Pith. sign in

REVIEW 4 major objections 8 minor 30 references

Rollout-LaSDI: Enhancing the long-term accuracy of Latent Space Dynamics

T0 review · 4 major / 8 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 Training latent-space ROMs to predict future frames reduces long-horizon error by 3x on 2D Burgers.

desk verdict A clean but incremental extension of GPLaSDI with a rollout loss; the 3x/2x error drop is plausible but not yet demonstrated on unseen parameters because the test grid overlaps the training set. read the letter →

arxiv 2509.08191 v1 pith:F7Q3MAIW submitted 2025-09-09 cs.LG

Robert Stephany , Youngsoo Choi This is my paper
classification cs.LG
keywords latentspacedynamicsidentificationreduced-ordermodelsrolloutlosslong-horizonpredictionnonuniformfinitedifferencesparametricPDEsautoencoderBurgersequation
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 argues that latent-space reduced-order models—fast surrogates that evolve a low-dimensional encoding of a PDE state—lose accuracy over long time horizons because their training objective only matches one-step dynamics, so small errors compound. To fix this, it introduces a rollout loss: during training, the model integrates its learned latent dynamics forward over randomly chosen future times and is penalized on the decoded difference from the true future state. On the 2D Burgers equation, adding this loss reduces the maximum relative error by 3x and the median error by 2x across a 121-parameter test grid, for both fixed and variable time steps, with no added inference cost. The paper also replaces the two-point derivative estimate used by earlier methods with a higher-order finite-difference stencil that works on nonuniform time series.

What carries the argument

Two mechanisms carry the argument. First, the rollout loss: it makes the encoder, decoder, and latent coefficients co-trained on multi-step prediction error by solving the latent ODE over random horizons up to an annealed maximum and comparing the decoded output with the true future FOM state; this directly penalizes the error accumulation that one-step dynamics losses miss. Second, a nonuniform three-point finite-difference stencil with one-sided variants: it estimates latent velocities with O(h^2) accuracy in linear time, replacing the two-point scheme when time steps are irregular. Both are training-side only, preserving the GPLaSDI inference pipeline of encode, integrate with RK4, decode

What would settle it

Re-run the 2D Burgers benchmark with and without rollout using the released code: compute the ratio of maximum relative error without rollout to with rollout over the 121-parameter grid, in both fixed- and variable-time-step settings. The central claim fails if the ratios are near 1 instead of near 3 (maximum) and 2 (median), or if inference wall-clock time increases noticeably.

Watch

Extended reading notes

Core claim

The paper's central claim is that training a latent-space ROM the way it will be used—rolling the latent ODE forward and penalizing decoded future states—is what fixes long-horizon drift, not a more expressive dynamics model. The rollout loss encodes each rollable frame, integrates the learned affine latent ODE with a differentiable RK4 solver to a random future time, decodes, and compares against a cubic-spline interpolation of the full-order solution. An annealed maximum horizon lets training start with short predictions and grow. Combined with an O(h^2) three-point finite-difference stencil for nonuniform time grids, this constitutes Rollout-LaSDI. On 2D Burgers with 121 test parameter co

Load-bearing premise

The paper assumes the latent dynamics are well approximated by a parameter-specific linear ODE, z-dot ≈ A_θ z + b_θ, that does not change over time; if the true latent dynamics are nonlinear or time-varying, neither the learned coefficients nor the rollout loss can represent them.

Editorial extensions

If this is right

  • Existing GPLaSDI-style ROMs can add the rollout loss during training without changing test-time code, gaining long-horizon accuracy at no inference cost.
  • ROMS trained on adaptively stepped, nonuniform time series no longer have to accept poor two-point derivative estimates; the new stencil gives O(h^2) accuracy for roughly the same cost.
  • The GP-based greedy parameter selection still works, so the method remains non-intrusive and can grow its training set to newly problematic parameter regions.
  • The reported 10^5 speedup over the full-order solver is maintained, since inference is unchanged.
  • For the 2D Burgers test family, maximum relative error drops by a factor of 3 and median by a factor of 2 across the 121-parameter grid, for both fixed and variable time steps.

