---
title: Large Wing Model (LWM) Overview
url: https://www.emergentmind.com/topics/large-wing-model-lwm
type: topic
---

# Large Wing Model (LWM) Overview

Searching arXiv for the specified paper and closely related wing-model/dataset papers.
arXiv search query: 2508.11827 Large Wing Model; 2512.12867 OptiWing3D; 2512.14397 SuperWing.
Large Wing Model (LWM) is a probabilistic machine-learning framework for finite-wing aerodynamics that predicts three-dimensional pressure coefficient distributions, $C_p(x,y)$, from a small, strictly experimental data set and propagates predictive uncertainty to sectional lift $c_l(y)$ and total lift $C_L$ [2508.11827]. In the paper titled "Large Wing Model" [2508.11827], the model is formulated as a modified deep kernel learning system in which a neural feature map embeds geometric and operating-condition inputs into a latent space, while a Gaussian Process (GP) models spanwise structure and uncertainty. The stated scope is rectangular or lightly tapered wings, low to moderate sweep, and attached, subcritical/transonic flows, with extrapolation aided by a physics-driven prior formed from two-dimensional pressure predictions produced by the Large Airfoil Model (LAM) [2508.11827].

## 1. Definition, scope, and nomenclature

Within the aerodynamic literature represented here, LWM denotes the "Large Wing Model" introduced for wing-surface pressure and lift prediction from experimental data [2508.11827]. Its target outputs are the mean and uncertainty of upper- and lower-surface $C_p(x,y)$, together with the mean and uncertainty of sectional lift $c_l(y)$ and total lift $C_L$, optionally followed by posterior conditioning to enforce known probabilistic lift descriptions [2508.11827]. The problem addressed is explicitly the prediction of three-dimensional pressure distributions over finite wings under data scarcity, including three-dimensional effects such as those induced by wing tip vortices [2508.11827].

The same acronym is used differently in other arXiv papers. In "Opus" [2412.00573], LWM stands for "Large Work Model," a workflow-centric model for Business Process Outsourcing. In "LWM-Temporal" [2603.10024], LWM stands for "Large Wireless Models," a family of wireless-channel foundation models. This acronymal overlap is a recurrent source of confusion; in the aerodynamic sense, however, LWM refers specifically to the finite-wing probabilistic predictor in [2508.11827].

Despite the name, the aerodynamic LWM is not presented as a generic large-scale wing generator. It is presented as a computationally efficient wing $C_p$ prediction model that facilitates rapid exploration of the wing design space, with uncertainty-aware inference and explicit lift propagation from the predicted pressure field [2508.11827].

## 2. Input representation and geometric parameterization

The model predicts pressure at query points expressed through a three-dimensional coordinate transform. The physical chordwise coordinate satisfies $x \in [0,c(y)]$, with normalized coordinate $x/c \in [0,1]$. A conformal-mapping-inspired representation is then introduced through
$$
\theta = \cos^{-1}(x/c),
$$
a linearly mapped $\hat{x} \in [-1,1]$, and $\hat{z}=\sin\theta \in [-1,1]$, together with the normalized semi-span coordinate $\eta = y/s \in [0,1]$ [2508.11827]. Upper and lower surfaces are distinguished by the sign of $\hat{z}$, with $\hat{z}>0$ on the upper surface and $\hat{z}<0$ on the lower surface, while leading and trailing edges have periodic identifications in $(\hat{x},\hat{z})$ [2508.11827].

The physical input vector is $u\in\mathbb{R}^{63}$ [2508.11827]. It contains 28 sampled $z/c$ values that encode the airfoil section geometry, operating conditions $\alpha$ and $M_\infty$, wing planform descriptors $s/c_{\text{root}}$, taper ratio $\lambda$, leading-edge sweep $\Lambda$, and the conformal query-point coordinates $\hat{x}$ and $\hat{z}$ [2508.11827]. The spanwise coordinate $\eta$ is intentionally excluded from the neural network input and is instead appended directly to the GP input so that the GP preserves spanwise topology and tip-boundary effects [2508.11827].

