Pith. sign in

REVIEW 4 major objections 5 minor 18 references

Divide and conquer complex flows. Part I: cluster and manifold-based local analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A fully automated method partitions complex flows into regions that share the same local dynamics, then builds a simple stochastic model for each region.

desk verdict Novel, promising pipeline for dynamics-based spatial partitioning, but the headline claim is supported only by selective, in-sample evidence; worth referee time, not yet citable. read the letter →

arxiv 2601.05117 v2 pith:HSUAS32A submitted 2026-01-08 physics.flu-dyn

Qihong L. Li-Hu , Guy Y. Cornejo Maceda , Andrea Ianiro , Stefano Discetti This is my paper · ORCID
classification physics.flu-dyn
keywords clusteringmanifoldlearningISOMAPcluster-basednetworkmodeldomaindecompositionreduced-ordermodelingfluidicpinballtransitionaljetflow
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 proposes a completely automated, data-driven way to divide a complex flow into spatial regions that share the same local dynamics, and to build a simple stochastic model for each region. The central idea is to treat each grid point by its vorticity time history, embed all points in a low-dimensional 'spatial manifold' using ISOMAP, and then cluster the points in that manifold so that proximity means dynamical similarity. Each resulting subdomain is modelled separately with a cluster-based network model. On a simulated fluidic pinball with two incommensurate forcing frequencies and on an experimental transitional jet, the method recovers local behaviours—cylinder-surface forcing dynamics, global vortex shedding, jet vortex pairing—that a global model blends together. If the claim holds, the framework offers an automated route from raw snapshots to interpretable, local flow descriptions useful for control and modelling.

What carries the argument

The load-bearing object is the spatial manifold: the low-dimensional ISOMAP embedding of all grid points, each represented by its full vorticity time series. ISOMAP replaces raw Euclidean distances with geodesic distances along a k-nearest-neighbour graph before multidimensional scaling, so two points end up close in manifold coordinates only if their dynamics are connected by a chain of similar histories. Clustering in this manifold—not in physical space—produces subdomains that may be disconnected in space, and each subdomain then gets its own cluster-based network model, a Markov chain with transition probability and transition time matrices whose centroids yield reconstructed flow sequen

What would settle it

Take a flow with two physically separate regions whose vorticity histories are identical up to a time shift or a reflection: if the Euclidean distance is large, the embedding would separate them despite identical dynamics. A concrete test is to run the pipeline on a synthetic domain with two independently oscillating patches with different frequencies and check that the resulting subdomains match the known patches; a mismatch would localise the failure to the distance/embedding choice.

Watch

Extended reading notes

Core claim

The central claim is that partitioning the flow domain by clustering points in a manifold built from their vorticity time series exposes local dynamics that a global analysis cannot. Concretely, the method (ST-CNM) first computes Euclidean distances between vorticity histories at every grid point, feeds them to ISOMAP to obtain a low-dimensional embedding, chooses the dimension by residual variance, and applies k-means++ in that embedding to obtain subdomains, with the number of subdomains chosen by a two-line fit criterion. Each subdomain is then described by a cluster-based network model, a Markov chain over temporal centroids with transition times. In the fluidic pinball, the partition se

Load-bearing premise

The method assumes that the Euclidean distance between two points' vorticity time series is a faithful measure of how similar their local dynamics are, and that the ISOMAP embedding preserves that similarity well enough for clustering.

Editorial extensions

If this is right

  • A dynamical atlas of a flow can be produced with no human labelling: the pipeline from raw snapshots to subdomains and local Markov models is automatic once the data are provided.
  • Local models can separate mechanisms that a global model merges, as in the jet where vortex shedding and vortex pairing appear as two distinct cycles within one subdomain.
  • The method flags regions that resist low-order modelling, such as the pinball's cylinder-gap region and the jet's turbulent mixing zones, as places needing higher resolution or more elaborate models.
  • Because the spatial partition is based on time-series similarity, the same framework can be applied to other scalar fields and to non-time-resolved data, since the distance computation does not require temporal resolution.

