betaFeedback

WorksMathematics

A Bayesian latent Gaussian process framework for aerodynamic uncertainty quantification

Abstract

Predicting the aerodynamic performance (e.g. lift, drag, and moment coefficients) of an aircraft is challenging – computational models are biased and direct simulations are prohibitive. A pragmatic way to overcome this limitation is by calibrating low-fidelity computational predictions with experimental measurements. This, however, requires calibrating against sparse measurements contaminated with uncertainty in both the control inputs and the measured aerodynamic response. We develop a methodology to address this problem based on Gaussian process surrogates and the classical Kennedy-O'Hagan calibration. A surrogate model learned on abundant-but-cheap low-fidelity data is calibrated with a sparse set of measurement data. Crucialy, we develop a Bayesian latent Gaussian process based approach that marginalizes the calibrated surrogate model over the input uncertainty, while also matching the marginal mean and variance of the measured output uncertainty. Once calibrated, our surrogate model predicts the uncertainty in aerodynamic coefficients with very high accuracy, including at extrapolative input settings. We validate our calibrated surrogate model predictions against measurement data with true uncertainty intervals to demonstrate that the model places 94.2−95.8%94.2-95.8\% of its predictive samples inside the released 95%95\% truth intervals, with endpoint cumulative probabilities very close to the nominal 0.025 and 0.975 levels.

Introduction

Computational aerodynamic tools are now used routinely in aircraft design, analysis, and certification studies, which makes uncertainty quantification (UQ) central to credible simulation-based prediction. Although the need for UQ is widely recognized, rigorous uncertainty propagation in computational fluid dynamics (CFD) remains difficult because aerodynamic predictions can depend on uncertain operating conditions, numerical approximations, surrogate-model error, and model-form discrepancy [26]. These challenges are especially pronounced when only sparse validation data are available. Verification, validation, and UQ (VVUQ) frameworks therefore emphasize the need to characterize parametric input uncertainty, numerical uncertainty, and model-form uncertainty together rather than treating them as separate afterthoughts [26].

Community benchmark problems provide a useful setting for developing and comparing such methods. The AIAA Fluid Dynamics Technical Committee’s Uncertainty Quantification Discussion Group (UQDG) has organized a series of UQ challenge problems for aerodynamics [10]. This paper addresses the Second Uncertainty Quantification Challenge Problem for Aerodynamics [12], which builds on the first challenge problem [11]. The first challenge asked whether a NACA 2412 airfoil with a 30% chord flap would satisfy specified lift and pitching-moment requirements under uncertain inputs. The second challenge shifts the focus from requirement satisfaction to predictive uncertainty: participants are given sparse truth data with epistemic uncertainty bounds and must predict uncertainty in aerodynamic coefficients at unseen operating conditions. This is the focus of the article and we present more detail as follows.

The UQ challenge problem

The second AIAA UQ challenge problem is anchored on the NACA 2412 airfoil at a nominal Reynolds number of 700,000, a 30% chord flap, and sea-level conditions; the resulting Mach number lies in the incompressible regime. The operating envelope is defined by angles of attack in [−5,10][-5, 10] degrees and flap angles in [−5,15][-5, 15] degrees, with epistemic uncertainty also assigned to the Reynolds number. Within this Reynolds number, angle-of-attack, and flap-angle parameter space, truth data are provided at seven operating conditions. At each of these locations, the aerodynamic coefficients – drag CdC_d, lift ClC_l, and quarter-chord pitching moment CmC_m – are supplied together with epistemic uncertainty intervals on both the input parameters and the coefficients; these data are summarized in Table 1.

α±0.02∘\alpha\pm0.02^\circβflap±0.1∘\beta_{\mathrm{flap}} \pm0.1^\circRe±3500\mathrm{Re} \pm3500Cl±0.009C_l \pm0.009Cd±0.0008C_d \pm0.0008Cm±0.008C_m \pm0.008
007×1057 \times10^50.2140.0121−0.047-0.047
507×1057 \times10^50.7370.0143−0.043-0.043
1007×1057 \times10^51.2090.0205−0.033-0.033
057×1057 \times10^50.5360.0131−0.094-0.094
0107×1057 \times10^50.8360.0157−0.137-0.137
557×1057 \times10^51.0410.0176−0.087-0.087
−5-5−5-57×1057 \times10^5−0.644-0.6440.0148−0.004-0.004

Table 1. Provided calibration data with epistemic uncertainty intervals

The computational model used to generate the initial low-fidelity data is XFOIL [16], a lightweight aerodynamic analysis tool appropriate for the operating conditions considered here. We use the XFOIL data to train surrogate models for the aerodynamic coefficients and then calibrate those surrogates with the truth data in Table 1. The resulting calibrated model is used to predict uncertainty in the aerodynamic coefficients at four unseen operating conditions, shown by the orange symbols in Figure 1. The released truth intervals for these four prediction locations [12] are used only for a posteriori assessment of the surrogate predictions.

Operating points for the UQ challenge problem

Figure 1. Operating points for the UQ challenge problem. The blue markers denote the seven calibration points where truth data are provided, and the orange markers denote the four prediction points where uncertainty estimates are required.

Background and literature review

XFOIL aerodynamics model for low-fidelity data generation

We use XFOIL v6.99 with a 100-panel baseline setup that accompanied the challenge problem. For cases at higher flap angles or angles of attack, where XFOIL may fail to converge, we incorporated a driver that first accumulates converged solutions at lower angles of attack for the same flap setting and then continues toward the desired condition. To train the XFOIL surrogate, we sample flap angle, angle of attack, and Reynolds number over the challenge envelope: angle of attack from −5∘-5^\circ to 10∘10^\circ, flap angle from −5∘-5^\circ to 15∘15^\circ, and Reynolds number from 665000 to 735000. A set of 100 training points is generated via Latin hypercube sampling [24], and XFOIL is used to compute the corresponding aerodynamic coefficients. These data define the simulator response that is emulated by the surrogate models reviewed next.

Gaussian process models

Gaussian process (GP) models are nonparametric probabilistic models that are particularly suited for small data setting in the presence of noise. GP models have been widely used for surrogate modeling in aerodynamics [35, 3], reliability analysis [33, 36, 7, 31], multidisciplinary analysis [43], fluid-structure interactions [1], and optimization [9, 32, 34]. Recently, Davis and Renganathan [15] applied them to aerodynamic UQ. Our approach builds on GP based regression and we now briefly review some of their background.

Let f:Rd→Rf : \mathbb{R}^{d} \to\mathbb{R} be a black-box function; we place a Gaussian process (GP) prior on ff,

f(x)∼GP(0,k(x,⋅)),f(\mathbf{x}) \sim\mathcal{GP}(0, k(\mathbf{x}, \cdot)),

where k(⋅,⋅):X×X→R+k(\cdot, \cdot) : \mathcal{X} \times\mathcal{X} \to\mathbb{R}_{+} denotes the covariance function (also known as the kernel). Observations from ff are modeled as yi=f(xi)+ϵiy_i = f(\mathbf{x}_i) + \epsilon_i for i=1,…,ni = 1, \ldots, n, where each ϵi\epsilon_i is assumed to be zero-mean Gaussian noise with unknown variance σ2\sigma^2.

Using the observed dataset Dn={xi,yi,σi2}i=1nD_n = \{\mathbf{x}_i, y_i, \sigma_i^2\}_{i=1}^n, we fit a posterior GP—parameterized by hyperparameters Ω\boldsymbol{\Omega}—to obtain the following conditional posterior distribution [30]:

Y(x)∣Dn,Ω∼GP(μn(x),σn2(x)),μn(x)=knT[Kn+τ−2I]−1yn,σn2(x)=k(x,x)−knT[Kn+τ−2I]−1kn,\begin{aligned} Y(\mathbf{x}) \mid D_n, \boldsymbol{\Omega} &\sim\mathcal{GP}\left(\mu_n(\mathbf{x}), \sigma_n^2(\mathbf{x})\right),\\ \mu_n(\mathbf{x}) &= \mathbf{k}_n^{\mathsf{T}}\left[\mathbf{K}_n+\tau^{-2}\mathbf{I}\right]^{-1}\mathbf{y}_n,\\ \sigma_n^2(\mathbf{x}) &= k(\mathbf{x},\mathbf{x})-\mathbf{k}_n^{\mathsf{T}}\left[\mathbf{K}_n+\tau^{-2}\mathbf{I}\right]^{-1}\mathbf{k}_n, \end{aligned}

where kn≡k(x,Xn)\mathbf{k}_n \equiv k(\mathbf{x},X_n) is the vector of covariances between the point x\mathbf{x} and all observation sites in DnD_n, Kn≡k(Xn,Xn)\mathbf{K}_n \equiv k(X_n,X_n) is the covariance matrix over the observed points, I\mathbf{I} is the identity matrix, and yn\mathbf{y}_n is the vector of corresponding outputs. The observation sites are denoted by Xn=[x1,…,xn]T∈Rn×dX_n = [\mathbf{x}_1,\ldots,\mathbf{x}_n]^{\mathsf{T}} \in\mathbb{R}^{n \times d}. Such “standard” GPs do not have a natural way of accommodating input uncertainty; they must be fit to some chosen value, typically the centroid, within the given epistemic range. To overcome this limitation, we introduce latent GPs (later in § 2.1), which treat the input as a random variable to rigorously account for uncertainty.

Literature review

GP surrogates provide a probabilistic representation of an expensive deterministic or noisy model response from a finite set of simulator evaluations [37, 38, 30, 18]. Their value for uncertainty quantification is twofold: the posterior mean provides a smooth response approximation, while the posterior covariance provides a local measure of surrogate uncertainty. These properties make GP surrogates a natural tool for aerodynamic UQ, where repeated evaluations of computational models may be expensive, convergence may be nonuniform across the input space, and uncertainty must be propagated through quantities of interest such as ClC_l, CdC_d, and CmC_m.