Reading between the lines

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

  • The result is demonstrated on one PDE; a natural next test is a problem with nonlinear or chaotic latent dynamics, where the random-horizon rollout may need its maximum horizon tied to the system's Lyapunov timescale.
  • If the improvement comes from matching training to deployment rather than from the linear-ODE representation, the rollout loss should transfer to other LaSDI variants, such as SINDy-based or weak-form versions, with no change to their coefficient models.
  • The nonuniform finite-difference stencil is a standalone numerical tool: it could improve latent-velocity estimates in any autoencoder-based dynamical-system pipeline with irregular sampling.
  • The annealed rollout horizon acts as an implicit curriculum from easy short-horizon tasks to hard long-horizon ones; an explicit schedule driven by validation error could be a cheap further improvement.
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 / 8 minor

Summary. The paper proposes two enhancements to the GPLaSDI reduced-order modeling framework: (1) a nonuniform three-point finite-difference scheme for estimating latent time derivatives, and (2) a rollout loss that trains the autoencoder and per-parameter affine latent dynamics to match future FOM states over randomly drawn horizons. The method is tested on a parameterized 2D Burgers equation with both fixed and variable time steps, in each case comparing a model trained with the rollout loss against one trained without it. The authors report that the rollout loss reduces the maximum relative error by 3x and the median error by 2x over 121 parameter values, while leaving inference cost unchanged because the rollout loss is used only during training.

Significance. If the empirical claim is correct, the rollout loss is a simple and inexpensive modification that directly targets error accumulation in latent-space ROMs, and the nonuniform finite-difference scheme is a useful practical improvement for adaptively sampled time series. The inference cost is unchanged by construction, and the central ablation is internally consistent: the with- and without-rollout models are otherwise described as identical. The finite-difference formula in Eq. (5) is standard and correct. However, the paper does not provide code, repeated runs, or a held-out parameter evaluation, and the only benchmark is the 2D Burgers equation. These omissions materially weaken the evidence for the central generalization claim.

major comments (4)
  1. [§3, Fig. 2 and Eq. (6)] The 121-point evaluation grid is the same grid from which the initial four training parameters and all greedily added parameters are drawn (red- and black-bordered squares in Fig. 2). Since the rollout loss in Eq. (6) directly penalizes future-state error for the training parameters, the reported 3x/2x improvement is in part in-sample and does not by itself establish that long-horizon accuracy improves for unseen θ, which is the target of a parameterized ROM. No error statistic is reported for the subset of parameters that were never used in training. Please add a held-out evaluation on parameters excluded from the training set and report trained vs. never-trained errors separately.
  2. [§3, greedy sampling and 'identical otherwise'] The paper states that the with- and without-rollout ROMs are trained identically otherwise, but greedy sampling selects new training parameters based on the current model's predictive variance. The two models will generally select different training sets, so the ablation compares different data as well as different losses. The manuscript does not report which parameters were selected in each run. Please either fix the training sets to be identical across conditions or report and analyze the selected sets.
  3. [§3, experimental procedure] Each condition is a single training run. There are no seeds, repeated trials, or error bars. Given the stochasticity of Adam and the adaptive greedy sampling, the reported 3x/2x error reductions could be within run-to-run variation. Please provide multiple independent runs (at least 3–5) and report the distribution of the max/median errors, or explain why the comparison is deterministic.
  4. [§2, rollout horizon schedule] The rollout horizon is described only as 'annealed from near zero during training,' with no schedule, final value, or random-seed details. Since Eq. (6) depends on Δtθ_max and the sampling of Δtθ, this missing information makes the exact training objective non-reproducible. Please specify the annealing schedule, the maximum horizon, and how the random rollout horizons are drawn.
minor comments (8)
  1. [Abstract/§2] The term 'high-order' overstates Eq. (5), which is second-order accurate. Suggest 'second-order' or 'higher-order than the two-point scheme.'
  2. [Abstract] 'Arbitrary time horizons' is stronger than what is tested; the experiments use T=2 and the rollout horizon appears capped at T. Please temper the claim.
  3. [§3] No exact numerical error values are given; Fig. 2 alone supports the 3x/2x statement. Please include a table with the median and max errors for each condition.
  4. [§2] The one-sided finite-difference stencils used at boundaries are mentioned but not written out; adding them would improve reproducibility.
  5. [§3] Only the 2D Burgers equation is tested. The authors should state this as a limitation and, if space permits, add a second parameterized PDE.
  6. [§3] The statement 'We will make our source code available after review to avoid revealing the authors' identities' is inappropriate in a journal submission; release code via a neutral repository or state availability in the final version.
  7. [Eq. (9)] The error metric normalizes by the temporal standard deviation of FOM components; the indices in the formula (i, j, k) are confusing and should be cleaned up.
  8. [Eq. (3)] The affine latent-dynamics assumption is inherited from GPLaSDI and is not validated on this problem. Adding a residual or correlation analysis would clarify whether the capacity assumption limits the method.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 3x/2x rollout improvement is partly in-sample because the 121-point 'test set' includes the training parameters that Eq. (6) was explicitly fit to.

