Pith. sign in

REVIEW 4 major objections 4 minor 56 references

Case Studies of Generative Machine Learning Models for Dynamical Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Physics-informed VAEs can synthesize valid flight trajectories from a few hundred examples.

desk verdict Split-VAE's advantage over S-VAE is not yet established—the comparison is confounded by extra clean data and architecture differences—but the Z-VAE results on the Zermelo problem are worth a second look. read the letter →

arxiv 2508.04459 v1 pith:DIYBFLRY submitted 2025-08-06 eess.SY cs.NEcs.SY

Nachiket U. Bapat , Randy C. Paffenroth , Raghvendra V. Cowlagi This is my paper
classification eess.SYcs.NEcs.SY
keywords generativemodelsvariationalautoencoderphysics-informedlearningHamiltonianinvarianceZermelonavigationmodelmismatchsmalltrainingdataaerospaceguidance
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

This paper asks whether generative AI models can turn the abundant but imperfect simulations used in aerospace design into synthetic data that matches real system behavior, despite having only hundreds of real trajectories to learn from. For two aircraft-guidance problems, minimum-time navigation in a wind field and minimum-exposure navigation in a threat field, it builds loss functions from an optimal-control invariant: the Hamiltonian stays zero along every optimal trajectory. The authors report that variational-autoencoder models trained this way generate trajectories that both satisfy the governing optimality equations and match the statistics of the training set, while GANs collapse and standard VAEs drift in small-data regimes. The central new mechanism is the Split-VAE, which partitions its latent space so that a small noisy dataset and an abundant noiseless model dataset share one subspace while noise-specific features occupy a narrow dedicated subspace.

What carries the argument

The Hamiltonian-invariance condition of optimal-control theory: for the two guidance problems, first-order necessary conditions imply the Hamiltonian function $H$ is constant (here, zero) along every optimal trajectory. The paper turns this algebraic identity into a differentiable training penalty added to the generator/decoder loss, so a neural network can be steered toward the manifold of optimal trajectories with far fewer examples than distribution matching alone would require. The Split-VAE is the second load-bearing mechanism: a latent space split into a shared component inferred from both noisy observed data and abundant noiseless model data, and a small noise-specific component infer

What would settle it

Train the standard S-VAE architecture on the identical combined dataset used for Split-VAE (200 observed trajectories plus 200 noiseless model trajectories), keeping latent size, layers, and hyperparameters fixed except for the removed partition; if its total-variance and Hamiltonian-residual metrics match Split-VAE's, the central architectural claim fails. A second check: in the threat problem, use observed trajectories with known misleading $\lambda$ values and see whether the Hamiltonian residual $\delta_1$ computed with the correct $\lambda$ distinguishes good from bad generations.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a physics-informed VAE can generate valid trajectories from a few hundred examples by penalizing Hamiltonian violation during training, and that when real data do not exactly obey the model equations, a split latent-space architecture trained jointly on noisy observed trajectories and noiseless model trajectories outperforms a standard VAE on both equation satisfaction and statistical fidelity. For the Zermelo problem the Hamiltonian-based VAE (Z-VAE) beats the standard VAE at both large and small training volumes, and all GAN variants fail due to mode collapse. For the threat-exposure problem, where observed trajectories are optimal under differe

Load-bearing premise

The Split-VAE's improvement is attributed to partitioning the latent space, but its comparison baseline S-VAE was trained only on the noisy observed data; the paper never trains a standard VAE on the same combined noisy-plus-noiseless dataset, so extra clean data alone might explain the gain.

Editorial extensions

If this is right

  • Physics-informed VAE training can reduce model mismatch in synthetic trajectory generation without requiring tens of thousands of samples; a few hundred trajectories suffice in the studied problems.
  • Generator outputs of the Z-VAE satisfy the Zermelo optimality equations (Hamiltonian zero, costate relations) far better than a standard VAE, while remaining statistically similar to the training set.
  • GAN-based physics-informed generators are not a reliable choice for these small-data trajectory problems because mode collapse defeats diversity, whereas VAE-based models avoid it.
  • When observed data do not exactly satisfy the governing equations, adding a plain Hamiltonian residual term to the VAE loss is not enough; the Split-VAE's latent-space architecture is what improves generalization.
  • The Split-VAE extends to high-dimensional LTI systems with unknown non-Gaussian process noise, where standard VAE-generated samples collapse in variance and fail to match the observed noise level.

