{"id":"cd916294-8c7e-4907-bf06-11635e9920f1","arxiv_id":"2506.13738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A local reduced-order model framework that quantizes the solution manifold into clusters, constructs a centroid-centered POD-Galerkin model per cluster, and switches between models via a change of basis improves stability, prediction horizon, and statistics for chaotic PDEs.","lead":"The authors split the chaotic attractor into clusters, build a separate low-dimensional Galerkin model for each cluster, and switch between models as the system moves. The method reports longer prediction horizons, better long-term statistics, and improved stability over a single global model on the Kuramoto-Sivashinsky and Kolmogorov flow test cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Local-basis continuity at cluster switches is unanalyzed; the hard switch rule (14) introduces uncontrolled state jumps whose effect on the reported prediction-horizon gains is not quantified.","rationale":"The reader identifies K-means clustering geometry and the local-subspace validity as the weakest assumption, which is the natural first guess. My stress-test pass agrees that the clustering premise is load-bearing, but the more fragile and less supported piece is the online switching dynamics: K-means assigns snapshots, but the method then makes a hard, discontinuous change of basis (Eq. 14) whenever the nearest centroid changes. The paper explicitly flags non-differentiability but dismisses it as 'not important' without any quantitative justification. In chaotic systems, small discontinuities in the reduced state can be amplified exponentially (the same mechanism that limits prediction horizons), so this is exactly where a hidden accuracy cost would live. The reported comparisons (Figs. 6, 7, 9, 10, 12) compare g-ROM to ql-ROM, conflating the local-subspace benefit with the switching cost; they do not test whether the switch rule itself is benign. My concrete test separates these mechanisms. It does not overturn the central claim; it strengthens the verdict as CONDITIONAL: the claim rests on the switch being harmless, and that premise needs a dedicated numerical check. The reader said the same overall verdict, so the verdict is unchanged, but the specific concern is different (switch mechanism rather than clustering geometry), hence partial agreement.","tokens_in":23847,"tokens_out":2287,"duration_ms":19654,"concrete_test":"I would settle this with the following diagnostic. For each test case in Section 4, run the ql-ROM twice: once with the hard switch rule (14) as in Algorithm 1, and once with the switch performed by projecting the current physical state u(t) onto the new cluster basis U_j and new centroid c_j (a 'reset' at each switch, using the same offline ingredients). Then compute the prediction-horizon metric (21), the cluster-affiliation PDF (20), and the energy spectrum for both variants. If the reset variant is not markedly worse than (14), the switch map is neither a bottleneck nor a contributor, and the reported gains are attributable to the local subspaces themselves.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that ql-ROMs 'significantly and consistently outperform' g-ROMs at equal degrees of freedom, with improved stability, prediction horizon, and long-term statistics. The load-bearing mechanism is the hard nearest-centroid switch with the projection-based change of basis in Equation (14) and the affiliation functions in Equations (2) and (3). The paper explicitly notes (footnote, Section 2.2.1) that the solution may be non-differentiable at cluster boundaries and defers smoothing to future work. This is not a cosmetic caveat: the change-of-basis map (14) is a composition of orthogonal projections plus an offset, which is generally discontinuous wherever the centroid affiliation changes, and the online trajectory is generated by integrating one local Galerkin model and then committing a discontinuous jump in the reduced coordinates at every switch. Nothing in the paper demonstrates that this switching error is small relative to the reported gains; the prediction-horizon comparisons (Section 4) measure the combined effect of local modeling and switching, but this is not a separate check on whether the switch is the bottleneck or a hidden source of inaccuracy. Because the method is data-driven in its partition but intrusive in its local models, the possibility remains that the observed stability and horizon improvements come mainly from the trivial centroid-shift and local-PCA splitting of a heterogeneous attractor, rather than from the stitched dynamics being a coherent approximation to the true flow. The paper also states (Footnote, Section 2.2.1) that non-differentiability 'is not important for the goal of this paper,' which is precisely the claim that needs testing. This is not an objection to the local-ROM idea itself, but a precise, in-scope gap: the central mechanism is asserted, not analyzed or quantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes quantized local reduced-order models (ql-ROMs) for spatiotemporally chaotic PDEs. The method partitions snapshot data into K clusters via