pith. sign in

arxiv: 2506.09207 · v4 · submitted 2025-06-10 · 💻 cs.LG · cs.NA· math.NA

mLaSDI: Multi-stage latent space dynamics identification

William Anderson , Seung Whan Chung , Robert Stephany , Youngsoo Choi This is my paper

Pith reviewed 2026-05-19 09:51 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NA
keywords multi-stage learninglatent space dynamics identificationreduced-order modelsresidual learningperiodic activation functionspartial differential equationsautoencodersdata-driven modeling
0
0 comments X

The pith

mLaSDI trains residual decoders in stages to recover high-frequency details in reduced-order PDE models while keeping latent dynamics interpretable.

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

The paper proposes mLaSDI as an extension to LaSDI for creating data-driven reduced-order models of partial differential equations. Standard LaSDI struggles because the autoencoder must both reconstruct data and enforce specified latent ODE dynamics at the same time, which limits performance on high-frequency or complex problems. mLaSDI addresses this by training in stages: first the autoencoder, then additional decoders that learn the residuals from earlier stages using periodic activations. This allows better reconstruction of fine details without altering the user-specified ODEs in the latent space. The authors prove that extra stages cannot increase the training residual and provide an error breakdown between autoencoder and dynamics parts.

Core claim

With mLaSDI, the initial autoencoder is trained, after which additional decoders are trained sequentially to map latent trajectories to residuals from previous stages. Combined with periodic activation functions, this staged residual learning recovers high-frequency content without sacrificing interpretability of the latent dynamics. An error decomposition separates autoencoder and latent dynamics contributions, and it is proven that additional training stages cannot increase the training residual.

What carries the argument

Sequential training of residual decoders on the outputs of prior stages, using periodic activations to capture high frequencies.

If this is right

  • Reconstruction and prediction errors drop significantly, often by an order of magnitude, on test problems like multiscale oscillations, wake flows, and Vlasov equations.
  • Training time decreases and hyperparameter tuning becomes less demanding compared to single-stage LaSDI.
  • The error decomposition allows independent assessment of autoencoder reconstruction quality versus latent dynamics accuracy.
  • User-specified ODEs remain intact, preserving the interpretability and flexibility of the latent space model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This method could be applied to other latent variable models in scientific machine learning to improve fidelity on oscillatory or turbulent systems without retraining the entire dynamics.
  • If the residual stages can be added without bound on training error, practitioners might continue refining models until desired accuracy is reached, subject only to overfitting risks.
  • The separation of concerns might enable hybrid models where latent dynamics are derived from first principles while residuals are learned from data.

Load-bearing premise

Sequential training of residual decoders will not degrade the previously learned user-specified ODE dynamics or introduce instabilities affecting generalization to unseen parameters.

What would settle it

Running the training on the unsteady wake flow example and finding that the training residual increases after the second stage would directly contradict the claim that additional stages cannot increase the training residual.

Figures

Figures reproduced from arXiv: 2506.09207 by Robert Stephany, Seung Whan Chung, William Anderson, Youngsoo Choi.

Figure 1
Figure 1. Figure 1: Schematic of mLaSDI. The first stage learns an autoencoder that reconstructs data and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: We apply mLaSDI to a wide range of problems including (a) our synthetic test problem [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Applying GPLaSDI and mLaSDI with 2 stages to [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative error for GPLaSDI and mLaSDI applied to toy problem [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative error for GPLaSDI and mLaSDI applied to unsteady wake flow. GPLaSDI is trained [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Applying mLaSDI with two stages to 1D-1V Vlasov equation using a wide range of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mesh used to solve unsteady wake flow, rotated 90 degrees clockwise. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relative errors applying mLaSDI with 2 stages to the 1D-1V Vlasov equation with Tanh [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Relative prediction error after training GPLaSDI for 50,000 iterations and applying to toy [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Applying GPLaSDI and mLaSDI with 2 stages to to problem [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Relative error for GPLaSDI and mLaSDI applied to toy problem [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Relative error after training GPLaSDI for 50,000 iterations and applying to unstead [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Relative error for GPLaSDI and mLaSDI applied to unsteady wake flow with Softplus [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Applying mLaSDI with two stages to 1D-1V Vlasov equation using a wide range of [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
read the original abstract