GP surrogates in model calibration. Model calibration refers to learning how much a model of a physical system falls short of predicting reality, so that the model can be corrected for improved predictive accuracy. The classical Bayesian calibration framework of Kennedy and O’Hagan models the physical response ζ(x)\zeta(\mathbf{x}) as the sum of a computer-model response and an additive model-discrepancy term [21]. GP models are a natural fit for model calibration under the Kennedy-O’Hagan (KOH) framework [21, 19, 20]. In a common notation,

yi=ζ(xi)+ϵi,ζ(x)=η(x,θ)+δ(x),(1)y_i = \zeta(\mathbf{x}_i) + \epsilon_i,\qquad\zeta(\mathbf{x}) = \eta(\mathbf{x},\theta) + \delta(\mathbf{x}), \tag*{(1)}

where η(x,θ)\eta(\mathbf{x},\theta) is the simulator evaluated at physical input x\mathbf{x} and calibration parameter θ\theta, δ(x)\delta(\mathbf{x}) is the model-form discrepancy, and ϵi\epsilon_i denotes observational error. In the original formulation, GP priors are used to represent the unknown response of the simulator and the discrepancy. When the simulator is inexpensive, η\eta may be evaluated directly during inference; when the simulator is expensive, a GP surrogate is first trained on a designed set of computer runs and then embedded within the calibration model. This surrogate-based interpretation is now standard in Bayesian calibration because it permits posterior inference over calibration parameters, discrepancy fields, and predictions without requiring repeated high-cost simulator calls.

Subsequent work extended the KOH construction to more complex data settings. Higdon et al. [19] demonstrated Bayesian calibration procedures that combine field observations with computer simulations for prediction and uncertainty quantification. The same line of work was extended to high-dimensional simulator output, where basis representations are combined with GP models to make calibration tractable for functional or spatially distributed quantities [20]. Bayarri et al. [4] developed a broader validation framework for computer models that separates calibration, discrepancy assessment, and prediction, emphasizing the need to account for both experimental uncertainty and simulator inadequacy. Modular Bayesian strategies, in which emulator training, calibration, and prediction are separated to improve stability or reduce feedback between model components, have also been proposed for computer-model analysis [23].

For computational fluid dynamics and aerodynamics, the KOH structure is particularly relevant because even carefully verified solvers may remain biased relative to physical truth due to turbulence modeling, transition modeling, geometry idealization, numerical discretization, or reduced-order physical assumptions. In such cases, a GP discrepancy model can be interpreted as a spatially correlated correction field over the aerodynamic input space. Multifidelity and model-form uncertainty studies in CFD have used related GP discrepancy constructions to combine low-fidelity simulations, higher-fidelity simulations, and available experimental data while propagating input uncertainty through the corrected predictive model [42]. This view is consistent with the present work: we train a surrogate model on XFOIL data and learn an additive correction model that represents the residual difference between XFOIL predictions and the supplied truth data.

A well-known limitation of the KOH framework is the potential lack of identifiability between calibration parameters and the discrepancy function. When δ(x)\delta(\mathbf{x}) is sufficiently flexible, it can absorb errors that might otherwise be attributed to θ\theta, leading to weakly identified or physically misleading calibration-parameter posteriors. This issue has been examined from several perspectives, including the role of discrepancy modeling in physical-parameter learning [8], model discrepancy and identifiability analysis [2], and asymptotic properties of calibration estimators for imperfect computer models [40, 41]. Alternative formulations, such as orthogonal or constrained discrepancy priors, have been proposed to reduce confounding between calibration parameters and discrepancy fields [28].

The present work uses the KOH idea primarily for predictive correction rather than for inference on tunable physical calibration parameters. The internal XFOIL settings are fixed, and the unknown term of interest is the discrepancy between the XFOIL-based aerodynamic response and the truth data. This distinction reduces the emphasis on physical-parameter identifiability and shifts the modeling objective toward calibrated prediction and uncertainty propagation. In this setting, the additive construction

g(x)=fX(x)+δ(x),(2)g(\mathbf{x}) = f_X(\mathbf{x}) + \delta(\mathbf{x}), \tag*{(2)}

where fXf_X is the GP surrogate of XFOIL and δ\delta is a GP discrepancy model, provides a direct surrogate-based analog of (1). Unlike common implementations that train the discrepancy only on centroid residuals, the proposed “distributional calibration” (more detail in Section 2.2) treats the truth-data intervals as finite-dimensional marginal constraints on the corrected GP. Thus, the truth data are used not only to correct the posterior mean, but also to calibrate the posterior variance of the corrected aerodynamic surrogate.

GP surrogates with uncertain inputs. Standard GP regression assumes that each input location is observed exactly. This assumption is restrictive in UQ settings where operating conditions, experimental settings, or geometric parameters are reported with measurement error, epistemic intervals, or incomplete knowledge. A direct response is to treat uncertain inputs probabilistically and integrate the GP prediction over the input distribution. Early work on GP prediction with uncertain inputs developed moment-matching and Gaussian approximation strategies for propagating input uncertainty through Bayesian kernel models, particularly for multi-step-ahead forecasting [17, 29]. Related methods derived analytic or approximate expressions for GP prediction under Gaussian input uncertainty, especially when squared-exponential kernels make the required integrals tractable [13]. McHutchon and Rasmussen [25] further showed that input noise can be interpreted, under local linearization, as a form of heteroscedastic output noise, leading to practical training procedures for GPs with noisy inputs.

A related but distinct class of models treats the input coordinates themselves as latent variables. In Gaussian process latent variable models (GP-LVMs), unknown latent coordinates are inferred jointly with the GP mapping, providing a probabilistic nonlinear dimension-reduction framework [22]. Bayesian extensions place priors on the latent coordinates and integrate over their uncertainty using variational inference or related approximations [39, 14]. Although GP-LVMs were originally developed for latent representations rather than physical input uncertainty, the same statistical idea is useful when the true physical input is unobserved but constrained by prior information. In the present aerodynamic setting, the latent variables are not abstract coordinates; they are physical quantities such as angle of attack, Reynolds number, and flap setting, constrained to lie within specified epistemic intervals.

Input uncertainty has also been considered directly in Bayesian calibration of time-consuming simulators. Perrin and Durantin [27] formulated calibration procedures that account for uncertain simulator inputs rather than assuming that field inputs are known exactly. Such approaches are important when discrepancies between simulation and observation can be caused either by model-form error or by uncertainty in the conditions at which the physical data were obtained. Collapsing an uncertain input interval to a midpoint can then bias the inferred discrepancy, underestimate predictive uncertainty, or produce an overconfident correction model.

The latent-input GP construction used in this work follows this general philosophy while targeting the epistemic interval structure of the AIAA UQ challenge data. For each observation, the recorded input is an interval or box UiU_i, and the true but unknown input x~i\tilde{x}_i is modeled as a latent variable supported on UiU_i. A prior density πi(x~i)\pi_i(\tilde{x}_i) encodes the epistemic information within the box, and the GP likelihood is evaluated at the latent locations rather than at fixed centroids. This formulation reduces to standard GP regression when each UiU_i collapses to a point, but it retains the ability to infer plausible input locations when the intervals have nonzero width. Computationally, the latent inputs can be inferred by full Bayesian sampling, variational inference, or maximum a posteriori optimization. The latter provides a practical compromise for the small truth-data regime considered here.

This latent-input treatment is complementary to the Kennedy–O’Hagan discrepancy model. The discrepancy GP accounts for model-form error between XFOIL and truth, while the latent-input GP accounts for epistemic uncertainty in the locations at which the truth data are observed. When the test inputs are also uncertain, prediction requires marginalization over the test input distribution in addition to conditioning on the learned latent training inputs. The resulting predictive uncertainty combines the GP posterior variance, the dispersion induced by uncertain inputs, and the calibrated discrepancy uncertainty. This is the main distinction between the present framework and a conventional surrogate-bias correction trained only at nominal input centroids.

Contributions

Together, these ideas lead to a Bayesian surrogate framework for aerodynamic UQ with epistemic uncertainty in both inputs and outputs. The main contributions of this paper are as follows.

  • We construct a Kennedy–O’Hagan-style additive correction model in which one GP surrogate represents the XFOIL response and a second GP represents the discrepancy between XFOIL and the supplied truth data.

  • We incorporate epistemic input intervals through a latent-input GP formulation, allowing the unknown true calibration inputs to be inferred within their prescribed uncertainty bounds rather than fixed at nominal centroids.

  • We introduce a distributional calibration step for the corrected GP so that the supplied truth intervals constrain both the posterior mean and the marginal predictive variance of the aerodynamic coefficients.

  • We demonstrate the predictive UQ capabilities of our method by comparing against the true uncertainty estimates at a set of prediction points.

  • Our software implementation is openly available at https://github.com/csdlpsu/uqchallenge2 for full reproducibility of our results.

The remainder of the manuscript develops these elements in sequence. § 2 presents the latent-input and distributionally calibrated GP methodology used in this work. § 3 describes the experimental setup and UQ results, and § 4 summarizes the main conclusions.

Methodology

Our overall approach involves three steps. First, using data generated by XFOIL, we learn a GP surrogate model for the aerodynamic coefficients CdC_d, ClC_l, and CmC_m as functions of angle of attack α\alpha, flap angle βflap\beta_{\mathrm{flap}}, and Reynolds number; the input domain is fixed as described above. Second, a Bayesian latent GP fitting step learns maximum a posteriori (MAP) estimates of the unknown input locations at the given 7 truth-data points. Third, a correction GP (δ\delta) is fit at the MAP estimates of the inputs and calibrated to match the given epistemic uncertainty intervals of the aerodynamic coefficients at the 7 truth points. We now present details of the Bayesian latent GP model fitting and the distributional model calibration approach, with their distinctions summarized in Figure 2.

Figure 2. Overview of the distinction between the first-moment and distributional calibration approaches.

Bayesian latent Gaussian process models

We now present a Gaussian process regression model for the common situation in which each recorded input is not known exactly but is instead provided as an epistemic interval (more generally, an axis–aligned box); in this work, we refer to this model class as “latent GPs”. Let d∈Nd \in\mathbb{N} denote the input dimension and suppose we observe pairs {(Ui,yi)}i=1n\{(\mathcal{U}_i,y_i)\}_{i=1}^{n}, where yi∈Ry_i \in\mathbb{R} is a scalar response and Ui∈∏j=1d[ℓji,uji]⊂Rd\mathcal{U}_i \in\prod_{j=1}^{d}[\ell_j^i,u_j^i] \subset\mathbb{R}^{d} is the set within which the unknown true input lies. We write x~i∈Rd\tilde{x}_i \in\mathbb{R}^{d} for this unknown true input, with x~i∈Ui\tilde{x}_i \in\mathcal{U}_i. To express the epistemic uncertainty in x~i\tilde{\mathbf{x}}_i, we place a prior density πi\pi_i on x~i\tilde{\mathbf{x}}_i supported on Ui\mathcal{U}_i; then, we estimate x~\tilde{\mathbf{x}} from data via Bayesian inference.