  1. fitted input called prediction [Section 3, Eq. (6), Eq. (9), and Figure 2]
    "The test set consists of an 11×11 grid (121 parameter combinations, with 11 values for both ν and ω). The training set begins with 4 parameters (red-bordered squares in Fig. 2) but dynamically grows (black-bordered squares) via greedy sampling as the ROM trains [6]. ... Figure 2 reports the results across all θ values."

    For every parameter that enters the training set (red/black-bordered), the Rollout loss in Eq. (6) directly minimizes the L1 difference between the decoded rollout and the FOM future frames. The evaluation metric in Eq. (9) is the same FOM-versus-decoded-rollout comparison, only normalized by frame standard deviation. Because the 121-parameter grid is called the test set but includes these training parameters, and because the reported 3×/2× reduction aggregates over all 121, a portion of the claimed improvement is the training objective evaluated on training points—i.e., a fitted input reported as a prediction. The paper never separates never-trained parameters from training parameters, so the out-of-sample component of the claim is not established.

full rationale

The core methodological novelty—an O(h^2) nonuniform finite-difference stencil and a rollout loss—is not circular in itself: the finite-difference formula is standard external mathematics, and the affine latent-dynamics ansatz (Eq. 3) is stated explicitly rather than smuggled in. The paper's reliance on the authors' prior GPLaSDI work for GP interpolation, greedy sampling, and the 10^5 speedup is a load-bearing component but not a circular conclusion: the central accuracy comparison is an ablation against a shared FOM baseline, not a derivation from the same authors' theorem. However, the main empirical claim is partially circular as an evaluation practice: the 'test set' of 121 parameter values includes the parameters on which the model was trained, and for those parameters the rollout loss directly supervises the exact quantity being measured. Thus the headline 3×/2× error reduction is, to an unknown degree, an in-sample fit rather than a demonstrated generalization gain. The paper does not report separate held-out errors, so the independent content of the claim remains unquantified. This is a partial circularity in the validation of the central quantitative claim, not in the mathematical derivation of the method.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The method depends on standard LaSDI and GPLaSDI assumptions (manifold hypothesis, affine latent dynamics, GP interpolation) and on a set of hand-chosen hyperparameters. The weakest unverified assumptions are the affine latent-dynamics model and the adequacy of cubic-spline rollout targets, both of which are load-bearing for the central claim.

free parameters (6)
  • Latent dimension L = 5
    Chosen by hand; sets the dimension of the latent dynamics. Larger or smaller L changes accuracy but is not varied.
  • Loss weights eta1..eta4 = {1, 1, 1, 0.001}
    Hand-chosen balance between reconstruction, latent dynamics, rollout, and coefficient regularization; no sensitivity study is reported.
  • Rollout horizon annealing schedule = unspecified, annealed from near zero
    Controls how long predictions are trained to stay accurate; this is a load-bearing training detail that is not described quantitatively.
  • Network architecture = [2601,250,100,100,100,5] with sin activations
    Choice of autoencoder capacity; no ablation is performed.
  • Training budget and greedy sampling settings = 17,500 epochs, Adam lr=1e-3, greedy every 2,500 epochs, 20 GP samples
    Hand-selected; greedy sampling schedule affects which parameters enter training and contributes to the test-set overlap.
  • Initial training parameter values = 4 red squares, values not listed in text
    Starting set for GPLaSDI greedy sampling; likely chosen manually and not stated, affecting the final ROM and the training set.
assumptions (6)
  • standard math The three-point nonuniform finite-difference formula (Eq. 5) gives O(h^2) derivative estimates for smooth latent trajectories.
    Valid via Taylor expansion and cited to a standard numerical analysis text; reasonable for smooth latent signals but may degrade on noisy latent time series.
  • domain assumption FOM states lie on or near a low-dimensional manifold, so an autoencoder can provide a chart and inverse.
    This is the manifold hypothesis invoked in Section 1; if it fails, the encoder-decoder cannot absorb complexity and the affine latent model will not represent the dynamics.
  • domain assumption Per-parameter latent dynamics are well approximated by the affine ODE in Eq. (3).
    Core modeling choice inherited from LaSDI and GPLaSDI; rollout training cannot correct representation error caused by this constraint.
  • domain assumption Cubic-spline interpolation of FOM solutions is a faithful target for the rollout loss in Eq. (6).
    Rollout loss compares against spline-interpolated future FOM frames; if the spline smooths or distorts the true solution, the training target is biased.
  • domain assumption Gaussian process interpolation of latent coefficients for unseen parameters accurately predicts dynamics.
    At test time the GP posterior mean supplies A_theta and b_theta; if GP extrapolation is poor, generalization to new parameters fails.
  • domain assumption The high-fidelity FOM solver is accurate enough to serve as ground truth for both training targets and error evaluation.
    All errors are measured against solver output, but the solver scheme, grid convergence, and numerical error are not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rollout-LaSDI: Enhancing the long-term accuracy of Latent Space Dynamics." pith.science (2026). https://pith.science/paper/F7Q3MAIW

