Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Adjoint-Based Aerodynamic Shape Optimization with a Manifold Constraint Learned by Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that constraining aerodynamic shape optimization to a diffusion-model-learned manifold of viable airfoils, with gradients chained from the discrete adjoint through the diffusion model's latent space, removes ad hoc…

desk verdict A clean gradient chain through a diffusion manifold, but the empirical claims outrun the evidence. read the letter →

arxiv 2507.23443 v1 pith:HAJYLXXR submitted 2025-07-31 cs.CE cs.LGmath.OC

classification cs.CEcs.LGmath.OC
keywords aerodynamicshapeoptimizationdiffusionmodelsdiscreteadjointmanifoldconstraintHicks-HenneparameterizationtransonicRANSautomaticdifferentiationDDIMsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's central claim is that aerodynamic shape optimization can be made more robust and more accurate by restricting the search to a smooth, low-dimensional manifold of viable airfoil shapes learned by a diffusion model from a corpus of existing designs. The method does not replace the design parameterization; it constrains the Hicks-Henne parameters to lie on the learned manifold, and derives gradients of the drag and lift objectives with respect to the manifold's latent coordinates by chaining the discrete adjoint of the RANS solver with backpropagation through the diffusion model. In transonic drag-minimization tests with lift and thickness constraints, the paper reports that this approach removes the need for ad hoc scaling and tuning, is insensitive to initialization and optimizer choice, and achieves lower drag than conventional or physics-scaled Hicks-Henne optimization, with the gap widening as constraints tighten. A sympathetic reader would take this as evidence that learned generative priors can act as effective, differentiable constraints for high-fidelity adjoint-based design.

What carries the argument

The load-bearing object is the learned manifold $X_G=\{x\in\mathbb{R}^{40}: x=G_\theta(z),\, z\in Z\}$, where $G_\theta$ is the deterministic DDIM generator from a 40-dimensional Gaussian latent space to Hicks-Henne parameters. The argument is carried by the chain-rule adjoint $\frac{dJ}{dz}=\frac{\partial J}{\partial x}\frac{\partial x}{\partial z}$: the discrete adjoint of the RANS solver supplies $\partial J/\partial x$, and reverse-mode automatic differentiation through the denoising network supplies $\partial x/\partial z$. The paper also uses a theoretical signature to validate the manifold picture: for a $k$-dimensional manifold in $d$-dimensional space, the Hessian of the log-density of the diffused distribution should have rank $d-k$, so the denoiser Jacobian's singular values should drop sharply; the paper reports such drops, with effective rank decreasing under tighter lift constraints.

What would settle it

Run each reported test case with a full-dimensional Hicks-Henne parameterization, using many restarts, careful scaling, and no frozen chord regions, with a much larger evaluation budget; if any such run finds a feasible design with drag below the diffusion-constrained optimum, then the claim that the learned manifold contains the near-optimal region is falsified. A complementary check: project each optimized diffusion design back to the training set and measure nearest-neighbor distance; if all optima sit essentially on training shapes, the manifold memorized the data instead of generalizing.

Watch

Extended reading notes

Core claim

On the paper's own terms: drag minimization over airfoil shapes can be reformulated as optimization over the latent variable $z$ of a deterministic denoising diffusion implicit model. The generator $x=G_\theta(z)$ maps $z$ to a 40-dimensional Hicks-Henne parameter vector, defining a low-rank set $X_G$ of aerodynamically viable shapes. Gradients are formed by chaining the discrete adjoint of the RANS solver, which gives $dJ/dx$, with backpropagation through the diffusion model, which gives $\partial x/\partial z$. The paper reports that this manifold-constrained formulation, tested on transonic RANS cases at Mach 0.8 with lift and thickness constraints, achieves lower drag counts than conventional and scaled Hicks-Henne baselines, requires no manual scaling, and performs robustly across two initializations and two off-the-shelf optimizers. It also reports a signature supporting the manifold picture: the Jacobian of the denoiser's score function at optimized shapes has sharply dropping singular values, with lower effective rank under tighter lift constraints.

