WorksComputer Science
Quantum Computing-Based Wasserstein Generative Adversarial Networks for Airfoil Parameterization
Abstract
Airfoil geometry parameterization is essential in aerodynamic design optimization. Optimization search in the large space suffers from low efficiency and modest feasible region, which could further incur failures to convergence. Generative adversarial networks (GAN), aiming to learn the statistical distribution of existing data, pave a new avenue for shape generation and airfoil parameterization. GAN manages to implicitly reduce design space by generating only smooth, realistic airfoils. Meanwhile, quantum computing has been emergingly exhibiting great potential for achieving superior performance in specific types of engineering problems but has not been introduced to airfoil parameterization. To fill this gap, this work develops quantum computing-based Wasserstein GAN (Q-WGAN) and investigates the performance for airfoil parameterization. The Q-WGAN adopts parameterized quantum circuits as the generator to promote the generative performance, while the discriminator remains classical. The developed Q-WGAN is demonstrated on a dataset of optimal airfoils parameterized by free-form deformation (FFD) control points, resulting from 992 airfoil aerodynamic design optimizations. Results show that Q-WGAN achieves 92.1% mean fitting accuracy on the optimal FFD control points and 99.2% accuracy on airfoil coordinates, comparable with the classical WGAN performance of 92.28% on FFD control points and 99.26% on airfoil coordinates. Q-WGAN achieves lower Jensen-Shannon divergence, Fréchet distance, and Wasserstein distance than the classical WGAN, indicating an improved agreement with the reference airfoil distribution. Therefore, this work presents a convincing evidence that Q-WGAN makes a viable option for parameterizing airfoils and possesses a great potential for other relevant engineering applications.
Introduction
Airfoil parameterization, a key component of aerodynamic shape optimization, plays a critical role in improving aircraft performance. Conventional airfoil parameterization methods, such as B-spline curves, free-form deformation (FFD), and PARSEC method, have exhibited success in airfoil shape design [1, 2, 3]. However, naively increasing the number and expanding the bounds of design variables to ensure geometric flexibility could incur extremely inefficient optimization, large portion of infeasible region, or even failed optimization convergence.
To tackle the above challenge, dimensional reduction has therefore become a viable strategy for constructing more compact aerodynamic design spaces. Traditional dimensionality reductions methods, such as principal component analysis, singular value decomposition, proper orthogonal decomposition, and active-subspace-type techniques have been investigated extensively, while more recent studies have increasingly adopted generative artificial intelligence [4, 5, 6, 7, 8, 9, 10, 11]. Chen et al. [12] introduced Bézier-GAN for compact airfoil parameterization. Du et al. [13] developed B-spline-based GAN and reported less than 1% fitting error towards the UIUC airfoil database. Wang et al. [14] proposed Airfoil GAN for latent-space airfoil reconstruction and optimization. Chen and Ramamurthy [3] developed FFD-GAN for compact generative parameterization of aerodynamic geometries. These work leveraged the feature of implicit dimensionality reduction of GAN, i.e., generating only realistic geometries and automatically removing unrealistic ones from the design space. Sisk and Du [15] further demonstrated the two-fold dimensionality reduction, i.e., both implicit and explicit reductions, in aerospace optimization, reducing a 41-variable takeoff-trajectory problem to three variables while retaining high fitting accuracy and feasible-space coverage.
Although GAN-based parameterization provide considerable flexibility, conventional GAN training can suffer from unstable convergence, vanishing gradients, and mode collapse [16, 17, 18]. Wasserstein GAN (WGAN) was introduced by Arjovsky et al. [19] to improve adversarial training by replacing the conventional divergence-based objective with the Wasserstein, or Earth-Mover, distance, while Gulrajani et al. [20] later proposed WGAN-GP using a gradient penalty to improve enforcement of the Lipschitz constraint. Wasserstein-based models have also been applied to airfoil generation. Yonekura et al. [21] used conditional WGAN-GP to generate smooth airfoils satisfying prescribed aerodynamic requirements, and related studies have incorporated Wasserstein objectives into conditional and hybrid generative models for aerodynamic inverse design.
