Pith. sign in

REVIEW 4 major objections 7 minor 64 references

Variational Rank Reduction Autoencoders for Generative Thermal Design

T0 review · 4 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a hybrid framework that couples a variational rank-reduction autoencoder, which builds an 8-dimensional continuous latent space via truncated SVD, with a DeepONet that predicts temperature gradients from those codes, ach

desk verdict Plausible incremental VRRAE+DeepONet surrogate for thermal design, but the underspecified per-minibatch SVD works against the central accuracy claim. read the letter →

arxiv 2509.08515 v1 pith:N3A5LIMO submitted 2025-09-10 cs.LG cs.AI

Alicia Tierz , Jad Mounayer , Beatriz Moya , Francisco Chinesta This is my paper
classification cs.LGcs.AI
keywords generativethermaldesignvariationalrank-reductionautoencodertruncatedsingularvaluedecompositionDeepONetlatentspacestructureoperatorlearningsurrogatemodelingtemperaturegradientprediction
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 aims to show that giving a variational autoencoder a structured latent space — one built by truncated singular value decomposition — removes the 'holes' and discontinuities that plague plain autoencoders, and that this structure makes a downstream operator network both more accurate and dramatically faster than traditional thermal solvers. The authors claim that their hybrid, VRRAE+DeepONet, predicts steady-state temperature gradients on plate geometries with the lowest normalized error and tightest variability among the tested combinations, while running at 0.0026 seconds per sample versus 0.273 seconds for Abaqus. A careful reader would care because the result suggests that the quality of the latent representation, not just the prediction head, is what limits physics surrogate models.

What carries the argument

The key mechanism is the truncated SVD applied inside the latent space of a variational autoencoder. For a batch of N input images, the encoder produces a latent matrix Y, which is factored as Y = USV^T and truncated to a fixed rank k* = 8; the retained singular coefficients become the mean of the variational posterior, while a neural network predicts the per-coefficient standard deviations. This enforces an ordered, orthonormal, low-rank coordinate system without adding loss terms, and it prevents posterior collapse because collapse can only occur toward structured, bounded values. The DeepONet then combines these 8 coefficients (branch net) with 2D coordinates (trunk net) via an inner prod

What would settle it

Take one geometry, encode it in many different minibatches and with several random seeds, and compare the 8D codes: if they are not identical up to sign flips or the singular vectors reorder, the latent manifold is not reproducible and the DeepONet mapping is not well-defined.

Watch

Extended reading notes

Core claim

The central claim is that inserting a truncated SVD into the latent space of a variational autoencoder, with the mean of the latent distribution fixed to the SVD coefficients, yields an 8-dimensional, continuous, interpretable geometry manifold that is largely free of non-physical 'holes'. A DeepONet fed with these codes as its branch input and spatial coordinates as its trunk input predicts temperature gradients with NMSE (5.54±2.02)×10^-7, a 21.2% improvement over the same DeepONet fed with a plain autoencoder latent space, and with tighter variability than alternative encoder–head combinations. The same structured manifold also makes linear interpolation between geometries produce physica

Load-bearing premise

The per-minibatch truncated SVD is assumed to produce a well-defined, consistent 8-coordinate latent system that stays the same for a given geometry across training batches and test samples.

Editorial extensions

If this is right

  • Replacing a plain autoencoder encoder with the VRRAE improves downstream operator learning: NMSE drops by 38.9% with a CNN head and 21.2% with a DeepONet head, showing that latent structure is a major driver of surrogate accuracy.
  • Because the DeepONet trunk depends only on coordinates, it can be evaluated once and reused across all geometries in a batch, cutting per-sample inference time to 0.0026 s versus 0.273 s for Abaqus.
  • Linear interpolation in the VRRAE latent space produces structurally valid four-hole geometries 86.8% of the time (versus 71.4% for AE), enabling meaningful design-space exploration.
  • The method generalizes the combination of generative models and operator networks to other geometry-to-field engineering problems where a compact, continuous parameterization is needed.
  • The KL-annealed variational training with a fixed SVD mean is a reproducible recipe for building interpretable latent manifolds without additional regularizers.