Reading between the lines

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

  • If a standard VAE trained on the same combined noisy-plus-clean dataset performs as well as Split-VAE, then the paper's architectural explanation would be unsupported; this ablation is the natural next experiment.
  • The Hamiltonian-zero loss is a generic invariant of optimal trajectories, so the same recipe could apply directly to other optimal-control-generated data, including energy-like invariants in conservative mechanical systems, whenever the output includes states and costates.
  • The NDRR result suggests the Split-VAE is effectively learning the noise distribution itself; one testable extension is to use the gap between generated and observed NDRR as a quantitative measure of residual model mismatch in digital-twin pipelines.
  • Because threat-field trajectories were evaluated assuming a known cost parameter $\lambda$ per trajectory, a harder and more realistic test would be to infer $\lambda$ from the data jointly and then check whether Hamiltonian residuals stay low.
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 / 4 minor

Summary. The paper studies physics-informed generative models for dynamical systems, with two aerospace-motivated case studies (Zermelo minimum-time navigation and minimum-threat-exposure navigation) and a high-dimensional LTI case study. It proposes two families of models: physics-informed GANs (Z-GAN1, Z-GAN2) and physics-informed VAEs, including a new Split-VAE architecture designed for training on data that do not exactly satisfy the governing equations. The main claim is that the proposed VAE variants synthesize data that are both statistically similar to the training data and close to satisfying the Hamiltonian-based necessary conditions, despite training sets of only a few hundred examples. The paper provides implementation details, performance tables, and code/dataset links.

Significance. If the claims hold, the paper makes a useful empirical contribution to physics-informed generative modeling in a low-data regime: it demonstrates that Hamiltonian invariance can be used as a training signal for VAEs, and it proposes an architecture intended to exploit abundant noiseless model data when observed data are scarce and imperfect. The Zermelo comparison is a genuine head-to-head between a standard VAE and a Hamiltonian-regularized VAE at the same N_D, and the inclusion of performance indices delta_2 and delta_3, which are not in the Z-VAE loss, is a thoughtful attempt to avoid purely circular evaluation. The public code and dataset links are a strength. However, the central architectural claim for the Split-VAE is not yet established because the comparison with S-VAE is confounded by data augmentation, network capacity, and normalization differences. The LTI results are also presented without repeated-seed variability, so the significance of the reported improvements is unclear.

major comments (4)
  1. [Sec. IV.C, Tables 6-8 and Table 13] The Split-VAE versus S-VAE comparison is not controlled. S-VAE is trained only on 200 noisy observed trajectories, while Split-VAE is trained on those 200 plus 200 noiseless model trajectories. No ablation trains a standard VAE on the identical combined dataset. Furthermore, Table 13 shows S-VAE uses a 64-d latent and deeper layers, while Split-VAE uses a 40-d latent (20+20) and fewer layers; Sec. IV.D adds layer normalization to Split-VAE but not S-VAE. Any of these differences, not the latent-space partition, could explain the gains. This is load-bearing because the paper's novel claim is the architectural benefit of the split latent space.
  2. [Eq. (21), Sec. III.C] The Split-VAE loss is not fully specified. As written, the reconstruction term appears to apply only to x in X, with the noiseless model trajectories contributing only through the KL term and the indicator function. If so, the model data never directly train the decoder to reconstruct trajectories, which weakens the stated interpretation that zeta2 captures shared features. The paper should define the loss for both data types precisely and test whether the noiseless data are acting as a regularizer rather than as a source of shared structure.
  3. [Sec. IV.B, Table 5] The Z-VAE versus S-VAE results are partially circular for delta_1, since the Hamiltonian residual is part of the Z-VAE training loss (Eq. 20). The paper correctly reports delta_2 and delta_3 as held-out, but at N_D=500 the Z-VAE delta_3 is worse than S-VAE (132.8 vs. 123.9), so the statement that the Z-VAE shows 'superior performance across most statistical measures' is too strong without a significance test or a per-metric qualification. The authors should distinguish metrics in the loss from truly held-out quantities and report variation over training seeds.
  4. [Sec. IV.C, Tables 7-8] The reported improvements are not uniformly favorable to Split-VAE. For lambda=10, S-VAE has a lower mean delta_1 (5.891 vs. 6.466) and lower standard deviation; for the first principal component variance in Table 8, Split-VAE is farther from the OTD (43.74 vs. 48.00, with OTD 129.4). The text says 'the majority of these metrics,' which is accurate but does not support the unqualified conclusion that Split-VAE consistently outperforms S-VAE. The paper should report per-metric breakdowns and, ideally, error bars over multiple training runs.
minor comments (4)
  1. [Throughout] Typos: 'Kullback-Liebler' should be 'Kullback-Leibler'; 'dicsussed' should be 'discussed'; 'the reader altogether unfamiliar' is awkward. Please copyedit.
  2. [Eq. (21)] The notation (x|x in X - G(E(x)))^2 is ambiguous; use an explicit indicator or separate expectations for X and Xs.
  3. [Table 7] Please clarify that mean delta_1 for lambda=10 is better for S-VAE; the current text says Split-VAE 'consistently achieves more desirable values for the majority of these metrics,' which obscures this exception.
  4. [Table 13] The table shows S-VAE with latent 64 and Split-VAE with latent 40; the text in Sec. IV.C says 'latent space size of 32' for both. This inconsistency should be resolved.