Reading between the lines

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

  • Beyond the paper: because the partition is feature-driven, using velocity components, pressure, or passive scalars as features would likely produce different subdomains; comparing features would test how much of the recovered structure is intrinsic to the flow rather than chosen by the vorticity feature.
  • Beyond the paper: the Euclidean distance between vorticity histories is a global similarity measure; a phase-invariant or spectral metric might separate regions more cleanly in flows with convection or phase delays, and would be a direct testable variant.
  • Beyond the paper: the local Markov models are built independently per subdomain, so a natural next step—already hinted at in the conclusions—is to couple subdomains through shared boundaries or cross-transition matrices; the current paper does not yet deliver such a global coupled model.
  • Beyond the paper: the automatic subdomain extraction could serve as a sensor-placement guide in flow control, since regions with distinct local dynamics are natural locations for probes or actuators.
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 / 5 minor

Summary. The paper introduces Space-Time Cluster-Based Network Models (ST-CNM), a data-driven framework to partition a fluid flow into spatial subdomains whose local dynamics are then modeled by cluster-based Markov models. The spatial partition is obtained by applying ISOMAP to vorticity time series at each grid point and clustering the resulting manifold coordinates with k-means; the number of spatial and temporal clusters is selected by the Two-Line Fit criterion. The method is demonstrated on two flows: a directly simulated fluidic pinball with two incommensurate periodic forcings and a planar PIV measurement of a transitional jet. The authors claim that ST-CNM automatically identifies local dynamics, such as vortex shedding and vortex pairing in the jet, that are not captured by global cluster-based models.

Significance. If substantiated, the method would be a useful unsupervised diagnostic for decomposing complex flows into interpretable, dynamically homogeneous regions, with potential applications in flow control and reduced-order modeling. The paper has several strengths: the pipeline is fully data-driven, the manifold coordinates are shown to correlate with physically interpretable quantities (dominant frequency, mean vorticity, fluctuation level), and the pinball results convincingly recover the known forcing frequencies in most subdomains. However, the central claim—that ST-CNM identifies local dynamics not captured by a global approach—is currently supported only by selective, mostly in-sample validation and by qualitative comparisons with global models that use different cluster counts. The significance is therefore conditional on a strengthened validation.

major comments (4)
  1. [§2.3, Eqs. (6)–(9)] The validation is in-sample. The transition probability and transition time matrices (Eqs. 6 and 7) and the cluster centroids are estimated from the full snapshot sequence, and the agreement is then measured by comparing the autocorrelation of a trajectory sampled from those same matrices with the autocorrelation of the same data (Eq. 9). This can substantially overstate model fidelity. The paper should use a train/test split (e.g., estimate the model on the first half of the time series and test on the second half), or compare the ensemble statistics of many model-generated trajectories with the data and report uncertainty. As written, the reported 'excellent overlap' is not evidence of predictive skill.
  2. [§3.3, Table 1 and §4.3, Table 2] The central examples are selected post hoc. In the pinball case, subdomains 3 and 8—the very regions highlighted as exhibiting complex local dynamics 'not obtained with a global approach'—are explicitly excluded from the autocorrelation comparison ('not shown here'), and Table 1 shows nontrivial discrepancies in their first-zero periods (4.19→4.36 and 8.64→9.13). In the jet case, only subdomains 5 and 6 are retained for analysis, while Table 2 shows that the other four subdomains have predicted periods differing from the original by factors of roughly 2–7 (e.g., subdomain 1: 72.14→10.47). The claim that ST-CNM generically identifies local dynamics requires a defined success criterion applied uniformly to all subdomains, not a retrospective selection of the subdomains that work.
  3. [§3.3, §4.3, Appendices B and D] The global-versus-local comparison is not quantitative and not controlled. The global CNM of the pinball uses 5 clusters while the local models use between 2 and 9 clusters; the global jet CNM uses 9 clusters while local models use up to 20. The comparison is made by visual inspection of transition networks and vorticity snapshots. As a result, the statement that local dynamics are 'not captured' by a global approach is an interpretation rather than a demonstrated fact. The paper should define a quantitative metric—for example, the spectral peaks of the local CNM reconstructions that are absent from the global CNM reconstruction, or the per-subdomain reconstruction error of the global model on the same masked data—and apply it to both methods under comparable settings.
  4. [§2.2, Eq. (2), §3.2] The spatial partition is built on the Euclidean distance between vorticity time series and on ISOMAP with a manually chosen number of neighbors, k=10. The paper asserts that 'this parameter is expected to have a limited influence' but provides no sensitivity test. Because the entire method rests on the assumption that Euclidean distance in vorticity time series and proximity in the ISOMAP embedding reflect dynamical similarity, the absence of a sensitivity analysis is a load-bearing gap. The authors should report how the spatial partition and the downstream local models change with k, with the ISOMAP embedding dimension, and ideally with alternative feature choices (e.g., velocity components). Without this, the claimed full automatization ('no meta-parameter tuning') is not established.
minor comments (5)
  1. [§2.3, Eqs. (6)–(7)] The indices in the transition probability and transition time matrices are confusing: the text says 'the transition probability from Sj to Si' but writes Q_ij with n_ij defined as the number of snapshots in Sj whose successor is in Si. Please clarify the row/column convention or change the notation so that Q_ij is unambiguously defined.
  2. [§3.2, Fig. 4] The interpretation of the manifold coordinates, especially γ4, is based on visual inspection of scatter plots. The text acknowledges that γ4 'needs further investigation.' Consider providing quantitative correlations or fitted trends to substantiate the stated relationships.
  3. [§4.1] The jet dataset subsampling is described as '5000 snapshots, subsampled by a factor of 4.' It should be stated explicitly whether the 5000 snapshots are already the subsampled set or the original set, and what the effective time step is after subsampling, since the nondimensional periods in Table 2 depend on this.
  4. [§4.3 and Appendix C.2] The 20-cluster model of subdomain 6 is described as useful 'for the sake of interpretation and not for the identification of limit cycles.' This caveat is appropriate, but the interpretation of vortex pairing is then used as support for the main claim. Please state clearly that the pairing interpretation is tentative and not validated by the 9-cluster model, whose two cycles differ in period by about 20%.
  5. [§2.2, TLF criterion] The Two-Line Fit criterion is used throughout to select the number of clusters, but the description is brief. In particular, the definition of the ratio R²/ε and the practical range of the sweep (e.g., up to 50 clusters) should be stated precisely enough for the method to be reproduced.