Meanwhile, advances in quantum computing-based machine learning have motivated quantum extensions of classical generative models. Lloyd and Weedbrook [22] introduced the theoretical framework for quantum generative adversarial networks (QGAN), while Situ et al. [23] and Zoufal et al. [24] demonstrated hybrid architectures in which parameterized quantum circuits acted as generators and classical neural networks performed discrimination. Huang et al. [25] experimentally demonstrated QGAN-based image generation and compared quantum and classical generative models. Chakrabarti et al. [26] extended Wasserstein adversarial learning to the quantum setting and reported improved robustness and scalability relative to earlier QGAN formulations in their quantum-data experiments. Reviews of QGAN research show that most existing quantum generative studies focus on quantum states, generic probability distributions, or image-generation benchmarks rather than aerodynamic shape representation [27, 28].
To fill the gap above, this work investigates the performance of quantum computing-based WGAN (Q-WGAN) on airfoil parameterization and compares performance with classical WGAN. The Q-WGAN constructs the generator networks using quantum circuits while keeping the discriminator networks classical to save quantum qubits (see Sec. II.B). The performance metrics include FFD and coordinate fitting accuracy, Jensen–Shannon divergence (JSD) [29], Kullback–Leibler (KL) divergence [30], Fréchet distance (FD) [31], Wasserstein distance, and t-distributed stochastic neighbor embedding (t-SNE) [32]. Please refer to Sec. III for more details.
The rest of this paper is organized as follows. Section II describes the classical WGAN and Q-WGAN, followed by performance metrics. Section III demonstrates and compares the performance of Q-WGAN on airfoil parameterization against classical WGAN. This paper ends with conclusions in Sec. IV.
Methodology
This section elaborates on the classical WGAN and Q-WGAN. This section also introduces the performance metrics, including fitting optimization accuracy and distribution metrics.
Classical Wasserstein Generative Adversarial Network
GAN (Figure 1(a)) consists of two competing neural networks: a generator that maps samples from a latent space to synthetic data; and a discriminator that attempts to distinguish generator-generated data from real samples. During adversarial training, both networks are updated iteratively until the generated distribution approaches the reference data distribution [16]. Conventional GAN training, however, can suffer from unstable gradients and mode collapse. WGAN addresses these issues by replacing the discriminator with a critic that estimates the Wasserstein-1 distance between the real and generated distributions.
Figure 1. Workflow and generator architecture of the Classical Wasserstein GAN.
Specifically, the generator is trained by minimizing the Wasserstein objective:
where is generator loss, refers to expectation over latent vectors sampled from the latent distribution , is Wasserstein discriminator function and is generator function.
The discriminator receives both real and generated data and assigns a scalar score to each sample. During training, it learns to assign higher scores to real samples than to generated samples. Unlike the discriminator of a conventional GAN, no sigmoid activation is applied to the wasserstein discriminator output. The wasserstein discriminator updates five times per generator update. The discriminator is trained by minimizing the following Wasserstein objective:
where is discriminator loss, refers to expectation over real samples drawn from the real-data distribution .
In this work, the classical WGAN is used to generate 18 FFD control points, which parameterize each airfoil. The classical generator is a fully connected neural network with layer dimensions . Leaky-ReLU activation functions are used in the hidden layers, together with batch normalization, while a hyperbolic tangent activation is applied to the output layer to match the normalized FFD-variable range (Figure 1(b)).
The original weight-clipping formulation of WGAN is adopted, with the discriminator parameters constrained to after each discriminator update. The discriminator is updated five times for each generator update, and both the generator and discriminator are trained using the Adam optimizer. Training is done with different combinations of learning rates and got the best result with a learning rate of for both generator and discriminator. The training dataset consists of 20 optimal airfoils from 20 airfoil aerodynamic design optimizations (Sec. III). The details of the classical WGAN are presented in Table 1
| Parameters | Generator | Discriminator |
| Input | 18-dimensional FFD vector | |
| Architecture | ||
| Trainable parameters | 679,442 | 675,841 |
| Learning rate | ||
| Hidden activation | LeakyReLU | LeakyReLU |
| Normalization | Batch normalization | N/A |
| Output | 18-dimensional FFD vector | Scalar score |
| Output activation | Tanh | Linear |
Table 1. Classical WGAN architecture and training parameters
Quantum Computing-Based Wasserstein Generative Adversarial Network
Q-WGAN extends the adversarial learning framework by replacing one or both classical neural networks with parameterized quantum circuits (PQC). In this work, the classical generator of the Classical WGAN is replaced by PQC while retaining a classical Wasserstein discriminator (Figure 2(a)), resulting in a hybrid Q-WGAN. A three-dimensional latent vector is supplied to a three-qubit quantum generator:
Figure 2. Workflow and generator architecture of the Q-WGAN.