Load-bearing premise

The method's gains all rest on the assumption that the set of shapes the diffusion model can produce, learned from 1,568 existing airfoil designs, contains shapes close to the true optimum for each transonic drag-minimization case; if the true best shape lies outside that learned set, no optimizer can find it.

Editorial extensions

If this is right

  • On the six transonic test problems, the diffusion-constrained method matches or beats conventional Hicks-Henne drag counts in every case, with the margin growing from a few counts at loose constraints to hundreds of counts at tight constraints.
  • The method works out of the box with two structurally different nonlinear optimizers, whereas the authors report they could not find a usable parameter set for the conventional parameterization on one of them.
  • Initialization robustness: starting from a symmetric airfoil farther from the constrained optimum, the method still reaches competitive or better designs, while the conventional baseline degrades noticeably.
  • The manifold constraint changes the optimization landscape: different optimizers converge to visibly different shapes with nearly identical lift, thickness, and drag, which the paper reads as evidence of a benign landscape and gives designers multiple acceptable solutions.
  • Integration cost is low because the diffusion model is trained offline; optimization-time use is inference and backpropagation through the generator.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural testable extension is to compare the diffusion-manifold results against a global or heavily restarted full-dimensional search; the paper's richness claim is central and currently supported only by beating local baselines.
  • The same latent-adjoint chaining should transfer to other PDE-constrained design problems if a differentiable generator of viable designs exists; the framework is parameterization-agnostic, so the key ingredient is a deterministic smooth generator, not necessarily diffusion-specific.
  • One could probe how far the manifold generalizes by testing initializations or target operating points far outside the training distribution; if performance degrades smoothly with distance, the prior is behaving like a regularizer rather than a hard filter.
  • Because the gap over baselines is largest for the most restrictive constraints, an implicit use of the manifold as a learned constraint may be effectively enforcing thickness and smoothness; measuring which constraints the manifold itself already enforces would clarify the mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a manifold-constrained aerodynamic shape optimization framework in which a diffusion model (DDIM with deterministic sampling) is trained on 1568 UIUC airfoils represented by 40 Hicks-Henne parameters, and the optimization is performed over the latent variable z of the generative map x = G(z). The objective gradient with respect to z is obtained by chaining the SU2 discrete adjoint gradient dJ/dx with backpropagation through the DDIM sampling process. The method is tested on transonic RANS drag-minimization problems at Mach 0.8 with lift and thickness constraints, using SLSQP and IPOPT and two different initial airfoils. The paper reports lower drag counts than unscaled and scaled Hicks-Henne baselines and claims robustness across initializations and optimizers without ad hoc scaling or tuning.

Significance. If the empirical claims hold, the paper would make a useful contribution by showing how a generative prior can be integrated into an adjoint-based design loop with a clean chain-rule derivation and without replacing the existing Hicks-Henne parameterization. The deterministic DDIM formulation and the explicit two-stage reverse-mode gradient computation are mathematically sound and practically appealing. The low-rank Jacobian analysis in Section III.B is a nice theoretical signature of an active manifold constraint, and the authors honestly state that it is necessary but not sufficient. However, the central claim that the learned manifold contains near-optimal transonic solutions is not yet supported by the experiments: all comparisons are against Hicks-Henne baselines, some of which violate the constraints, and no comparison with a well-tuned full-dimensional optimum is provided. The contribution is therefore promising but needs substantially stronger empirical validation.