Circularity Check

2 steps flagged · score 6.0 of 10

In-sample, selective validation: local 'predictions' are sampled from transition matrices fitted to the same data, and the subdomains that fail validation are excluded from the evidence yet used as the headline result.

  1. fitted input called prediction [Section 2.3 (Eqs. 6–9); applied in Sections 3.3 and 4.3]
    "The direct transition probability Q and transition time matrices T are then computed. ... The flow field data is then reconstructed by sampling a sequence of clusters from the transition probability matrix ... Similar to Li et al. [2021], the validity of the local cluster-based network models is verified with the autocorrelation function of the velocity field"

    Q and T are fitted by counting transitions and residence times in the exact same snapshot sequence against which R(τ) is evaluated. A 'predicted' trajectory is a Monte-Carlo sample from these fitted matrices, and its time axis is set by the fitted T (Eqs. 7–8), so the reconstructed dominant period is largely determined by the fitted residence times. The autocorrelation match is therefore an in-sample consistency check, not an independent prediction; the 'local dynamics' extracted from the same data are used to validate themselves.

  2. other [Section 3.3 and Section 4.3]
    "An analysis of the autocorrelation of the original data and the prediction data shows an excellent overlap for all cases except subdomains 3 and 8 (not shown here). ... Results show that only Subdomains 5 and 6 present similar periods (first zero crossing point), thus the analysis and discussion of the results are only assessed for these two subdomains."

    The subdomains that fail the in-sample autocorrelation check are precisely the ones used to support the headline claim ('The proposed methodology isolates regions of the flow, revealing complex local dynamics that cannot be obtained with a global approach'). Pinball subdomains 3 and 8, and in the jet all subdomains except 5 and 6, are excluded from the quantitative validation after the fact. This post-hoc selection means the central claim is not tested on the cases that matter; combined with step 1, the validation is both in-sample and selective.

full rationale

The spatial-manifold decomposition itself is a legitimate unsupervised step: ISOMAP on vorticity time series followed by k-means does not depend on the target claim, and the paper does not invoke any uniqueness theorem or load-bearing self-citation. However, the demonstration that ST-CNM 'identifies local dynamics not captured by a global approach' rests on the local CNMs' ability to reproduce the original dynamics. This validation is circular/in-sample: Q and T are estimated from the same trajectory used to compute the autocorrelation, and the 'predicted' trajectory is generated by sampling those fitted matrices, so agreement is a consistency check. The paper further restricts the analysis to the subdomains that pass this check (pinball: all but 3 and 8; jet: only 5 and 6), while the headline complex dynamics (pinball 3 and 8; jet vortex pairing) come from the excluded or qualitatively-only cases. The comparison to global CNMs is qualitative and uses different cluster counts. Thus the central claim is only partially supported and has a built-in in-sample component; score 6.