Accurately solving partial differential equations (PDEs) is essential across many scientific disciplines. However, high-fidelity solvers can be computationally prohibitive, motivating the development of reduced-order models (ROMs). Recently, Latent Space Dynamics Identification (LaSDI) was proposed as a data-driven, non-intrusive ROM framework. LaSDI compresses the training data via an autoencoder and learns user-specified ordinary differential equations (ODEs), governing the latent dynamics, enabling rapid predictions for unseen parameters. While LaSDI has produced effective ROMs for numerous problems, the autoencoder must simultaneously reconstruct the training data and satisfy the imposed latent dynamics, which are often competing objectives that limit accuracy, particularly for complex or high-frequency phenomena. To address this limitation, we propose multi-stage Latent Space Dynamics Identification (mLaSDI). With mLaSDI, we train LaSDI sequentially in stages. After training the initial autoencoder, we train additional decoders which map the latent trajectories to residuals from previous stages. This staged residual learning, combined with periodic activation functions, enables recovery of high-frequency content without sacrificing interpretability of the latent dynamics. We further provide an error decomposition separating autoencoder and latent dynamics contributions, and prove that additional training stages cannot increase the training residual. Numerical experiments on a multiscale oscillating system, unsteady wake flow, and the 1D-1V Vlasov equation demonstrate that mLaSDI achieves significantly lower reconstruction and prediction errors, often by an order of magnitude, while requiring less training time and reduced hyperparameter tuning compared to standard LaSDI.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes mLaSDI as a multi-stage extension of Latent Space Dynamics Identification (LaSDI) for reduced-order modeling of PDEs. After an initial LaSDI stage that learns an autoencoder and user-specified latent ODE, additional decoders are trained sequentially on residuals using periodic activation functions to recover high-frequency content. The authors present an error decomposition separating autoencoder and latent-dynamics contributions and prove that further stages cannot increase the training residual. Experiments on a multiscale oscillating system, unsteady wake flow, and 1D-1V Vlasov equation report order-of-magnitude reductions in reconstruction and prediction errors relative to standard LaSDI, with reduced training time and hyperparameter tuning.

Significance. If the central claims hold, mLaSDI provides a practical route to higher-accuracy data-driven ROMs for multiscale and high-frequency problems while retaining the interpretability of user-specified latent dynamics. The staged residual learning and periodic activations directly address a known tension in LaSDI between reconstruction fidelity and imposed dynamics. The error decomposition and non-increase proof supply useful theoretical structure, and the reported gains on three distinct test problems suggest the method could be broadly applicable in scientific computing.

major comments (2)
  1. [Theoretical analysis] The non-increase proof for training residuals (abstract and theoretical section) is internal to the training objective and does not establish that the composite model preserves the original user-specified latent ODE under long-term integration or parameter extrapolation. This preservation is load-bearing for the interpretability claim across the multiscale, wake, and Vlasov cases.
  2. [§4] §4 (numerical experiments): the reported error reductions are presented for the composite model, but no ablation isolates whether the added residual decoders alter the previously learned latent trajectories or introduce instabilities when the full model is integrated forward for unseen parameters.
minor comments (2)
  1. [Method] Notation for the residual decoders and their periodic activations should be introduced with explicit equations rather than descriptive text only.
  2. [Figures] Figure captions for the Vlasov and wake examples should state the exact time horizons and parameter ranges used for the prediction-error metrics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped us improve the clarity of our theoretical contributions and experimental validation. We address each major comment in turn below.

read point-by-point responses

  1. Referee: The non-increase proof for training residuals (abstract and theoretical section) is internal to the training objective and does not establish that the composite model preserves the original user-specified latent ODE under long-term integration or parameter extrapolation. This preservation is load-bearing for the interpretability claim across the multiscale, wake, and Vlasov cases.

    Authors: We agree that the non-increase result applies specifically to the training residual. However, the preservation of the user-specified latent ODE is ensured by the structure of mLaSDI: the latent dynamics are identified and fixed during the first stage, and subsequent stages train residual decoders that take the latent states (obtained by integrating the fixed ODE) as input. Thus, the latent trajectories are identical to those of standard LaSDI for any integration length or parameter value. The interpretability claim rests on this architectural choice rather than the residual non-increase proof. We have revised Section 3 to include an explicit statement clarifying that the latent ODE remains unchanged across stages. revision: yes

  2. Referee: §4 (numerical experiments): the reported error reductions are presented for the composite model, but no ablation isolates whether the added residual decoders alter the previously learned latent trajectories or introduce instabilities when the full model is integrated forward for unseen parameters.

