manuscript must supply concrete numerical MAE values, baseline model architectures and training details, train/validation/test split sizes, and error bars or statistical tests for the reported improvements. The abstract supplies none of these, and without them the magnitude, robustness, and statistical significance of the performance claims cannot be evaluated.

minor comments (2)

[Figures] Figure captions and axis labels should explicitly state the number of diffusion samples used to compute the Local/Global Reliability Indices and the exact definition of the spread metric.
[Notation] Notation for the conditioning variables (Mach, AoA, control deflections) should be introduced once and used consistently in equations, text, and figure legends.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and constructive comments. We address each major comment below, providing clarifications and committing to revisions where appropriate.

read point-by-point responses

Referee: PCA reparameterization section: The assertion that the non-truncated PCA constitutes a reversible linear reparameterization that 'preserves all information needed to reconstruct the original unstructured pressure field without introducing artifacts' is not supported when the number of training samples N is smaller than the number of surface nodes D (typical for CFD meshes). The learned basis has rank at most N, so any test-field component orthogonal to the training subspace is projected away prior to diffusion; the inverse transform therefore cannot recover the true field. This directly undermines the claim of improved reconstruction of strong-gradient features (suction peaks, shocks, control-surface jumps), which may partly reflect the projection rather than the diffusion architecture or signal-aware loss.

Authors: We agree that the PCA reparameterization is limited to the subspace spanned by the training samples, with rank at most N when N < D. Our description of it as a 'reversible linear reparameterization' refers to the exact invertibility of the PCA transform within that subspace, without truncation of components. We acknowledge that this projection can affect reconstruction of features not well-represented in the training data. However, for the transonic wing pressure fields in our dataset, the dominant variations are captured by the leading modes, and the orthogonal residuals are minimal for test conditions within the sampled range. The improvements over baselines are measured on the same projected representation, so they are attributable to the diffusion model and loss rather than the projection alone. We will revise the manuscript to explicitly discuss this limitation of the PCA approach and its implications for sharp feature recovery. revision: partial
Referee: Signal-aware objective derivation: Propagating the reconstruction loss through the diffusion process to produce a timestep-dependent weighting is presented as improving fidelity in strong-gradient regions, but the manuscript does not demonstrate that this weighting is materially different from a standard diffusion objective or that it specifically targets gradient-rich areas. A quantitative ablation (signal-aware vs. standard loss on identical PCA inputs) is needed to establish that the claimed fidelity gains are attributable to this mechanism.

Authors: We will add a quantitative ablation study in the revised manuscript comparing the signal-aware training objective to the standard diffusion loss (such as the simplified loss) using identical PCA inputs, network architecture, and training setup. This will include metrics on overall MAE as well as localized errors in high-gradient regions like suction peaks and shocks to demonstrate the specific benefits. The derivation shows that the weighting varies with timestep and signal strength, but the ablation will confirm its material difference and impact. revision: yes
Referee: Results section: The manuscript must supply concrete numerical MAE values, baseline model architectures and training details, train/validation/test split sizes, and error bars or statistical tests for the reported improvements. The abstract supplies none of these, and without them the magnitude, robustness, and statistical significance of the performance claims cannot be evaluated.

Authors: We agree with this assessment and will substantially expand the results section and abstract in the revision. We will report specific MAE values for the proposed model and all baselines, provide full details on baseline architectures and training procedures, specify the train/validation/test split sizes, and include error bars from repeated trainings or statistical tests to assess significance of the improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained on external data

full rationale

The paper trains a conditional diffusion model on external CFD datasets for the NASA CRM wing. The PCA reparameterization is presented as a standard, non-truncated linear preprocessing step to enable a fully connected architecture on unstructured meshes; it is not derived from or defined in terms of the diffusion outputs. The signal-aware loss is obtained by propagating a reconstruction loss through the diffusion process, which is an explicit methodological derivation rather than a self-referential fit. No equations reduce the claimed MAE reductions or improved recovery of shocks/suction peaks to quantities defined solely by the model's own fitted parameters or self-citations. The sampling-spread diagnostics are post-hoc analyses of the trained model against held-out data. The framework therefore remains independent of its own predictions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that the pressure field admits an exact reversible PCA representation and that the conditional diffusion process can be trained to match the true data distribution under the given flight parameters.

axioms (2)

domain assumption Principal component analysis supplies a complete, lossless, and reversible linear reparameterization of the unstructured surface pressure field
Explicitly invoked to enable a fully connected architecture without truncation.
domain assumption The conditional distribution of pressure fields given Mach number, angle of attack, and control-surface deflections is learnable by a denoising diffusion model
Core modeling assumption underlying the entire surrogate.

pith-pipeline@v0.9.0 · 5515 in / 1462 out tokens · 31831 ms · 2026-05-10T15:43:40.862429+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

[1]

doi:10.2514/1.J063440. C. Wang, H. Xiang, H. Jiang, B. Li, Y. Peng, Z. Fan, M. Zhang, J. Li, AeroDiT: Diffusion transformers for Reynolds-Averaged Navier–Stokes simulations of airfoil flows, Physics of Fluids 37 (2025) 124120. doi:10. 1063/5.0256147. K.Ogbuagu,S.Maleki,G.Bruni,S.Krishnababu, Foildiff: A hybrid diffusion transformer model for airfoil flow ...

work page doi:10.2514/1.j063440 2025
[2]

doi:10.2514/6.2014-0080. J. Tran, K. Fukami, K. Inada, D. Umehara, Y. Ono, K. Ogawa, K. Taira, Aerodynamics-guided machine learning for design optimization of electric vehicles, Communications Engineering 3 (2024). doi:10.1038/ s44172-024-00322-0. V. Francés-Belda, A. Solera-Rico, J. Nieto-Centenero, E. Andrés, C. Sanmiguel Vila, R. Castellanos, Toward ae...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.2514/6.2014-0080 2014

[1] [1]

doi:10.2514/1.J063440. C. Wang, H. Xiang, H. Jiang, B. Li, Y. Peng, Z. Fan, M. Zhang, J. Li, AeroDiT: Diffusion transformers for Reynolds-Averaged Navier–Stokes simulations of airfoil flows, Physics of Fluids 37 (2025) 124120. doi:10. 1063/5.0256147. K.Ogbuagu,S.Maleki,G.Bruni,S.Krishnababu, Foildiff: A hybrid diffusion transformer model for airfoil flow ...

work page doi:10.2514/1.j063440 2025

[2] [2]

doi:10.2514/6.2014-0080. J. Tran, K. Fukami, K. Inada, D. Umehara, Y. Ono, K. Ogawa, K. Taira, Aerodynamics-guided machine learning for design optimization of electric vehicles, Communications Engineering 3 (2024). doi:10.1038/ s44172-024-00322-0. V. Francés-Belda, A. Solera-Rico, J. Nieto-Centenero, E. Andrés, C. Sanmiguel Vila, R. Castellanos, Toward ae...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.2514/6.2014-0080 2014