Assumptions & free parameters 6 free parameters · 8 assumptions · 0 invented entities

The method introduces no new physical entities, forces, or dimensions. It relies on a sequence of data-driven modeling choices: the feature (vorticity), the metric (L2), the manifold-learning algorithm (ISOMAP with k neighbors), the clustering algorithm (k-means with TLF-selected cluster counts), and the Markov model per subdomain. These are heuristics and domain assumptions rather than derived or externally verified principles.

free parameters (6)
  • k (ISOMAP nearest neighbors) = 10 for the pinball; not specified for the jet
    Sets the k-nearest-neighbor graph used to estimate geodesic distances in ISOMAP (Section 2.2, step 1). The authors claim limited influence but provide no sensitivity study.
  • ISOMAP embedding dimension = 4 (pinball), 3 (jet)
    Chosen as the first minimum of the residual variance Rv(n) (Eq. 3). Data-driven but still a modeling choice that affects the manifold.
  • Number of spatial subdomains Nc = 9 (pinball), 6 (jet)
    Determined by the TLF criterion applied to WCSS (Eq. 4). TLF is a heuristic, not a principled selection method.
  • Number of temporal subsequences per subdomain = 2-9 for pinball subdomains; 9 for jet subdomain 6, 10 for jet subdomain 5; 20 in appendix
    Selected via TLF for each local CNM. The choice affects the resolution and interpretability of the cycles.
  • k-means replicates = Default for most; ~1000 for jet subdomain 5
    Increased to ensure convergence of cluster assignments. This is a manual intervention that contradicts the 'fully automated' claim.
  • Snapshot subsampling for the jet = 5000 snapshots, every 4th frame from 60000
    Subsampled to reduce computational load, assuming converged statistics. This changes temporal resolution and may affect the manifold.
assumptions (8)
  • domain assumption Manifold hypothesis: high-dimensional vorticity time-series data at spatial points lie near a low-dimensional manifold.
    Invoked in Section 1: 'Under the manifold hypothesis, high-dimensional flow data can be assumed to lie near a lower-dimensional structure.' This justifies using ISOMAP.
  • domain assumption Euclidean L2 distance between vorticity time series measures dynamical similarity.
    Eq. (2) defines di,j as the integral of squared vorticity difference over time. The whole spatial manifold construction rests on this metric.
  • standard math ISOMAP's geodesic distances, estimated via k-nearest-neighbor graph, preserve the manifold structure.
    This is the standard assumption of ISOMAP (Tenenbaum et al., 2000), used in Section 2.2. The k parameter must be chosen to avoid short-circuiting.
  • domain assumption k-means clustering in ISOMAP coordinates yields spatially coherent subdomains with similar dynamics.
    Section 2.2: k-means partitions the spatial manifold, and the affiliation function partitions the flow field. The Voronoi tessellation of the manifold is assumed to correspond to dynamics-based regions.
  • domain assumption Vorticity alone is a sufficient feature for local dynamics.
    Section 2.2: 'a vorticity time series is selected as the feature for each point as it synthesizes both velocity components into a single quantity.' This is a modeling choice that is not tested against including velocity components.
  • domain assumption The flows are statistically stationary, and the CNM Markov assumption holds.
    The analysis uses post-transient data and builds a Markov chain with transition probabilities fitted from data (Eq. 6). This assumes stationarity and that transitions depend only on the current cluster.
  • ad hoc to paper The TLF criterion is a valid method for selecting the number of spatial and temporal clusters.
    The Two-Line Fit method (from Brindise and Vlachos, 2017) is used repeatedly to pick Nc. It is a heuristic, and its reliability for this purpose is not independently established.
  • ad hoc to paper Incommensurable forcing frequencies fc1=sqrt(pi) f0 and fc2=pi f0 create a usefully complex multi-frequency flow.
    Section 3.1: the π and √π coefficients are chosen so that the local dynamics are incommensurable. This is an experimental design choice tailored to demonstrate the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Divide and conquer complex flows. Part I: cluster and manifold-based local analysis." pith.science (2026). https://pith.science/paper/HSUAS32A

@misc{pith2026260105117,
  author       = {Pith},
  title        = {Pith review of: Divide and conquer complex flows. Part I: cluster and manifold-based local analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HSUAS32A}},
  note         = {Machine review of arXiv:2601.05117}
}
read the original abstract