Recall that we place a GP prior on ff, and observe its noisy response according to

yi=f(x~i)+εi,εi∼iidN(0,σ2).y_i = f(\tilde{\mathbf{x}}_i) + \varepsilon_i,\qquad\varepsilon_i \mathrel{\mathop{\sim}^{\mathrm{iid}}} \mathcal{N}(0,\sigma^2).

Let X~=[x~1,…,x~n]T∈Rn×d\tilde{X} = [\tilde{\mathbf{x}}_1,\ldots,\tilde{\mathbf{x}}_n]^{\mathsf{T}} \in\mathbb{R}^{n \times d} and y=(y1,…,yn)T∈Rn\mathbf{y} = (y_1,\ldots,y_n)^{\mathsf{T}} \in\mathbb{R}^n. Then, the marginal likelihood of y\mathbf{y} given X~\tilde{X}, θ\theta, and σ2\sigma^2 is multivariate normal,

y∣X~,θ,σ2∼N(μ(X~),Kθ(X~,X~)+σ2In),(3)\mathbf{y} \mid\tilde{X},\theta,\sigma^2 \sim\mathcal{N}\left(\boldsymbol{\mu}(\tilde{X}), K_\theta(\tilde{X},\tilde{X})+\sigma^2 I_n\right), \tag*{(3)}

where μ(X~)=(μ(x~1),…,μ(x~n))T\boldsymbol{\mu}(\tilde{X}) = (\mu(\tilde{\mathbf{x}}_1),\ldots,\mu(\tilde{\mathbf{x}}_n))^{\mathsf{T}}, Kθ(X~,X~)∈Rn×nK_\theta(\tilde{X},\tilde{X}) \in\mathbb{R}^{n \times n} has (i,j)(i,j) entry kθ(x~i,x~j)k_\theta(\tilde{\mathbf{x}}_i,\tilde{\mathbf{x}}_j), and InI_n is the n×nn \times n identity matrix.

In order to respect the box constraints x~i∈Ui\tilde{\mathbf{x}}_i \in\mathcal{U}_i during gradient-based inference, it is convenient to reparameterize each latent input with an unconstrained vector ui∈Rd\mathbf{u}_i \in\mathbb{R}^d and a smooth, invertible map into the box. A simple choice is the elementwise logistic (sigmoid) map. Writing ℓi=(ℓi1,…,ℓid)T\boldsymbol{\ell}_i = (\ell_i^1,\ldots,\ell_i^d)^{\mathsf{T}} and upi=(ui1,…,uid)T\mathbf{u}\mathbf{p}_i = (u_i^1,\ldots,u_i^d)^{\mathsf{T}} for the lower and upper corners of Ui\mathcal{U}_i, and σ(t)=1/(1+e−t)\sigma(t) = 1/(1+e^{-t}) for the logistic function applied coordinate-wise defines

x~i(ui)=ℓi+σ(ui)⊙(upi−ℓi),(4)\tilde{\mathbf{x}}_i(\mathbf{u}_i) = \boldsymbol{\ell}_i + \sigma(\mathbf{u}_i) \odot(\mathbf{u}\mathbf{p}_i-\boldsymbol{\ell}_i), \tag*{(4)}

where ⊙\odot denotes the Hadamard (elementwise) product. The transformation (4) guarantees x~i(ui)∈Ui\tilde{\mathbf{x}}_i(\mathbf{u}_i) \in\mathcal{U}_i for all ui∈Rd\mathbf{u}_i \in\mathbb{R}^d, and the chain rule carries gradients ∂/∂ui\partial/\partial\mathbf{u}_i through to the marginal likelihood in (3).

The full Bayesian target is the joint posterior density of the unknowns given the data:

p(X~,θ,σ2∣y)∝p(y∣X~,θ,σ2) p(X~) p(θ)p(σ2),(5)p(\tilde{X},\theta,\sigma^2 \mid\mathbf{y}) \propto p(\mathbf{y} \mid\tilde{X},\theta,\sigma^2)\,p(\tilde{X})\,p(\theta)p(\sigma^2), \tag*{(5)}

where p(X~)=∏i=1nπi(x~i)p(\tilde{X}) = \prod_{i=1}^{n} \pi_i(\tilde{\mathbf{x}}_i) encodes the epistemic beliefs on the inputs. One may explore (5) by Markov chain Monte Carlo (MCMC), such as Hamiltonian Monte Carlo (HMC) [5], or by variational inference (VI) [6]. These methods propagate uncertainty in both the inputs and the hyperparameters, but they are computationally intensive for large nn because each evaluation of the Gaussian likelihood involves a linear solve with Kθ(X~,X~)+σ2InK_\theta(\tilde{X},\tilde{X})+\sigma^2 I_n.

A pragmatic alternative is maximum a posteriori (MAP) estimation of X~\tilde{X} together with the hyperparameters. Let u=(u1,…,un)\mathbf{u} = (\mathbf{u}_1,\ldots,\mathbf{u}_n) collect the unconstrained variables and write X~(u)\tilde{X}(\mathbf{u}) for the corresponding latent inputs via (4). The negative log-posterior objective is

J(u,θ,σ2)=12(y−μ(X~(u)))T[Kθ(X~(u),X~(u))+σ2In]−1(y−μ(X~(u)))+12log⁡det⁡[Kθ(X~(u),X~(u))+σ2In]+n2log⁡(2π)−log⁡p(X~(u))−log⁡p(θ)−log⁡p(σ2).(6)\begin{aligned} \mathcal{J}(\mathbf{u},\theta,\sigma^2) ={}& \frac{1}{2}\bigl(\mathbf{y}-\boldsymbol{\mu}(\tilde{X}(\mathbf{u}))\bigr)^{\mathsf{T}}\left[K_\theta(\tilde{X}(\mathbf{u}),\tilde{X}(\mathbf{u}))+\sigma^2 I_n\right]^{-1}\bigl(\mathbf{y}-\boldsymbol{\mu}(\tilde{X}(\mathbf{u}))\bigr) \\ &+ \frac{1}{2}\log\det\left[K_\theta(\tilde{X}(\mathbf{u}),\tilde{X}(\mathbf{u}))+\sigma^2 I_n\right] + \frac{n}{2}\log(2\pi) - \log p(\tilde{X}(\mathbf{u})) \\ &- \log p(\theta) - \log p(\sigma^2). \tag*{(6)} \end{aligned}

The first quadratic term in (6) measures data misfit under the GP prior, the log-determinant penalizes model complexity by the marginal predictive covariance volume at the training locations, and the final three terms add the contributions of the input and hyperparameter priors. A stable and effective optimization strategy alternates between two differentiable steps: with the latent inputs fixed, one maximizes the marginal likelihood over (θ,σ2)(\theta,\sigma^2) using standard GP training; with (θ,σ2)(\theta,\sigma^2) fixed, one takes a few gradient steps in u\mathbf{u} to reduce J\mathcal{J}. The logistic transformation (4) ensures the box constraints are maintained throughout and allows gradients to flow to u\mathbf{u} by the chain rule. In practice it is numerically advantageous, once the alternating steps have converged, to rebuild the GP at the final X~\widetilde{X} and refit (θ,σ2)(\theta, \sigma^2) before computing predictions.

For a fixed test input x⋆∈Rd\mathbf{x}_{\star} \in\mathbb{R}^{d}, the GP posterior is Gaussian. Writing k⋆=Kθ(X~,x⋆)∈Rn\mathbf{k}_{\star} = K_{\theta}(\widetilde{X}, \mathbf{x}_{\star}) \in\mathbb{R}^{n} for the vector of covariances between the training inputs and the test input, and k⋆⋆=kθ(x⋆,x⋆)k_{\star\star} = k_{\theta}(\mathbf{x}_{\star}, \mathbf{x}_{\star}) for the prior variance at the test input, the posterior mean and variance of f(x⋆)f(\mathbf{x}_{\star}) are

μ~(x⋆)=μ(x⋆)+k⋆T[Kθ(X~,X~)+σ2In]−1(y−μ(X~)),\widetilde{\mu}(\mathbf{x}_{\star}) = \mu(\mathbf{x}_{\star}) + \mathbf{k}_{\star}^{\mathsf{T}}\left[K_{\theta}(\widetilde{X}, \widetilde{X}) + \sigma^{2}I_{n}\right]^{-1}\left(\mathbf{y} - \mu(\widetilde{X})\right),
s2(x⋆)=k⋆⋆−k⋆T[Kθ(X~,X~)+σ2In]−1k⋆.(7)s^{2}(\mathbf{x}_{\star}) = k_{\star\star} - \mathbf{k}_{\star}^{\mathsf{T}}\left[K_{\theta}(\widetilde{X}, \widetilde{X}) + \sigma^{2}I_{n}\right]^{-1}\mathbf{k}_{\star}. \tag*{(7)}

Each symbol in (7) is defined in terms of the learned latent inputs X~\widetilde{X} and the fitted hyperparameters. When the test input is itself epistemically uncertain, specified by a target box U⋆\mathcal{U}_{\star} and an associated belief density π⋆\pi_{\star} on U⋆\mathcal{U}_{\star}, one reports a prediction that is marginalized over this uncertainty. The law of total expectation and the law of total variance yield

μ‾⋆=Ex⋆∼π⋆[μ~(x⋆)],s‾⋆2=Ex⋆∼π⋆[s2(x⋆)]+Var⁡x⋆∼π⋆(μ~(x⋆)),(8)\overline{\mu}_{\star} = \mathbb{E}_{\mathbf{x}_{\star}\sim\pi_{\star}}\left[\widetilde{\mu}(\mathbf{x}_{\star})\right], \qquad\overline{s}^{2}_{\star} = \mathbb{E}_{\mathbf{x}_{\star}\sim\pi_{\star}}\left[s^{2}(\mathbf{x}_{\star})\right] + \operatorname{Var}_{\mathbf{x}_{\star}\sim\pi_{\star}}\left(\widetilde{\mu}(\mathbf{x}_{\star})\right), \tag*{(8)}

where the first term in s‾⋆2\overline{s}^{2}_{\star} represents the average conditional uncertainty of the GP at a fixed input and the second term quantifies the additional dispersion due to the epistemic spread of the input itself. When π⋆\pi_{\star} collapses to a point mass at a known x⋆\mathbf{x}_{\star}, the second term vanishes, and (8) reduces to (7). The expectations in (8) are efficiently approximated by low–order tensor product Gaussian quadrature in low dimensions, by sparse grids in moderate dimensions, or by stratified Monte Carlo sampling when a non–uniform π⋆\pi_{\star} is required. In our work, we draw iid samples from π⋆\pi_{\star} and take a Monte Carlo average to compute the expectation and variance in (8). We show an illustration of the latent GP on a simple 1D function in Figure 3 – the horizontal lines indicate the input epistemic intervals and the orange circle is the MAP estimate of the input.