major comments (4)
  1. [Section I.B.3 and Section IV] The claim that the learned manifold contains optimal solutions x* with very high probability is not empirically established. The experiments compare the diffusion-manifold method only against Hicks-Henne baselines started from the same initialization; there is no comparison against a full-dimensional HHM optimum obtained by multi-start, nor against any reference optimum for these transonic problems. Section III.B explicitly concedes that the low-rank Jacobian observations are necessary but not sufficient to establish the manifold assumptions. Consequently, the drag values in Tables 1-4 could be the best within a restrictive subset of the design space rather than near the true constrained optimum. Please add a check such as continuing optimization from the DM optimum in the full Hicks-Henne space, or a multi-start HHM search, and report the resulting performance gap.
  2. [Tables 1-3] Several comparisons include Hicks-Henne designs that violate the stated constraints. For example, Table 1 row (0.50, 0.120) reports HHM Cl = 0.4902 and tc = 0.1127 with Cd = 921.7 against required lower bounds Cl = 0.50 and tc = 0.120, and Table 2 contains multiple HHM rows with tc below 0.120 or Cl below 0.50. Because these baselines are infeasible, the reported drag advantage of the diffusion-manifold method is not a like-for-like feasible comparison, and the claim of superior aerodynamic performance is not supported. The selection tolerance in Eq. (44) applies to the chosen solution, but the tables should separate feasible from infeasible rows or provide a constraint-handling version of HHM with comparable feasibility.
  3. [Section IV and Tables 1-4] No mesh-convergence study, flow/adjoint convergence tolerances, or run-to-run variability is reported, while some reported differences are as small as 1.3 drag counts (e.g., DM 218.7 vs. scaled HHM 222.0 for RAE (0.30, 0.105), and DM 286.7 vs. HHM 286.8 for (0.40, 0.105)). Without a grid-refinement study and a clear description of how the 'best optimal solution' is selected among candidates satisfying the epsilon tolerance in Eq. (44), it is unclear whether the reported margins are above numerical noise. Please add grid-convergence data, report the convergence criteria for the SU2 forward and adjoint solves, and clarify the selection protocol, including the number of candidate runs or restarts.
  4. [Section IV and Eq. (4)] The latent variable z is defined on a bounded set Z in Eq. (4), but the manuscript does not state whether bounds on z are supplied to SLSQP or IPOPT. If the optimizers are allowed to move z outside Z, the actual search set is G(R^d) rather than X_G, and the low-rank manifold analysis of Section III.B does not apply to the optimized points. Please specify the bounds used for the latent variables, or justify that unconstrained z remains in the relevant region of the trained distribution.
minor comments (5)
  1. [Section II.C] The notation in Eqs. (33)-(37) uses ∂J/∂x for what is actually the total derivative computed by the SU2 discrete adjoint, while Eq. (5) uses dJ/dx for the total derivative and ∂J/∂x for the partial derivative; please clarify the notation to avoid confusing the two.
  2. [Section III.A] The inverse problem in Eq. (39) that fits 40 Hicks-Henne parameters to the UIUC airfoils is not validated: no fitting error is reported, and it is unclear whether all 1568 airfoils were fit successfully. Please report the distribution of fitting errors and describe how the normalized HH parameters generated by the diffusion model are unnormalized before being passed to SU2.
  3. [Various] There are several typos and incomplete sentences, including 'siginifantly' (Section IV.D), 'test seires' and 'infeior' (Section IV.B), 'marketed' instead of 'marked' (Section IV.A), and 'Hickse-Henne' (Section III.A). The captions of Figures 3-5 are also incomplete ('for optimized shape with and constraint').
  4. [Table 3] The NACA0012 row for (0.40, 0.120) reports Cl = 0.5000, which appears inconsistent with the constraint value 0.40 and with the neighboring rows; please verify this entry.
  5. [References] Reference [8] appears to point to a Computers & Fluids URL while citing a Structural and Multidisciplinary Optimization paper; please correct the link. Also, the paper would benefit from a data/code availability statement, since the trained diffusion model and the exact SU2 configuration are needed for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the diffusion-manifold optimization chain is a genuine composition of an externally trained generative model with an external CFD adjoint, and the unverified manifold-richness premise is a validity gap, not a circular definition.