This representation defines a 15-dimensional active space. The neural feature map produces 14 latent variables,
$$
y=f_\theta(u)\in\mathbb{R}^{14},
$$
and the GP operates on $(y,\eta)\in\mathbb{R}^{15}$ [2508.11827]. The modeling scope reported in the paper is primarily rectangular wings with $\lambda \approx 1$, low to moderate sweep $\Lambda \approx 0$--$10^\circ$, and aspect ratios $s/c_{\text{root}} \approx 2$--$6$, over angles of attack $\alpha \in [-7^\circ,12^\circ]$ and Mach numbers $M_\infty \in [0,0.8]$ [2508.11827].

## 3. Deep-kernel-learning architecture and physics-driven prior

The core architecture is a modified deep kernel learning model [2508.11827]. The neural feature extractor has architecture
$$
[63\!-\!1000\!-\!1000\!-\!500\!-\!50\!-\!14],
$$
and its weights are learned jointly with GP hyperparameters by marginal likelihood maximization [2508.11827]. On the probabilistic side, the GP prior is written as
$$
f(y,\eta)\sim \mathrm{GP}(m(y,\eta),k((y,\eta),(y',\eta'))),
$$
with observations corrupted by noise
$$
\Sigma_s=\sigma_{\mathrm{expt}}^2 I + \sigma_{\mathrm{add}}^2 I,
$$
where the second term captures digitization and unreported additional noise [2508.11827].

The covariance structure is a product kernel over the 14 latent dimensions and the spanwise coordinate [2508.11827]. Each latent dimension uses a Matérn-$5/2$ kernel, while the spanwise dimension uses a Matérn-$3/2$ kernel with spatially varying hyperparameters. The spanwise hyperparameters are blended through a logistic transition,
$$
v(\eta)=v_{\mathrm{inboard}}+\frac{v_{\mathrm{outboard}}-v_{\mathrm{inboard}}}{1+e^{-R(\eta-\eta_{\mathrm{crit}})}},
$$
with calibrated values $\eta_{\mathrm{crit}}=0.70$ and $R=15$ [2508.11827]. The spanwise length scale $\ell_\eta$ decreases from approximately $0.5958$ inboard to approximately $0.1967$ outboard, while the signal variance changes slightly from approximately $2.7299$ to approximately $2.7210$ [2508.11827]. The stated purpose of this nonstationarity is to allow the model to represent localized distortions near the tip, including trailing-edge pressure undulations associated with tip vortices [2508.11827].

A distinctive component is the physics-driven prior supplied by the Large Airfoil Model [2508.11827]. LAM provides a probabilistic two-dimensional airfoil pressure distribution, and the LWM prior is formed by averaging LAM posteriors at four angles
$$
\alpha_n=\alpha_{\mathrm{geo}}(n-1)/3,\qquad n=1,\dots,4,
$$
thereby spanning the geometric angle of attack to $0^\circ$ [2508.11827]. The three-dimensional wing pressure field is then modeled as
$$
C_p(u,\eta)=\mu^{\mathrm{LAM}}(u,\eta)+\Delta_{3D}(u,\eta),
$$
where $\Delta_{3D}$ is a latent-space GP residual [2508.11827]. The paper attributes several practical effects to this prior: reduced posterior variance, especially near the trailing edge; elimination of non-physical loops; and strong extrapolation across unseen airfoil sections [2508.11827].

## 4. Pressure, lift, uncertainty, and posterior conditioning

The aerodynamic outputs follow standard nondimensional definitions. The pressure coefficient is
$$
C_p=\frac{p-p_\infty}{\tfrac{1}{2}\rho_\infty U_\infty^2},
$$
the sectional lift coefficient at spanwise station $y$ is
$$
c_l(y)=\frac{1}{c(y)}\int_0^{c(y)}\big(C_{p,\ell}(x,y)-C_{p,u}(x,y)\big)\,dx,
$$
and the total wing lift coefficient is
$$
C_L=\frac{1}{S}\int_{-b/2}^{b/2} c_l(y)c(y)\,dy.
$$
For rectangular wings, this reduces under $\eta=y/s$ to
$$
C_L=\int_0^1 c_l(\eta)\,d\eta
$$
[2508.11827].