Circularity Check

1 steps flagged · score 2.0 of 10

One performance metric for Z-VAE duplicates a training-loss term; otherwise the derivation is self-contained.

  1. fitted input called prediction [Sec. III.B Eq. (20) and Sec. IV.A Eq. (22)]
    "For the Z-VAE we consider the loss function L_ZVAE(θ,φ) := E_{x∈X} [(x−G_θ(E_φ(x)))^2 + α1 L_sim(μ,Σ) + α2 ||H[G_θ(E_φ(x))]||^2] (20) ... δ1 := ||H[G_θ(z)]||^2, δ2 := ||tan u[G_θ(z)] − p2[G_θ(z)]/p1[G_θ(z)]||^2, δ3 := ||p1[G_θ(z)] + cos u[G_θ(z)]/ν[G_θ(z)]||^2 + ||p2[G_θ(z)] + sin u[G_θ(z)]/ν[G_θ(z)]||^2. (22)"

    The Z-VAE training loss (20) includes the term α2||H[G_θ(E_φ(x))]||^2, and the performance index δ1 in (22) is the same Hamiltonian residual ||H[G_θ(z)]||^2 evaluated on generated samples. Thus a low δ1 for Z-VAE is partly a check that the generator minimized its own penalty term; the metric is not an independent test of physics satisfaction. The comparison of Z-VAE vs. S-VAE on δ1 is therefore partly forced by construction. This is only partial: δ2 and δ3 are not present in the Z-VAE loss, and statistical-moment comparisons provide independent evidence of the model's quality.

full rationale

The paper's central derivation is not circular: the physics-informed losses are explicit design choices, and the main comparisons against S-VAE and S-GAN are backed by multiple independent metrics (statistical moments, δ2, δ3, total variance, NDRR). The only notable circular feature is the Z-VAE's δ1 metric, which is the same Hamiltonian residual that appears as a penalty in its own loss function; low δ1 is therefore partly a reflection of the training objective rather than a discovered property. However, the Z-VAE also improves on δ2 and δ3, which are not in its loss, and the Split-VAE (for the minimum-threat and LTI problems) does not use a Hamiltonian penalty at all, so its physics satisfaction is externally measured. The Split-VAE comparison is confounded by differing training data, latent dimensions, and layer normalization, but that is an experimental-design limitation rather than a circularity. No load-bearing self-citation or imported uniqueness theorem is present.

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

The central claim depends on the optimal-control Hamiltonian conditions (standard theory), on the assumption that training data come from the same 'perfect' model, on NN approximation capability, and on the unverified architectural hypothesis that latent-space splitting improves physics-consistent generation. The loss weights and data-mix ratios are tuned free parameters.

free parameters (3)
  • alpha1 (physics/regularization loss weight) = not reported
    Scales Hamiltonian penalty in Z-GAN and Z-VAE losses; chosen by hand, no values given.
  • alpha2 (additional penalty weight) = not reported
    Scales tan(u)=p2/p1 term in Z-GAN2 and KL terms in Split-VAE; tuned by experiments.
  • Split-VAE observed-to-model trajectory ratio = 200:200 (N_D=200 observed, 200 model)
    Authors state finding the optimal combination was key; ratio is a tuned hyperparameter.
assumptions (5)
  • domain assumption Hamiltonian invariance H=0 along optimal trajectories for Zermelo and minimum-threat problems
    Used to define training losses (Eqs. 4, 10); relies on data being optimal trajectories as per first-order necessary conditions [46].
  • domain assumption Model trajectories identical to observed trajectories (Zermelo case)
    Explicitly stated in Sec. II.A: 'the model is perfect'. All training data are numerical solutions of the same model equations.
  • domain assumption Observed and model trajectories are optimal, possibly with different cost parameter lambda (min-threat case)
    Sec. II.B: 'we do assume that the model and observed trajectories are optimal, but the optimality objective functions may slightly differ.'
  • standard math Neural networks are universal function approximators
    Assumed in Sec. I (refs. 17,18).
  • ad hoc to paper Partitioning the latent space into shared (zeta2) and noisy-specific (zeta1) components improves generalization
    Core architectural hypothesis of Split-VAE (Sec. III.C, Eq. 21); motivated heuristically, no independent support outside this paper.
invented entities (1)
  • Split latent space partition (zeta1, zeta2)
    purpose: Separate noisy deviations (zeta1) from shared physics features (zeta2) in the Split-VAE
    A model architecture element introduced in Sec. III.C; its benefit is claimed based on the paper's own comparisons, which are confounded by extra training data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Case Studies of Generative Machine Learning Models for Dynamical Systems." pith.science (2026). https://pith.science/paper/DIYBFLRY