Each latent variable is encoded on the corresponding qubit using successive and rotations, where, rotates qubit by an angle of and rotates by an angle of . The encoding transforms the classical latent variables into the quantum state before application of the trainable variational circuit.
The variational portion of the generator consists of repeated quantum layers. Each layer applies trainable , , and rotations to every qubit, followed by a ring of controlled-NOT gates to introduce entanglement between the qubits. For three qubits, the entangling pattern is , , . In each connection, the first qubit acts as the control qubit and the second as the target qubit. This closed-loop structure allows information encoded in one qubit to influence the states of the other qubits, enabling the circuit to represent correlations among the latent variables. The combination of trainable single-qubit rotations and ring entanglement is repeated across the variational layers of the quantum generator.
Therefore, each variational layer contains nine trainable quantum parameters, corresponding to three rotation parameters for each of the three qubits. A 12-layer circuit is used for the final Q-WGAN comparison (Figure 2(b)). Consequently, the number of trainable quantum parameters of final quantum circuit is
After the variational circuit, the expectation value of the Pauli-(Z) operator is measured on each qubit using Eq. (5), producing the three-dimensional quantum output vector using Eq. (6). Because the desired airfoil representation contains 18 FFD variables, the three measured expectation values are subsequently passed through a classical fully connected layer from Eq. (7), where and .
where is generated or reconstructed 18-dimensional FFD vector, means classical output-layer weight matrix, refers to quantum measurement output vector. The hyperbolic tangent activation maps the generated values to the normalized FFD-variable range. The classical output layer contributes trainable parameters. Thus, the 12-layer Q-WGAN generator contains a total of trainable parameters.
The generated 18-dimensional FFD vectors and the real FFD samples are evaluated by a classical Wasserstein discriminator. As in the classical WGAN, no sigmoid activation is applied to the discriminator output because the discriminator produces an unrestricted scalar score rather than a probability. The Q-WGAN is trained using the same Wasserstein objectives,
where denotes the hybrid quantum generator. The original weight-clipping formulation of WGAN is used, with the discriminator weights restricted to . The Q-WGAN is trained with different combinations of learning rates and got the best result with a generator learning rate of and a discriminator learning rate of . Both networks are optimized using the Adam optimizer [33].
The quantum circuit is implemented using PennyLane with the default.qubit simulator. Analytic expectation values are used by setting the number of measurement shots to None. Therefore, the expectation values are evaluated directly from the simulated quantum state rather than estimated using finite-shot sampling. The architecture and training parameters of the Q-WGAN are summarized in Table 2.
| Generator | Discriminator | ||
| Input | Input | 18-dimensional FFD vector | |
| Input encoding | and | Architecture | |
| Entanglement | CNOT | Hidden activation | LeakyReLU () |
| Variational gates | , , | Discriminator updates per generator update | 5 |
| Circuit depth | 12 | Optimizer | Adam |
| Quantum measurement | , , | Batch size | 4 |
| Classical output layer | Loss function | Wasserstein loss | |
| Output activation | Tanh | Output activation | Linear |
| Classical-head parameters | 72 | Weight-clipping range | [-0.01, 0.01] |
| Trainable parameters | 180 | Trainable parameters | 3329 |
| Output | 18-dimensional FFD vector | Output | Scalar score |
| Learning rate | Learning rate |
Table 2. Q-WGAN architecture and training parameters.
Latent-Space Optimization and Reconstruction Accuracy
To test whether a trained generator maintains sufficient variability compared with the training data, this work conducts fitting optimizations to reconstruct real airfoil data. Differential evolution (DE) [34], a population-based global optimization algorithm, is employed to search the bounded latent space (i.e., in this work). DE does not require derivatives with respect to the latent variables and can explore multiple regions of the latent space simultaneously.
The latent-search objective is defined using the relative squared error,
where is the th real FFD control point of a target airfoil, is the th generated FFD control point. DE minimizes with respect to the three latent variables. The optimization bounds are for each latent variable, with a population size of 150, a maximum of 10,000 iterations, a convergence tolerance of , and final solution polishing enabled. Because DE is population based, a single DE optimization is performed for each target airfoil rather than using multiple independent latent-space restarts.