Reading between the lines

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

  • The paper does not check whether the per-minibatch SVD produces a consistent basis across training batches and test samples; a reader could test reproducibility by encoding the same geometry in different minibatches or with different random seeds and comparing the 8D codes.
  • The rank k* = 8 is fixed; an adaptive-rank variant that chooses truncation per batch or per sample could improve the trade-off between reconstruction fidelity and downstream operator accuracy.
  • The structural consistency metric — counting shapes and checking area tolerance — could be applied to other shape-constrained design problems, not just four cooling holes.
  • If the latent basis is unstable, the claimed continuity and interpretability of the VRRAE manifold, and hence the reliability of the DeepONet mapping, would be threatened even if the reported test errors remain 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 / 7 minor

Summary. The paper proposes a hybrid framework combining a Variational Rank-Reduction Autoencoder (VRRAE) with a DeepONet to predict steady-state temperature gradients on 2D plates with internal cooling features. The VRRAE computes per-batch truncated SVD latent codes (rank 8) and uses the resulting coefficients as the mean of a variational posterior; the DeepONet maps these 8D codes (branch) and spatial coordinates (trunk) to temperature gradients. The central claims are that the VRRAE latent space is more structured, continuous, and interpolable than a plain autoencoder, and that the VRRAE+DeepONet achieves the lowest MSE/NMSE with a reported >100x speedup over Abaqus. Results are based on 100,000 synthetic geometries, with thermal fields computed for a 5,000-geometry subset.

Significance. If the central claims hold, the paper demonstrates a useful integration of a structured generative latent space with operator learning for many-query thermal design. The 2x2 ablation (encoder x prediction head) is a sensible experimental design, and reporting mean +/- standard deviation on four error metrics is a positive feature. The paper also makes an explicit inference-efficiency claim that is relevant for engineering practice. However, the manuscript is not self-contained: the VRRAE implementation and deeper analysis are delegated to companion papers [22,23], no code is released, and the evaluation lacks a standard VAE baseline and statistical significance tests. The main methodological risk is the unaddressed basis-alignment problem in the per-batch SVD latent code, which is load-bearing for the central accuracy claim.

major comments (4)
  1. [3.2, 3.3, Table 2] The central modeling premise is the per-batch truncated SVD latent code. Section 3.2 defines Y = USV^T and truncates to rank k*=8, but SVD singular vectors are unique only up to sign, and can rotate when singular values are close or degenerate. The paper provides no sign convention, fixed reference basis, Procrustes alignment, or consistency check across training batches, and it does not explain how a single test geometry is mapped to an 8D code at inference (a single vector has no 8-mode SVD). Without a fixed projection basis, the branch-net inputs are not well-defined, so the NMSE numbers in Table 2 are not a reliable evaluation of the proposed pipeline. This must be fixed, e.g., by fixing U from a training-set SVD or by explicit alignment, with a consistency experiment.
  2. [3.3 vs 3.4] There is a direct contradiction in the Dirichlet boundary conditions. Section 3.3 states that the outer plate boundary is set to 100 degC and the cooling holes to 0 degC, and the same is repeated in the PDE description. Section 3.4, Dataset Design, states that the outer edges are held at 20 degC to mimic an external bath, while internal inclusions are fixed at 0 degC. These two specifications produce different temperature fields; all ground-truth labels and Figure 4 depend on which one was actually used. The simulation protocol must be stated unambiguously and consistently.
  3. [4.2, Table 2] The paper claims that VRRAE+DeepONet attains the lowest NMSE/MSE with tighter variability, but the differences versus VRRAE+CNN are small (NMSE 5.54e-7 vs 5.68e-7; MSE 3.12e-3 vs 3.17e-3) and within one standard deviation. The number of independent runs is not stated, and no paired significance test is provided. As reported, the best-head claim is not statistically supported. Please add run counts and appropriate tests, or moderate the claim.
  4. [2, 4.1, Table 1] The motivation in Sections 1 and 2 emphasizes VAE posterior collapse and blur, and the VRRAE is introduced as an improvement over VAEs, but the evaluation only compares AE vs VRRAE (with CNN or DeepONet heads). No standard VAE baseline appears in Table 1 or Table 2. Consequently, the claim that the SVD-induced structure is responsible for the observed gains cannot be separated from the effect of variational sampling. A VAE row (or VAE+DeepONet) should be added, or the claims should be limited to the AE comparison.
minor comments (7)
  1. [4.1] The structural consistency metric is defined by counting detected figures and requiring a 5% area tolerance, but the detection algorithm and the exact 'expected target range' from the training set are not specified. Please provide details so the metric is reproducible.
  2. [Eq. (1)] Equation (1) uses symbols bar_alpha_sigma and bar_alpha_mu without explicit definitions, and the KL term notation is hard to parse. Also, Section 5 mentions 'setting the latent mean function to the identity', which is not clearly explained in Section 3.2.
  3. [3.4] The encoder is described as three convolutional layers (channels 32,64,128) while the decoder has four transposed layers (256,128,32,8); please clarify the dimensionality flow for 128x128 inputs, especially the final output size.
  4. [3.3 vs 3.4] The ground-truth solver is described as 'MATLAB PDE Toolbox' in Section 3.3 and as a 'finite-difference solver implemented in MATLAB' in Section 3.4. Unify the description.
  5. [5] The Abaqus timing comparison (0.273 s vs 0.0026 s per sample) lacks setup details: the same geometry, mesh, hardware, and accuracy tolerance are not specified. Please report the exact comparison protocol.
  6. [References [22,23]] The paper relies on companion references for the VRRAE implementation and 'deeper analysis'. Since no code is released, please make the manuscript more self-contained, even if it means adding an appendix with the SVD/projection equations and training details.
  7. [Throughout] There are several typos: 'demostrating' (Section 1), 'laten matriz' (Section 3.2), inconsistent spacing in 'V AEs', and the Figure 4 colorbar caption gives units but no numeric scale. Please copyedit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy and speed claims are empirical benchmarks computed from independently generated FEM reference fields and wall-clock measurements, not reductions of the outputs to fitted inputs.