Illustration of the Bayesian latent GP approach for fitting data with epistemic uncertainty in inputs and outputs

Figure 3. Illustration of the Bayesian latent GP approach for fitting data with epistemic uncertainty in inputs and outputs. The orange circles represent MAP estimates of the unknown input location x~\tilde{\mathbf{x}}; the horizontal lines represent the epistemic intervals that express prior uncertainty over x\mathbf{x}.

The latent-input construction seamlessly reduces to standard GP regression in the limit that each box UiU_i shrinks to a singleton. For wider boxes there is an intuitive confounding between large kernel lengthscales and diffuse input locations; this is mitigated by weakly informative priors on the lengthscales (for example, log-normal priors centered on the physical scales of the inputs) and by modest shrinkage of x~i\widetilde{\mathbf{x}}_i toward the box midpoints through the prior p(X~)p(\widetilde{X}).

The latent GP provides a MAP estimate of x~\widetilde{\mathbf{x}}, which is then used to fit standard GPs at the MAP estimates of the inputs for downstream UQ. Unlike existing approaches that calibrate with the moments of the given truth data, we calibrate with the full distribution of the given data, as explained next.

Distributional calibration of the correction Gaussian process

The purpose of the distributional calibration method is to use the available truth-data bounds not only to correct the posterior mean, but also to calibrate the posterior uncertainty of the corrected model. In particular, at each truth-data input location (that is, the MAP estimates from the latent GP), the calibrated corrected GP is constructed so that μcal(xi)=ci\mu_{\mathrm{cal}}(\mathbf{x}_i) = c_i and μcal(xi)±z0.975σcal(xi)=[Li,Ui]\mu_{\mathrm{cal}}(\mathbf{x}_i) \pm z_{0.975}\sigma_{\mathrm{cal}}(\mathbf{x}_i) = [L_i, U_i], where [Li,Ui][L_i, U_i] is the given epistemic truth interval, cic_i is its centroid, and z0.975=1.959963984…z_{0.975} = 1.959963984\ldots is the standard normal quantile associated with a two-sided pointwise 95% Gaussian interval. The same construction is applied separately to each scalar aerodynamic response: CdC_d, ClC_l, and CmC_m.

For convenience, we denote the MAP estimate of x~\tilde{\mathbf{x}} as xT\mathbf{x}^{T} and introduce the following notation. Let the XFOIL training data be DX={(xjX,yjX)}j=1NXD_X = \{(\mathbf{x}^{X}_j, y^{X}_j)\}_{j=1}^{N_X}, and let the truth-data bounds be DT={(xiT,[Li,Ui])}i=1NTD_T = \{(\mathbf{x}^{T}_i, [L_i, U_i])\}_{i=1}^{N_T}. Define the matrix of truth-data input locations as XT=[x1T,…,xNTT]⊤\mathbf{X}_T = [\mathbf{x}^{T}_1, \ldots, \mathbf{x}^{T}_{N_T}]^{\top}. For each truth-data interval, define the centroid and half-width as ci=(Li+Ui)/2c_i = (L_i + U_i)/2 and hi=(Ui−Li)/2h_i = (U_i - L_i)/2. As previously mentioned, we treat the given epistemic intervals as 95% Gaussian credible intervals; then, the target marginal variance for the composite GP at xiT\mathbf{x}^{T}_i is

vi⋆=(hiz0.975)2,v_i^\star= \left(\frac{h_i}{z_{0.975}}\right)^2,

which follows from the fact that the half-width hi=z0.975σcalh_i = z_{0.975}\sigma_{\mathrm{cal}}. More generally, for a two-sided pointwise probability level pp,

zp=Φ−1(1+p2),vi⋆=(hizp)2,z_p = \Phi^{-1}\left(\frac{1+p}{2}\right), \qquad v_i^\star= \left(\frac{h_i}{z_p}\right)^2,

where Φ−1\Phi^{-1} is the inverse standard normal cumulative distribution function. Stack the truth centroids into c=[c1,…,cNT]⊤\mathbf{c} = [c_1, \ldots, c_{N_T}]^{\top}, and denote the target covariance matrix at the truth-data locations by VT\mathbf{V}_T. The simplest choice is a diagonal target covariance,

VT=diag⁡(v1⋆,…,vNT⋆),(9)\mathbf{V}_T = \operatorname{diag}(v_1^\star, \ldots, v_{N_T}^\star), \tag*{(9)}

which assumes independence of the individual truth measurements. Because the composite GP is an additive combination of two GPs, the resulting composite GP g0(x)=fX(x)+δ0(x)g_0(\mathbf{x}) = f_X(\mathbf{x}) + \delta_0(\mathbf{x}) is also a GP under the mild assumption that fXf_X and δ0\delta_0 are independent. Then,

g0(x)∼GP(m0(x),k0(x,x′)),g_0(\mathbf{x}) \sim\mathcal{GP}\left(m_0(\mathbf{x}), k_0(\mathbf{x},\mathbf{x}')\right),

with

m0(x)=μX(x)+mδ(x),k0(x,x′)=kX(x,x′)+kδ(x,x′).m_0(\mathbf{x}) = \mu_X(\mathbf{x}) + m_\delta(\mathbf{x}), \qquad k_0(\mathbf{x},\mathbf{x}') = k_X(\mathbf{x},\mathbf{x}') + k_\delta(\mathbf{x},\mathbf{x}').

A “first-moment” calibration uses the centroid mismatch

di=ci−μX(xiT)d_i = c_i - \mu_X(\mathbf{x}_i^T)

as training data for the correction GP. This enforces a mean correction but does not use the interval width Ui−LiU_i - L_i. The distributional calibration method instead treats the truth bounds as a target marginal distribution for the corrected GP at the truth-data input locations.

Let

gT=g0(XT)=[g0(x1T),…,g0(xNTT)]T.\mathbf{g}_T = g_0(X_T) = \left[g_0(\mathbf{x}_1^T), \ldots, g_0(\mathbf{x}_{N_T}^T)\right]^{\mathsf{T}}.

Under the uncalibrated corrected GP,

gT∼N(m0T,KTT),where m0T=m0(XT),KTT=k0(XT,XT).\mathbf{g}_T \sim\mathcal{N}\left(\mathbf{m}_{0T}, \mathbf{K}_{TT}\right), \quad\text{where } \mathbf{m}_{0T} = m_0(X_T), \qquad\mathbf{K}_{TT} = k_0(X_T,X_T).

The idea is to replace gT∼N(m0T,KTT)\mathbf{g}_T \sim\mathcal{N}\left(\mathbf{m}_{0T}, \mathbf{K}_{TT}\right) with the distributional calibration target:

gT∼N(c,VT).(10)\mathbf{g}_T \sim\mathcal{N}\left(\mathbf{c}, \mathbf{V}_T\right). \tag*{(10)}

The overall calibrated distribution is defined by replacing the prior finite-dimensional marginal distribution of gT\mathbf{g}_T with the target distribution in (10), while preserving the original conditional distribution of the process away from XTX_T. Equivalently,

pcal(g∣DX,DT)=∫p0(g∣gT=z,DX)qT(z) dz,p_{\mathrm{cal}}\left(g \mid D_X,D_T\right) = \int p_0\left(g \mid\mathbf{g}_T = \mathbf{z},D_X\right)q_T(\mathbf{z})\,d\mathbf{z},

where

qT(z)=N(z;c,VT).q_T(\mathbf{z}) = \mathcal{N}\left(\mathbf{z};\mathbf{c},\mathbf{V}_T\right).

That is, we first draw a sample z\mathbf{z} from the target truth distribution N(c,VT)\mathcal{N}\left(\mathbf{c},\mathbf{V}_T\right). Then, conditional on gT=z\mathbf{g}_T = \mathbf{z}, we draw the rest of the function using the original GP conditional distribution. This changes the marginal distribution at XTX_T, but leaves the original spatial covariance structure intact away from those points. Therefore, crucially, we calibrate our surrogate model with both the centroid and epistemic intervals of the given 7 truth points.

Closed-form calibrated mean and covariance

The calibrated composite GP has closed form. For any prediction input x\mathbf{x}, define kxT=k0(x,XT)=[k0(x,x1T),…,k0(x,xNTT)]\mathbf{k}_{xT} = k_0(\mathbf{x},X_T) = \left[k_0(\mathbf{x},\mathbf{x}_1^T),\ldots,k_0(\mathbf{x},\mathbf{x}_{N_T}^T)\right]. Similarly, kTx′=k0(XT,x′)\mathbf{k}_{Tx'} = k_0(X_T,\mathbf{x}'). Since the covariance matrix KTT\mathbf{K}_{TT} is positive definite, the inverse below is evaluated through a Cholesky solve. The calibrated corrected GP is

gcal(x)∼GP(mcal(x),kcal(x,x′)),g_{\mathrm{cal}}(\mathbf{x}) \sim\mathcal{GP}\left(m_{\mathrm{cal}}(\mathbf{x}), k_{\mathrm{cal}}(\mathbf{x},\mathbf{x}')\right),

with mean

mcal(x)=m0(x)+kxTKTT−1(c−m0T),(11)m_{\mathrm{cal}}(\mathbf{x}) = m_0(\mathbf{x}) + \mathbf{k}_{x_T}\mathbf{K}_{TT}^{-1}\left(\mathbf{c} - \mathbf{m}_{0T}\right), \tag*{(11)}

and covariance

kcal(x,x′)=k0(x,x′)−kxTKTT−1kTx′+kxTKTT−1VTKTT−1kTx′.k_{\mathrm{cal}}(\mathbf{x},\mathbf{x}') = k_0(\mathbf{x},\mathbf{x}') - \mathbf{k}_{x_T}\mathbf{K}_{TT}^{-1}\mathbf{k}_{Tx'} + \mathbf{k}_{x_T}\mathbf{K}_{TT}^{-1}\mathbf{V}_T\mathbf{K}_{TT}^{-1}\mathbf{k}_{Tx'}.