The final reconstruction performance is reported using relative accuracy for FFD control points and reconstructed airfoil coordinates:
where is the norm operator. This two-stage evaluation distinguishes between reconstruction quality in the reduced FFD design space and reconstruction quality of the resulting physical airfoil geometry. The same DE procedure and accuracy definitions are applied to both the classical WGAN and Q-WGAN, enabling a fair comparison between the two types of models.
Distribution Evalution Metrics
In addition to reconstruction accuracy, the quality of the learned generative distributions is evaluated to assess the generative performance. The resulting generated distributions are compared with the real data using JSD, KL divergence, Wasserstein distance, FD, and t-SNE. Lower values of JSD, Wasserstein distance, FD, and KL divergence indicate greater agreement between the generated and reference distributions.
The Jensen–Shannon divergence provides a symmetric measure of similarity between two probability distributions. For a reference distribution () and generated distribution (), the intermediate distribution is defined as . and the Jensen–Shannon divergence is calculated as
where KL denotes the Kullback–Leibler divergence. In this study, the JSD is evaluated independently for each of the FFD control points through histogram-based probability distributions, and the mean JSD value across all FFD variables is reported as the overall JSD.
Kullback–Leibler divergence quantifies the directional difference between the reference and generated probability distributions. It is defined as
The index denotes the discrete histogram bin used to estimate the probability distributions Because KL divergence is asymmetric, both directions are evaluated:
and,
The first direction emphasizes regions of the reference distribution that are insufficiently represented by the generator, whereas the reverse direction is more sensitive to generated probability mass occurring in regions with limited reference support.
Wasserstein distance represents the minimum amount of probability mass that must be transported, multiplied by the distance over which it is moved, to transform the generated distribution into the reference distribution. For the cumulative distribution functions () and (), the first-order Wasserstein distance can be expressed as
The one-dimensional Wasserstein distance is calculated independently for each FFD control point, and the mean Wasserstein distance is used provide an overall measure of distributional discrepancy.
Fréchet distance is used to compare the overall statistical structure of the generated and reference samples by considering their means and covariance matrices. Assuming the reference and generated FFD samples are characterized by (()) and (()), respectively, the Fréchet distance is calculated as
where and represent the corresponding mean vectors and covariance matrices. Unlike the individual-feature JSD and Wasserstein calculations, the FD incorporates relationships among the FFD variables through the covariance matrices. Moreover, t-SNE is employed to visualize high-dimensional joint structure in lower dimensions. Unlike the preceding scalar metrics, t-SNE is used only as a visualization tool and is not interpreted as a quantitative distance measure. Before applying t-SNE, each FFD variable is standardized using statistics obtained from the baseline dataset. The same transformation is then applied to the Q-WGAN and classical WGAN samples. The standardized reference and generated samples are subsequently embedded into a common two-dimensional space. This visualization enables qualitative assessment of whether generated samples occupy neighborhoods similar to those represented by the baseline airfoils and whether the two generative models learn different regions or structures of the original design space.
Results and Discussion
Airfoil Dataset and Reference Problem Formulation
The airfoil dataset used in this study is generated from the aerodynamic design problem summarized in Table 3. The reference formulation considers airfoil shape optimization over a Mach-number range of [0.6, 0.7] and a Reynolds-number range of , with an allowable FFD-point variation of . The coefficient of drag is minimized while aerodynamic and geometric constraints are imposed on lift coefficient, pitching-moment coefficient, airfoil volume, thickness. The reference problem is formulated using the MACH-Aero [35] aerodynamic shape optimization framework, which employs FFD through pyGeo [36] for geometric parameterization and uses ADflow [37] for aerodynamic analysis within a gradient-based optimization framework.
| Variable | Description | Quantity | |
| Minimize: | Coefficient of drag | 1 | |
| With respect to: | FFD control points | 20 | |
| Angle of attack | 1 | ||
| Subject to: | Lift coefficient | 1 | |
| Pitch moment coefficient | 1 | ||
| Volume constraint | 1 | ||
| Thickness constraint | 100 |
Table 3. Optimization variables, objectives, and constraints
This work also imposes leading- and trailing-edge constraints in which the corresponding upper and lower FFD control points move with equal magnitude in opposite directions. These constraints keep the airfoil chord staying on the -axis between (0, 0) and (1, 0). In addition, the 20 independent FFD control points are reduced because the paired leading-edge variables and paired trailing-edge variables are not independent. Each pair is represented by a single independent design variable, resulting in the 18-dimensional FFD representation used throughout this study. These 18 independent FFD control points, therefore, constitute the airfoil dataset used for training and evaluating GAN models.