@misc{pith2026250804459,
  author       = {Pith},
  title        = {Pith review of: Case Studies of Generative Machine Learning Models for Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIYBFLRY}},
  note         = {Machine review of arXiv:2508.04459}
}
read the original abstract

Systems like aircraft and spacecraft are expensive to operate in the real world. The design, validation, and testing for such systems therefore relies on a combination of mathematical modeling, abundant numerical simulations, and a relatively small set of real-world experiments. Due to modeling errors, simplifications, and uncertainties, the data synthesized by simulation models often does not match data from the system's real-world operation. We consider the broad research question of whether this model mismatch can be significantly reduced by generative artificial intelligence models (GAIMs). Unlike text- or image-processing, where generative models have attained recent successes, GAIM development for aerospace engineering applications must not only train with scarce operational data, but their outputs must also satisfy governing equations based on natural laws, e.g., conservation laws. The scope of this paper primarily focuses on two case studies of optimally controlled systems that are commonly understood and employed in aircraft guidance, namely: minimum-time navigation in a wind field and minimum-exposure navigation in a threat field. We report GAIMs that are trained with a relatively small set, of the order of a few hundred, of examples and with underlying governing equations. By focusing on optimally controlled systems, we formulate training loss functions based on invariance of the Hamiltonian function along system trajectories. We investigate three GAIM architectures, namely: the generative adversarial network (GAN) and two variants of the variational autoencoder (VAE). We provide architectural details and thorough performance analyses of these models. The main finding is that our new models, especially the VAE-based models, are able to synthesize data that satisfy the governing equations and are statistically similar to the training data despite small volumes of training data.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 28 canonical work pages

  1. [1]

    Optimal and Autonomous Control Using Reinforcement Learning: A Survey,

    Kiumarsi, B., Vamvoudakis, K. G., Modares, H., and Lewis, F. L., “Optimal and Autonomous Control Using Reinforcement Learning: A Survey,”IEEE Transactions on Neural Networks and Learning Systems, Vol. 29, No. 6, 2018, pp. 2042–2062. doi:10.1109/TNNLS.2017.2773458

    work page arXiv 2018

  2. [2]

    Asurveyofdeeplearningapplicationstoautonomousvehiclecontrol,

    Kuutti,S.,Bowden,R.,Jin,Y.,Barber,P.,andFallah,S.,“Asurveyofdeeplearningapplicationstoautonomousvehiclecontrol,” IEEE Transactions on Intelligent Transportation Systems, Vol. 22, No. 2, 2021, pp. 712–733. doi:10.1109/TITS.2019.2962338

    arXiv 2021

  3. [3]

    doi:10.2514/6.2022-2103, URLhttps://arc.aiaa

    Gupta, S., Durak, U., Ellis, O., and Torens, C.,From Operational Scenarios to Synthetic Data: Simulation-Based Data Generation for AI-Based Airborne Systems, AIAA 2022-2103, 2022. doi:10.2514/6.2022-2103, URLhttps://arc.aiaa. org/doi/pdf/10.2514/6.2022-2103

    work page doi:10.2514/6.2022-2103 2022

  4. [4]

    doi:10.2514/6.2023-1814, URLhttps://arc.aiaa.org/doi/abs/10.2514/6.2023-1814

    Sisson, N., and Moncayo, H.,Machine Learning Based Architecture for Generation of Synthetic Flight Test Data, AIAA 2023-1814, 2024. doi:10.2514/6.2023-1814, URLhttps://arc.aiaa.org/doi/abs/10.2514/6.2023-1814

    work page doi:10.2514/6.2023-1814 2023

  5. [5]

