pith. sign in

arxiv: 2605.19076 · v1 · pith:W62X5XUJnew · submitted 2026-05-18 · 💻 cs.LG · physics.flu-dyn

The impact of observation density on Bayesian inversion of latent dynamics in shock-dominated flows

Bipin Tiwari , Muhammad Abid , Omer San This is my paper

Pith reviewed 2026-05-20 11:34 UTC · model grok-4.3

classification 💻 cs.LG physics.flu-dyn
keywords Bayesian inversionreduced-order modelingshock-dominated flowsconvolutional autoencoderuncertainty quantificationSod shock tubelatent dynamicsNo-U-Turn Sampler
0
0 comments X

The pith

Increasing the density of observations in Bayesian inversion of shock-tube initial states contracts posterior uncertainty by roughly three-quarters.

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

The paper builds a reduced-order surrogate for efficient Bayesian recovery of unknown initial conditions in compressible flows with shocks. It trains a convolutional autoencoder to map high-dimensional density and pressure fields into a 32-dimensional latent space and pairs it with a learned operator that advances those latent states forward in time. This surrogate is inserted into a No-U-Turn Sampler so that posterior distributions over the four unknown initial scalars can be explored with thousands of cheap evaluations. The central result is that raising the number of noisy final-time observation points produces markedly narrower posteriors, cutting the average standard deviation by about 78 percent for density and 76 percent for pressure.

Core claim

The authors show that, within their autoencoder-based latent dynamical model of the Sod shock tube, denser sets of final-time observations drive a large contraction in the posterior uncertainty of the recovered left and right density and pressure values. The framework recovers the rarefaction wave, contact discontinuity, and shock front with good fidelity once the latent dimension reaches 32 and roughly 250 training trajectories are available, and the NUTS sampler then yields well-calibrated uncertainty estimates whose spread shrinks sharply as more observation locations are added.

What carries the argument

A convolutional autoencoder that compresses flow snapshots into a 32-dimensional nonlinear latent representation, combined with a learned latent-space forward operator that predicts the final-time latent state from an encoded initial state.

If this is right

  • The surrogate enables thousands of forward evaluations inside the sampler at a cost that full-order models cannot sustain.
  • A latent dimension of 32 supplies enough capacity to reconstruct the main shock-tube features while keeping the inverse problem tractable.
  • Two hundred fifty high-fidelity trajectories are already sufficient to train an accurate reconstruction.
  • The same workflow can be applied to other initial-state recovery tasks once the autoencoder is retrained on the target flow family.

Where Pith is reading between the lines

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

  • Sensor networks for compressible-flow experiments could be designed by placing new probes where they most reduce the remaining posterior variance.
  • The same latent-dynamics approach may transfer to inverse problems in other nonlinear wave systems such as detonations or supersonic inlets.
  • If the latent dimension is increased beyond 32, the extra cost of sampling must be weighed against any further uncertainty reduction.

Load-bearing premise

The autoencoder and its learned latent forward operator must reproduce the essential nonlinear wave interactions of the shock tube well enough that the Bayesian posteriors remain trustworthy.

What would settle it

Re-running the same inversion with the original high-fidelity WENO solver in place of the autoencoder surrogate and finding that the resulting posterior standard deviations differ by more than a few percent would indicate the surrogate is not faithful enough.

Figures

Figures reproduced from arXiv: 2605.19076 by Bipin Tiwari, Muhammad Abid, Omer San.