full rationale

The paper's central claims are empirical and self-contained. VRRAE+DeepONet accuracy is supported by test-set MSE/NMSE values in Table 2, computed against reference temperature-gradient fields generated by a MATLAB finite-difference/FEM solver, not by an equation that re-inserts the fitted latent values into the target. The speedup claim compares wall-clock timings (Abaqus 0.273 s/sample vs DeepONet 0.0026 s/sample) and is not derived from the model's parameters. The VRRAE's truncated SVD is an architectural mechanism, and the claim that it yields a structured latent space is tested directly through the interpolation metric in Section 4.1 and through downstream DeepONet error; it is not assumed by definition. The paper does cite the authors' companion works [22,23] for the VRRAE concept and 'deeper analysis', but these citations are background support: the VRRAE is trained, evaluated, and compared within this manuscript, so the benchmark does not reduce to the cited works. The per-minibatch SVD sign/basis-ambiguity identified by the skeptic is a real validity/reproducibility risk, but it is not circularity: inconsistent latent coordinates would make the reported errors unreliable, not make the claim true by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is borrowed, and no ansatz is smuggled in solely by citation. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central modeling choice rests on a hand-set SVD rank and KL weight, plus an unverified assumption that minibatch SVD yields a stable 8D latent manifold. No new physical entities are introduced.

free parameters (4)
  • KL weight beta = 0.2 (annealed)
    Selected by empirical tuning in Section 3.4; controls the balance between reconstruction and KL regularization in Eq. 1.
  • truncated rank k* = 8
    Set by hand in Section 3.2; defines the latent dimension and the DeepONet branch input size.
  • DeepONet width p = 128
    Picked based on validation performance and computational considerations in Sections 3.3 and 3.4.
  • Structural validity thresholds = exactly 4 figures; 5% area tolerance
    Hand-defined evaluation thresholds in Section 4.1; changing them changes the reported interpolation and random sampling scores.
assumptions (4)
  • standard math Universal Approximation Theorem for Operators
    Invoked in Section 3.3 to justify DeepONet capacity; does not guarantee finite-sample accuracy.
  • domain assumption Ground-truth FEM/FD solutions are accurate and represent the physical heat equation
    Sections 3.3 and 3.4; no solver convergence or mesh independence verification is reported.
  • domain assumption The 8D truncated-SVD latent code preserves geometry information relevant to temperature gradients
    Core premise for feeding VRRAE codes into DeepONet; empirically asserted, not proven.
  • ad hoc to paper SVD coefficients from different minibatches form a consistent latent coordinate system during training and inference
    Section 3.2; SVD singular vectors are sign/rotation non-unique and no alignment mechanism is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Variational Rank Reduction Autoencoders for Generative Thermal Design." pith.science (2026). https://pith.science/paper/N3A5LIMO

@misc{pith2026250908515,
  author       = {Pith},
  title        = {Pith review of: Variational Rank Reduction Autoencoders for Generative Thermal Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N3A5LIMO}},
  note         = {Machine review of arXiv:2509.08515}
}
read the original abstract

Generative thermal design for complex geometries is fundamental in many areas of engineering, yet it faces two main challenges: the high computational cost of high-fidelity simulations and the limitations of conventional generative models. Approaches such as autoencoders (AEs) and variational autoencoders (VAEs) often produce unstructured latent spaces with discontinuities, which restricts their capacity to explore designs and generate physically consistent solutions. To address these limitations, we propose a hybrid framework that combines Variational Rank-Reduction Autoencoders (VRRAEs) with Deep Operator Networks (DeepONets). The VRRAE introduces a truncated SVD within the latent space, leading to continuous, interpretable, and well-structured representations that mitigate posterior collapse and improve geometric reconstruction. The DeepONet then exploits this compact latent encoding in its branch network, together with spatial coordinates in the trunk network, to predict temperature gradients efficiently and accurately. This hybrid approach not only enhances the quality of generated geometries and the accuracy of gradient prediction, but also provides a substantial advantage in inference efficiency compared to traditional numerical solvers. Overall, the study underscores the importance of structured latent representations for operator learning and highlights the potential of combining generative models and operator networks in thermal design and broader engineering applications.