    15 Sample outputs of the Split-VAE for 200 training samples for𝜆= 5

    Sprockhoff, J., Gupta, S., Durak, U., and Krueger, T.,Scenario-Based Synthetic Data Generation for an AI-based System Using Fig. 15 Sample outputs of the Split-VAE for 200 training samples for𝜆= 5. 33 Fig. 16 Sample outputs of the Split-VAE for 200 training samples for𝜆= 10. a Flight Simulator, AIAA 2024-1462, 2024. doi:10.2514/6.2024-1462, URLhttps://arc...

    work page doi:10.2514/6.2024-1462 2024

  6. [6]

    97 – 155

    Jategaonkar, R.,Flight Vehicle System Identification: A Time Domain Methodology, Progress in Aeronautics and Astronautics, AIAA, Reston, VA, USA, 2006, pp. 97 – 155. doi:10.2514/4.102790

    doi:10.2514/4.102790 2006

  7. [7]

    Ioannou, P., and Sun, J.,Robust Adaptive Control, Dover Publications, Inc., Mineola, NY, USA, 2012

    work page 2012

  8. [8]

    doi:10.1137/1.9780898719376

    Hovakimyan, N., and Cao, C.,L1 adaptive control theory: Guaranteed robustness with fast adaptation, SIAM, Philadelphia, PA, USA, 2010. doi:10.1137/1.9780898719376

    work page doi:10.1137/1.9780898719376 2010

Show all 56 references
  1. [9]

    Anintroductiontodeepreinforcementlearning,

    François-Lavet,V.,Henderson,P.,Islam,R.,Bellemare,M.G.,andPineau,J.,“Anintroductiontodeepreinforcementlearning,” Foundations and Trends®in Machine Learning, Vol. 11, No. 3-4, 2018, pp. 219–354. doi:10.1561/2200000071

    2018 doi

  2. [10]

    I.,Synthetic Data for Deep Learning, Springer Optimization and Its Applications, Springer, Cham, Switzerland,

    Nikolenko, S. I.,Synthetic Data for Deep Learning, Springer Optimization and Its Applications, Springer, Cham, Switzerland,

  3. [11]

    Hierarchical Text-Conditional Image Generation with CLIP Latents,

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M., “Hierarchical Text-Conditional Image Generation with CLIP Latents,” , 2022. doi:10.48550/arXiv.2204.06125, URLhttps://arxiv.org/abs/2204.06125

    2022 doi arXiv

  4. [12]

    AnempiricalevaluationofGitHubcopilot’scodesuggestions,

    Nguyen,N.,andNadi,S.,“AnempiricalevaluationofGitHubcopilot’scodesuggestions,” Proceedingsofthe19thInternational Conference on Mining Software Repositories, 2022, pp. 1–5. doi:10.1145/3524842.3528470

    2022

  5. [13]

    Face generation and editing with stylegan: A survey,

    Melnik,A.,Miasayedzenkau,M.,Makaravets,D.,Pirshtuk,D.,Akbulut,E.,Holzmann,D.,Renusch,T.,Reichert,G.,andRitter, H., “Face generation and editing with stylegan: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence,

  6. [14]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., and Song, S., “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, 2023, p. 02783649241273668. doi: 10.48550/arXiv.2303.04137

    2023 doi arXiv

  7. [15]

    Diffusion models for generating ballistic spacecraft trajectories,

    Presser, T., Dasgupta, A., Erwin, D., and Oberai, A., “Diffusion models for generating ballistic spacecraft trajectories,”arXiv preprint arXiv:2405.11738, 2024. doi:10.48550/arXiv.2405.11738. 34

    2024 doi arXiv

  8. [16]

    An Example of Synthetic Data Generation for Control Systems using GenerativeAdversarialNetworks: ZermeloMinimum-TimeNavigation,

    Bapat, N. U., Paffenroth, R., and Cowlagi, R. V., “An Example of Synthetic Data Generation for Control Systems using GenerativeAdversarialNetworks: ZermeloMinimum-TimeNavigation,” Proceedingsofthe2024AmericanControlConference (ACC), Toronto, Canada, 2024. doi:10.23919/ACC60939...

    2024

  9. [17]

    Multilayer feedforward networks are universal approximators,

    Hornik, K., Stinchcombe, M., and White, H., “Multilayer feedforward networks are universal approximators,”Neural Networks, Vol.2,No.5,1989,pp.359–366. doi:10.1016/0893-6080(89)90020-8,URL https://www.sciencedirect.com/science/ article/pii/0893608089900208

    1989

  10. [18]

    Multilayerfeedforwardnetworkswithanonpolynomialactivationfunction can approximate any function,

    Leshno,M.,Lin,V.Y.,Pinkus,A.,andSchocken,S.,“Multilayerfeedforwardnetworkswithanonpolynomialactivationfunction can approximate any function,”Neural Networks, Vol. 6, No. 6, 1993, pp. 861–867. doi:10.1016/S0893-6080(05)80131-5, URL https://www.sciencedirect.com/science/article/...