Figure 1
Figure 1. Figure 1: Mach 1.44 flow field around a Genesis sample return capsule model, showing the major shock-dominated [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of a blunt body exposed to freestream disturbances in supersonic flow. Incoming disturbances [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Conceptual diagram of an inverse problem framework. The simulator (forward model) is queried repeatedly [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the Sod shock tube problem after the diaphragm is removed, showing the resulting rarefaction [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ensemble of initial conditions generated by Latin hypercube sampling for the Sod shock tube simulations. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Let X denote a high-dimensional flow state of size (3, 1000). The encoder, denoted by Φ, maps this state to a compact latent representation defined as z = Φ(X), z ∈ R Nz where the optimal latent dimension was found to be Nz = 32 which will be discussed in result section. The encoder applies successive one-dimensional convolutional layers with stride two, progressively reducing the spatial resolution while … view at source ↗
Figure 6
Figure 6. Figure 6: Data preprocessing and one-dimensional convolutional autoencoder architecture used for reduced-order [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training workflow for the AE-ROM forward operator. The pretrained encoder is frozen and used to map [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Candidate initial-state parameters are drawn from the prior distribution, propagated through the AE-ROM [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Schematic of the AE-ROM-enabled Bayesian inversion framework. Candidate initial-state parameters are [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of latent dimension on AE-ROM reconstruction accuracy. (a) Total validation MSE decreases rapidly [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of AE-ROM reconstructions for latent dimensions 4 and 32. (Top Row) [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Data efficiency analysis for the AE-ROM with fixed latent dimension of [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Training convergence of the latent-space forward operator. The MSE decreases steadily over 200 epochs, [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Forward prediction performance of the AE-ROM on a held-out test case at [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Synthetic observation configurations used for Bayesian inversion. The first row shows density measurements [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Posterior probability density functions of the inferred initial density and pressure parameters for increasing [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Effect of observation density on posterior predictive accuracy and uncertainty. (a) RMSE between the [PITH_FULL_IMAGE:figures/full_fig_p018_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Spatial posterior predictive reconstructions of the inferred initial density (top row) and pressure (bottom [PITH_FULL_IMAGE:figures/full_fig_p019_17.png] view at source ↗
read the original abstract

Inferring unknown initial states in shock-dominated compressible flows from sparse and noisy measurements is a challenging ill-posed inverse problem due to nonlinear wave interactions and limited sensing. In this work, we develop a non-intrusive reduced-order modeling framework for efficient Bayesian initial-state inversion with uncertainty quantification. The framework combines a convolutional autoencoder with a learned latent-space forward operator. The autoencoder compresses high-dimensional flow fields into a compact nonlinear latent representation, while the forward operator predicts final-time latent states from encoded initial conditions. This AE-ROM surrogate enables rapid forward evaluations and is embedded within a No-U-Turn Sampler (NUTS) for posterior exploration. The framework is demonstrated using 500 high-fidelity Sod shock tube simulations generated through Latin hypercube sampling and solved using a fifth-order WENO scheme. The inverse problem seeks to recover unknown left and right density and pressure states from sparse noisy observations of final-time density and pressure fields. Results show that the AE-ROM accurately reconstructs key shock-tube structures, including the rarefaction wave, contact discontinuity, and shock front. A latent dimension of 32 provides an effective balance between reconstruction accuracy and reduced-space compactness, while 250 training simulations are sufficient for accurate reconstruction. Increasing observation density significantly contracts posterior uncertainty, reducing the mean posterior standard deviation by approximately 78% for density and 76% for pressure. Overall, the proposed framework provides a computationally efficient and uncertainty-aware approach for inverse analysis of shock-dominated flows, with potential extensions to multidimensional compressible-flow and digital-twin applications.

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 develops a non-intrusive reduced-order modeling framework that combines a convolutional autoencoder (latent dimension 32) with a learned latent-space forward operator, trained on 250 high-fidelity Sod shock-tube simulations. This surrogate is embedded inside NUTS sampling to perform Bayesian inversion of unknown left/right density and pressure initial states from sparse noisy final-time observations. The central empirical result is that increasing observation density contracts posterior uncertainty, reducing mean posterior standard deviation by ~78% for density and ~76% for pressure while accurately reconstructing the rarefaction, contact, and shock structures.

Significance. If the surrogate operator error remains small relative to observation noise and does not distort shock sensitivity, the work supplies a computationally tractable route to uncertainty-aware inversion for nonlinear compressible flows. The explicit quantification of uncertainty contraction with observation density, together with the use of 500 Latin-hypercube simulations and a fifth-order WENO solver, offers a concrete, reproducible demonstration that could support extensions to multidimensional problems and digital-twin settings.

major comments (2)
  1. [Abstract / Results] Abstract and results section: the reported 78% and 76% reductions in mean posterior standard deviation are obtained by embedding the learned latent forward operator inside NUTS; however, no test-set prediction error, shock-speed error, or discontinuity-capturing metric for this operator is supplied. Without such quantification it is impossible to rule out that the apparent contraction is inflated by surrogate smoothing or bias that grows with observation density.
  2. [Method / Results] Method and results: the framework fixes the autoencoder and latent operator after training and then performs standard NUTS; yet no comparison of posteriors obtained with the surrogate versus a modest number of high-fidelity evaluations at posterior samples is presented, leaving open whether operator error propagates into the reported uncertainty contraction.
minor comments (2)
  1. [Results] The manuscript does not state whether cross-validation or hold-out error propagation through the surrogate was performed when reporting reconstruction accuracy.
  2. [Method] Notation for the latent forward operator and the precise definition of 'observation density' should be clarified with an equation or table entry.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below, indicating where revisions have been made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Results] Abstract and results section: the reported 78% and 76% reductions in mean posterior standard deviation are obtained by embedding the learned latent forward operator inside NUTS; however, no test-set prediction error, shock-speed error, or discontinuity-capturing metric for this operator is supplied. Without such quantification it is impossible to rule out that the apparent contraction is inflated by surrogate smoothing or bias that grows with observation density.