For a finite prediction set X⋆=[x1⋆,…,xN⋆⋆]T\mathbf{X}_\star= [\mathbf{x}_1^\star,\ldots,\mathbf{x}_{N_\star}^\star]^\mathsf{T}, the corresponding matrix form is

mcal,⋆=m0,⋆+K⋆TKTT−1(c−m0T),\mathbf{m}_{\mathrm{cal},\star} = \mathbf{m}_{0,\star} + \mathbf{K}_{\star T}\mathbf{K}_{TT}^{-1}\left(\mathbf{c} - \mathbf{m}_{0T}\right),

and

Kcal,⋆⋆=K⋆⋆−K⋆TKTT−1KT⋆+K⋆TKTT−1VTKTT−1KT⋆.\mathbf{K}_{\mathrm{cal},\star\star} = \mathbf{K}_{\star\star} - \mathbf{K}_{\star T}\mathbf{K}_{TT}^{-1}\mathbf{K}_{T\star} + \mathbf{K}_{\star T}\mathbf{K}_{TT}^{-1}\mathbf{V}_T\mathbf{K}_{TT}^{-1}\mathbf{K}_{T\star}.

We provide details of the derivation in Section A.2.

Verification of the calibration conditions

Evaluating Eq. (11) at the truth-data locations gives

mcal(XT)=m0T+KTTKTT−1(c−m0T)=c.\mathbf{m}_{\mathrm{cal}}(\mathbf{X}_T) = \mathbf{m}_{0T} + \mathbf{K}_{TT}\mathbf{K}_{TT}^{-1}\left(\mathbf{c} - \mathbf{m}_{0T}\right) = \mathbf{c}.

Thus the calibrated corrected GP posterior mean exactly matches the centroid of the truth-data epistemic interval at every truth-data input location. Similarly, evaluating calibrated variance at X⋆=XT\mathbf{X}_\star= \mathbf{X}_T gives

Kcal(XT,XT)=KTT−KTTKTT−1KTT+KTTKTT−1VTKTT−1KTT=VT.\mathbf{K}_{\mathrm{cal}}(\mathbf{X}_T,\mathbf{X}_T) = \mathbf{K}_{TT} - \mathbf{K}_{TT}\mathbf{K}_{TT}^{-1}\mathbf{K}_{TT} + \mathbf{K}_{TT}\mathbf{K}_{TT}^{-1}\mathbf{V}_T\mathbf{K}_{TT}^{-1}\mathbf{K}_{TT} = \mathbf{V}_T.

Therefore, for the diagonal target covariance in Eq. (9),

σcal2(xiT)=vi⋆=(Ui−Li2z0.975)2.\sigma_{\mathrm{cal}}^2(\mathbf{x}_i^T) = v_i^\star= \left(\frac{U_i-L_i}{2z_{0.975}}\right)^2.

It follows that

mcal(xiT)−z0.975σcal(xiT)=ci−z0.975hiz0.975=ci−hi=Li,m_{\mathrm{cal}}(\mathbf{x}_i^T) - z_{0.975}\sigma_{\mathrm{cal}}(\mathbf{x}_i^T) = c_i - z_{0.975}\frac{h_i}{z_{0.975}} = c_i-h_i=L_i,
mcal(xiT)+z0.975σcal(xiT)=ci+z0.975hiz0.975=ci+hi=Ui.m_{\mathrm{cal}}(\mathbf{x}_i^T) + z_{0.975}\sigma_{\mathrm{cal}}(\mathbf{x}_i^T) = c_i + z_{0.975}\frac{h_i}{z_{0.975}} = c_i+h_i=U_i.

Hence the pointwise 95% Gaussian bounds of the calibrated corrected GP coincide exactly with the epistemic truth bounds at the truth-data locations. The full methodology is summarized in Algorithm 1.

Monte Carlo simulation to propagate uncertainty

After distributional calibration, each scalar aerodynamic coefficient q∈{Cd,Cl,Cm}q \in\{C_d,C_l,C_m\} is represented by a calibrated corrected Gaussian process gq,cal(x)∼GP(mq,cal(x),kq,cal(x,x′))g_{q,\mathrm{cal}}(\mathbf{x}) \sim\mathcal{GP}\left(m_{q,\mathrm{cal}}(\mathbf{x}), k_{q,\mathrm{cal}}(\mathbf{x},\mathbf{x}')\right). For notational compactness, write mq(x)=mq,cal(x)m_q(\mathbf{x}) = m_{q,\mathrm{cal}}(\mathbf{x}) and vq(x)=kq,cal(x,x)v_q(\mathbf{x}) = k_{q,\mathrm{cal}}(\mathbf{x},\mathbf{x}). At a fixed input condition, the calibrated predictive distribution is therefore q⋆∣x⋆,DX,DT∼N(mq(x⋆),vq(x⋆))q_\star\mid\mathbf{x}_\star, D_X, D_T \sim\mathcal{N}\left(m_q(\mathbf{x}_\star),v_q(\mathbf{x}_\star)\right).

Algorithm 1 Latent input distributional calibration and Monte Carlo propagation

Require: XFOIL data DXq={(xjX,yjX,q)}j=1NXD_X^q = \{(x_j^X,y_j^{X,q})\}_{j=1}^{N_X} and truth data DTq={(UiT,[Liq,Uiq])}i=1NTD_T^q = \{(U_i^T,[L_i^q,U_i^q])\}_{i=1}^{N_T} for q∈Q={Cd,Cl,Cm}q \in\mathcal{Q} = \{C_d,C_l,C_m\}; target input box U⋆U_\star with belief density π⋆\pi_\star; level p=0.95p=0.95; sample size NMCN_{\mathrm{MC}}; jitter η>0\eta>0.

Ensure: Predicted empirical PDFs, CDFs, and interval mass for each qq.

  1. for q∈Qq \in\mathcal{Q} do

  2. Fit the XFOIL surrogate fXq(x)∣DXq∼GP(μXq(x),kXq(x,x′))f_X^q(\mathbf{x}) \mid D_X^q \sim\mathcal{GP}(\mu_X^q(\mathbf{x}),k_X^q(\mathbf{x},\mathbf{x}')).

  3. Infer latent truth inputs by MAP: XT,q=[x~1,q,…,x~NT,q]⊤\mathbf{X}_{T,q}=[\tilde{\mathbf{x}}_{1,q},\ldots,\tilde{\mathbf{x}}_{N_T,q}]^\top.

  4. Choose an uncalibrated discrepancy GP δq(x)∼GP(mδq(x),kδq(x,x′))\delta_q(\mathbf{x}) \sim\mathcal{GP}(m_{\delta}^q(\mathbf{x}),k_{\delta}^q(\mathbf{x},\mathbf{x}')).

  5. Convert output intervals to target moments: ciq=(Liq+Uiq)/2c_i^q=(L_i^q+U_i^q)/2, hiq=(Uiq−Liq)/2h_i^q=(U_i^q-L_i^q)/2, vi⋆,q=(hiq/zp)2v_i^{\star,q}=(h_i^q/z_p)^2, with zp=Φ−1((1+p)/2)z_p=\Phi^{-1}((1+p)/2). Set cq=(c1q,…,cNTq)⊤\mathbf{c}_q=(c_1^q,\ldots,c_{N_T}^q)^\top and VT,q=diag⁡(v1⋆,q,…,vNT⋆,q)V_{T,q}=\operatorname{diag}(v_1^{\star,q},\ldots,v_{N_T}^{\star,q}).

  6. Compute m0Tq=mδq(XT,q)\mathbf{m}_{0T}^q=m_\delta^q(\mathbf{X}_{T,q}) and K~TTq=kδq(XT,q,XT,q)+ηI\widetilde{K}_{TT}^q=k_\delta^q(\mathbf{X}_{T,q},\mathbf{X}_{T,q})+\eta I. Solve K~TTqαq=cq−m0Tq\widetilde{K}_{TT}^q\boldsymbol{\alpha}_q=\mathbf{c}_q-\mathbf{m}_{0T}^q.

  7. for r=1,…,NMCr=1,\ldots,N_{\mathrm{MC}} do

  8. Draw x⋆(r)∼π⋆\mathbf{x}_\star^{(r)}\sim\pi_\star on U⋆U_\star and ξ(r)∼N(0,1)\xi^{(r)}\sim\mathcal{N}(0,1).

  9. Set krTq=kδq(x⋆(r),XT,q)\mathbf{k}_{rT}^q=k_\delta^q(\mathbf{x}_\star^{(r)},\mathbf{X}_{T,q}) and solve K~TTqbqr=(krTq)⊤\widetilde{K}_{TT}^q\mathbf{b}_q^r=(\mathbf{k}_{rT}^q)^\top.

  10. Evaluate the calibrated predictive moments mrq=m0q(x⋆(r))+krTqαqm_r^q=m_{0}^q(\mathbf{x}_\star^{(r)})+\mathbf{k}_{rT}^q\boldsymbol{\alpha}_q and vrq=kδq(x⋆(r),x⋆(r))−krTqbqr+(bqr)⊤VT,qbqrv_r^q=k_\delta^q(\mathbf{x}_\star^{(r)},\mathbf{x}_\star^{(r)})-\mathbf{k}_{rT}^q\mathbf{b}_q^r+(\mathbf{b}_q^r)^\top V_{T,q}\mathbf{b}_q^r.

  11. Draw the posterior predictive sample q⋆(r)=mrq+max⁡(vrq,0)ξ(r)q_\star^{(r)}=m_r^q+\sqrt{\max(v_r^q,0)}\xi^{(r)}.

  12. end for

  13. Report empirical PDF, CDF, and interval mass.

  14. end for

For the 4 given prediction points of interest, recall that we have an epistemic uncertainty interval, to which we assign a belief density π⋆\pi_\star supported on that interval. Propagating this input uncertainty through the calibrated GP gives the marginal predictive distribution

p(q⋆∣DX,DT,U⋆)=∫x∈U⋆N(q⋆;mq(x),vq(x))π⋆(x) dx.p(q_\star\mid D_X,D_T,U_\star)=\int_{\mathbf{x}\in U_\star}\mathcal{N}(q_\star;m_q(\mathbf{x}),v_q(\mathbf{x}))\pi_\star(\mathbf{x})\,\mathrm{d}\mathbf{x}.

Although the conditional prediction at any fixed input is Gaussian, this marginal distribution is generally non-Gaussian because both the mean and variance of the calibrated GP vary over the uncertain input region.

To estimate the full predictive distribution of the quantities of interest, we draw independent samples x⋆(r)∼π⋆\mathbf{x}_\star^{(r)}\sim\pi_\star, r=1,…,NMCr=1,\ldots,N_{\mathrm{MC}}, and evaluate the calibrated GP at each sample. Then, we additionally draw ξ(r)∼N(0,1)\xi^{(r)}\sim\mathcal{N}(0,1) and form posterior predictive samples as

q⋆(r)=mq(r)+vq(r)ξ(r).q_\star^{(r)}=m_q^{(r)}+\sqrt{v_q^{(r)}}\xi^{(r)}.