    1993 doi

  11. [19]

    Generative adversarial networks,

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y., “Generative adversarial networks,”Communications of the ACM, Vol. 63, No. 11, 2020, pp. 139–144. doi:10.48550/arXiv.1406.2661

    2020 doi arXiv

  12. [20]

    An introduction to variational autoencoders,

    Kingma, D. P., and Welling, M., “An introduction to variational autoencoders,”Foundations and Trends®in Machine Learning, Vol. 12, No. 4, 2019, pp. 307–392. doi:10.48550/arXiv.1906.02691

    2019 doi arXiv

  13. [21]

    A Brief Introduction to Generative Models,

    Lamb, A., “A Brief Introduction to Generative Models,” , 2021. doi:10.48550/arXiv.2103.00265, URLhttps://arxiv.org/ abs/2103.00265

    2021 doi arXiv

  14. [22]

    Generative adversarial networks: An overview,

    Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., and Bharath, A. A., “Generative adversarial networks: An overview,”IEEE Signal Processing Magazine, Vol. 35, No. 1, 2018, pp. 53–65. doi:10.48550/arXiv.1710.07035

    2018 doi arXiv

  15. [23]

    NVAE: A deep hierarchical variational autoencoder,

    Vahdat, A., and Kautz, J., “NVAE: A deep hierarchical variational autoencoder,”Advances in Neural Information Processing Systems, Vol. 33, 2020, pp. 19667–19679. doi:10.48550/arXiv.2007.03898

    2020 doi arXiv

  16. [24]

    TrajVAE:Avariationalautoencodermodelfortrajectorygeneration,

    Chen,X.,Xu,J.,Zhou,R.,Chen,W.,Fang,J.,andLiu,C.,“TrajVAE:Avariationalautoencodermodelfortrajectorygeneration,” Neurocomputing, Vol. 428, 2021, pp. 332–339. doi:10.1016/j.neucom.2020.03.120

    2021 doi

  17. [25]

    Physically Interpretable Feature Learning of Supercritical Airfoils Based on Variational Autoencoders,

    Li, R., Zhang, Y., and Chen, H., “Physically Interpretable Feature Learning of Supercritical Airfoils Based on Variational Autoencoders,”AIAA Journal, Vol. 60, No. 11, 2022, pp. 6168–6182. doi:10.2514/1.J061673, URLhttps://doi.org/10. 2514/1.J061673

    2022 doi

  18. [26]

    Anomaly detection for time series using vae-lstm hybrid model,

    Lin, S., Clark, R., Birke, R., Schönborn, S., Trigoni, N., and Roberts, S., “Anomaly detection for time series using vae-lstm hybrid model,”ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Ieee, 2020, pp. 4322–4326. doi:10.110...

    2020

  19. [27]

    Collaborative variational deep learning for healthcare recommendation,

    Deng, X., and Huangfu, F., “Collaborative variational deep learning for healthcare recommendation,”IEEE Access, Vol. 7, 2019, pp. 55679–55688. doi:10.1109/ACCESS.2019.2913468

    2019

  20. [28]

    StyleHEAT: One-Shot High-Resolution Editable Talking Face Generation via Pre-trained StyleGAN,

    Yin, F., Zhang, Y., Cun, X., Cao, M., Fan, Y., Wang, X., Bai, Q., Wu, B., Wang, J., and Yang, Y., “StyleHEAT: One-Shot High-Resolution Editable Talking Face Generation via Pre-trained StyleGAN,”Computer Vision – ECCV 2022, edited by 35 S. Avidan, G. Brostow, M. Cissé, G. M. Fa...

    2022 doi arXiv

  21. [29]

    Time-series Generative Adversarial Networks,

    Yoon, J., Jarrett, D., and van der Schaar, M., “Time-series Generative Adversarial Networks,”Advances in Neural Infor- mation Processing Systems, Vol. 32, edited by H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, Curran Associates, Inc., 2019...