Figures

Figures reproduced from arXiv: 2509.08515 by the authors.

Figure 1
Figure 1. Schematic illustrating the architecture combining a Variational Rank Reduction Autoencoder and a DeepONet. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Randomly generated geometrical samples for each one of the selected models. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Example of an interpolation (linear, in the latent space) between two samples. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Linear interpolation in the latent space between two geometries (Sample 1 and Sample 2). The first row [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 5 linked inside Pith

  1. [1]

    Fan Chen, Min Yang, and Wentao Yan. Data-driven prognostic model for temperature field in additive manufac- turing based on the high-fidelity thermal-fluid flow simulation.Computer Methods in Applied Mechanics and Engineering, 392:114652, 2022

    2022

  2. [2]

    Real-time 2d temperature field pre- diction in metal additive manufacturing using physics-informed neural networks.arXiv preprint arXiv:2401.02403, 2024

    Pouyan Sajadi, Mostafa Rahmani Dehaghani, Yifan Tang, and G Gary Wang. Real-time 2d temperature field pre- diction in metal additive manufacturing using physics-informed neural networks.arXiv preprint arXiv:2401.02403, 2024

    arXiv 2024

  3. [3]

    Nanxu Pan, Xin Ye, Peng Xia, and Guangshun Zhang. The temperature field prediction and estimation of ti-al alloy twin-wire plasma arc additive manufacturing using a one-dimensional convolution neural network.Applied Sciences, 14(2):661, 2024

    2024

  4. [4]

    A study of simulation of the urban space 3d temperature field at a community scale based on high-resolution remote sensing and cfd.Remote Sensing, 14(13):3174, 2022

    Hongyuan Huo and Fei Chen. A study of simulation of the urban space 3d temperature field at a community scale based on high-resolution remote sensing and cfd.Remote Sensing, 14(13):3174, 2022

    2022

  5. [5]

    Estimating urban spatial temperatures considering anthropogenic heat release factors focusing on the mobility characteristics.Sustainable Cities and Society, 85:104073, 2022

    Yesuel Kim, Hwasoo Yeo, and Youngchul Kim. Estimating urban spatial temperatures considering anthropogenic heat release factors focusing on the mobility characteristics.Sustainable Cities and Society, 85:104073, 2022

    2022

  6. [6]

    Machine learning for urban heat island (uhi) analysis: Predicting land surface temperature (lst) in urban environments.Urban Climate, 55:101962, 2024

    Ghazaleh Tanoori, Ali Soltani, and Atoosa Modiri. Machine learning for urban heat island (uhi) analysis: Predicting land surface temperature (lst) in urban environments.Urban Climate, 55:101962, 2024

    2024

  7. [7]

    3-d fem analysis of the temperature field and the thermal stress for plastics thermalforming.Journal of Materials Processing Technology, 97(1-3):35–43, 2000

    YH Song, KF Zhang, ZR Wang, and FX Diao. 3-d fem analysis of the temperature field and the thermal stress for plastics thermalforming.Journal of Materials Processing Technology, 97(1-3):35–43, 2000

    2000

  8. [8]

    Chaquet and Pedro Galán del Sastre

    Jose M. Chaquet and Pedro Galán del Sastre. Temperature fields calculation in heat exchangers using the finite element method.Finite Elements in Analysis and Design, 249:104385, 2025

    2025

Show all 64 references
  1. [9]

    Evolutionary opti- mization methods for high-dimensional expensive problems: A survey.IEEE/CAA Journal of Automatica Sinica, 11(5):1092–1105, 2024

    MengChu Zhou, Meiji Cui, Dian Xu, Shuwei Zhu, Ziyan Zhao, and Abdullah Abusorrah. Evolutionary opti- mization methods for high-dimensional expensive problems: A survey.IEEE/CAA Journal of Automatica Sinica, 11(5):1092–1105, 2024

    2024

  2. [10]