This work is a two-part study on the description and prediction of complex fluid flows through the partitioning of the flow domain. In this first part, we propose a framework for a global description of the dynamics of complex flows via clustered spatial representations of the flow, isolating and identifying local dynamics, retrieving different \acp{ST-CNM}. The key enabler is the partitioning of the domain based on a nonlinear manifold learning approach, in which spatial points are clustered based on the similarity of their dynamics, as observed in their compact embedding in manifold coordinates. The method receives as input time-resolved flow fields. The spatial manifold is computed through isometric mapping applied to the vorticity time histories at each spatial location. An unsupervised clustering method, applied in the manifold space, partitions the full flow domain into subdomains. The dynamics of each subdomain are then described with cluster-based modelling. The method is demonstrated on two flow-field datasets obtained with a direct numerical simulation of a fluidic pinball under periodic forcing and with two-dimensional particle image velocimetry measurements of a transitional jet flow. The spatial manifold-based flow partitioning identifies regions with similar dynamics in an automated way. For both cases, \ac{ST-CNM} identifies local dynamics that are not captured by a global approach. In particular, vortex shedding and vortex pairing dynamics are isolated in the jet flow experiment. The proposed fully automated domain partitioning method will benefit the structural description of controlled flows and unveil the actuation mechanisms at play.

Figures

Figures reproduced from arXiv: 2601.05117 by the authors.

Figure 1
Figure 1. Schematics of the space-time cluster-based network model methodology. (a) Starting point is time-resolved [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Fluidic pinball configuration. (a) Computational grid. (b) Vorticity field of a sample snapshot of the flow [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Spatial manifold for the fluidic pinball. (a-f) Projection of the manifold on the first four manifold coordinates. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Intepretation of the spatial manifold coordinates [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Local interpretation of the fluidic pinball under control. The flow is partitioned into [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Description of 5-subsequence subdomains. Depiction of the centroids, transition and snapshots for (a) subdomain 7 and (b) subdomain 9. The circular arrows indicate the average rotation speed of the cylinder for each subsequence. A full circle corresponds to a periphera…
Figure 7
Figure 7. Figure 7: Description of subdomain 3. (a) Transition probability matrix; (b) Transition time matrix. c. t. u. stands for convective time units. The transition probability matrix suggests the dynamics include two cyclic trajectories (S1 → S2 → S3 → S4 and S8 → S5 → S6 → S7). The …
Figure 8
Figure 8. Figure 8: Similar to Figure [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Schematics of the planar PIV setup for the jet flow experiment in water. (a) Top view and (b) front view. The [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Spatial manifold of the jet flow data. (a) Residual variance versus the dimensionality of the spatial recon [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Local interpretation of the experimental water jet flow. The flow is partitioned into [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Description of subdomain 6. a) Transition probability matrix. b) Transition time matrix. The transition probability matrix suggests that dynamics include two periodic trajectories (S1 → S2 → S4 → S5 → S6 and S7 → S8 → S9). The corresponding nondimensional vorticity sn…
Figure 13
Figure 13. Figure 13: Description of subdomain 5. (a) Transition probability matrix. (b) Transition time matrix c. t. u. stands for convective time units. Transition probability matrix suggests the dynamics include a periodic trajectory (S2 → S3 → S4 → S5 → S6 → S7). The period of the cycl…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 7 canonical work pages

  1. [7]

    doi: 10.1016/j.jcp.2017.10.001

    ISSN 0021-9991. doi: 10.1016/j.jcp.2017.10.001. L. Gkimisis, N. Aretz, M. Tezzele, T. Richter, P . Benner, and K. E. Willcox. Non-intrusive reduced-order modeling for dynamical systems with spatially localized features. Comput. Methods Appl. Mech. Eng. , 444:118115,

    work page doi:10.1016/j.jcp.2017.10.001 2017

  2. [11]

    doi: 10.1017/jfm.2020.725. J. L. Callaham, J. V . Koch, B. W. Brunton, J. N. Kutz, and S. L. Brunton. Learning dominant physical processes with data-driven balance models. Nat. Commun., 12:1016,

    work page doi:10.1017/jfm.2020.725 2020

  3. [12]

    doi: 10.1038/s41467-021-21331-z. K.-E. Otmani, G. Ntoukas, O. A. Mariño, and E. Ferrer. Toward a robust detection of viscous and turbulent flow regions using unsupervised machine learning. Phys. Fluids, 35(2):027112, 02

    doi:10.1038/s41467-021-21331-z

  4. [13]

    doi: 10.1007/s00366-024-02062-3. A. Parente, J. C. Sutherland, L. Tognotti, and P . J. Smith. Identification of low-dimensional manifolds in turbulent flames. P . Combust. Inst., 32(1):1579–1586,

    work page doi:10.1007/s00366-024-02062-3

  5. [15]

    doi: 10.1017/jfm.2019.959. G. Y . Cornejo Maceda, Y . Li, F. Lusseyran, M. Morzyski, and B. R. Noack. Stabilization of the fluidic pinball with gradient-enriched machine learning control. J. Fluid Mech., 917:A42,

    doi:10.1017/jfm.2019.959 2019

  6. [16]