The complete dataset contains 992 airfoil designs. For the present study, both the classical WGAN and Q-WGAN learn a mapping from a three-dimensional latent space to this 18-dimensional FFD design space. The physical airfoil geometry is recovered by applying the corresponding FFD deformation to the baseline airfoil, producing the final coordinate representation. Model performance is therefore evaluated in both the 18-dimensional FFD design space and the physical coordinate space. From the complete dataset, 20 representative airfoils are selected using conditioned Latin hypercube sampling (cLHS) [38], and the same subset is used for training both models to ensure a fair comparison.
Reconstruction and Fitting Performance
The structures and training details of both classical WGAN and Q-WGAN are shown in Sec. II. Table 4 summarizes the reconstruction performance of the classical WGAN configuration and the investigated Q-WGAN circuit depths. The classical WGAN achieves an FFD fitting accuracy of 92.29%, while the 12-layer Q-WGAN achieves 92.10%. The corresponding airfoil-coordinate fitting accuracy are 99.26% and 99.20%, respectively. These results indicate that the Q-WGAN is able to reproduce the target airfoil geometries with reconstruction performance comparable with the classical WGAN despite using a fundamentally different generator architecture.
| FFD control points (%) | Airfoil coordinates (%) | |
| Classical WGAN | 92.29 | 99.26 |
| Q-WGAN | 92.10 | 99.20 |
Table 4. Fitting accuracy of classical WGAN and Q-WGAN
An important observation is the difference between the FFD-control-point and airfoil-coordinate fitting accuracies. Although the FFD-control-point accuracy is approximately 92%, the corresponding reconstructed airfoils achieve approximately 99% coordinate accuracy. This indicates that deviations in individual FFD design variables do not necessarily produce equally large deviations in the resulting physical airfoil geometry.
Representative reconstructed airfoils are shown in Figure 3. The reconstructed airfoils by classical WGAN and Q-WGAN are compared against arbitrary reference airfoils. Both models closely reproduce the overall airfoil profile, with only minor differences visible in the reconstructed geometry.

Figure 3. Reconstructed airfoils using classical WGAN and Q-WGAN.
Distribution Fidelity
Besides the ability to reconstruct reference airfoils, the trained generative models are evaluated based on distribution metrics, namely, JSD, FD, Wasserstein distance, and KL divergence. Note that lower values of these metrics indicate closer agreement between the learned and reference distributions.
Table 5 shows that Q-WGAN achieves a lower mean JSD of 0.496 than that of the classical WGAN. Similarly, Q-WGAN obtains , which is 47.2% lower than the counterpart of the classical WGAN. Additionally, the mean Wasserstein distance of Q-WGAN is also lower than that of classical Q-WGAN, presenting a 52.6% reduction. These three metrics unanimously and consistently indicate that Q-WGAN-generated data distribution more closely matches the baseline distribution than the classical WGAN.
| Model | JSD | FD | Wasserstein |
| Classical WGAN | 0.505 | ||
| Quantum WGAN | 0.496 |
Table 5. Distribution metrics comparison for classical WGAN and Q-WGAN
Table 6 presents that the KL-divergence results provide a more mixed comparison because KL divergence is directional. For the baseline-to-generated direction, the mean KL divergence of the Q-WGAN model is higher than the classical WGAN. At the individual-variable level, each model achieves the lower KL divergence for 9 of the 18 FFD variables. Therefore, despite the lower KL divergence by the classical WGAN, neither model consistently outperforms the other across all individual FFD variables.
| classical WGAN | Q-WGAN | |
| Reference Generated | 1.35 | 4.74 |
| Lower KL divergence | 9/18 | 9/18 |
Table 6. KL divergence comparison for baseline-to-generated direction
Table 7 exhibits that for the generated-to-baseline direction, the Q-WGAN achieves a slightly lower mean KL divergence compared with the classical WGAN. In this setup, the Q-WGAN produces lower KL divergence values for 11 of the 18 FFD variables, while the classical WGAN produces the lower value for 7 variables. Therefore, the Q-WGAN and classical WGAN performance are still comparable with each other.
| classical WGAN | Q-WGAN | |
| Generated → Reference | 2.31 | 2.27 |
| Lower KL divergence | 7/18 | 11/18 |
Table 7. KL divergence comparison for generated-to-baseline direction
The difference between the two KL directions is consistent with the asymmetric nature of KL divergence. The baseline-to-generated form penalizes regions of the reference distribution that receive insufficient probability coverage from the generated distribution, whereas the reverse direction is more sensitive to generated probability appearing in regions with limited baseline support. Because the reference set contains only 20 cLHS-selected airfoils, histogram-based KL estimates are also sensitive to bin occupancy and should therefore be interpreted together with the other distribution metrics rather than as an isolated measure of generator quality.