full rationale

Walking the claimed derivation chain, I find no step in which a prediction or first-principles result is equivalent to its inputs by construction. The central gradient formula (Eqs. 33-37) is the exact chain rule dJ/dz = (dJ/dx)(dx/dz), where dJ/dx is the SU2 discrete adjoint (external open-source solver) and dx/dz is backpropagation through the DDIM generator; neither term is fitted to the reported drag values, and the optimization results do not feed back into the diffusion training. The manifold X_G is defined from the generator (Eq. 4), but the claim that it contains optimal designs is an empirical hypothesis, not a definition. The low-rank Jacobian analysis in Section III.B is explicitly self-limiting: the authors state the observations are 'necessary if the manifold constraint is active and are not sufficient to establish the validity of the assumptions (A1) and (A2),' so the paper itself avoids turning this check into a circular proof. The only self-citations (SU2 adjoint works by Gauger's group) are tool citations backed by a widely used, externally validated open-source implementation; they are not used to justify the manifold-richness premise. What remains is a genuine validity risk: Section I.B.3 asserts the manifold 'contains optimal solutions x* with very high probability' and points to Section IV, but Section IV only demonstrates existence of low-drag solutions within the manifold, not containment of the unconstrained optimum, and the HHM baseline is sometimes infeasible. That is an unsupported assumption or weak-baseline comparison, not a circular reduction, and it does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on the diffusion model having learned a useful low-dimensional manifold from independent airfoil data, plus standard assumptions about RANS adjoint accuracy and Hicks-Henne expressiveness. No parameter is fitted to the optimization outcomes themselves, but several modeling choices are hand-picked and not fully documented.

free parameters (3)
  • Diffusion training hyperparameters (learning rate, batch size, training steps) = lr=1e-5, batch=2, steps=140,000
    Chosen by hand because the dataset is small; no ablation shows sensitivity of the optimization results to these values.
  • DDIM sampling step count
    The number of sampling steps used in the deterministic generator G is not reported; it affects the map and gradient quality.
  • Normalization of Hicks-Henne parameters to [0,1]
    A fixed preprocessing choice that is a form of scaling embedded in the model; the paper claims no scaling is needed.
assumptions (5)
  • domain assumption A1: The population distribution of aerodynamically viable shapes is locally supported on a compact smooth manifold of dimension k<d.
    Stated in Section III.B; used to derive the low-rank Jacobian prediction and to justify why restricting to X_G should preserve design quality.
  • ad hoc to paper A2: The diffusion model achieves perfect score matching (zero training loss).
    Stated in Section III.B; unrealistic in practice and acknowledged by the authors; used only for the theoretical signature, not for the gradient computation.
  • domain assumption The UIUC airfoil database is a representative sample of aerodynamically viable shapes relevant to transonic drag minimization.
    Section III.A; the whole manifold richness claim depends on this.
  • domain assumption The 40-parameter Hicks-Henne representation is expressive enough to contain good transonic optima and to represent training airfoils accurately.
    Section III.A inverse fitting; fitting error is not reported.
  • domain assumption SU2 discrete adjoint with the Spalart-Allmaras model provides accurate gradients for the objective and constraints.
    Section II.A; the paper relies on the solver's correctness without reporting verification in this paper.
invented entities (1)
  • Learned manifold X_G generated by deterministic DDIM map
    purpose: Restricts the design space to shapes consistent with the training distribution, acting as an implicit constraint in the optimization problem.
    Its existence and quality are only evidenced by the training data and by the low-rank Jacobian check on the same model; there is no external benchmark showing it contains true constrained optima, and the training set may overlap with test initializations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adjoint-Based Aerodynamic Shape Optimization with a Manifold Constraint Learned by Diffusion Models." pith.science (2026). https://pith.science/paper/HAJYLXXR

@misc{pith2026250723443,
  author       = {Pith},
  title        = {Pith review of: Adjoint-Based Aerodynamic Shape Optimization with a Manifold Constraint Learned by Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HAJYLXXR}},
  note         = {Machine review of arXiv:2507.23443}
}
read the original abstract

We introduce an adjoint-based aerodynamic shape optimization framework that integrates a diffusion model trained on existing designs to learn a smooth manifold of aerodynamically viable shapes. This manifold is enforced as an equality constraint to the shape optimization problem. Central to our method is the computation of adjoint gradients of the design objectives (e.g., drag and lift) with respect to the manifold space. These gradients are derived by first computing shape derivatives with respect to conventional shape design parameters (e.g., Hicks-Henne parameters) and then backpropagating them through the diffusion model to its latent space via automatic differentiation. Our framework preserves mathematical rigor and can be integrated into existing adjoint-based design workflows with minimal modification. Demonstrated on extensive transonic RANS airfoil design cases using off-the-shelf and general-purpose nonlinear optimizers, our approach eliminates ad hoc parameter tuning and variable scaling, maintains robustness across initialization and optimizer choices, and achieves superior aerodynamic performance compared to conventional approaches. This work establishes how AI generated priors integrates effectively with adjoint methods to enable robust, high-fidelity aerodynamic shape optimization through automatic differentiation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 18 canonical work pages