@misc{pith2026250908191,
  author       = {Pith},
  title        = {Pith review of: Rollout-LaSDI: Enhancing the long-term accuracy of Latent Space Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7Q3MAIW}},
  note         = {Machine review of arXiv:2509.08191}
}
read the original abstract

Solving complex partial differential equations is vital in the physical sciences, but often requires computationally expensive numerical methods. Reduced-order models (ROMs) address this by exploiting dimensionality reduction to create fast approximations. While modern ROMs can solve parameterized families of PDEs, their predictive power degrades over long time horizons. We address this by (1) introducing a flexible, high-order, yet inexpensive finite-difference scheme and (2) proposing a Rollout loss that trains ROMs to make accurate predictions over arbitrary time horizons. We demonstrate our approach on the 2D Burgers equation.

Figures

Figures reproduced from arXiv: 2509.08191 by the authors.

Figure 1
Figure 1. Rolling out ⃗uθ(t), the discretization of uθ(t, X). From left to right, we begin with ⃗uθ(t), a numerical approximation of the FOM solution, uθ(t, X) (black circles represent the spatial nodes). We encode ⃗uθ(t), use the encoding as the latent dynamics’ initial condition. We solve the dynamics over a random time horizon, then decode the final latent state to predict the future FOM state. We train our ROM using Adam … view at source ↗
Figure 2
Figure 2. ) but dynamically grows (black-bordered squares) via greedy sampling as the ROM trains [6]. At test time, the final ROM and learned GPs predict the FOM solution for each θ: we encode the FOM initial condition5 , set latent coefficients to the GP posterior mean, integrate the latent dynamics on (0, 2], and decode to obtain the predicted FOM time series. Because Rollout-LaSDI only uses LRollout during training, and be… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 6 canonical work pages

  1. [1]

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

    Robert Anderson, Julian Andrej, Andrew Barker, Jamie Bramwell, Jean-Sylvain Camier, Jakub Cerveny, Veselin Dobrev, Yohann Dudouit, Aaron Fisher, Tzanio Kolev, et al. Mfem: A modular finite element methods library.Computers & Mathematics with Applications, 81:42–74, 2021

    2021

  2. [2]

    mlasdi: Multi-stage latent space dynamics identification.arXiv preprint arXiv:2506.09207, 2025

    William Anderson, Seung Whan Chung, and Youngsoo Choi. mlasdi: Multi-stage latent space dynamics identification.arXiv preprint arXiv:2506.09207, 2025

    arXiv 2025

  3. [3]

    High- performance finite elements with mfem.The International Journal of High Performance Computing Applications, 38(5):447–467, 2024

    Julian Andrej, Nabil Atallah, Jan-Phillip Bäcker, Jean-Sylvain Camier, Dylan Copeland, Veselin Dobrev, Yohann Dudouit, Tobias Duswald, Brendan Keith, Dohyun Kim, et al. High- performance finite elements with mfem.The International Journal of High Performance Computing Applications, 38(5):447–467, 2024

    work page 2024

  4. [4]

    The proper orthogonal decomposition in the analysis of turbulent flows.Annual review of fluid mechanics, 25(1):539–575, 1993

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

    1993

  5. [5]

    Data-driven autoencoder numerical solver with uncertainty quantification for fast physical simulations.arXiv preprint arXiv:2312.01021, 2023

    Christophe Bonneville, Youngsoo Choi, Debojyoti Ghosh, and Jonathan L Belof. Data-driven autoencoder numerical solver with uncertainty quantification for fast physical simulations.arXiv preprint arXiv:2312.01021, 2023

    arXiv 2023

  6. [6]

    Christophe Bonneville, Youngsoo Choi, Debojyoti Ghosh, and Jonathan L Belof. GPLaSDI: Gaussian process-based interpretable latent space dynamics identification through deep autoen- coder.Computer Methods in Applied Mechanics and Engineering, 418:116535, 2024

    work page 2024

  7. [7]

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

    Christophe Bonneville, Xiaolong He, April Tran, Jun Sur Park, William Fries, Daniel A Messenger, Siu Wun Cheung, Yeonjong Shin, David M Bortz, Debojyoti Ghosh, 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

    arXiv 2024

  8. [8]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

Show all 30 references
  1. [9]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