The empirical distribution of {q⋆(r)}r=1NMC\{q_\star^{(r)}\}_{r=1}^{N_{\mathrm{MC}}} is then used to construct predictive densities, cumulative distribution functions, and credible intervals. For example, the reported 95% uncertainty interval is taken as the empirical 2.5% and 97.5% quantiles of these samples. The same procedure is applied independently to the calibrated GP models for CdC_d, ClC_l, and CmC_m.

Experiments

Our primary goal with the experiments is to estimate the uncertainty at the 4 prediction points (orange circles in Figure 1) in the form of probability density functions and cumulative distribution functions. The true uncertainties at these points were reported in [12 Figs. 2–5] as epistemic bounds. We treat these bounds as 95% confidence bounds, consistent with how we treated the bounds on the given truth data. The XFOIL GP is trained on 100 Latin hypercube designs and then calibrated using the given 7 truth points. Once calibrated, the uncertainty estimates are obtained from 10,000 samples drawn uniformly at random from the input uncertainty bounds and propagated through the calibrated model as described in the previous section.

Model validation

We validate our GP surrogates in two stages. In the first stage, we validate the XFOIL-data-fit GP posterior mean predictions against a held-out set of 100 XFOIL data points; this is shown in the top row of Figure 4. In the second stage, the composite GP posterior mean predictions are validated against the 7 validation points; this is included as the bottom row of Figure 4. This figure establishes accuracy in both the XFOIL GP and the overall calibrated GP we fit.

Validation plots for the XFOIL-data-fit and composite Gaussian processes

Figure 4. Top row: validation of our XFOIL-data-fit GP on a held-out test set. Bottom row: validation of our composite GP on the 7 data points in Table 1; the squares and triangles denote the vanilla and latent GP results, respectively.

UQ results

Figures 5 to 8 compare the Monte Carlo predictive distributions obtained from the first-moment-calibrated model and the distribution-calibrated model at the four prediction locations. In each panel, the red vertical lines and shaded region denote the released epistemic bounds for the corresponding quantity of interest, while the colored histograms show samples from the calibrated predictive distribution after propagation of input uncertainty. Thus, a well-calibrated result should do more than place its mean near the center of the red interval; it should also assign approximately the correct amount of probability mass to that interval.

The distribution-calibrated results satisfy this criterion consistently. Across all four prediction locations and all three aerodynamic coefficients, the green predictive distributions are centered near the provided epistemic interval centroids. This agreement is visible in the individual histogram comparisons: for ClC_l, CdC_d, and CmC_m, the green samples span nearly the same range as the truth intervals, including at the off-design point α=7∘\alpha= 7^\circ, βflap=15∘\beta_{\mathrm{flap}} = 15^\circ, where extrapolation from the calibration data is most demanding. Quantitatively, the empirical mass of the green samples lying inside the supplied truth interval remains close to the intended 95% level for every QoI, ranging from about 94.2% to 95.8% over the twelve coefficient/location combinations.

The first-moment calibrated results, shown in orange, are less reliable because the calibration constrains only the centroid of the truth data and not the epistemic width of the truth intervals. As a result, the orange distributions are often shifted, too narrow, or both. For example, at α=7∘\alpha= 7^\circ, βflap=15∘\beta_{\mathrm{flap}} = 15^\circ, the first-moment calibrated CdC_d distribution lies mostly to the right of the released truth interval, while the ClC_l distribution places a substantial fraction of its mass below the interval. In several other cases, especially for CdC_d and CmC_m near α=0∘\alpha= 0^\circ and α=2∘\alpha= 2^\circ, the orange distributions fall inside the red interval but are much narrower than the truth bounds; this gives apparent coverage of the interval edges without representing the released epistemic uncertainty. Over all twelve comparisons, the fraction of orange samples inside the truth interval varies widely, from roughly 11% to 100%, which indicates that first-moment calibration does not produce uniformly calibrated epistemic bounds.

Figure 9 gives the same comparison in cumulative form. If the released bounds are interpreted as 95% intervals, the left and right red lines should intersect a calibrated CDF near probability levels 0.025 and 0.975, respectively. The green CDFs follow this behavior closely, whereas the orange CDFs are typically steeper and displaced relative to the red interval. This confirms that the improvement from distributional calibration is not only a correction of the posterior mean but also a correction of the full marginal predictive distribution of each QoI.

Overview of empirical CDFs for the four prediction settings and three aerodynamic coefficients

Figure 9. Overview of empirical CDFs for the four prediction settings and three aerodynamic coefficients. The orange curves show the first-moment-calibrated samples, the green curves show the distribution-calibrated samples, and the red vertical lines mark the provided epistemic interval edges.

Table 2 summarizes this calibration check quantitatively. The three reported quantities are complementary: the interval mass measures how much of the predictive distribution lies within the released epistemic bounds, while F(L)F(L) and F(U)F(U) measure whether the lower and upper truth edges occur at the intended lower and upper tail probabilities. For a central 95% epistemic interval, the target values are therefore 95% mass, F(L)=0.025F(L) = 0.025, and F(U)=0.975F(U) = 0.975. The distribution-calibrated predictions are close to these targets across all twelve QoI/location pairs. In particular, the mass inside [L,U][L,U] remains between 94.2% and 95.8%, and the endpoint CDF values stay close to the nominal 0.025 and 0.975 levels.

α\alphaβflap\beta_{\mathrm{flap}}QoIMass in [L,U][L,U] (%)F(L)F(L)F(U)F(U)
TruthFirst-momentDistributionTruthFirst-momentDist.TruthFirst-momentDistribution
−3∘-3^\circ−3∘-3^\circClC_l95.073.394.20.0250.0000.0290.9750.7330.971
−3∘-3^\circ−3∘-3^\circCdC_d95.0100.094.30.0250.0000.0240.9751.0000.967
−3∘-3^\circ−3∘-3^\circCmC_m95.0100.094.50.0250.0000.0220.9751.0000.967
0∘0^\circ0∘0^\circClC_l95.038.295.50.0250.0000.0190.9750.3820.974
0∘0^\circ0∘0^\circCdC_d95.0100.094.40.0250.0000.0300.9751.0000.974
0∘0^\circ0∘0^\circCmC_m95.0100.095.80.0250.0000.0200.9751.0000.978
2∘2^\circ2∘2^\circClC_l95.063.094.90.0250.0450.0250.9750.6750.974
2∘2^\circ2∘2^\circCdC_d95.0100.094.90.0250.0000.0230.9751.0000.972
2∘2^\circ2∘2^\circCmC_m95.0100.094.90.0250.0000.0260.9751.0000.975
7∘7^\circ15∘15^\circClC_l95.027.495.80.0250.7260.0220.9751.0000.980
7∘7^\circ15∘15^\circCdC_d95.011.095.10.0250.0000.0330.9750.1100.984
7∘7^\circ15∘15^\circCmC_m95.0100.095.50.0250.0000.0230.9751.0000.978

Table 2. Empirical calibration of the prediction distributions against the released epistemic intervals. The truth target corresponds to a 95% interval, so F(L)=0.025F(L)=0.025, F(U)=0.975F(U)=0.975, and the interval mass is 95%.

The same table also clarifies why matching only the first moment is insufficient. Several first-moment calibrated predictions have 100% of their samples inside the truth interval, but their F(L)F(L) and F(U)F(U) values are 0 and 1, respectively; these cases are over-concentrated relative to the released epistemic uncertainty, not genuinely distribution calibrated. Other cases fail because the predictive distribution is displaced relative to the truth interval: for example, at α=7∘\alpha= 7^\circ, βflap=15∘\beta_{\mathrm{flap}} = 15^\circ, the first-moment calibrated ClC_l result gives F(L)=0.726F(L) = 0.726, meaning that most of the predicted mass lies below the lower truth edge, while the corresponding CdC_d result places only 11% of the samples inside the truth interval. By contrast, the distribution-calibrated results at the same off-design point retain near-nominal interval mass and endpoint probabilities for all three QoIs. Thus, Table 2 supports the visual conclusion from Figure 9 that distributional calibration corrects the predictive uncertainty itself, not merely the location of the predictive mean.

We tabulate the learned latent input locations for each of the 7 validation points in Table 3 to Table 5; most of the learned latent inputs x~\tilde{x} tend to lie closer to one of the bounds.

Pointsα\alphaβflap\beta_{\mathrm{flap}}Re
LbUbx~\tilde{x}LbUbx~\tilde{x}LbUbx~\tilde{x}
1−0.020.020.019962−0.100.100.099801696500703500700000
24.985.024.980000−0.100.100.099807696500703500700000
39.9810.029.980000−0.100.100.099807696500703500700000
4−0.020.020.0199614.905.105.099900696500703500700000
5−0.020.020.0199599.9010.109.900200696500703500700000
64.985.024.9800004.905.104.900200696500703500700000
7−5.02−4.98−4.980000−5.10−4.90−4.900200696500703500700000

Table 3. Learned latent input locations x~\tilde{x} for ClC_l via the latent GP

Pointsα\alphaβflap\beta_{\mathrm{flap}}Re
LbUbx~\tilde{x}LbUbx~\tilde{x}LbUbx~\tilde{x}
1−0.020.020.019961−0.100.100.099805696500703500700000
24.985.024.980000−0.100.100.099803696500703500700000
39.9810.029.980000−0.100.100.099802−0.100.100.099802
4−0.020.020.0199604.905.104.900200696500703500700000
5−0.020.020.0199609.9010.109.900200696500703500700000
64.985.024.9800004.905.104.900200696500703500700000
7−5.02−4.98−4.980000−5.10−4.90−4.900200696500703500700000

Table 5. Learned latent input locations x~\tilde{x} for CmC_m via the latent GP

Conclusions

This work developed a calibrated Bayesian surrogate framework for the second AIAA UQ challenge problem, where aerodynamic coefficients must be predicted at unseen operating conditions while accounting for epistemic uncertainty in both the inputs and outputs. The method combines three ingredients: a GP surrogate for the XFOIL response, a Kennedy–O’Hagan-style GP discrepancy model for the residual between XFOIL and truth data, and a latent-input formulation that infers the unknown physical calibration locations within their prescribed epistemic intervals. The principal additional contribution is a distributional calibration step that uses the supplied output intervals as constraints on the corrected predictive distribution, thereby calibrating not only the posterior mean but also the marginal predictive variance of each aerodynamic coefficient.