    Gary Wang

    Songqing Shan and G. Gary Wang. Survey of modeling and optimization strategies to solve high-dimensional de- sign problems with computationally-expensive black-box functions.Structural and Multidisciplinary Optimization, 41(2):219–241, March 2010

    2010

  3. [11]

    Ali Kashefi and Tapan Mukerji. Physics-informed pointnet: A deep learning solver for steady-state incompressible flows and thermal fields on multiple sets of irregular geometries.Journal of Computational Physics, 468:111510, 2022

    2022

  4. [12]

    Junyan He, Seid Koric, Diab Abueidda, Ali Najafi, and Iwona Jasiuk. Geom-deeponet: A point-cloud-based deep operator network for field predictions on 3d parameterized geometries.Computer Methods in Applied Mechanics and Engineering, 429:117130, 2024

    2024

  5. [13]

    Hua Su, An Wang, Tianyi Zhang, Tian Qin, Xiaoping Du, and Xiao-Hai Yan. Super-resolution of subsurface temperature field from remote sensing observations based on machine learning.International Journal of Applied Earth Observation and Geoinformation, 102:102440, 2021

    2021

  6. [14]

    A study on improving temperature field prediction accuracy using a surrogate model assisted by airflow field.Results in Engineering, 25:104544, 2025

    Yue Zhang, Jiashuo Song, Shirui Cheng, and Xinru Wang. A study on improving temperature field prediction accuracy using a surrogate model assisted by airflow field.Results in Engineering, 25:104544, 2025. 12 Rank Reduction Autoencoders for GenerativeTECHNICALREPORT

    2025

  7. [15]

    A physics- informed machine learning approach for temperature field prediction in metallic additive manufacturing.Journal of Industrial Information Integration, page 100899, 2025

    Zhonghao Chen, Haochen Mu, Fengyang He, Lei Yuan, Hongtao Zhu, Ninshu Ma, and Zengxi Pan. A physics- informed machine learning approach for temperature field prediction in metallic additive manufacturing.Journal of Industrial Information Integration, page 100899, 2025

    2025

  8. [16]

    Juan Diego Toscano, Theo Käufer, Zhibo Wang, Martin Maxey, Christian Cierpka, and George Em Karniadakis. Inferring turbulent velocity and temperature fields and their statistics from lagrangian velocity measurements using physics-informed kolmogorov-arnold networks.arXiv prepr...

    2024 arXiv

  9. [17]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, March 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, March 2021

    2021

  10. [18]

    Physically interpretable airfoil parameterization using variational autoencoder-based generative modeling

    Yu-Eop Kang, Dawoon Lee, and Kwanjung Yee. Physically interpretable airfoil parameterization using variational autoencoder-based generative modeling. InAIAA SCITECH 2024 Forum, page 0685, 2024

    2024

  11. [19]

    A generative design method of airfoil based on conditional variational autoencoder.Engineering Applications of Artificial Intelligence, 139:109461, 2025

    Xu Wang, Weiqi Qian, Tun Zhao, Hai Chen, Lei He, Haisheng Sun, and Yuan Tian. A generative design method of airfoil based on conditional variational autoencoder.Engineering Applications of Artificial Intelligence, 139:109461, 2025

    2025

  12. [20]

    Using a generative adversarial network for the inverse design of soft morphing composite beams.Engineering Applications of Artificial Intelligence, 133:108527, 2024

    Tomaž Brzin and Miha Brojan. Using a generative adversarial network for the inverse design of soft morphing composite beams.Engineering Applications of Artificial Intelligence, 133:108527, 2024

    2024

  13. [21]

    Representation learning: A review and new perspectives

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013

    2013

  14. [22]

    Rank reduction autoencoders, 2025

    Jad Mounayer, Sebastian Rodriguez, Chady Ghnatios, Charbel Farhat, and Francisco Chinesta. Rank reduction autoencoders, 2025

    2025

  15. [23]

    Variational rank reduction autoencoder.arXiv preprint arXiv:2505.09458, 2025

    Jad Mounayer, Alicia Tierz, Jerome Tomezyk, Chady Ghnatios, and Francisco Chinesta. Variational rank reduction autoencoder.arXiv preprint arXiv:2505.09458, 2025

    2025

  16. [24]

    Deep generative models in engineering design: A review.Journal of Mechanical Design, 144(7):071704, 2022

    Lyle Regenwetter, Amin Heyrani Nobari, and Faez Ahmed. Deep generative models in engineering design: A review.Journal of Mechanical Design, 144(7):071704, 2022

    2022

  17. [25]

    Deep generative models through the lens of the manifold hypothesis: A survey and new connections.arXiv preprint arXiv:2404.02954, 2024