    Authors: We agree that explicit quantification of the surrogate operator accuracy is required to support the reported posterior contraction. In the revised manuscript we have added a dedicated subsection on surrogate validation that reports: (i) mean-squared prediction error on a 250-simulation test set for both the autoencoder reconstruction and the latent forward operator, (ii) shock-speed error measured as the L1 deviation of the captured shock location relative to the high-fidelity WENO solution, and (iii) a discontinuity-capturing metric given by the integrated squared gradient error across the rarefaction, contact, and shock fronts. These metrics confirm that the operator error remains at least an order of magnitude smaller than the observation noise level and does not exhibit systematic bias that increases with observation density. Consequently the observed uncertainty contraction is not an artifact of surrogate smoothing. revision: yes

  2. Referee: [Method / Results] Method and results: the framework fixes the autoencoder and latent operator after training and then performs standard NUTS; yet no comparison of posteriors obtained with the surrogate versus a modest number of high-fidelity evaluations at posterior samples is presented, leaving open whether operator error propagates into the reported uncertainty contraction.

    Authors: We acknowledge that a direct head-to-head comparison would further strengthen confidence in the results. Because full high-fidelity WENO evaluations at every NUTS sample would be prohibitive, we have performed a targeted validation: ten representative samples drawn from the posterior (spanning the 5th to 95th percentiles of the marginals) were re-evaluated with the original fifth-order WENO solver. The surrogate-based posterior means and standard deviations agree with the high-fidelity results to within the observation noise for both density and pressure fields. These comparisons are now reported in a new figure and accompanying table in the revised Results section, demonstrating that operator error does not materially inflate or deflate the reported uncertainty contraction. revision: yes

Circularity Check

0 steps flagged

No circularity: uncertainty contraction is an empirical outcome from standard sampling with fixed surrogate

full rationale

The paper trains a convolutional autoencoder (latent dim 32) and latent-space forward operator on 250-500 independent high-fidelity WENO Sod shock-tube simulations generated via Latin hypercube sampling, then fixes this surrogate and embeds it inside a standard NUTS sampler to run Bayesian inversions for varying observation densities. The central quantitative claim (mean posterior std dev reduced ~78% for density and ~76% for pressure) is obtained by comparing the resulting posterior samples across those densities; this comparison does not reduce by construction to the training loss, the latent dimension choice, or any self-citation. No self-definitional step, fitted-input-renamed-as-prediction, or load-bearing self-citation appears in the derivation chain. The framework remains self-contained against external benchmarks because the training data and the observation-density experiments are generated separately from the high-fidelity solver.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on standard assumptions from reduced-order modeling and Bayesian inference; the main free parameters are empirically selected hyperparameters whose values are justified by reconstruction performance rather than derived from first principles.

free parameters (2)
  • latent dimension = 32
    Selected as 32 to balance reconstruction accuracy and compactness; directly affects the surrogate quality used in inversion.
  • number of training simulations = 250
    250 simulations stated as sufficient for accurate reconstruction; chosen after testing on the 500-simulation dataset.
axioms (2)
  • domain assumption The autoencoder preserves key nonlinear structures (rarefaction, contact, shock) in the latent representation.
    Invoked when claiming the surrogate enables reliable inversion of initial states.
  • domain assumption The learned latent-space forward operator faithfully approximates the high-fidelity WENO dynamics for the sampled initial conditions.
    Required for the rapid forward evaluations inside the NUTS sampler to produce valid posteriors.

pith-pipeline@v0.9.0 · 5811 in / 1695 out tokens · 56611 ms · 2026-05-20T11:34:05.689121+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.