    2019

  22. [30]

    TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation,

    Desai, A., Freeman, C., Wang, Z., and Beaver, I., “TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation,” , 2021. doi:10.48550/arXiv.2111.08095, URLhttps://arxiv.org/abs/2111.08095

    2021 doi arXiv

  23. [31]

    RL-CycleGAN: Reinforcement learning aware simulation-to-real,

    Rao, K., Harris, C., Irpan, A., Levine, S., Ibarz, J., and Khansari, M., “RL-CycleGAN: Reinforcement learning aware simulation-to-real,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11157–11166. doi:10.48550/arXiv.2006.09001

    2020 doi arXiv

  24. [32]

    Tutorial on Amortized Optimization,

    Amos, B., “Tutorial on Amortized Optimization,”Foundations and Trends®in Machine Learning, Vol. 16, No. 5, 2023, pp. 592–732. doi:10.1561/2200000102, URLhttp://dx.doi.org/10.1561/2200000102

    2023 doi

  25. [33]

    Semi-amortized variational autoencoders,

    Kim, Y., Wiseman, S., Miller, A., Sontag, D., and Rush, A., “Semi-amortized variational autoencoders,”International Conference on Machine Learning, PMLR, 2018, pp. 2678–2687. doi:10.48550/arXiv.1802.02550

    2018 doi arXiv

  26. [34]

    Iterative amortized inference,

    Marino, J., Yue, Y., and Mandt, S., “Iterative amortized inference,”International Conference on Machine Learning, PMLR, 2018, pp. 3403–3412. doi:10.48550/arXiv.1807.09356

    2018 doi arXiv

  27. [35]

    DC3: A learning method for optimization with hard constraints,

    Donti, P. L., Rolnick, D., and Kolter, J. Z., “DC3: A learning method for optimization with hard constraints,”arXiv preprint arXiv:2104.12225, 2021. doi:10.48550/arXiv.2104.12225

    2021 doi arXiv

  28. [36]

    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. doi:10.1...

    2019 doi

  29. [37]

    nn-PINNs: Non-Newtonian physics-informed neural networks for complex fluid modeling,

    Mahmoudabadbozchelou, M., Karniadakis, G. E., and Jamali, S., “nn-PINNs: Non-Newtonian physics-informed neural networks for complex fluid modeling,”Soft Matter, Vol. 18, No. 1, 2022, pp. 172–185. doi:10.1039/D1SM01298C

    2022 doi

  30. [38]

    Physics-informedmachinelearninganduncertainty quantification for mechanics of heterogeneous materials,

    Bharadwaja,B.,Nabian,M.A.,Sharma,B.,Choudhry,S.,andAlankar,A.,“Physics-informedmachinelearninganduncertainty quantification for mechanics of heterogeneous materials,”Integrating Materials and Manufacturing Innovation, Vol. 11, No. 4, 2022, pp. 607–627. doi:10.48550/arXiv.2202.10423

    2022 doi arXiv

  31. [39]

    Multiple Case Physics-Informed Neural Network for Biomedical Tube Flows,

    Wong, H. S., Chan, W. X., Li, B. H., and Yap, C. H., “Multiple Case Physics-Informed Neural Network for Biomedical Tube Flows,” , 2023. doi:10.48550/arXiv.2309.15294, URLhttps://arxiv.org/abs/2309.15294

    2023 doi arXiv

  32. [40]

    A physics-informed Transformer model for vehicle trajectory prediction on highways,

    Geng, M., Li, J., Xia, Y., and Chen, X. M., “A physics-informed Transformer model for vehicle trajectory prediction on highways,”Transportation research part C: emerging technologies, Vol. 154, 2023, p. 104272. doi:10.1016/j.trc.2023.104272. 36

    2023

  33. [41]

    Physics-informed neural networks to model and control robots: A theoretical and experimental investigation,

    Liu, J., Borja, P., and Della Santina, C., “Physics-informed neural networks to model and control robots: A theoretical and experimental investigation,”Advanced Intelligent Systems, Vol. 6, No. 5, 2024, p. 2300385. doi:10.48550/arXiv.2305.05375

    2024 doi arXiv

  34. [42]

    Physics-informed generative adversarial networks for stochastic differential equations,

    Yang, L., Zhang, D., and Karniadakis, G. E., “Physics-informed generative adversarial networks for stochastic differential equations,”SIAM Journal on Scientific Computing, Vol. 42, No. 1, 2020, pp. A292–A317. doi:10.48550/arXiv.1811.02033

    2020 doi arXiv

  35. [43]

    Generative ensemble regression: Learning particle dynamics from observations of ensembles with physics-informed deep generative models,

    Yang, L., Daskalakis, C., and Karniadakis, G. E., “Generative ensemble regression: Learning particle dynamics from observations of ensembles with physics-informed deep generative models,”SIAM Journal on Scientific Computing, Vol. 44, No. 1, 2022, pp. B80–B99. doi:10.48550/arXi...