    Gabriel Loaiza-Ganem, Brendan Leigh Ross, Rasa Hosseinzadeh, Anthony L Caterini, and Jesse C Cresswell. Deep generative models through the lens of the manifold hypothesis: A survey and new connections.arXiv preprint arXiv:2404.02954, 2024

    2024 arXiv

  18. [26]

    Auto-encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013

    2013

  19. [27]

    An indirect design representation for topology optimization using variational autoencoder and style transfer

    Tinghao Guo, Danny J Lohan, Ruijin Cang, Max Yi Ren, and James T Allison. An indirect design representation for topology optimization using variational autoencoder and style transfer. In2018 AIAA/ASCE/AHS/ASC structures, structural dynamics, and materials conference, page 0804, 2018

    2018

  20. [28]

    Research on multi-heat source arrangement optimization based on equivalent heat source method and reconstructed variational autoencoder.Scientific Reports, 14(1):21208, 2024

    Yide Yang, Mali Gong, and Jianshe Ma. Research on multi-heat source arrangement optimization based on equivalent heat source method and reconstructed variational autoencoder.Scientific Reports, 14(1):21208, 2024

    2024

  21. [29]

    Gaussian process prior variational autoencoders.Advances in neural information processing systems, 31, 2018

    Francesco Paolo Casale, Adrian Dalca, Luca Saglietti, Jennifer Listgarten, and Nicolo Fusi. Gaussian process prior variational autoencoders.Advances in neural information processing systems, 31, 2018

    2018

  22. [30]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    2014

  23. [31]

    An adaptive artificial neural network-based generative design method for layout designs.International Journal of Heat and Mass Transfer, 184:122313, 2022

    Chao Qian, Ren Kai Tan, and Wenjing Ye. An adaptive artificial neural network-based generative design method for layout designs.International Journal of Heat and Mass Transfer, 184:122313, 2022

    2022

  24. [32]

    A continuous genetic algorithm designed for the global optimization of multimodal functions.Journal of Heuristics, 6(2):191–213, 2000

    Rachid Chelouah and Patrick Siarry. A continuous genetic algorithm designed for the global optimization of multimodal functions.Journal of Heuristics, 6(2):191–213, 2000

    2000

  25. [33]

    Thermodynamics- informed super-resolution of scarce temporal dynamics data.Computer Methods in Applied Mechanics and Engineering, 430:117210, 2024

    Carlos Bermejo-Barbanoj, Beatriz Moya, Alberto Badías, Francisco Chinesta, and Elías Cueto. Thermodynamics- informed super-resolution of scarce temporal dynamics data.Computer Methods in Applied Mechanics and Engineering, 430:117210, 2024

    2024

  26. [34]

    Jichao Bao, Liangping Li, and Arden Davis. Variational autoencoder or generative adversarial networks? a comparison of two deep learning methods for flow and transport data assimilation.Mathematical Geosciences, 54(6):1017–1042, 2022

    2022

  27. [35]

    Physics-integrated variational autoencoders for robust and interpretable generative modeling.Advances in Neural Information Processing Systems, 34:14809–14821, 2021

    Naoya Takeishi and Alexandros Kalousis. Physics-integrated variational autoencoders for robust and interpretable generative modeling.Advances in Neural Information Processing Systems, 34:14809–14821, 2021

    2021

  28. [36]

    Symplectic encoders for physics-constrained variational dynamics inference.Scientific Reports, 13(1):2643, 2023

    Kiran Bacsa, Zhilu Lai, Wei Liu, Michael Todd, and Eleni Chatzi. Symplectic encoders for physics-constrained variational dynamics inference.Scientific Reports, 13(1):2643, 2023. 13 Rank Reduction Autoencoders for GenerativeTECHNICALREPORT

    2023

  29. [37]

    Pi-vae: Physics-informed variational auto-encoder for stochastic differential equations.Computer Methods in Applied Mechanics and Engineering, 403:115664, 2023

    Weiheng Zhong and Hadi Meidani. Pi-vae: Physics-informed variational auto-encoder for stochastic differential equations.Computer Methods in Applied Mechanics and Engineering, 403:115664, 2023

    2023

  30. [38]

    Generating required motor rotor shape by physics-guided vae/wgan-gp.Results in Engineering, page 106181, 2025

    Kazuo Yonekura, Masayuki Tamura, Yoshihisa Kondo, Keisuke Suzuki, Shun Ogawa, and Katsuyuki Suzuki. Generating required motor rotor shape by physics-guided vae/wgan-gp.Results in Engineering, page 106181, 2025