The individual distributions of the 18 FFD variables are shown in Figure 4. The classical WGAN exhibits stronger concentration near the boundaries of several FFD-variable ranges, whereas the Q-WGAN generally produces a broader distribution of samples within the interior of the baseline range. This behavior is consistent with the lower JSD, FD, and Wasserstein distance of Q-WGAN, observed in the quantitative comparison. However, the individual plots do not fully represent the multivariate relationships among the 18 FFD variables; the joint structure of the generated design spaces is therefore examined separately using t-SNE in the following subsection.

Figure 4. Distribution showcase for classical WGAN and Q-WGAN.
Joint Distribution Analysis Using t-SNE
To examine the joint structure of the learned airfoil design spaces, t-SNE is applied to the FFD representation, classical WGAN generated, and Q-WGAN generated samples. Specifically, the FFD variables are first standardized using statistics obtained from the baseline dataset, and the same transformation is applied to the generated samples. The three datasets are then embedded together into common t-SNE spaces for qualitative comparison of their local neighborhood structures (Figure 5).

Figure 5. t-SNE visualization of the airfoil parameterization in 2D and 3D spaces.
Figure 5(a) shows that the baseline airfoils are distributed across regions occupied by both models. Several baseline points appear embedded within or close to neighborhoods populated by Q-WGAN samples, while other baseline points lie near classical WGAN-generated regions. This indicates that neither generative model reproduces the complete baseline structure identically. However, the Q-WGAN appears to retain stronger local overlap with several of the baseline neighborhoods, whereas the classical WGAN shows greater dispersion into regions that are less densely populated by the baseline samples.
The resulting 3D embedding is shown in Figure 5(b). The classical WGAN and Q-WGAN samples occupy noticeably different regions and form distinct structures in the embedded design space, indicating that the two generators learn different representations of the original FFD distribution. The Q-WGAN samples are concentrated more strongly through the central region of the embedding, while the classical WGAN samples form broader and more separated branches extending across the t-SNE space.
These observations are consistent with the quantitative distribution metrics presented in the previous subsection. The Q-WGAN achieves lower JSD, FD, and Wasserstein distance than the classical WGAN, suggesting closer overall agreement with the baseline distribution under the selected configurations. At the same time, the t-SNE visualization reveals that the two models differ also in the structure of the design spaces they generate.
Conclusion
This study developed a hybrid quantum Wasserstein generative adversarial networks (Q-WGAN), consisting of quantum generator networks and classical discriminator networks, for airfoil parameterization. Both models were trained using only 20 optimal airfoils and mapped a three-dimensional latent space to the free-form deformation (FFD) control-point space. The trained Q-WGAN was then evaluated using reconstruction accuracy, distribution-based metrics, and t-distributed stochastic neighbor embedding visualization, and compared against classical Wasserstein generative adversarial networks (WGAN).
The results showed that both classical WGAN and Q-WGAN models were able to reconstruct reference airfoil shapes with high accuracy. The Q-WGAN achieved almost the same fitting accuracies on FFD control points and airfoil coordinates. Moreover, in the distribution metric comparisons, the Q-WGAN obtained lower Jensen–Shannon divergence, Fréchet distance, and Wasserstein distance, representing a closer match with the baseline distribution. The Kullback-Leibler divergence values were mixed depending on the direction of comparison, showing a comparable performance between classical WGAN and Q-WGAN. The t-distributed stochastic neighbor embedding analysis also showed that the classical WGAN and Q-WGAN learned different structures in the design space, with the Q-WGAN showing stronger overlap with the baseline distribution.
Overall, the results indicated that the Q-WGAN can provide reconstruction performance comparable to the classical WGAN while using a much smaller generator architecture. At the same time, the distribution analyses suggested that the Q-WGAN captured distribution characteristics of the baseline design space using a highly limited training dataset. These results would support further applications of Q-WGAN for aerodynamic shape optimization.