    2022 doi arXiv

  36. [44]

    Navier–stokes generative adversarial network: A physics-informed deep learning model for fluid flow generation,

    Wu, P., Pan, K., Ji, L., Gong, S., Feng, W., Yuan, W., and Pain, C., “Navier–stokes generative adversarial network: A physics-informed deep learning model for fluid flow generation,”Neural Computing and Applications, Vol. 34, No. 14, 2022, pp. 11539–11552. doi:10.1007/s00521-0...

    2022 doi

  37. [45]

    Hierarchical dynamic wake modeling of wind turbine based on physics-informed generative deep learning,

    Wang, Q., Ti, Z., Yang, S., Yang, K., Wang, J., and Deng, X., “Hierarchical dynamic wake modeling of wind turbine based on physics-informed generative deep learning,”Applied Energy, Vol. 378, 2025, p. 124812. doi:10.1016/j.apenergy.2024.124812, URL https://www.sciencedirect.co...

    2025

  38. [46]

    E., and Ho, Y.-C.,Applied optimal control: optimization, estimation and control, Taylor & Francis, New York, NY, USA, 1975

    Bryson, A. E., and Ho, Y.-C.,Applied optimal control: optimization, estimation and control, Taylor & Francis, New York, NY, USA, 1975

    1975

  39. [47]

    Tackling mode collapse in multi-generator GANs with orthogonal vectors,

    Li, W., Fan, L., Wang, Z., Ma, C., and Cui, X., “Tackling mode collapse in multi-generator GANs with orthogonal vectors,” Pattern Recognition, Vol. 110, 2021, p. 107646. doi:10.1016/j.patcog.2020.107646, URLhttps://www.sciencedirect. com/science/article/pii/S0031320320304490

    2021

  40. [48]

    Multi-level variational autoencoder: Learning disentangled representations from grouped observations,

    Bouchacourt, D., Tomioka, R., and Nowozin, S., “Multi-level variational autoencoder: Learning disentangled representations from grouped observations,”Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 32, 2018. doi:10.48550/arXiv. 1705.08841

    2018 doi arXiv

  41. [49]

    Pytorch: An imperative style, high-performance deep learning library,

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., “Pytorch: An imperative style, high-performance deep learning library,”Advances in neural information processing systems, Vol. 32, 2019. doi:10.4855...

    2019 doi arXiv

  42. [50]

    An Empirical Study on Generalizations of the ReLU Activation Function,

    Banerjee, C., Mukherjee, T., and Pasiliao, E., “An Empirical Study on Generalizations of the ReLU Activation Function,” Proceedings of the 2019 ACM Southeast Conference, Association for Computing Machinery, New York, NY, USA, 2019, p. 164–167. doi:10.1145/3299815.3314450, URL1...

    2019

  43. [51]

    Animagedifferencemetricbasedonsimulationofimagedetailvisibilityandtotalvariation,

    Pedersen,M.,“Animagedifferencemetricbasedonsimulationofimagedetailvisibilityandtotalvariation,” ColorandImaging Conference, Vol. 22, Society for Imaging Science and Technology, 2014, pp. 37–42. doi:10.2352/CIC.2014.22.1.art00005, URL https://doi.org/10.2352/CIC.2014.22.1.art00005. 37

    2014 doi

  44. [52]

    Understanding GANs: Fundamentals, variants, training challenges, applications, and open problems,

    Ahmad, Z., Jaffri, Z. u. A., Chen, M., and Bao, S., “Understanding GANs: Fundamentals, variants, training challenges, applications, and open problems,”Multimedia Tools and Applications, 2024, pp. 1–77. doi:10.1007/s11042-024-19361-y

    2024 doi

  45. [53]

    Training generative adversarial networks with limited data,

    Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., and Aila, T., “Training generative adversarial networks with limited data,”Advances in neural information processing systems, Vol. 33, 2020, pp. 12104–12114. doi:10.48550/arXiv.2006.06676

    2020 doi arXiv

  46. [54]

    BEGANv3: avoidingmodecollapseinGANsusingvariationalinference,

    Park,S.-W.,Huh,J.-H.,andKim,J.-C.,“BEGANv3: avoidingmodecollapseinGANsusingvariationalinference,” Electronics, Vol. 9, No. 4, 2020, p. 688. doi:10.3390/electronics9040688. 38

    2020 doi

  47. [2021]

    doi:10.1007/978-3-030-75178-4

    doi

  48. [2024]

    doi:10.1109/TPAMI.2024.3350004

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.