    2025

  31. [39]

    Alberto Solera-Rico, Carlos Sanmiguel Vila, Miguel Gómez-López, Yuning Wang, Abdulrahman Almashjary, Scott TM Dawson, and Ricardo Vinuesa.β-variational autoencoders and transformers for reduced-order modelling of fluid flows.Nature Communications, 15(1):1361, 2024

    2024

  32. [40]

    Topological autoencoders

    Michael Moor, Max Horn, Bastian Rieck, and Karsten Borgwardt. Topological autoencoders. InInternational conference on machine learning, pages 7045–7054. PMLR, 2020

    2020

  33. [41]

    Towards extraction of orthogonal and parsimonious non-linear modes from turbulent flows.Expert Systems with Applications, 202:117038, 2022

    Hamidreza Eivazi, Soledad Le Clainche, Sergio Hoyas, and Ricardo Vinuesa. Towards extraction of orthogonal and parsimonious non-linear modes from turbulent flows.Expert Systems with Applications, 202:117038, 2022

    2022

  34. [42]

    Stefania Fresca and Andrea Manzoni. Pod-dl-rom: Enhancing deep learning-based reduced order models for nonlinear parametrized pdes by proper orthogonal decomposition.Computer Methods in Applied Mechanics and Engineering, 388:114181, 2022

    2022

  35. [43]

    A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes.Journal of Scientific Computing, 87(2):61, 2021

    Stefania Fresca, Luca Dede’, and Andrea Manzoni. A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes.Journal of Scientific Computing, 87(2):61, 2021

    2021

  36. [44]

    A graph convolutional autoencoder approach to model order reduction for parametrized pdes.Journal of Computational Physics, 501:112762, 2024

    Federico Pichi, Beatriz Moya, and Jan S Hesthaven. A graph convolutional autoencoder approach to model order reduction for parametrized pdes.Journal of Computational Physics, 501:112762, 2024

    2024

  37. [45]

    Latent neural operator for solving forward and inverse pde problems.Advances in Neural Information Processing Systems, 37:33085–33107, 2024

    Tian Wang and Chuang Wang. Latent neural operator for solving forward and inverse pde problems.Advances in Neural Information Processing Systems, 37:33085–33107, 2024

    2024

  38. [46]

    Arda Mavi, Ali Can Bekar, Ehsan Haghighat, and Erdogan Madenci. An unsupervised latent/output physics- informed convolutional-lstm network for solving partial differential equations using peridynamic differential operator.Computer Methods in Applied Mechanics and Engineering, ...

    2023

  39. [47]

    Investigation and implementation of model order reduction technique for large scale dynamical systems.Archives of Computational Methods in Engineering, 29(5):3087– 3108, 2022

    Santosh Kumar Suman and Awadhesh Kumar. Investigation and implementation of model order reduction technique for large scale dynamical systems.Archives of Computational Methods in Engineering, 29(5):3087– 3108, 2022

    2022

  40. [48]

    Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders.Physics of Fluids, 33(3), 2021

    Romit Maulik, Bethany Lusch, and Prasanna Balaprakash. Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders.Physics of Fluids, 33(3), 2021

    2021

  41. [49]

    Discovering governing equations from partial measurements with deep delay autoencoders.Proceedings of the Royal Society A, 479(2276):20230422, 2023

    Joseph Bakarji, Kathleen Champion, J Nathan Kutz, and Steven L Brunton. Discovering governing equations from partial measurements with deep delay autoencoders.Proceedings of the Royal Society A, 479(2276):20230422, 2023

    2023

  42. [50]

    Thermody- namically consistent latent dynamics identification for parametric systems.arXiv preprint arXiv:2506.08475, 2025

    Xiaolong He, Yeonjong Shin, Anthony Gruber, Sohyeon Jung, Kookjin Lee, and Youngsoo Choi. Thermody- namically consistent latent dynamics identification for parametric systems.arXiv preprint arXiv:2506.08475, 2025

    2025 arXiv

  43. [51]

    Physics-informed geometry-aware neural operator.Computer Methods in Applied Mechanics and Engineering, 434:117540, 2025

    Weiheng Zhong and Hadi Meidani. Physics-informed geometry-aware neural operator.Computer Methods in Applied Mechanics and Engineering, 434:117540, 2025

    2025

  44. [52]

    Deep learning of thermodynamics-aware reduced-order models from data.Computer Methods in Applied Mechanics and Engineering, 379:113763, 2021

    Quercus Hernandez, Alberto Badias, David Gonzalez, Francisco Chinesta, and Elias Cueto. Deep learning of thermodynamics-aware reduced-order models from data.Computer Methods in Applied Mechanics and Engineering, 379:113763, 2021