    The framework combines a convolutional autoencoder with a learned latent-space forward operator... latent dimension of 32... forward operator predicts final-time latent states from encoded initial conditions... embedded within a No-U-Turn Sampler (NUTS)

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

47 extracted references · 47 canonical work pages · 2 internal anchors

  1. [1]

    J. D. Anderson, Jr., Hypersonic and High-Temperature Gas Dynamics, AIAA Education Series, AIAA, 2019. doi:10.2514/4.105142

    work page doi:10.2514/4.105142 2019

  2. [2]

    J. J. Bertin, R. M. Cummings, Aerodynamics for Engineers, Pearson, 2021.doi:10.1017/9781009105842

    work page doi:10.1017/9781009105842 2021

  3. [3]

    W. H. Heiser, D. T. Pratt, Hypersonic Airbreathing Propulsion, AIAA Education Series, AIAA, 2018

    work page 2018

  4. [4]

    S. Han, B. J. Lee, M. Ahn Furudate, H. Nagai, Numerical investigation of the dynamic instability of a reentry capsule in transonic flow, AIAA Journal 62 (2) (2024) 449–459.doi:10.2514/1.J063428

    work page doi:10.2514/1.j063428 2024

  5. [5]

    E. F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics, Springer, 2009. doi:10.1007/ 978-3-540-49834-6

    work page 2009

  6. [6]

    R. J. LeVeque, Finite V olume Methods for Hyperbolic Problems, Cambridge University Press, 2012. doi: 10.1017/CBO9780511791253

    work page doi:10.1017/cbo9780511791253 2012

  7. [7]

    Fedorov, Transition and stability of high-speed boundary layers, Annual Review of Fluid Mechanics 43 (1) (2011) 79–95.doi:10.1146/annurev-fluid-122109-160750

    A. Fedorov, Transition and stability of high-speed boundary layers, Annual Review of Fluid Mechanics 43 (1) (2011) 79–95.doi:10.1146/annurev-fluid-122109-160750

    work page doi:10.1146/annurev-fluid-122109-160750 2011

  8. [8]

    Z. Wang, M. Shi, X. Tang, H. Lv, L. Xu, Effect of a roughness element on the receptivity of a hypersonic boundary layer over a blunt cone due to pulse entropy disturbance with a single frequency, Entropy 20 (6) (2018) 404. doi:10.3390/e20060404

    work page doi:10.3390/e20060404 2018

  9. [9]

    D. S. Dolling, Fifty years of shock-wave/boundary-layer interaction research: What next?, AIAA Journal 39 (8) (2001) 1517–1531.doi:10.2514/2.1476

    work page doi:10.2514/2.1476 2001

  10. [10]

    P. A. Jumars, J. H. Trowbridge, E. Boss, L. Karp-Boss, Turbulence-plankton interactions: A new cartoon (2009). doi:10.1111/j.1439-0485.2009.00288.x

    work page doi:10.1111/j.1439-0485.2009.00288.x 2009

  11. [11]

    Tarantola, Inverse Problem Theory and Methods for Model Parameter Estimation, SIAM, 2005

    A. Tarantola, Inverse Problem Theory and Methods for Model Parameter Estimation, SIAM, 2005. doi:10. 1137/1.9780898717921

    work page 2005

  12. [12]

    J. P. Kaipio, E. Somersalo, Statistical and Computational Inverse Problems, Springer, 2005. doi:10.1007/ b138659

    work page 2005

  13. [13]

    R. C. Aster, B. Borchers, C. H. Thurber, Parameter Estimation and Inverse Problems, Elsevier, 2018. doi: 10.1016/C2015-0-02458-3

    work page doi:10.1016/c2015-0-02458-3 2018

  14. [14]

    Anhichem, S

    M. Anhichem, S. Timme, J. Castagna, A. J. Peace, M. Maina, Bayesian machine learning for predicting wing pressure distributions at transonic flow conditions, AIAA Journal 63 (10) (2025) 4126–4142. doi:10.2514/1. J064617

    work page doi:10.2514/1 2025

  15. [15]

    S. Pan, K. Duraisamy, Data-driven discovery of closure models, SIAM Journal on Applied Dynamical Systems 17 (4) (2018) 2381–2413.doi:10.1137/18M1177263

    work page doi:10.1137/18m1177263 2018

  16. [16]

    Reich, C

    S. Reich, C. Cotter, Probabilistic Forecasting and Bayesian Data Assimilation, Cambridge University Press, 2015. doi:10.1017/CBO9781107706804

    work page doi:10.1017/cbo9781107706804 2015

  17. [18]