The empirical results show that this distribution-level correction is essential for reliable uncertainty quantification. Across the four withheld prediction locations and three quantities of interest, the distribution-calibrated model places between 94.2% and 95.8% of its predictive samples inside the released 95% truth intervals, with endpoint CDF values close to the nominal 0.025 and 0.975 levels. This behavior holds even at the more extrapolative operating point α=7∘\alpha=7^\circ, βflap=15∘\beta_{\mathrm{flap}}=15^\circ. In contrast, first-moment calibration often produces predictive distributions that are shifted, over-concentrated, or both, with interval masses ranging from 11% to 100%. These comparisons indicate that matching the center of the truth data is not sufficient; the released epistemic widths must also inform the calibrated uncertainty model.

Several extensions are natural. A fully Bayesian treatment of the latent inputs, GP hyperparameters, and discrepancy field could better quantify posterior uncertainty than the MAP-based implementation used here, particularly with sparse truth data. The

Figure 5. Prediction bounds at α=−3∘\alpha= -3^\circ and βflap=−3∘\beta_{\mathrm{flap}} = -3^\circ. Each row shows one aerodynamic coefficient, with the first-moment-calibrated result on the left and the distribution-calibrated result on the right.

Prediction bounds at alpha = 0 degrees and beta subscript flap = 0 degrees, comparing first-moment-calibrated and distribution-calibrated results for aerodynamic coefficients

Figure 6. Prediction bounds at α=0∘\alpha= 0^\circ and βflap=0∘\beta_{\mathrm{flap}} = 0^\circ. Each row shows one aerodynamic coefficient, with the first-moment-calibrated result on the left and the distribution-calibrated result on the right.

Prediction bounds for aerodynamic coefficients under first-moment and distribution calibration

Figure 7. Prediction bounds at α=2∘\alpha= 2^\circ and βflap=2∘\beta_{\mathrm{flap}} = 2^\circ. Each row shows one aerodynamic coefficient, with the first-moment-calibrated result on the left and the distribution-calibrated result on the right.

Prediction bounds at $\alpha = 7^\circ$ and $\beta_{\mathrm{flap}} = 15^\circ$

Figure 8. Prediction bounds at α=7∘\alpha= 7^\circ and βflap=15∘\beta_{\mathrm{flap}} = 15^\circ. Each row shows one aerodynamic coefficient, with the first-moment-calibrated result on the left and the distribution-calibrated result on the right.

distributional calibration strategy could also be extended to correlated multi-output models so that joint uncertainty in ClC_l, CdC_d, and CmC_m is represented directly. Finally, the framework provides a natural basis for active learning: new XFOIL runs or truth-data acquisitions could be selected where the calibrated predictive distribution is most uncertain, most poorly constrained by the interval information, or most influential for downstream aerodynamic decisions.

Acknowledgments

This work was funded by the Penn State Institute of Computational and Data Sciences (ICDS) Junior Researcher Collaboration grant. The authors also acknowledge Penn State ICDS (RRID:SCR_025154) for providing access to computational research infrastructure within the Roar Core Facility (RRID:SCR_026424).

Appendix

MAP estimates of latent inputs

Pointsα\alphaβflap\beta_{\mathrm{flap}}Re
LbUbx~\tilde{x}LbUbx~\tilde{x}LbUbx~\tilde{x}
1−0.020.020.019961−0.100.100.099806696500703500700000
24.985.024.980000−0.100.100.099803696500703500700000
39.9810.029.980000−0.100.100.099802696500703500700000
4−0.020.020.0199604.905.104.900200696500703500700000
5−0.020.020.0199609.9010.109.900200696500703500700000
64.985.024.9800004.905.104.900200696500703500700000
7−5.02−4.98−4.980000−5.10−4.90−4.900200696500703500700000

Table 4. Learned latent input locations x~\tilde{x} for CdC_d via the latent GP

Derivation of the calibrated mean and covariance

Let the uncalibrated corrected Gaussian process be

g0(x)∼GP(m0(x),k0(x,x′)).g_0(\mathbf{x}) \sim\mathcal{GP}\left(m_0(\mathbf{x}), k_0(\mathbf{x},\mathbf{x}')\right).

At the truth-data locations

XT=[x1T,…,xNTT]T,X_T = [\mathbf{x}_1^T,\ldots,\mathbf{x}_{N_T}^T]^T,

define

gT=g0(XT)=[g0(x1T),…,g0(xNTT)]T.\mathbf{g}_T = g_0(X_T) = [g_0(\mathbf{x}_1^T),\ldots,g_0(\mathbf{x}_{N_T}^T)]^T.

Under the uncalibrated corrected GP,

gT∼N(m0T,KTT),m0T=m0(XT),KTT=k0(XT,XT).\mathbf{g}_T \sim\mathcal{N}(\mathbf{m}_{0T}, K_{TT}), \qquad\mathbf{m}_{0T} = m_0(X_T), \qquad K_{TT} = k_0(X_T,X_T).

The distributional calibration target replaces this finite-dimensional marginal by

gT∼qT=N(c,VT).\mathbf{g}_T \sim q_T = \mathcal{N}(\mathbf{c}, V_T).

Equivalently, the calibrated process is defined by

pcal(g∣DX,DT)=∫p0(g∣gT=z,DX) qT(z) dz,qT(z)=N(z;c,VT).p_{\mathrm{cal}}(g \mid D_X,D_T) = \int p_0(g \mid\mathbf{g}_T = \mathbf{z},D_X)\,q_T(\mathbf{z})\,\mathrm{d}\mathbf{z}, \qquad q_T(\mathbf{z}) = \mathcal{N}(\mathbf{z};\mathbf{c},V_T).

For any prediction input x\mathbf{x}, define

kxT=k0(x,XT)=[k0(x,x1T),…,k0(x,xNTT)],k_{xT} = k_0(\mathbf{x},X_T) = [k_0(\mathbf{x},\mathbf{x}_1^T),\ldots,k_0(\mathbf{x},\mathbf{x}_{N_T}^T)],

and similarly

kTx′=k0(XT,x′).k_{Tx'} = k_0(X_T,\mathbf{x}').

The joint Gaussian distribution of g0(x)g_0(\mathbf{x}) and gT\mathbf{g}_T implies the standard conditional mean

E0[g0(x)∣gT=z,DX]=m0(x)+kxTKTT−1(z−m0T),\mathbb{E}_0[g_0(\mathbf{x}) \mid\mathbf{g}_T = \mathbf{z},D_X] = m_0(\mathbf{x}) + k_{xT}K_{TT}^{-1}(\mathbf{z}-\mathbf{m}_{0T}),

and conditional covariance

Cov⁡0[g0(x),g0(x′)∣gT=z,DX]=k0(x,x′)−kxTKTT−1kTx′.\operatorname{Cov}_0[g_0(\mathbf{x}),g_0(\mathbf{x}') \mid\mathbf{g}_T = \mathbf{z},D_X] = k_0(\mathbf{x},\mathbf{x}') - k_{xT}K_{TT}^{-1}k_{Tx'}.

Taking expectation over z∼qT=N(c,VT)\mathbf{z} \sim q_T = \mathcal{N}(\mathbf{c},V_T), the calibrated mean is

mcal(x)=EqT[E0[g0(x)∣gT=z,DX]]=EqT[m0(x)+kxTKTT−1(z−m0T)]=m0(x)+kxTKTT−1(c−m0T).\begin{aligned} m_{\mathrm{cal}}(\mathbf{x}) &= \mathbb{E}_{q_T}\left[\mathbb{E}_0[g_0(\mathbf{x}) \mid\mathbf{g}_T = \mathbf{z},D_X]\right] \\ &= \mathbb{E}_{q_T}\left[m_0(\mathbf{x}) + k_{xT}K_{TT}^{-1}(\mathbf{z}-\mathbf{m}_{0T})\right] \\ &= m_0(\mathbf{x}) + k_{xT}K_{TT}^{-1}(\mathbf{c}-\mathbf{m}_{0T}). \end{aligned}

Therefore,

mcal(x)=m0(x)+kxTKTT−1(c−m0T).m_{\mathrm{cal}}(\mathbf{x}) = m_0(\mathbf{x}) + k_{xT}K_{TT}^{-1}(\mathbf{c}-\mathbf{m}_{0T}).

For the covariance, use the law of total covariance:

kcal(x,x′)=EqT[Cov⁡0[g0(x),g0(x′)∣gT=z,DX]]+Cov⁡qT(E0[g0(x)∣gT=z,DX],E0[g0(x′)∣gT=z,DX]).\begin{aligned} k_{\mathrm{cal}}(\mathbf{x},\mathbf{x}') &= \mathbb{E}_{q_T}\left[\operatorname{Cov}_0[g_0(\mathbf{x}),g_0(\mathbf{x}') \mid\mathbf{g}_T = \mathbf{z},D_X]\right] \\ &\quad+ \operatorname{Cov}_{q_T}\left(\mathbb{E}_0[g_0(\mathbf{x}) \mid\mathbf{g}_T = \mathbf{z},D_X],\mathbb{E}_0[g_0(\mathbf{x}') \mid\mathbf{g}_T = \mathbf{z},D_X]\right). \end{aligned}

The first term is

k0(x,x′)−kxTKTT−1kTx′,k_0(\mathbf{x},\mathbf{x}') - k_{xT}K_{TT}^{-1}k_{Tx'},

since the conditional covariance does not depend on z\mathbf{z}. The second term is

Cov⁡qT(kxTKTT−1z,kx′TKTT−1z)=kxTKTT−1VTKTT−1kTx′.\operatorname{Cov}_{q_T}\left(k_{xT}K_{TT}^{-1}\mathbf{z},k_{x'T}K_{TT}^{-1}\mathbf{z}\right) = k_{xT}K_{TT}^{-1}V_TK_{TT}^{-1}k_{Tx'}.

Hence

kcal(x,x′)=k0(x,x′)−kxTKTT−1kTx′+kxTKTT−1VTKTT−1kTx′.k_{\mathrm{cal}}(\mathbf{x},\mathbf{x}') = k_0(\mathbf{x},\mathbf{x}') - k_{xT}K_{TT}^{-1}k_{Tx'} + k_{xT}K_{TT}^{-1}V_TK_{TT}^{-1}k_{Tx'}.