The model produces posterior mean and variance through the usual GP predictive equations. With covariance matrices $K$, $K_*$, and $K_{**}$ and noise covariance $\Sigma_s$, the posterior mean and variance are
$$
m_* = K_*^T (K+\Sigma_s)^{-1}y,\qquad
\sigma_*^2=\mathrm{diag}\!\left(K_{**}-K_*^T(K+\Sigma_s)^{-1}K_*\right),
$$
with numerically stable Cholesky forms used in practice [2508.11827].

Uncertainty propagation to $c_l(\eta)$ and $C_L$ is not performed analytically in the final system because the nonlinear feature map $f_\theta$ breaks linear-Gaussian closure [2508.11827]. Instead, the paper uses Monte Carlo propagation: it draws $N_s=100{,}000$ samples from the full posterior predictive distribution of $C_p$, integrates each sample chordwise to obtain $c_l^{(i)}(\eta)$, integrates spanwise to obtain $C_L^{(i)}$, and then computes sample means, variances, and $95\%$ confidence intervals as $\text{mean}\pm1.96\times\text{std}$ under approximate Gaussianity of aggregated outputs [2508.11827].

An additional mechanism constrains the posterior predictive space when a target lift distribution is known. If, for example,
$$
C_L \sim \mathcal{N}(\mu_{C_{L,c}},\sigma^2_{C_{L,c}}),
$$
the paper applies rejection sampling with acceptance probability
$$
P_{\mathrm{accept}}(C_L)=\frac{p(C_L)}{M q(C_L)},
$$
where $q(C_L)$ is the empirical proposal derived from unconditioned posterior samples and $M=\max_{C_L}[p(C_L)/q(C_L)]$ normalizes the ratio [2508.11827]. The stated effect is reduced variance in $c_l$ and $C_L$, together with improved pressure predictions consistent with integrated lift constraints [2508.11827].

## 5. Experimental database, validation, and computational characteristics

The training database compiled for LWM comprises 16 unique wings, 54 operating conditions, and 23,723 individual pressure measurements [2508.11827]. The wing airfoils include NACA 0012, 0015, 0023, RAE 101, NACA 64A-105, a 5% thick supercritical airfoil, and a "NACA 6-series-like" profile [2508.11827]. Source data were gathered from government reports, technical notes, and papers, with tables extracted via OCR using ExtractTable and graphs digitized with WebPlotDigitizer; the resulting data were added to ASPIRE and released with code repositories for the Large Wing Model and Large Airfoil Model [2508.11827].

Three rectangular-wing test cases are reported [2508.11827]. For a NACA 0012 wing at $\alpha=8.85^\circ$, $M_\infty=0.13$, and $s/c_{\text{root}}=2.95$, the model achieved $\mathrm{MAE}(C_p)=0.044$, $\mathrm{MAE}_{\mathrm{enclosed}}=0.014$, and $C_L=0.659\pm0.005$ versus an experimental value of $0.667$, corresponding to $-1.259\%$ error [2508.11827]. For a second NACA 0012 case at $\alpha=2.01^\circ$, $M_\infty=0.60$, and $s/c_{\text{root}}=2.0$, it achieved $\mathrm{MAE}(C_p)=0.062$, $\mathrm{MAE}_{\mathrm{enclosed}}=0.029$, with predicted $C_L=0.126\pm0.002$ and no reference value available [2508.11827]. For a NACA 0015 wing excluded from training, at $\alpha=4^\circ$, $M_\infty=0.17$, and $s/c_{\text{root}}=3.3$, it achieved $\mathrm{MAE}(C_p)=0.052$, $\mathrm{MAE}_{\mathrm{enclosed}}=0.016$, and $C_L=0.359\pm0.003$ versus an experimental value of $0.353$, corresponding to $+1.700\%$ error; posterior conditioning further reduced $\mathrm{MAE}_{\mathrm{enclosed}}$ to $0.013$ [2508.11827].