Future work will focus on evaluating the method using larger and more diverse airfoil datasets, studying the effect of training-set size, and extending the framework toward aerodynamic optimization tasks with additional physical and performance constraints.
References
- [1]Li, J., Du, X., and Martins, J. R., “Machine learning in aerodynamic shape optimization,” Progress in Aerospace Sciences, Vol. 134, 2022, p. 100849.DOI
- [2]Chen, W., Chiu, K., and Fuge, M., “Aerodynamic design optimization and shape exploration using generative adversarial networks,” AIAA Scitech 2019 forum, 2019, p. 2351.DOI
- [3]Chen, W., and Ramamurthy, A., “Deep generative model for efficient 3D airfoil parameterization and generation,” AIAA Scitech 2021 Forum, 2021, p. 1690.DOI
- [4]Yasong, Q., Junqiang, B., Nan, L., and Chen, W., “Global aerodynamic design optimization based on data dimensionality reduction,” Chinese Journal of Aeronautics, Vol. 31, No. 4, 2018, pp. 643–659.DOI
- [5]Viswanath, A., Forrester, A. I., and Keane, A., “Dimension reduction for aerodynamic design optimization,” AIAA journal, Vol. 49, No. 6, 2011, pp. 1256–1266.DOI
- [6]Alswaitti, M., Siddique, K., Jiang, S., Almoush, W., and Alrosan, A., “Dimensionality reduction, modelling, and optimization of multivariate problems based on machine learning,” Symmetry, Vol. 14, No. 7, 2022, p. 1282.DOI
- [7]Jun, T., Gang, S., Liqiang, G., and Xinyu, W., “Application of a PCA-DBN-based surrogate model to robust aerodynamic design optimization,” Chinese Journal of Aeronautics, Vol. 33, No. 6, 2020, pp. 1573–1588.DOI
- [8]Cinquegrana, D., and Iuliano, E., “Investigation of adaptive design variables bounds in dimensionality reduction for aerodynamic shape optimization,” Computers & Fluids, Vol. 174, 2018, pp. 89–109.DOI
- [9]Tejero, F., MacManus, D., Hueso-Rebassa, J., Sanchez-Moreno, F., Goulous, I., and Sheaf, C., “Aerodynamic optimisation of civil aero-engine nacelles by dimensionality reduction and multi-fidelity techniques,” International Journal of Numerical Methods for Heat & Fluid Flow, Vol. 33, No. 4, 2023, pp. 1336–1353.
- [10]Wu, X., Ma, L., and Zuo, Z., “High-dimensional aerodynamic shape optimization framework using geometric domain decomposition and data-driven support strategy for wing design,” Aerospace Science and Technology, Vol. 149, 2024, p. 109152.DOI
- [11]Bo, P., Zhang, Y., Junlin, L., Xudong, W., Chang, M., and Junqiang, B., “Data-driven surrogate model for aerodynamic design using separable shape tensor method,” Chinese Journal of Aeronautics, Vol. 37, No. 9, 2024, pp. 41–58.
- [12]Chen, W., Chiu, K., and Fuge, M. D., “Airfoil design parameterization and optimization using bézier generative adversarial networks,” AIAA journal, Vol. 58, No. 11, 2020, pp. 4723–4735.arxiv.org/abs/2006.12496
- [13]Du, X., He, P., and Martins, J. R., “A B-spline-based generative adversarial network model for fast interactive airfoil aerodynamic optimization,” AIAA scitech 2020 forum, 2020, p. 2128.DOI
- [14]Wang, Y., Shimada, K., and Farimani, A. B., “Airfoil GAN: Encoding and synthesizing airfoils for aerodynamic-aware shape optimization,” arXiv preprint arXiv:2101.04757, 2021.DOI
- [15]Sisk, S., and Du, X., “Physics-Constrained Generative Adversarial Networks for Dimensionality Reduction in Optimization,” AIAA SCITECH 2026 Forum, 2026, p. 2016.DOI
- [16]Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y., “Generative adversarial nets,” Advances in neural information processing systems, Vol. 27, 2014.
- [17]Wu, X., Shi, C., Li, X., He, J., Wu, X., Lv, J., and Zhou, J., “PWGAN: Wasserstein GANs with perceptual loss for mode collapse,” Proceedings of the ACM Turing Celebration Conference-China, 2019, pp. 1–7.DOI
- [18]Chen, Y., and Hou, X., “An improvement based on Wasserstein GAN for alleviating mode collapsing,” 2020 international joint conference on neural networks (IJCNN), IEEE, 2020, pp. 1–7.