    23 arXiv Template A P REPRINT L

    doi: 10.1017/jfm.2021.301. 23 arXiv Template A P REPRINT L. Franceschelli, C. E. Willert, M. Raiola, and S. Discetti. An assessment of event-based imaging velocimetry for efficient estimation of low-dimensional coordinates in turbulent flows. Exp. Therm. Fluid Sci. , 164:111425,

    work page doi:10.1017/jfm.2021.301 2021

  7. [17]

    doi: 10.1017/jfm.2024.593. R. W. Floyd. Algorithm 97: shortest path. Commun. ACM, 5(6):345–345,

    work page doi:10.1017/jfm.2024.593 2024

  8. [2003]

    doi: 10.1017/S0022112003006694. C. Bogey and C. Bailly. Large eddy simulations of transitional round jets: influence of the reynolds number on flow development and energy dissipation. Phys. Fluids, 18(6),

    doi:10.1017/s0022112003006694

Show all 18 references
  1. [2010]

    doi:10.1017/S0022112010001217. M. A. Mendez, M. Balabane, and J.-M. Buchlin. Multi-scale proper orthogonal decomposition of complex fluid flows. J. Fluid Mech., 870:9881036,

    doi

  2. [2014]

    doi:10.1017/jfm.2014.355. H. Li, D. Fernex, R. Semaan, J. Tan, M. Morzyski, and B. R. Noack. Cluster-based network model. J. Fluid Mech. , 906:A21,

    2014 doi

  3. [2018]

    doi: 10.1017/jfm.2018.283. P . J. Schmid. Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech. , 656:528,

    2018 doi

  4. [2019]

    doi: 10.1017/jfm.2019.212. E. Kaiser, B. R. Noack, L. Cordier, A. Spohn, M. Segond, M. Abel, G. Daviller, J. Östh, S. Krajnovi, and R. K. Niven. Cluster-based reduced-order modelling of a mixing layer. J. Fluid Mech. , 754:365414,

    2019 doi

  5. [2020]

    URL https://global-sci.com/index.php/ cicp/article/view/6911

    doi: 10.4208/cicp.OA-2020-0164. URL https://global-sci.com/index.php/ cicp/article/view/6911. Z. Hu, A. D. Jagtap, G. E. Karniadakis, and K. Kawaguchi. When do extended physics-informed neural networks (XPINNs) improve generalization? SIAM J. Sci. Comput. , 44(5):A3158–A3182,

    2020 doi

  6. [2021]

    22 arXiv Template A P REPRINT D

    doi: 10.1017/jfm.2020.785. 22 arXiv Template A P REPRINT D. Fernex, B. R. Noack, and R. Semaan. Cluster-based network modelingFrom snapshots to complex dynamical systems. Sci. Adv., 7(25):eabf5006,

    2020 doi

  7. [2022]

    doi: 10.1137/21M1447039. B. Li, Z. Y ang, X. Zhang, G. He, B.-Q. Deng, and L. Shen. Using machine learning to detect the turbulent region in flow past a circular cylinder. J. Fluid Mech., 905:A10,

    doi

  8. [2023]

    doi: 10.1063/5.0163833

    ISSN 1070-6631. doi: 10.1063/5.0163833. J. B. Tenenbaum, V . de Silva, and J. C. Langford. A global geometric framework for nonlinear dimensionality reduc- tion. Science, 290(5500):2319–2323,

    doi

  9. [2024]

    doi: 10.1017/jfm.2024.428. A. Colanera and L. Magri. Towards extreme event prediction of turbulent flows with quantized local reduced-order models. arXiv preprint arXiv:2511.04586,

    2024

  10. [2025]

    doi: 10.1016/j.cma.2025.118115

    ISSN 0045-7825. doi: 10.1016/j.cma.2025.118115. A. D. Jagtap and G. E. Karniadakis. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential. Commun. Comput. Phys. , 28(5)...

    2025

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.