  1. [1]

    Investigation into aerodynamic shape optimization of planar and nonplanar wings,

    Koo, D., and Zingg, D. W., “Investigation into aerodynamic shape optimization of planar and nonplanar wings,”AIAA Journal, Vol. 56, No. 1, 2018, pp. 250–263

  2. [2]

    Dynamic geometry control for robust aerodynamic shape optimization,

    Streuber, G. M., and Zingg, D. W., “Dynamic geometry control for robust aerodynamic shape optimization,”AIAA Aviation 2021 Forum, 2021, p. 3031

  3. [3]

    Sensitivity-based geometric parameterization for aerodynamic shape optimization,

    Wu, E., Mader, C., and Martins, J. R., “Sensitivity-based geometric parameterization for aerodynamic shape optimization,” AIAA AVIATION 2022 Forum, 2022, p. 3931. 22 (a) Bounds(𝐶𝑙,𝑡𝑐)=(0.30, 0.105) (b) Bounds(𝐶𝑙,𝑡𝑐)=(0.30, 0.120) (c) Bounds(𝐶𝑙,𝑡𝑐)=(0.40, 0.105) (d) Bounds(𝐶𝑙,𝑡𝑐)=(0.40, 0.120) (e) Bounds(𝐶𝑙,𝑡𝑐)=(0.50, 0.105) (f) Bounds(𝐶𝑙,𝑡𝑐)=(0.50, 0.120)...

  4. [4]

    Gradient-limiting shape control for efficient aerodynamic optimization,

    Kedward, L., Allen, C. B., and Rendall, T., “Gradient-limiting shape control for efficient aerodynamic optimization,”AIAA Journal, Vol. 58, No. 9, 2020, pp. 3748–3764

  5. [5]

    Aerostructural design exploration of a wing in transonic flow,

    Bons, N. P., and Martins, J. R., “Aerostructural design exploration of a wing in transonic flow,”Aerospace, Vol. 7, No. 8, 2020, p. 118

  6. [6]

    SU2 Tutorials,

    SU2 development team, “SU2 Tutorials,”https://su2code.github.io/tutorials/, n.d. Accessed: 2025-07-04

  7. [7]

    Proper Orthogonal Decomposition, surrogate modelling and evolutionary optimization in aerodynamic design,

    Iuliano, E., and Quagliarella, D., “Proper Orthogonal Decomposition, surrogate modelling and evolutionary optimization in aerodynamic design,”Computers & Fluids, Vol. 84, 2013, pp. 327–350. https://doi.org/https://doi.org/10.1016/j.compfluid. 2013.06.007, URL https://www.sciencedirect.com/science/article/pii/S0045793013002223

  8. [8]

    Improving kriging surrogates of high-dimensional design models by PartialLeastSquaresdimensionreduction,

    Bouhlel, M. A., Bartoli, N., Otsmane, A., and Morlier, J., “Improving kriging surrogates of high-dimensional design models by PartialLeastSquaresdimensionreduction,” StructuralandMultidisciplinaryOptimization ,Vol.53,2016,p.935–952. https://doi. org/https://doi.org/10.1007/s00158-015-1395-9, URL https://www.sciencedirect.com/science/article/pii/S0045793013002223

Show all 27 references
  1. [9]

    Denoising diffusion probabilistic models,

    Ho, J., Jain, A., and Abbeel, P., “Denoising diffusion probabilistic models,”Advances in neural information processing systems, Vol. 33, 2020, pp. 6840–6851

  2. [10]

    Diffusion models encode the intrinsic dimension of data manifolds,

    Stanczuk, J. P., Batzolis, G., Deveney, T., and Schönlieb, C.-B., “Diffusion models encode the intrinsic dimension of data manifolds,”Forty-first International Conference on Machine Learning, 2024

  3. [11]

    Diffairfoil: An efficient novel airfoil sampler based on latent space diffusion model for aerodynamic shape optimization,