k-means, constructs a centroid-centered local POD-Galerkin ROM for each cluster, and integrates whichever local model corresponds to the nearest centroid, transforming coordinates at cluster switches by Eq. (14). The approach is tested on the Kuramoto-Sivashinsky equation (bursting and chaotic regimes) and two-dimensional Kolmogorov flow (quasiperiodic and turbulent regimes), reporting improvements over global POD-Galerkin ROMs in numerical stability, short-term prediction horizon, and long-term statistics such as energy spectra and probability distributions, at equal online dimension r.","tokens_in":24117,"tokens_out":7281,"duration_ms":74764,"significance":"The proposed framework is simple, physically interpretable, and combines data-driven phase-space partitioning with intrusive projection, so if the reported gains hold under a clean evaluation protocol it would be a useful addition to the local-ROM literature. The paper covers four regimes across two PDEs, compares against a natural global-ROM baseline, and provides a clear algorithmic description (Algorithm 1) with precomputed transition maps. The diagnostic breadth (prediction horizon, spectra, PDFs, cluster-affiliation statistics) is a genuine strength. However, the manuscript does not provide code or data, and the evaluation currently relies on test-set-based hyperparameter selection and on a single realization per regime, both of which limit the strength of the central claim until addressed.","major_comments":[{"comment":"The number of modes r is selected using reconstruction error on the test dataset: Section 2.3 states that 'the number of modes r is selected for the root mean squared error (16) of the test dataset', and Section 4.2 repeats this for r = 100 and r = 400. The same test dataset is then used to report the prediction horizons and errors, e.g., Eq. (21) and Eq. (22) and Figure 9. This is a test-set leakage: r is a hyperparameter chosen to fit the test snapshots, so the reported out-of-sample improvements are optimistically biased and do not support the claim of a controlled comparison. Please select r (and K, where relevant) using only training or validation data, and report all evaluation metrics on a held-out test interval that was not used for any model choice; alternatively, provide a sensitivity analysis showing that the reported conclusions are stable over a range of r.","section":"§2.3, Eq. (16) and §4.1--§4.2"},{"comment":"The hard nearest-centroid switch induced by Eq. (14) is the central mechanism, but its effect on trajectory continuity is not analyzed. If the state before a switch is c_i + U_i a_i, then after applying Eq. (14) the reconstructed physical state is c_j + U_j U_j^H (c_i - c_j + U_i a_i), so the physical trajectory jumps by (I - U_j U_j^H)(c_j - c_i - U_i a_i), which is generally nonzero when the local POD subspaces differ. The footnote only states that the solution may be non-differentiable and defers smoothing to future work, but the reported prediction-horizon gains are measured on trajectories that include these jumps. Without an estimate of the jump magnitude or a sensitivity test, it is unclear whether the switch is a minor bookkeeping operation or a hidden source of error. Please quantify the switching error (e.g., accumulate ||u_r(t+) - u_r(t-)|| at every switch, report the number and timing of switches, and test a short-overlap or hysteresis variant) and present it separately from the local-model error.","section":"§2.2.1, Eq. (14), Algorithm 1, and footnote"},{"comment":"The claim that ql-ROMs 'significantly and consistently outperform' g-ROMs is supported by a single test realization per regime. For chaotic systems, prediction horizons and long-time statistics fluctuate across initial conditions and test intervals, so the reported factors of 2--3 in prediction horizon, and the qualitative spectrum and PDF comparisons, may not be systematic. Please report results from multiple initial conditions or multiple disjoint test windows, and provide a variability measure such as the standard deviation or interquartile range of T_ph (Eq. (21)) and of the error metrics, so that the consistency claim can be assessed statistically.","section":"§4.1, §4.2"},{"comment":"The hyperparameter-selection protocol is not consistently described. Section 2.3 says r is selected from the reconstruction error of the test dataset, while the caption of Figure 6 states that the basis and centroids were constructed using only the training dataset and that the MSE is evaluated on the test dataset. Please clarify in the text whether the reconstruction-error curves in Figures 6(a) and 8(a) are computed on training, validation, or test data, and ensure that the procedure used to choose r and K is compatible with the out-of-sample evaluation protocol required for the prediction metrics.","section":"§2.3 and Figure 6(a) caption"}],"minor_comments":[{"comment":"In the sentence 'the second termo of (18)', 'termo' should be 'term'.","section":"§2.3"},{"comment":"Equation (15) uses beta_v(u_r(t)) for the reconstruction, while Algorithm 1 uses beta_c(t) during integration; please align the notation and clarify how beta_c is evaluated during prediction when the time index