    Wogrin, A

    S. Wogrin, A. Singh, D. Allaire, O. Ghattas, K. Willcox, From data to decisions: A real-time measurement– inversion–prediction–steering framework for hazardous events and health monitoring, in: Handbook of Dynamic Data Driven Applications Systems: V olume 2, Springer, 2023, pp. 195–227. doi:10.1007/ 978-3-031-27986-7_8. 20 APREPRINT

    work page 2023

  18. [19]

    M. G. Kapteyn, J. V . R. Pretorius, K. E. Willcox, A probabilistic graphical model foundation for enabling predictive digital twins at scale, Nature Computational Science 1 (5) (2021) 337–347. doi:10.1038/ s43588-021-00069-0

    work page 2021

  19. [20]

    Rasheed, O

    A. Rasheed, O. San, T. Kvamsdal, Digital twin: Values, challenges and enablers from a modeling perspective, IEEE Access 8 (2020) 21980–22012.doi:10.48550/arXiv.1910.01719

    work page doi:10.48550/arxiv.1910.01719 2020

  20. [21]

    Shu, Essentially non-oscillatory and weighted essentially non-oscillatory schemes, Acta Numerica 29 (2020) 701–762.doi:10.1017/S0962492920000057

    C.-W. Shu, Essentially non-oscillatory and weighted essentially non-oscillatory schemes, Acta Numerica 29 (2020) 701–762.doi:10.1017/S0962492920000057

    work page doi:10.1017/s0962492920000057 2020

  21. [22]

    Pirozzoli, Numerical methods for high-speed flows, Annual Review of Fluid Mechanics 43 (1) (2011) 163–194

    S. Pirozzoli, Numerical methods for high-speed flows, Annual Review of Fluid Mechanics 43 (1) (2011) 163–194. doi:10.1146/annurev-fluid-122109-160718

    work page doi:10.1146/annurev-fluid-122109-160718 2011

  22. [23]

    H. Le, P. Moin, J. Kim, Direct numerical simulation of turbulent flow over a backward-facing step, Journal of Fluid Mechanics 330 (1997) 349–374.doi:10.1017/S0022112096003941

    work page doi:10.1017/s0022112096003941 1997

  23. [24]

    R. C. Smith, Uncertainty Quantification: Theory, Implementation, and Applications, SIAM, 2024. doi:10.1137/ 1.9781611973228

    work page 2024

  24. [25]

    Benner, S

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

    work page doi:10.1137/130932715 2015

  25. [26]

    C. W. Rowley, S. T. M. Dawson, Model reduction for flow analysis and control, Annual Review of Fluid Mechanics 49 (2017) 387–417.doi:10.1146/annurev-fluid-010816-060042

    work page doi:10.1146/annurev-fluid-010816-060042 2017

  26. [27]

    Quarteroni, A

    A. Quarteroni, A. Manzoni, F. Negri, Reduced Basis Methods for Partial Differential Equations: An Introduction, Springer, 2015.doi:10.1007/978-3-319-15431-2

    work page doi:10.1007/978-3-319-15431-2 2015

  27. [28]

    Tang, W.-Z

    Y .-Q. Tang, W.-Z. Fang, C.-Y . Zheng, W.-Q. Tao, Applications of POD-Based reduced order model to the rapid prediction of velocity and temperature in data centers, Applied Thermal Engineering 263 (2025) 125310. doi:10.1016/j.applthermaleng.2024.125310

    work page doi:10.1016/j.applthermaleng.2024.125310 2025

  28. [29]

    C. Dai, D. Yang, C. Zhang, P. Ding, C. Duan, J. Song, A joint parameter-state estimation framework using POD- Galerkin projected reduced order model and 3dvar data assimilation, Computers & Fluids (2025) 106815doi: 10.1016/j.compfluid.2025.106815

    work page doi:10.1016/j.compfluid.2025.106815 2025

  29. [30]

    Carlberg, M

    K. Carlberg, M. Barone, H. Antil, Galerkin v. least-squares petrov–galerkin projection in nonlinear model reduction, Journal of Computational Physics 330 (2017) 693–734.doi:10.1016/j.jcp.2016.10.033

    work page doi:10.1016/j.jcp.2016.10.033 2017

  30. [31]

    Ohlberger, S

    M. Ohlberger, S. Rave, Reduced basis methods: Success, limitations and future challenges, arXiv preprint arXiv:1511.02021 (2015).doi:10.48550/arXiv.1511.02021

    work page doi:10.48550/arxiv.1511.02021 2015

  31. [32]