    2021

  45. [53]

    Physics perception in sloshing scenes with guaranteed thermodynamic consistency.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2136–2150, 2022

    Beatriz Moya, Alberto Badias, David Gonzalez, Francisco Chinesta, and Elias Cueto. Physics perception in sloshing scenes with guaranteed thermodynamic consistency.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2136–2150, 2022

    2022

  46. [54]

    Physics-informed neural ode (pinode): embedding physics into models using collocation points.Scientific Reports, 13(1):10166, 2023

    Aleksei Sholokhov, Yuying Liu, Hassan Mansour, and Saleh Nabi. Physics-informed neural ode (pinode): embedding physics into models using collocation points.Scientific Reports, 13(1):10166, 2023

    2023

  47. [55]

    Latentpinns: Generative physics-informed neural networks via a latent representation learning.Artificial Intelligence in Geosciences, page 100115, 2025

    Mohammad H Taufik and Tariq Alkhalifah. Latentpinns: Generative physics-informed neural networks via a latent representation learning.Artificial Intelligence in Geosciences, page 100115, 2025

    2025

  48. [56]

    Learning two-phase microstructure evolution using neural operators and autoencoder architectures.npj Computational Materials, 8(1):190, 2022

    Vivek Oommen, Khemraj Shukla, Somdatta Goswami, Rémi Dingreville, and George Em Karniadakis. Learning two-phase microstructure evolution using neural operators and autoencoder architectures.npj Computational Materials, 8(1):190, 2022

    2022

  49. [57]

    Physics-enhanced machine learning: a position paper for dynamical systems investigations

    Alice Cicirello. Physics-enhanced machine learning: a position paper for dynamical systems investigations. In Journal of Physics: Conference Series, volume 2909, page 012034. IOP Publishing, 2024. 14 Rank Reduction Autoencoders for GenerativeTECHNICALREPORT

    2024

  50. [58]

    Physics-guided deep markov models for learning nonlinear dynamical systems with uncertainty.Mechanical Systems and Signal Processing, 178:109276, 2022

    Wei Liu, Zhilu Lai, Kiran Bacsa, and Eleni Chatzi. Physics-guided deep markov models for learning nonlinear dynamical systems with uncertainty.Mechanical Systems and Signal Processing, 178:109276, 2022

    2022

  51. [59]

    Extracting interpretable physical parameters from spatiotemporal systems using unsupervised learning.Physical Review X, 10(3):031056, 2020

    Peter Y Lu, Samuel Kim, and Marin Soljaˇci´c. Extracting interpretable physical parameters from spatiotemporal systems using unsupervised learning.Physical Review X, 10(3):031056, 2020

    2020

  52. [60]

    Solving inverse-pde problems with physics-aware neural networks.Journal of Computational Physics, 440:110414, 2021

    Samira Pakravan, Pouria A Mistani, Miguel A Aragon-Calvo, and Frederic Gibou. Solving inverse-pde problems with physics-aware neural networks.Journal of Computational Physics, 440:110414, 2021

    2021

  53. [61]

    Discovering sparse interpretable dynamics from partial observations.Communications Physics, 5(1):206, 2022

    Peter Y Lu, Joan Ariño Bernad, and Marin Solja ˇci´c. Discovering sparse interpretable dynamics from partial observations.Communications Physics, 5(1):206, 2022

    2022

  54. [62]

    Learning proper orthogonal decomposition of complex dynamics using heavy-ball neural odes.Journal of Scientific Computing, 95(2):54, 2023

    Justin Baker, Elena Cherkaev, Akil Narayan, and Bao Wang. Learning proper orthogonal decomposition of complex dynamics using heavy-ball neural odes.Journal of Scientific Computing, 95(2):54, 2023

    2023

  55. [63]

    Reduced basis approximations of parameterized dynamical partial differential equations via neural networks.Foundations of Data Science, 7(SAND-2025-04099J), 2025

    Peter Sentz, Kristian Beckwith, Eric C Cyr, Luke N Olson, and Ravi Patel. Reduced basis approximations of parameterized dynamical partial differential equations via neural networks.Foundations of Data Science, 7(SAND-2025-04099J), 2025

    2025

  56. [64]

    Nonlinear model reduction for operator learning.arXiv preprint arXiv:2403.18735, 2024

    Hamidreza Eivazi, Stefan Wittek, and Andreas Rausch. Nonlinear model reduction for operator learning.arXiv preprint arXiv:2403.18735, 2024. 15

    2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.