    Wei, Z., Dufour, E. R., Pelletier, C., Fua, P., and Bauerheim, M., “Diffairfoil: An efficient novel airfoil sampler based on latent space diffusion model for aerodynamic shape optimization,”AIAA AVIATION FORUM AND ASCEND 2024, 2024, p. 3755

  4. [12]

    Airfoil Diffusion: Denoising Diffusion Model For Conditional Airfoil Generation,

    Graves, R., and Farimani, A. B., “Airfoil Diffusion: Denoising Diffusion Model For Conditional Airfoil Generation,”arXiv preprint arXiv:2408.15898, 2024

  5. [13]

    Generative Aerodynamic Design with Diffusion Probabilistic Models,

    Wagenaar, T., Mancini, S., and Mateo-Gabín, A., “Generative Aerodynamic Design with Diffusion Probabilistic Models,”arXiv preprint arXiv:2409.13328, 2024

  6. [14]

    An Introduction to the Adjoint Approach to Design,

    Giles, M., and Pierce, N., “An Introduction to the Adjoint Approach to Design,”Flow, Turbulence and Combustion, Vol. 65,

  7. [15]

    Albring,T.,Sagebaum,M.,andGauger,N., AConsistentandRobustDiscreteAdjointSolverfortheSU2Framework—Validation and Application, 2016, pp. 77–86. https://doi.org/10.1007/978-3-319-27279-5_7

  8. [16]

    High-Performance Derivative Computations using CoDiPack,

    M. Sagebaum, N. G., T. Albring, “High-Performance Derivative Computations using CoDiPack,”ACM Transactions on Mathematical Software (TOMS), Vol. 45, No. 4, 2019. URL https://dl.acm.org/doi/abs/10.1145/3356900

  9. [17]

    Numerical sensitivity analysis for aerodynamic optimization: A survey of approaches,

    Peter, J. E., and Dwight, R. P., “Numerical sensitivity analysis for aerodynamic optimization: A survey of approaches,” Computers & Fluids, Vol. 39, No. 3, 2010, pp. 373–391. https://doi.org/https://doi.org/10.1016/j.compfluid.2009.09.013, URL https://www.sciencedirect.com/sci...

  10. [18]

    D.,Optimal shape design using an unsteady continuous adjoint approach, Stanford University, 2014

    Economon, T. D.,Optimal shape design using an unsteady continuous adjoint approach, Stanford University, 2014

  11. [19]

    Diffusion models: A comprehensive survey of methods and applications,

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., and Yang, M.-H., “Diffusion models: A comprehensive survey of methods and applications,”ACM computing surveys, Vol. 56, No. 4, 2023, pp. 1–39

  12. [20]

    Denoising diffusion implicit models,

    Song, J., Meng, C., and Ermon, S., “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502, 2020

  13. [21]

    Efficient aerodynamic design using the discrete adjoint method in SU2,

    Albring, T. A., Sagebaum, M., and Gauger, N. R., “Efficient aerodynamic design using the discrete adjoint method in SU2,” 17th AIAA/ISSMO multidisciplinary analysis and optimization conference, 2016, p. 3518

  14. [22]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    Raissi, M., Perdikaris, P., and Karniadakis, G. E., “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Journal of Computational physics, Vol. 378, 2019, pp. 686–707

  15. [23]

    A software package for sequential quadratic programming,

    Kraft, D., “A software package for sequential quadratic programming,”Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt, 1988

  16. [24]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,

    Wächter, A., and Biegler, L. T., “On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,”Mathematical programming, Vol. 106, 2006, pp. 25–57

  17. [25]

    Das Adjungiertenverfahren in der aerodynamischen Formoptimierung,

    Gauger, N. R., “Das Adjungiertenverfahren in der aerodynamischen Formoptimierung,” Ph.D. thesis, Braunschweig, Techn. Univ., Diss., 2003, 2003

  18. [26]

    Diffusion model for data-driven black-box optimization,

    Li, Z., Yuan, H., Huang, K., Ni, C., Ye, Y., Chen, M., and Wang, M., “Diffusion model for data-driven black-box optimization,” arXiv preprint arXiv:2403.13219, 2024. 24

  19. [2000]

    https://doi.org/10.1023/A:1011430410075

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.