    Authors: The referee is correct that an explicit ablation was not presented. By construction, the residual decoders do not alter the latent trajectories, as these are generated exclusively by the user-specified ODE from the first stage; the decoders only refine the reconstruction from those fixed latent states. To address this, we have added an ablation subsection in §4 that confirms the latent trajectories are unchanged and reports forward integration results for unseen parameters, showing stable behavior with no introduced instabilities in any of the three test problems. revision: yes

Circularity Check

0 steps flagged

New staged residual procedure and non-increase proof are self-contained; minor LaSDI citation is not load-bearing

full rationale

The derivation introduces an explicit sequential training procedure (initial LaSDI followed by residual decoders on fixed latent trajectories) together with a fresh error decomposition and a direct proof that additional stages cannot increase the training residual. These steps are constructed and proven internally without any equation reducing a claimed prediction to a fitted parameter or to a prior self-citation. The LaSDI reference is used only to define the baseline stage and does not carry the central claims about high-frequency recovery or interpretability preservation. The approach therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The work inherits the LaSDI assumptions on latent ODEs and adds the staged residual procedure; no new physical entities are postulated.

free parameters (2)
  • Number of training stages
    Chosen by user or via tuning to balance accuracy and cost.
  • Periodic activation hyperparameters
    Frequencies and amplitudes in periodic activations require selection.
axioms (2)
  • domain assumption Latent dynamics can be adequately modeled by user-specified ODEs
    Core assumption carried over from the original LaSDI framework.
  • ad hoc to paper Residuals between stages can be learned independently by additional decoders without destabilizing prior latent dynamics
    Introduced to justify the sequential training procedure.

pith-pipeline@v0.9.0 · 5828 in / 1460 out tokens · 56959 ms · 2026-05-19T09:51:37.651189+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    This staged residual learning, combined with periodic activation functions, enables recovery of high-frequency content without sacrificing interpretability of the latent dynamics. We further provide an error decomposition separating autoencoder and latent dynamics contributions, and prove that additional training stages cannot increase the training residual.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · 1 internal anchor

  1. [1]

    Aldirany, R

    Z. Aldirany, R. Cottereau, M. Laforest, and S. Prudhomme. Multi-level neural networks for accurate solutions of boundary-value problems. Computer Methods in Applied Mechanics and Engineering, 419:116666, 2024. ISSN 0045-7825. doi: https://doi.org/10.1016/j.cma.2023. 116666

    work page doi:10.1016/j.cma.2023 2024

  2. [2]

    Anderson et al

    R. Anderson et al. Mfem: A modular finite element methods library. Computers & Mathematics with Applications, 81:42–74, 2021

    work page 2021

  3. [3]

    Benner, S

    P. Benner, S. Gugercin, and K. Willcox. A survey of projection-based model reduction methods for parametric dynamical systems. SIAM Review , 57(4):483–531, 2015. doi: 10.1137/130932715

    work page doi:10.1137/130932715 2015

  4. [4]

    Benner, P

    P. Benner, P. Goyal, B. Kramer, B. Peherstorfer, and K. Willcox. Operator inference for non- intrusive model reduction of systems with non-polynomial nonlinear terms. Computer Methods in Applied Mechanics and Engineering, 372:113433, 2020

    work page 2020

  5. [5]

    Berkooz, P

    G. Berkooz, P. Holmes, and J. L. Lumley. The proper orthogonal decomposition in the analysis of turbulent flows. Annual review of fluid mechanics, 25(1):539–575, 1993

    work page 1993

  6. [6]

    Bonneville, Y

    C. Bonneville, Y . Choi, D. Ghosh, and J.L. Belof. Gplasdi: Gaussian process-based interpretable latent space dynamics identification through deep autoencoder. Computer Methods in Applied Mechanics and Engineering, 418:116535, 2024. ISSN 0045-7825. doi: https://doi.org/10.1016/ j.cma.2023.116535

    work page arXiv 2024

  7. [7]

    Bonneville et al

    C. Bonneville et al. A comprehensive review of latent space dynamics identification algorithms for intrusive and non-intrusive reduced-order-modeling. arXiv preprint arXiv:2403.10748 , 2024

    work page arXiv 2024

  8. [8]

    Brunton, Joshua L

    S.L. Brunton, J.L. Proctor, and J.N. Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016. doi: 10.1073/pnas.1517384113

    work page doi:10.1073/pnas.1517384113 2016

  9. [9]

    Calder, C

    M. Calder, C. Craig, D. Culley, R. de Cani, C. A. Donnelly, R. Douglas, B. Edmonds, J. Gas- coigne, N. Gilbert, C. Hargrove, F. Hinds, D. C. Lane, D. Mitchell, G. Pavey, D. Robertson, B. Rosewell, S. Sherwin, M. Walport, and A. Wilson. Computational modelling for decision- making: where, why, what, who and how. Royal Society Open Science, 5(6):172096, 201...