References

  1. [1]Srishti Adhikary, Venkat Narayanaswamy, and Ashwin Renganathan. Adaptive multitask gaussian process surrogate models for supersonic fluid-structure interactions. In AIAA AVIATION 2026 Forum, page 4781, 2026. doi: https://doi.org/10.2514/6.2026-4781.
  2. [2]Paul D. Arendt, Daniel W. Apley, and Wei Chen. Quantification of model uncertainty: Calibration, model discrepancy, and identifiability. Journal of Mechanical Design, 134(10):100908, 2012. doi: https://doi.org/10.1115/1.4007390.
  3. [3]S Ashwin Renganathan, Romit Maulik, Stefano Letizia, and Giacomo Valerio Iungo. Data-driven wind turbine wake modeling via probabilistic machine learning. Neural Computing and Applications, 34(8):6171–6186, 2022. doi: https://doi.org/10.1007/s00521-021-06799-6.
  4. [4]M. J. Bayarri, James O. Berger, Rui Paulo, Jerry Sacks, John A. Cafeo, James Cavendish, Chin-Hsu Lin, and Jian Tu. A framework for validation of computer models. Technometrics, 49(2):138–154, 2007. doi: https://doi.org/10.1198/004017007000000092.
  5. [5]Michael Betancourt. A conceptual introduction to hamiltonian monte carlo. arXiv preprint arXiv:1701.02434, 2017.arxiv.org/abs/1701.02434
  6. [6]David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational inference: A review for statisticians. Journal of the American statistical Association, 112(518):859–877, 2017. doi: https://doi.org/10.1080/01621459.2017.1285773.
  7. [7]Annie S Booth, S Ashwin Renganathan, and Robert B Gramacy. Contour location for reliability in airfoil simulation experiments using deep gaussian processes. The Annals of Applied Statistics, 19(1):191–211, 2025. doi: https://doi.org/10.1214/24-AOAS1951.
  8. [8]Jenny Brynjarsdóttir and Anthony O’Hagan. Learning about physical parameters: The importance of model discrepancy. Inverse Problems, 30(11):114007, 2014. doi: https://doi.org/10.1088/0266-5611/30/11/114007.
  9. [9]Kade Carlson and Ashwin Renganathan. Multiobjective aerodynamic design optimization of the nasa common research model. Aerospace Science and Technology, page 111120, 2025. doi: https://doi.org/10.1016/j.ast.2025.111120.
  10. [10]Andrew W. Cary, John A. Schaefer, Earl P. N. Duque, Manas S. Khurana, and Erin C. DeCarlo. Overview of challenges in performing uncertainty quantification for fluids engineering problems. In AIAA SciTech Forum, 2022. doi: https://doi.org/10.2514/6.2022-2357.
  11. [11]Andrew W Cary, John A Schaefer, Erin C DeCarlo, Earl P Duque, and Manas Khurana. Overview of fluid dynamics uncertainty quantification challenge problem and results. In AIAA SCITECH 2024 Forum, page 0705, 2024. doi: https://doi.org/10.2514/6.2024-0705.
  12. [12]Andrew W Cary, Markus Rumpfkeil, Peter O Hristov, and John A Schaefer. Summary of the second aiaa uncertainty quantification challenge problem for aerodynamics. In AIAA SCITECH 2026 Forum, page 0091, 2026. doi: https://doi.org/10.2514/6.2026-0091.
  13. [13]Patrick Dallaire, Camille Besse, and Brahim Chaib-draa. Learning gaussian process models from uncertain data. In Neural Information Processing, volume 5863 of Lecture Notes in Computer Science, pages 433–440. Springer, 2009. doi: https://doi.org/10.1007/978-3-642-10677-4_49.
  14. [14]Andreas C. Damianou, Michalis K. Titsias, and Neil D. Lawrence. Variational inference for latent variables and uncertain inputs in gaussian processes. Journal of Machine Learning Research, 17:1–62, 2016.
  15. [15]Geoffrey Davis and Ashwin Renganathan. Uncertainty quantification via latent gaussian process surrogates for the second aiaa fluid dynamics uq challenge problem. In AIAA SCITECH 2026 Forum, page 0095, 2026.
  16. [16]Mark Drela. Xfoil: An analysis and design system for low reynolds number airfoils. In Low Reynolds number aerodynamics, pages 1–12. Springer, 1989.
  17. [17]Agathe Girard, Carl Rasmussen, Joaquin Q Candela, and Roderick Murray-Smith. Gaussian process priors with uncertain inputs application to multiple-step ahead time series forecasting. Advances in neural information processing systems, 15, 2002.
  18. [18]Robert B. Gramacy. Surrogates: Gaussian Process Modeling, Design, and Optimization for the Applied Sciences. Chapman and Hall/CRC, Boca Raton, FL, 2020. doi: https://doi.org/10.1201/9780367815493.
  19. [19]Dave Higdon, Marc Kennedy, James C. Cavendish, John A. Cafeo, and Robert D. Ryne. Combining field data and computer simulations for calibration and prediction. SIAM Journal on Scientific Computing, 26(2):448–466, 2004. doi: https://doi.org/10.1137/S1064827503426693.
  20. [20]Dave Higdon, James Gattiker, Brian Williams, and Maria Rightley. Computer model calibration using high-dimensional output. Journal of the American Statistical Association, 103(482):570–583, 2008. doi: https://doi.org/10.1198/016214507000000888.
  21. [21]Marc C. Kennedy and Anthony O’Hagan. Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(3):425–464, 2001. doi: https://doi.org/10.1111/1467-9868.00294.
  22. [22]Neil D. Lawrence. Probabilistic non-linear principal component analysis with gaussian process latent variable models. Journal of Machine Learning Research, 6:1783–1816, 2005.
  23. [23]Fei Liu, Maria J. Bayarri, and James O. Berger. Modularization in bayesian analysis, with emphasis on analysis of computer models. Bayesian Analysis, 4(1):119–150, 2009. doi: https://doi.org/10.1214/09-BA404.
  24. [24]Wei-Liem Loh. On latin hypercube sampling. The annals of statistics, 24(5):2058–2080, 1996. doi: https://doi.org/10.1214/aos/1069362310.
  25. [25]Andrew McHutchon and Carl Edward Rasmussen. Gaussian process training with input noise. In Advances in Neural Information Processing Systems, volume 24, pages 1341–1349, 2011.
  26. [26]William L. Oberkampf and Christopher J. Roy. Verification and validation in scientific computing. Cambridge University Press, 2010.
  27. [27]Guillaume Perrin and Cédric Durantin. Taking into account input uncertainties in the bayesian calibration of time-consuming simulators. Journal de la Société Française de Statistique, 160(2):24–46, 2019.
  28. [28]Matthew Plumlee. Bayesian calibration of inexact computer models. Journal of the American Statistical Association, 112(519):1274–1285, 2017. doi: https://doi.org/10.1080/01621459.2016.1211016.
  29. [29]Joaquin Quiñonero-Candela, Agathe Girard, Jan Larsen, and Carl Edward Rasmussen. Propagation of uncertainty in bayesian kernel models—application to multiple-step ahead forecasting. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 2, pages 701–704. IEEE, 2003. doi: https://doi.org/10.1109/ICASSP.2003.1202463.
  30. [30]Carl E. Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning. MIT Press, 2006.
  31. [31]Ashwin Renganathan. Efficient reliability analysis with multifidelity gaussian processes and normalizing flows. In AIAA SCITECH 2024 Forum, page 0576, 2024. doi: https://doi.org/10.2514/6.2024-0576.
  32. [32]Ashwin Renganathan and Kade Carlson. qpots: Efficient batch multiobjective bayesian optimization via pareto optimal thompson sampling. In International Conference on Artificial Intelligence and Statistics, pages 4051–4059. PMLR, 2025.
  33. [33]Ashwin Renganathan, Vishwas Rao, and Ionel Navon. Multifidelity gaussian processes for failure boundary and probability estimation. In AIAA Scitech 2022 Forum, page 0390, 2022. doi: https://doi.org/10.2514/6.2022-0390.
  34. [34]S Ashwin Renganathan, Jeffrey Larson, and Stefan M Wild. Lookahead acquisition functions for finite-horizon time-dependent bayesian optimization and application to quantum optimal control. arXiv preprint arXiv:2105.09824, 2021.arxiv.org/abs/2105.09824
  35. [35]S Ashwin Renganathan, Romit Maulik, and Jai Ahuja. Enhanced data efficiency using deep neural networks and gaussian processes for aerodynamic design optimization. Aerospace Science and Technology, 111:106522, 2021. doi: https://doi.org/10.1016/j.ast.2021.106522.
  36. [36]S Ashwin Renganathan, Vishwas Rao, and Ionel M Navon. Camera: A method for cost-aware, adaptive, multifidelity, efficient reliability analysis. Journal of Computational Physics, 472:111698, 2023. doi: https://doi.org/10.1016/j.jcp.2022.111698.
  37. [37]Jerome Sacks, William J Welch, Toby J Mitchell, and Henry P Wynn. Design and analysis of computer experiments. Statistical science, 4(4):409–423, 1989. doi: https://doi.org/10.1214/ss/1177012413.
  38. [38]Thomas J Santner, Brian J Williams, William I Notz, and Brain J Williams. The design and analysis of computer experiments, volume 1. Springer, 2003. doi: https://doi.org/10.1007/978-1-4939-8847-1.
  39. [39]Michalis K. Titsias and Neil D. Lawrence. Bayesian gaussian process latent variable model. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 844–851, 2010.
  40. [40]Rui Tuo and C. F. Jeff Wu. Efficient calibration for imperfect computer models. The Annals of Statistics, 43(6):2331–2352, 2015. doi: https://doi.org/10.1214/15-AOS1314.
  41. [41]Rui Tuo and C. F. Jeff Wu. A theoretical framework for calibration in computer models: Parametrization, estimation and convergence properties. SIAM/ASA Journal on Uncertainty Quantification, 4(1):767–795, 2016. doi: https://doi.org/10.1137/151005841.
  42. [42]Jian-Xun Wang, Christopher J. Roy, and Heng Xiao. Propagation of input uncertainty in presence of model-form uncertainty: A multifidelity approach for computational fluid dynamics applications. ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part B: Mechanical Engineering, 4(1):011002, 2018. doi: https://doi.org/10.1115/1.4037452.
  43. [43]Kail Yuan and Ashwin Renganathan. Remal: Residual equilibrium manifold active learning for surrogate-based multidisciplinary design analysis. arXiv preprint arXiv:2606.13245, 2026.arxiv.org/abs/2606.13245

Paper details

vlnc:W_uye5wmzv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents