Feature-preserving Latent-EnKF for Data Assimilation of Flows with Shocks

Hangchuan Hu; Hemanth Chandravamsi; Ponkrshnan Thiagarajan; Tamer A. Zaki

REVIEW 2 major objections 1 minor 26 references

Feature-preserving Latent-EnKF for Data Assimilation of Flows with Shocks

T0 review · 2 major / 1 minor · reviewed 2026-06-27 · grok-4.3

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

Pith's one-line read Performing the ensemble Kalman filter update in a learned latent space preserves sharp shocks and discontinuities in compressible flow data assimilation.

desk verdict The latent-EnKF with shared decoder removes per-member training but rests on an untested claim that shocks form a smooth Gaussian manifold in latent space. read the letter →

arxiv 2606.12559 v1 pith:UF5ENBDE submitted 2026-06-10 physics.comp-ph cs.LGcs.NAmath.NAphysics.flu-dyn

Hemanth Chandravamsi , Hangchuan Hu , Ponkrshnan Thiagarajan , Tamer A. Zaki This is my paper · ORCID

classification physics.comp-phcs.LGcs.NAmath.NAphysics.flu-dyn

keywords dataassimilationensembleKalmanfilterlatentspacecompressibleflowsshocksfeaturepreservationdiscontinuities

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

Standard ensemble Kalman filters produce spurious oscillations when assimilating data into flows with shocks because uncertainty in shock location creates multimodal ensemble statistics that violate Gaussian assumptions. This paper shows that mapping the ensemble to a learned low-dimensional latent space allows the analysis step to occur where features have smooth manifold representations. The updated latent states are then decoded back to the physical domain using a single shared decoder. This approach recovers shocks and contact discontinuities accurately in test cases like the Sod shock tube and Mach 2 cylinder interaction without the oscillations seen in conventional methods. It also removes the need for member-specific training and positivity flooring required in earlier latent approaches.

What carries the argument

A learned low-dimensional latent space in which discontinuities admit smooth manifold representations, enabling the EnKF analysis to be performed there before decoding to physical space with a shared decoder.

What would settle it

If the Sod shock tube experiment produces spurious oscillations or fails to recover sharp shock features in the analysis state, the claim that the latent space representation preserves features would be falsified.

Watch

Extended reading notes

Core claim

The feature-preserving latent-EnKF performs the ensemble update in a learned low-dimensional latent space, where shock and flow features admit a smooth manifold representation, thereby preserving sharp features during EnKF analysis. The updated latent state is mapped back to physical state through a shared decoder for all ensemble members. The algorithm eliminates the member-specific ordered training and positivity flooring used in prior approaches. Numerical experiments on a Sod shock tube and Mach 2 shock interaction with a 2D cylinder, using sparse and noisy observations, show accurate feature recovery of shocks and contact discontinuities without spurious oscillations.

Load-bearing premise

Shock and flow features admit a smooth manifold representation in the learned low-dimensional latent space, allowing the EnKF analysis step to be performed there while preserving sharp features upon decoding.

Editorial extensions

If this is right

Shocks and contact discontinuities are recovered accurately from sparse noisy observations.
No large-scale spurious oscillations appear in the analysis state for the tested flows.
Member-specific ordered training is no longer required for the ensemble.
Positivity flooring constraints are eliminated from the workflow.

Reading between the lines

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

The latent space approach may extend to other discontinuous flow problems such as multiphase or reacting flows.
Assimilation in latent space could lower computational demands for high-resolution simulations.
The manifold smoothness could be verified by checking continuity of latent representations for incrementally shifted shock positions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The main thing here is a latent-space EnKF that runs the analysis step in a learned low-dimensional space and decodes with one shared network for the whole ensemble. This sidesteps the member-specific ordered training and positivity fixes that earlier feature-preserving methods needed.

The approach is straightforward on paper: encode the ensemble, do the usual EnKF update where the features are smoother, then decode. The benchmarks are the usual ones—Sod tube and Mach 2 cylinder—so at least the test cases are standard.

The soft spot is exactly the one the stress-test note flags. Nothing in the abstract shows that shock locations or strengths actually produce a low-dimensional smooth manifold once encoded, or that the latent ensemble stays close enough to Gaussian for the Kalman update to be reliable. If the latent coordinates for shock position stay multimodal, the decoded fields could still oscillate or smear. The paper would need to show at least some diagnostics on latent statistics or reconstruction error across the ensemble to make the central claim convincing.

No numbers, error bars, or direct comparisons appear in the description, which makes it difficult to judge whether the method actually reduces spurious oscillations more than simpler fixes.

This is for people already working on data assimilation for compressible CFD who have run into the multimodal problem with standard EnKF. A reader who wants to try a latent-space route would find the algorithmic description useful, but would still need the full results and checks before adopting it.

It is worth sending to referees. The idea is clear, the problem is real, and the shared-decoder angle is distinct enough from prior work that a proper review can sort out whether the manifold assumption holds in practice.

Referee Report

2 major / 1 minor

Summary. The paper introduces a feature-preserving latent-EnKF for sequential data assimilation in compressible flows containing shocks and discontinuities. The method performs the EnKF analysis step entirely in a learned low-dimensional latent space under the assumption that shock location, strength, and contact discontinuities admit a smooth manifold representation there; the updated latent state is then decoded to physical space via a shared decoder. This is claimed to eliminate member-specific ordered training and positivity flooring required in prior approaches while avoiding spurious oscillations. Numerical experiments are described on the Sod shock tube and a Mach 2 shock-cylinder interaction using sparse noisy observations, with the abstract asserting accurate feature recovery without oscillations.

Significance. If the central claim holds, the approach would provide a practical route to applying EnKF to discontinuous flows without ad-hoc fixes, which is relevant for computational fluid dynamics and data assimilation applications. The algorithmic simplification (shared decoder, no per-member ordering) and testing on standard benchmarks are positive features.

major comments (2)

The central assumption that shock and flow features admit a smooth manifold representation in the learned latent space (allowing Gaussian/EnKF assumptions to hold) is not supported by any analysis of the latent ensemble statistics, manifold dimensionality, or Gaussianity after encoding. This assumption is load-bearing for the claim that the method eliminates spurious oscillations and prior training constraints.
Abstract: the numerical experiments are described only at a high level with no quantitative metrics (e.g., L2 errors, shock-location errors), error bars, baseline comparisons to standard EnKF or prior latent methods, or implementation details (latent dimension, autoencoder architecture, training procedure). This prevents assessment of whether the claimed feature preservation is achieved in practice.

minor comments (1)

The abstract and description would benefit from explicit statements of the latent-space dimension and the form of the observation operator used in the experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: The central assumption that shock and flow features admit a smooth manifold representation in the learned latent space (allowing Gaussian/EnKF assumptions to hold) is not supported by any analysis of the latent ensemble statistics, manifold dimensionality, or Gaussianity after encoding. This assumption is load-bearing for the claim that the method eliminates spurious oscillations and prior training constraints.

Authors: We agree that the manuscript does not provide direct statistical analysis of the latent ensemble (e.g., dimensionality estimates or normality tests). The claim is supported indirectly by the absence of oscillations in the decoded physical-space results on the Sod and shock-cylinder cases. To strengthen the presentation, the revised manuscript will add a new subsection with latent-space diagnostics, including pairwise scatter plots of encoded ensemble members and a brief assessment of local linearity around shock features. revision: yes
Referee: Abstract: the numerical experiments are described only at a high level with no quantitative metrics (e.g., L2 errors, shock-location errors), error bars, baseline comparisons to standard EnKF or prior latent methods, or implementation details (latent dimension, autoencoder architecture, training procedure). This prevents assessment of whether the claimed feature preservation is achieved in practice.

Authors: The abstract follows the conventional high-level format. The full manuscript reports L2 errors, shock-location accuracy, and comparisons in Section 4, with architecture and training details in Section 3. We will revise the abstract to incorporate concise quantitative indicators (e.g., reported L2 errors and latent dimension) while remaining within length limits. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; derivation is self-contained algorithmic contribution

full rationale

The paper introduces a latent-space EnKF algorithm whose central step is the explicit design choice to perform ensemble updates in a learned latent representation (under the stated manifold assumption) followed by shared decoding. No equations, predictions, or first-principles results are shown that reduce by construction to fitted inputs, self-definitions, or self-citation chains. The method is presented and validated on standard benchmark problems (Sod tube, cylinder shock interaction) without renaming known results or smuggling ansatzes via prior self-citations. The reader's assessment of score 2 is consistent with the absence of any load-bearing circular step.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The method rests on the domain assumption that a learned latent representation converts discontinuous physical states into a smooth manifold suitable for Gaussian EnKF updates.

assumptions (1)

domain assumption Shock and flow features admit a smooth manifold representation in the learned low-dimensional latent space
This premise is invoked to justify performing the ensemble update in latent space while preserving sharp features upon decoding.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Feature-preserving Latent-EnKF for Data Assimilation of Flows with Shocks." pith.science (2026). https://pith.science/paper/UF5ENBDE

@misc{pith2026260612559,
  author       = {Pith},
  title        = {Pith review of: Feature-preserving Latent-EnKF for Data Assimilation of Flows with Shocks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UF5ENBDE}},
  note         = {Machine review of arXiv:2606.12559}
}

read the original abstract

The ensemble Kalman filter (EnKF) is widely adopted for sequential data assimilation, but fails for solutions with discontinuities, such as shocks in compressible flows. Uncertainty in shock location induces multimodal ensemble statistics that violate the Gaussian assumptions underlying the EnKF, producing large-scale spurious oscillations in the analysis state. We introduce a feature-preserving latent-EnKF that performs the ensemble update in a learned low-dimensional latent space, where shock and flow features admit a smooth manifold representation, thereby preserving sharp features during EnKF analysis. The updated latent state is mapped back to physical state through a shared decoder for all ensemble members. The algorithm eliminates the member-specific ordered training and positivity flooring used in prior approaches. Numerical experiments on a Sod shock tube and Mach 2 shock interaction with a 2D cylinder, using sparse and noisy observations, show accurate feature recovery of shocks and contact discontinuities without spurious oscillations.

Figures

Figures reproduced from arXiv: 2606.12559 by the authors.

**Figure 2.** Sod shock tube at the first DA cycle, t = 0.025, with p, ρ, u in the three rows. Left: forecast ensemble. Middle: analysis from the standard EnKF in physical space. Right: analysis from the feature-space Latent-EnKF. The physical-space update produces oscillations near the shock, contact, and rarefaction, whereas the feature-space update contracts the ensemble onto the truth while preserving the sharp features. ense… view at source ↗

**Figure 3.** Space-time evolution of the Sod shock tube under Latent-EnKF assimilation. Rows correspond to p, ρ, u. Left: analytical solution. Middle: forecast trajectory initialized from the decoded post-analysis latent mean Dθ(z¯ a,k, x) at each DA cycle and advanced between observation times using the numerical solver. Right: absolute error relative to the truth 1 3 6 8 No. of DA cycles 0.00 0.05 0.10 0.15 0.20 RMSE / Spread … view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Per-cycle RMSE (red squares) and ensemble spread (blue circles) for [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Shock-cylinder interaction under feature-space Latent-EnKF. Columns show the initial state, forecast, and analyses over successive DA cycles. Row 1: truth, with density ρ (top half) and instantaneous |∇ρ| (bottom half). Row 2: observations dk + ηi, given by noisy, time…

**Figure 6.** Figure 6: time (t = 2.5) assimilated flow field for the cylinder case. Top row: decoded ensemble-mean primitive fields p, ρ, u, and v. Bottom row: absolute error relative to the truth, normalized by the maximum of each corresponding field, with a common color scale for direct co…

**Figure 7.** Figure 7: RMSE and ensemble spread across DA cycles for the cylinder case with a 10% biased prior. Results are shown for [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 3 canonical work pages

[1]

Evensen, Data assimilation: the ensemble Kalman filter, Springer, 2009

G. Evensen, Data assimilation: the ensemble Kalman filter, Springer, 2009

2009
[2]

T. A. Zaki, M. Wang, Data assimilation and flow estimation, in: Data Driven Analysis and Modeling of Turbulent Flows, Elsevier, 2025, pp. 129–181

2025
[3]

T. A. Zaki, Turbulence from an observer perspective, Annual Review of Fluid Mechanics 57 (Volume 57,
[4]

M. Asch, M. Bocquet, M. Nodet, Data Assimilation: Methods, Algorithms, and Applications, SIAM, 2016

2016
[5]

Houba, A

T. Houba, A. Edoh, R. Munipalli, M. E. Harvazinski, Sequential data assimilation in flows with shocks, in: AIAA SCITECH 2024 Forum, 2024, paper 0587

2024
[6]

Morzfeld, D

M. Morzfeld, D. Hodyss, Gaussian approximations in filters and smoothers for data assimilation, Tellus A: Dynamic Meteorology and Oceanography 71 (2019) 1600344

2019
[7]

X.-H. Zhou, L. Beronilla, M. K. Sleeman, H. Hu, M. Morzfeld, A. M. Stuart, T. A. Zaki, Neural ensemble kalman filter: Data assimilation for compressible flows with shocks, arXiv preprint arXiv:2602.23461 (2026)

work page Pith review arXiv 2026
[8]

A. Edoh, T. Houba, R. Munipalli, M. E. Harvazinski, Sequential ensemble Kalman filtering of compressible flows with shocks: enforcing positivity, in: AIAA SCITECH 2025 Forum, 2025, paper 0918

2025

Show all 26 references

[9]

J. J. Hansen, D. Brouzet, M. Ihme, A normal-score ensemble Kalman filter for 1D shock waves, in: AIAA SCITECH 2024 Forum, 2024, paper 1022

2024
[10]

H. Zhou, J. J. Gomez-Hernandez, H.-J. H. Franssen, L. Li, An approach to handling non-Gaussianity of parameters and state variables in ensemble Kalman filtering, Advances in Water Resources 34 (2011) 844–864

2011
[11]

T. Li, A. Gelb, Y. Lee, A structurally informed data assimilation approach for nonlinear partial differential equations, Journal of Computational Physics 519 (2024) 113396

2024
[12]

T. Li, A. Gelb, Y. Lee, Structurally informed data assimilation in two dimensions, arXiv preprint arXiv:2510.06369 (2025)

2025
[13]

H. Fan, L. Bai, B. Fei, Y. Xiao, K. Chen, Y. Liu, Y. Qu, F. Ling, P. Gentine, Physically consistent global atmospheric data assimilation with machine learning in latent space, Science Advances 12 (1) (2026) eaea4248

2026
[14]

Y. Chen, D. Sanz-Alonso, R. Willett, Reduced-order autodifferentiable ensemble kalman filters, Inverse Problems 39 (12) (2023) 124001

2023
[15]

X. T. Tong, Y. Wang, L. Yan, Latent autoencoder ensemble Kalman filter for nonlinear data assimilation, arXiv preprint arXiv:2603.06752 (2026)

2026 arXiv
[16]

Z. Chen, H. Zhang, Learning implicit fields for generative shape modeling, in: Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5939–5948

2019
[17]

H. Fan, Y. Liu, Z. Huo, Y. Liu, Y. Shi, Y. Li, A novel latent space data assimilation framework with autoencoder-observation to latent space (ae-o2l) network. part i: The observation-only analysis method, Monthly Weather Review 153 (8) (2025) 1335–1348

2025
[18]

Pasmans, Y

I. Pasmans, Y. Chen, T. Sebastian Finn, M. Bocquet, A. Carrassi, Ensemble kalman filter in latent space using a variational autoencoder pair, Quarterly Journal of the Royal Meteorological Society 152 (775) (2026) e70070. 7

2026
[19]

J. J. Park, P. Florence, J. Straub, R. Newcombe, S. Lovegrove, DeepSDF: learning continuous signed distance functions for shape representation, in: Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 165–174

2019
[20]

P. D. Lax, X.-D. Liu, Solution of two-dimensional riemann problems of gas dynamics by positive schemes, SIAM Journal on Scientific Computing 19 (2) (1998) 319–340

1998
[21]

Jiang, C.-W

G.-S. Jiang, C.-W. Shu, Efficient implementation of weighted eno schemes, Journal of computational physics 126 (1) (1996) 202–228

1996
[22]

E. F. Toro, M. Spruce, W. Speares, Restoration of the contact surface in the hll-riemann solver, Shock waves 4 (1) (1994) 25–34

1994
[23]

Chandravamsi, A

H. Chandravamsi, A. S. Chamarthi, N. Hoffmann, S. H. Frankel, On the application of gradient based reconstruction for flow simulations on generalized curvilinear and dynamic mesh domains, Computers & Fluids 258 (2023) 105859

2023
[24]

Gottlieb, C.-W

S. Gottlieb, C.-W. Shu, E. Tadmor, Strong stability-preserving high-order time discretization methods, SIAM review 43 (1) (2001) 89–112

2001
[25]

G. Ke, S. J. Grauer, T. A. Zaki, Observability of initial states in one-dimensional inviscid flows with shocks, in: AIAA SCITECH 2026 Forum, 2026, paper 0493

2026
[26]

Chaudhuri, A

A. Chaudhuri, A. Hadjadj, A. Chinnayya, On the use of immersed boundary methods for shock/obstacle interactions, Journal of Computational Physics 230 (5) (2011) 1731–1748. 8

2011

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.

[1] [1]

Evensen, Data assimilation: the ensemble Kalman filter, Springer, 2009

G. Evensen, Data assimilation: the ensemble Kalman filter, Springer, 2009

2009

[2] [2]

T. A. Zaki, M. Wang, Data assimilation and flow estimation, in: Data Driven Analysis and Modeling of Turbulent Flows, Elsevier, 2025, pp. 129–181

2025

[3] [3]

T. A. Zaki, Turbulence from an observer perspective, Annual Review of Fluid Mechanics 57 (Volume 57,

[4] [4]

M. Asch, M. Bocquet, M. Nodet, Data Assimilation: Methods, Algorithms, and Applications, SIAM, 2016

2016

[5] [5]

Houba, A

T. Houba, A. Edoh, R. Munipalli, M. E. Harvazinski, Sequential data assimilation in flows with shocks, in: AIAA SCITECH 2024 Forum, 2024, paper 0587

2024

[6] [6]

Morzfeld, D

M. Morzfeld, D. Hodyss, Gaussian approximations in filters and smoothers for data assimilation, Tellus A: Dynamic Meteorology and Oceanography 71 (2019) 1600344

2019

[7] [7]

X.-H. Zhou, L. Beronilla, M. K. Sleeman, H. Hu, M. Morzfeld, A. M. Stuart, T. A. Zaki, Neural ensemble kalman filter: Data assimilation for compressible flows with shocks, arXiv preprint arXiv:2602.23461 (2026)

work page Pith review arXiv 2026

[8] [8]

A. Edoh, T. Houba, R. Munipalli, M. E. Harvazinski, Sequential ensemble Kalman filtering of compressible flows with shocks: enforcing positivity, in: AIAA SCITECH 2025 Forum, 2025, paper 0918

2025