    work page doi:10.1098/rsos.172096 2018

  10. [10]

    Champion, B

    K. Champion, B. Lusch, J. N. Kutz, and S. L. Brunton. Data-driven discovery of coordinates and governing equations. Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019

    work page 2019

  11. [11]

    Cranmer et al

    M. Cranmer et al. Discovering symbolic models from deep learning with inductive biases. Advances in Neural Information Processing Systems, 33:17429–17442, 2020

    work page 2020

  12. [12]

    R. M. Cummings, W. H. Mason, S. A. Morton, and D. R. McDaniel. Applied computational aerodynamics: A modern engineering approach, volume 53. Cambridge University Press, 2015

    work page 2015

  13. [13]

    A. N. Diaz, Y . Choi, and M. Heinkenschloss. A fast and accurate domain decomposition nonlinear manifold reduced order model. Computer Methods in Applied Mechanics and Engineering, 425:116943, 2024

    work page 2024

  14. [14]

    Fries, X.iaolong He, and Y .oungsoo Choi

    W.illiam D. Fries, X.iaolong He, and Y .oungsoo Choi. Lasdi: Parametric latent space dynamics identification. Computer Methods in Applied Mechanics and Engineering, 399:115436, 2022. ISSN 0045-7825. doi: https://doi.org/10.1016/j.cma.2022.115436

    work page doi:10.1016/j.cma.2022.115436 2022

  15. [15]

    Geuzaine and J.F

    C. Geuzaine and J.F. Remacle. Gmsh: A 3-d finite element mesh generator with built-in pre- and post-processing facilities. International Journal for Numerical Methods in Engineering, 79 (11):1309–1331, 2009. doi: https://doi.org/10.1002/nme.2579

    work page doi:10.1002/nme.2579 2009

  16. [16]

    X. He, Y . Choi, W.D. Fries, J.L. Belof, and J. Chen. glasdi: Parametric physics-informed greedy latent space dynamics identification. Journal of Computational Physics, 489:112267, 2023. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2023.112267

    work page doi:10.1016/j.jcp.2023.112267 2023

  17. [17]

    X. He, A. Tran, D.M. Bortz, and Y . Choi. Physics-informed active learning with simultaneous weak-form latent space dynamics identification. International Journal for Numerical Methods in Engineering, 126(1):e7634, 2025. doi: https://doi.org/10.1002/nme.7634

    work page doi:10.1002/nme.7634 2025

  18. [18]

    G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006

    work page 2006

  19. [19]

    Howard, S.H

    A.A. Howard, S.H. Murphy, S.E. Ahmed, and P. Stinis. Stacked networks improve physics- informed training: Applications to neural networks and deep operator networks. Foundations of Data Science, 7(1):134–162, 2025. doi: 10.3934/fods.2024029

    work page doi:10.3934/fods.2024029 2025

  20. [20]

    Hypar repository

    HyPar. Hypar repository. https://bitbucket.org/deboghosh/hypar

    work page

  21. [21]

    Issan and B

    O. Issan and B. Kramer. Predicting solar wind streams from the inner-heliosphere to earth via shifted operator inference. arXiv preprint arXiv:2203.13372, 2022

    work page arXiv 2022

  22. [22]

    Jiang and C.-W

    G.-S. Jiang and C.-W. Shu. Efficient implementation of weighted ENO schemes. Journal of Computational Physics, 126(1):202–228, 1996. doi: 10.1006/jcph.1996.0130

    work page doi:10.1006/jcph.1996.0130 1996

  23. [23]

    Jones, C

    D. Jones, C. Snider, A. Nassehi, J. Yon, and B. Hicks. Characterising the digital twin: A systematic literature review. CIRP Journal of Manufacturing Science and Technology , 29: 36–52, 2020. ISSN 1755-5817. doi: https://doi.org/10.1016/j.cirpj.2020.02.002

    work page doi:10.1016/j.cirpj.2020.02.002 2020

  24. [24]

    Y . Kim, Y . Choi, D. Widemann, and T. Zohdi. A fast and accurate physics-informed neural network reduced order model with shallow masked autoencoder. Journal of Computational Physics, 451:110841, 2022

    work page 2022

  25. [25]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  26. [26]

    Messenger and D.M

    D.A. Messenger and D.M. Bortz. Weak sindy: Galerkin-based data-driven model selection. Multiscale Modeling & Simulation, 19(3):1474–1497, 2021. doi: 10.1137/20M1343166

    work page doi:10.1137/20m1343166 2021

  27. [27]