    S. L. Brunton, B. R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annual Review of Fluid Mechanics 52 (1) (2020) 477–508.doi:10.1146/annurev-fluid-010719-060214

    work page doi:10.1146/annurev-fluid-010719-060214 2020

  32. [33]

    Duraisamy, G

    K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data, Annual Review of Fluid Mechanics 51 (1) (2019) 357–377.doi:10.1146/annurev-fluid-010518-040547

    work page doi:10.1146/annurev-fluid-010518-040547 2019

  33. [34]

    M. Abid, O. San, Simr-no: A spectrally-informed multi-resolution neural operator for turbulent flow super- resolution, arXiv preprint arXiv:2603.28073 (2026).doi:10.48550/arXiv.2603.28073

    work page doi:10.48550/arxiv.2603.28073 2026

  34. [35]

    M. Abid, O. San, Spectral embedding via chebyshev bases for robust DeepONet approximation, arXiv preprint arXiv:2512.09165 (2025).doi:10.48550/arXiv.2512.09165

    work page doi:10.48550/arxiv.2512.09165 2025

  35. [36]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, A. Courville, Deep Learning, MIT Press, 2016,http://www.deeplearningbook. org

    work page 2016

  36. [37]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444. doi:10.1038/ nature14539

    work page 2015

  37. [38]

    Loyal, B

    R. Maulik, B. Lusch, P. Balaprakash, Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders, Physics of Fluids 33 (3) (2021) 037106. doi:10.1063/5. 0039986

    work page doi:10.1063/5 2021

  38. [39]

    A. T. Mohan, D. V . Gaitonde, A deep learning based approach to reduced order modeling for turbulent flow control using lstm neural networks, arXiv preprint arXiv:1804.09269 (2018).doi:10.48550/arXiv.1804.09269

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1804.09269 2018

  39. [40]

    T. J. Sullivan, Introduction to uncertainty quantification (2015).doi:10.1007/978-3-319-23395-6

    work page doi:10.1007/978-3-319-23395-6 2015

  40. [41]

    Gelman, J

    A. Gelman, J. B. Carlin, H. S. Stern, D. B. Dunson, A. Vehtari, D. B. Rubin, Bayesian Data Analysis, 3rd Edition, Chapman and Hall/CRC, 2013.doi:10.1201/b16018

    work page doi:10.1201/b16018 2013

  41. [42]

    D. S. Sivia, J. Skilling, Data Analysis: A Bayesian Tutorial, Oxford University Press, 2006. doi:10.1093/oso/ 9780198568315.001.0001. 21 APREPRINT

    work page doi:10.1093/oso/ 2006

  42. [43]

    Brooks, A

    S. Brooks, A. Gelman, G. L. Jones, X.-L. Meng (Eds.), Handbook of Markov Chain Monte Carlo, CRC Press, 2011.doi:10.1201/b10905

    work page doi:10.1201/b10905 2011

  43. [44]

    M. D. Hoffman, A. Gelman, The No-U-Turn sampler: Adaptively setting path lengths in Hamiltonian Monte Carlo, Journal of Machine Learning Research 15 (47) (2014) 1593–1623, http://jmlr.org/papers/v15/ hoffman14a.html

    work page 2014

  44. [45]

    A Conceptual Introduction to Hamiltonian Monte Carlo

    M. Betancourt, A conceptual introduction to Hamiltonian Monte Carlo, arXiv preprint arXiv:1701.02434 (2017). doi:10.48550/arXiv.1701.02434

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1701.02434 2017

  45. [46]

    T. Cui, Y . M. Marzouk, K. Willcox, Data-driven model reduction for the Bayesian solution of inverse problems, International Journal for Numerical Methods in Engineering 105 (5) (2016) 329–353. doi:10.1002/nme.4748

    work page doi:10.1002/nme.4748 2016

  46. [47]

    Ghadami, B

    A. Ghadami, B. I. Epureanu, Data-driven prediction in dynamical systems: Recent developments, Philosophical Transactions of the Royal Society A 380 (2229) (2022) 20210213.doi:10.1098/rsta.2021.0213

    work page doi:10.1098/rsta.2021.0213 2022

  47. [48]

    Kassem, S

    A. Kassem, S. A. S. Romeo, B. Tiwari, L. Musharrat, O. San, K. Kara, Discovery of governing equations for reduced-order aerodynamic modes of atmospheric entry capsules, Aerospace Science and Technology 178 (2026) 112454.doi:10.1016/j.ast.2026.112454. 22

    work page doi:10.1016/j.ast.2026.112454 2026