    Steven L Brunton, Joshua L Proctor, and J Nathan 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

    2016

  2. [10]

    Numerical analysis, brooks, 1997

    Richard L Burden and J Douglas Faires. Numerical analysis, brooks, 1997

    1997

  3. [11]

    Data-driven discovery of coordinates and governing equations.Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019

    Kathleen Champion, Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Data-driven discovery of coordinates and governing equations.Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019

    2019

  4. [12]

    Latent space dynamics identification for interface tracking with application to shock-induced pore collapse.arXiv preprint arXiv:2507.10647, 2025

    Seung Whan Chung, Christopher Miller, Youngsoo Choi, Paul Tranquilli, H Keo Springer, and Kyle Sullivan. Latent space dynamics identification for interface tracking with application to shock-induced pore collapse.arXiv preprint arXiv:2507.10647, 2025

    2025 arXiv

  5. [13]

    LaSDI: Parametric latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 399:115436, 2022

    William D Fries, Xiaolong He, and Youngsoo Choi. LaSDI: Parametric latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 399:115436, 2022. 5

    2022

  6. [14]

    Alexander N Gorban and Ivan Yu Tyukin. Blessing of dimensionality: mathematical foun- dations of the statistical physics of data.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 376(2118):20170237, 2018

    2018

  7. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  8. [16]

    gLaSDI: Parametric physics-informed greedy latent space dynamics identification.Journal of Computational Physics, 489:112267, 2023

    Xiaolong He, Youngsoo Choi, William D Fries, Jonathan L Belof, and Jiun-Shyan Chen. gLaSDI: Parametric physics-informed greedy latent space dynamics identification.Journal of Computational Physics, 489:112267, 2023

    2023

  9. [17]

    Physics-informed active learning with simultaneous weak-form latent space dynamics identification.International Journal for Numerical Methods in Engineering, 126(1):e7634, 2025

    Xiaolong He, April Tran, David M Bortz, and Youngsoo Choi. Physics-informed active learning with simultaneous weak-form latent space dynamics identification.International Journal for Numerical Methods in Engineering, 126(1):e7634, 2025

    2025

  10. [18]

    Openfoam: A c++ library for complex physics simulations

    Hrvoje Jasak, Aleksandar Jemcov, Zeljko Tukovic, et al. Openfoam: A c++ library for complex physics simulations. InInternational workshop on coupled methods in numerical dynamics, volume 1000, pages 1–20. Dubrovnik, Croatia), 2007

    2007

  11. [19]

    Sparse identification of nonlinear dynamics for model predictive control in the low-data limit.Proceedings of the Royal Society A, 474 (2219):20180335, 2018

    Eurika Kaiser, J Nathan Kutz, and Steven L Brunton. Sparse identification of nonlinear dynamics for model predictive control in the low-data limit.Proceedings of the Royal Society A, 474 (2219):20180335, 2018

    2018

  12. [20]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

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

    2014 arXiv

  13. [21]

    Imagenet classification with deep convolutional neural networks.Communications of the ACM, 60(6):84–90, 2017

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks.Communications of the ACM, 60(6):84–90, 2017

    2017

  14. [22]

    tlasdi: Thermodynamics-informed latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 429:117144, 2024

    Jun Sur Richard Park, Siu Wun Cheung, Youngsoo Choi, and Yeonjong Shin. tlasdi: Thermodynamics-informed latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 429:117144, 2024

    2024

  15. [23]

    Universal differential equations for scientific machine learning.arXiv preprint arXiv:2001.04385, 2020

    Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. Universal differential equations for scientific machine learning.arXiv preprint arXiv:2001.04385, 2020

    2001 arXiv

  16. [24]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

    2019

  17. [25]

    Gaussian processes in machine learning

    Carl Edward Rasmussen. Gaussian processes in machine learning. InSummer school on machine learning, pages 63–71. Springer, 2003

    2003

  18. [26]

    Data-driven discovery of partial differential equations.Science advances, 3(4):e1602614, 2017

    Samuel H Rudy, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Data-driven discovery of partial differential equations.Science advances, 3(4):e1602614, 2017

    2017

  19. [27]

    Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    Peter J Schmid. Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    2010

  20. [28]

    Cambridge university press, 2007

    Jos Thijssen.Computational physics. Cambridge university press, 2007

    2007

  21. [29]

    Weak-form latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 427:116998, 2024

    April Tran, Xiaolong He, Daniel A Messenger, Youngsoo Choi, and David M Bortz. Weak-form latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 427:116998, 2024

    2024

  22. [30]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 6

    2017

Pith tools

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