Across these test cases, the paper states that the error in $C_L$ did not exceed $1.7\%$ and that the model effectively captures three-dimensional effects such as those induced by wing tip vortices [2508.11827]. It also reports a comparison against RANS using Spalart–Allmaras with 1.1M points, stating that LWM better captured the magnitude and extent of tip-vortex distortions in the outboard trailing-edge region, whereas RANS underpredicted them [2508.11827].

Training uses exact GP inference, which is reported as feasible because of the modest data size, with optimization by Adam from an initial learning rate of $10^{-3}$, halved every 1000 epochs, for approximately 3000 epochs; an SWA-like procedure restores the best weights from three runs [2508.11827]. Inference precomputes $(K+\Sigma_s)^{-1}$ and is dominated by cross-covariance evaluation, reported as $22.15$ s on a single NVIDIA A100 GPU for approximately 25k surface points and approximately $54.11$ s on a Xeon E5-2660 CPU [2508.11827].

## 6. Broader LWM landscape, related datasets, and limitations

The 2025 LWM paper defines a specific probabilistic framework for finite-wing pressure and lift prediction, but later aerodynamic work uses the term more broadly to describe data-driven wing surrogates and generative systems [2508.11827]. This suggests a widening usage of "Large Wing Model" from an uncertainty-aware predictor toward a larger ecosystem of aerodynamic representation learning, inverse design, and multi-fidelity modeling.

"OptiWing3D" [2512.12867] describes a dataset explicitly positioned as enabling a Large Wing Model. It provides 1552 CFD evaluations yielding 776 successful 3D wing optimizations, initialized from distinct extruded airfoil cross-sections and paired in large part with 2D counterparts optimized under identical conditions, thereby forming a multi-fidelity dataset [2512.12867]. The associated baseline is a constraint-aware conditional latent diffusion model that predicts optimized wings from flow conditions, with a reported test-set shape MSE of $3.1\times10^{-5}\pm2.1\times10^{-5}$, normalized Vendi score of $0.803\pm0.075$, and volume constraint satisfaction of $99.74\%\pm0.52\%$ on $N=77$ test wings [2512.12867]. In that setting, LWM is framed not as a GP-based pressure predictor but as a possible learned model for optimized wing geometry conditioned on flow and constraints.

"SuperWing" [2512.14397] provides a different pathway. It is an open dataset of 4,239 parameterized wing geometries and 28,856 RANS flow-field solutions, with prepared surface meshes and labels suitable for transformer-based training [2512.14397]. In the benchmark reported there, a Vision Transformer predicts surface fields on a common $256\times128$ reference mesh and achieves approximately $0.329\%$ normalized MAE for $C_p$, $0.245\%$ for $C_{f,\tau}$, $0.281\%$ for $C_{f,z}$, and an integrated drag error of $2.48\times10^{-4}$, or 2.48 drag counts [2512.14397]. The same paper reports strong zero-shot generalization to DLR-F6 and NASA CRM without fine-tuning [2512.14397]. Relative to [2508.11827], this reflects a shift from sparse experimental data and explicit GP uncertainty to large-scale mesh-based supervised learning over transonic swept wings.

The original LWM also states clear limitations [2508.11827]. Data coverage is sparse and biased toward rectangular wings and low-speed regimes; high sweep, strong taper, twist, and spanwise-varying airfoil sections are under-represented [2508.11827]. Strong three-dimensional separation, stall, and highly compressible or shock-dominated regimes are outside the primary scope, and tip geometry is only implicitly rounded rather than explicitly modeled [2508.11827]. The model further assumes a uniform airfoil section along the span, and exact GP inference imposes memory limits that motivate stochastic-variational or sparse-GP extensions for larger datasets [2508.11827]. A plausible implication is that the probabilistic architecture of [2508.11827] and the broader datasets of [2512.12867] and [2512.14397] are complementary rather than redundant: one emphasizes data efficiency and calibrated uncertainty under strict experimental scarcity, while the others emphasize geometric diversity, multi-fidelity pairing, and large-scale supervised or generative training.

Source: https://www.emergentmind.com/topics/large-wing-model-lwm