    D. Noble. The rise of computational biology. Nature Reviews Molecular Cell Biology, 3(6): 459–463, 2002

    work page 2002

  28. [28]

    J. Sur R. Park, S.W. Cheung, Y . Choi, and Yeonjong Shin. tlasdi: Thermodynamics-informed latent space dynamics identification. Computer Methods in Applied Mechanics and Engineering, 429:117144, 2024. ISSN 0045-7825. doi: https://doi.org/10.1016/j.cma.2024.117144

    work page doi:10.1016/j.cma.2024.117144 2024

  29. [29]

    A. T. Patera, G. Rozza, et al. Reduced basis approximation and a posteriori error estimation for parametrized partial differential equations. 2007

    work page 2007

  30. [30]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and É. Duchesnay. Scikit-learn: Machine learning in python. J. Mach. Learn. Res., 12(null):2825–2830, November 2011. ISSN 1532-4435. 11

    work page 2011

  31. [31]

    E. Qian, B. Kramer, B. Peherstorfer, and K. Willcox. Lift & learn: Physics-informed machine learning for large-scale nonlinear dynamical systems. Physica D: Nonlinear Phenomena, 406: 132401, 2020

    work page 2020

  32. [32]

    C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. The MIT Press, 11 2005. ISBN 9780262256834. doi: 10.7551/mitpress/3206.001.0001

    work page doi:10.7551/mitpress/3206.001.0001 2005

  33. [33]

    M. G. Safonov and R. Chiang. A schur method for balanced-truncation model reduction. IEEE Transactions on Automatic Control, 34(7):729–733, 1989

    work page 1989

  34. [34]

    Schmidt and H

    M. Schmidt and H. Lipson. Distilling free-form natural laws from experimental data. Science, 324(5923):81–85, 2009

    work page 2009

  35. [35]

    Thijssen

    J. Thijssen. Computational physics. Cambridge university press, 2007

    work page 2007

  36. [36]

    A. Tran, X. He, D. A. Messenger, Y . Choi, and D. M. Bortz. Weak-form latent space dynamics identification. Computer Methods in Applied Mechanics and Engineering , 427:116998, Jul 2024

    work page 2024

  37. [37]

    Vasileska, S

    D. Vasileska, S. M. Goodnick, and G. Klimeck. Computational Electronics: semiclassical and quantum device modeling and simulation. CRC press, 2017

    work page 2017

  38. [38]

    Vincent, H

    P. Vincent, H. Larochelle, I. Lajoie, Y . Bengio, and P. Manzagol. Stacked denoising autoen- coders: Learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research, 11(110):3371–3408, 2010

    work page 2010

  39. [39]

    Wang and C

    Y . Wang and C. Lai. Multi-stage neural networks: Function approximator of machine precision. Journal of Computational Physics, 504:112865, 2024. ISSN 0021-9991. doi: https://doi.org/10. 1016/j.jcp.2024.112865

    work page arXiv 2024

  40. [40]

    Multi-stage convolutional autoencoder network for hyperspectral unmixing

    Yang Yu, Yong Ma, Xiaoguang Mei, Fan Fan, Jun Huang, and Hao Li. Multi-stage convolutional autoencoder network for hyperspectral unmixing. International Journal of Applied Earth Observation and Geoinformation, 113:102981, 2022. ISSN 1569-8432. doi: https://doi.org/10. 1016/j.jag.2022.102981

    work page arXiv 2022

  41. [41]

    Zabalza, J

    J. Zabalza, J. Ren, J. Zheng, H. Zhao, C. Qing, Z. Yang, P. Du, and S. Marshall. Novel segmented stacked autoencoder for effective dimensionality reduction and feature extraction in hyperspectral imaging. Neurocomputing, 185:1–10, 2016. ISSN 0925-2312. doi: https: //doi.org/10.1016/j.neucom.2015.11.044. 12 Table 1: Hyperparameters and training time of aut...

    work page doi:10.1016/j.neucom.2015.11.044 2016

  42. [42]

    We run full-order simulations for parameter values T ∈ [0.9, 1.1] and k ∈ [1.0, 1.2], where the parameter ranges are discretized by ∆T = ∆k = 0.01

    and RK4 time integration scheme with timestep ∆t = 0.005. We run full-order simulations for parameter values T ∈ [0.9, 1.1] and k ∈ [1.0, 1.2], where the parameter ranges are discretized by ∆T = ∆k = 0.01. To generate data, we sample the solution at every timestep from a uniform64×64 grid in the space-velocity field to obtain 251 snapshots of our state ve...

    work page