- [19]Arjovsky, M., Chintala, S., and Bottou, L., “Wasserstein generative adversarial networks,” International conference on machine learning, Pmlr, 2017, pp. 214–223.
- [20]Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., and Courville, A., “Improved training of wasserstein gans,” Advances in neural information processing systems, Vol. 30, 2017.arxiv.org/abs/1704.00028
- [21]Yonekura, K., Tomori, Y., and Suzuki, K., “Airfoil shape generation and feature extraction using the conditional vae-wgan-gp,” AI, Vol. 5, No. 4, 2024, pp. 2092–2103.
- [22]Lloyd, S., and Weedbrook, C., “Quantum generative adversarial learning,” Physical review letters, Vol. 121, No. 4, 2018, p. 040502.DOI
- [23]Situ, H., He, Z., Wang, Y., Li, L., and Zheng, S., “Quantum generative adversarial network for generating discrete distribution,” Information Sciences, Vol. 538, 2020, pp. 193–208.DOI
- [24]Zoufal, C., Lucchi, A., and Woerner, S., “Quantum generative adversarial networks for learning and loading random distributions,” npj Quantum Information, Vol. 5, No. 1, 2019, p. 103.
- [25]Huang, H.-L., Du, Y., Gong, M., Zhao, Y., Wu, Y., Wang, C., Li, S., Liang, F., Lin, J., Xu, Y., et al., “Experimental quantum generative adversarial networks for image generation,” Physical Review Applied, Vol. 16, No. 2, 2021, p. 024051.DOI
- [26]Chakrabarti, S., Yiming, H., Li, T., Feizi, S., and Wu, X., “Quantum Wasserstein generative adversarial networks,” Advances in Neural Information Processing Systems, Vol. 32, 2019.arxiv.org/abs/1911.00111
- [27]Kushwaha, V., Nandi, G., et al., “Study of prevention of mode collapse in generative adversarial network (GAN),” 2020 IEEE 4th Conference on Information & Communication Technology (CICT), IEEE, 2020, pp. 1–6.DOI
- [28]Ngo, T. A., Nguyen, T., and Thang, T. C., “A survey of recent advances in quantum generative adversarial networks,” Electronics, Vol. 12, No. 4, 2023, p. 856.DOI
- [29]Lin, J., “Divergence measures based on the Shannon entropy,” IEEE Transactions on Information theory, Vol. 37, No. 1, 1991, pp. 145–151.DOI
- [30]Kullback, S., and Leibler, R. A., “On information and sufficiency,” The annals of mathematical statistics, Vol. 22, No. 1, 1951, pp. 79–86.DOI
- [31]Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S., “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, Vol. 30, 2017.
- [32]Van der Maaten, L., and Hinton, G., “Visualizing data using t-SNE,” Journal of machine learning research, Vol. 9, No. 11, 2008.
- [33]Kingma, D. P., and Ba, J., “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.arxiv.org/abs/1412.6980
- [34]Storn, R., and Price, K., “Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces,” Journal of global optimization, Vol. 11, No. 4, 1997, pp. 341–359.DOI
- [35]Martins, J. R., “Aerodynamic design optimization: Challenges and perspectives,” Computers & Fluids, Vol. 239, 2022, p. 105391.DOI
- [36]Hajdik, H. M., Yildirim, A., Wu, E., Brelje, B. J., Seraj, S., Mangano, M., Anibal, J. L., Johnson, E., Adler, E. J., Mader, C. A., et al., “pyGeo: A geometry package for multidisciplinary design optimization,” Journal of Open Source Software, Vol. 8, No. 87, 2023, p. 5319.DOI
- [37]Mader, C. A., Kenway, G. K., Yildirim, A., and Martins, J. R., “ADflow: An open-source computational fluid dynamics solver for aerodynamic and multidisciplinary optimization,” Journal of Aerospace Information Systems, Vol. 17, No. 9, 2020, pp. 508–527.DOI
- [38]Minasny, B., and McBratney, A. B., “A conditioned Latin hypercube method for sampling in the presence of ancillary information,” Computers & geosciences, Vol. 32, No. 9, 2006, pp. 1378–1388.DOI