refers to the reduced trajectory.","section":"§2.2.1, Algorithm 1"},{"comment":"Equation (16) is called 'root mean squared error' but defines a pointwise residual vector; consider defining a scalar RMSE and reserving r_m for the residual vector to avoid confusion.","section":"Eq. (16)"},{"comment":"The vertical axes in these figures are labeled with values but not with the quantity, which appears to be Delta BIC / Delta K; please add explicit axis labels.","section":"Figures 8(b) and 13"},{"comment":"The paper does not state whether code or data will be made available; for a computational physics journal, please add a data and code availability statement.","section":"General"},{"comment":"The sentence 'The ql-ROMs significantly and consistently outperforms g-ROMs' has a subject-verb agreement error; it should be 'outperform'.","section":"§5, Conclusions"}],"recommendation":"major_revision","confidential_remarks":"This is a promising and clearly written contribution, and I do not see grounds for rejection. However, the test-set leakage in hyperparameter selection and the unquantified switching jumps are load-bearing issues for the central performance claims, and both need to be resolved before publication. The editor may also wish to ask the authors to place the work more explicitly in the context of cluster-based network models and local PCA approaches, since those are the most closely related existing methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, clearly written method paper. The ql-ROM combination — K-means quantization of the attractor, per-cluster centroid-centered POD-Galerkin models, and a nearest-centroid switch with a projection-based change of basis — is new in that exact form, and the demonstrations across four regimes (KS bursting/chaotic, Kolmogorov quasiperiodic/turbulent) show consistent gains over global ROMs at equal DoF: stability at lower mode counts, prediction horizons roughly 2–3x longer, and better long-term spectra and PDFs. That is worth a serious referee.\n\nThe strengths are real. The derivation is standard Galerkin projection, so the reduced equations are interpretable and the overhead is minimal. The change-of-basis formula (14) is a simple orthogonal projection plus centroid shift, easy to implement offline. The authors cite the relevant local-ROM and cluster-modeling literature and are upfront about the hard switch (footnote in Sec. 2.2.1).\n\nThe soft spots are mostly about evaluation, not conception. First, r is selected from reconstruction error on the test set (Sec. 2.3, Eq. 16) and the same test set is used for the prediction metrics. That is a real leakage problem. It is not fatal because reconstruction error is not the same as prediction horizon, but the selection should be done on a validation set or with nested CV. Second, the switching mechanism is asserted, not analyzed. The paper admits the solution may be non-differentiable at cluster boundaries and waves it away; but the switch is exactly where the reduced trajectory is re-initialized by projection, and nothing quantifies how much error that jump injects or whether the local bases alone (without the hard switch, or with soft weighting) would give similar results. The reported gains include the switch, so you cannot tell if it is a feature or a bug. Third, for chaotic systems, single-trajectory error metrics are not very persuasive; ensemble-averaged prediction horizons and statistics over many initial conditions would be more convincing. And there is no comparison with other local ROM baselines, so the value-added over, say, cluster-based network models or local PCA is unclear. No code or data is released, which makes the detailed numbers hard to check.\n\nNone of this sinks the paper. The method is plausible and the experiments are broad. But the claims currently outrun the evidence in three ways: leakage, unquantified switching, and lack of baseline comparisons.\n\nWho is this for? Anyone building projection-based ROMs for chaotic or turbulent flows, and people in data assimilation or control who need cheap stable surrogate models. It will likely become a useful reference for local ROM strategies.\n\nMy recommendation: send it to peer review, and push for a revision that fixes the test-set leakage, adds at least one comparison against a local baseline (or a clear argument why g-ROM is the right baseline), and ideally reports ensemble statistics. The switch analysis could be a future paper; the empirical evidence may be enough if the other issues are addressed.","headline":"A clean and clearly presented local-ROM idea with genuine promise, but test-set leakage and an unquantified hard switch make the evaluation weaker than the claims; still worth peer review.","tokens_in":24706,"tokens_out":5485,"would_cite":true,"duration_ms":49677,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that partitioning a chaotic attractor into clusters and building a centroid-centered POD-Galerkin reduced-order model in each cluster yields numerically stable models that predict further ahead and reproduce long-term…","keywords":["quantized local reduced-order modeling","K-means clustering","proper orthogonal decomposition","Galerkin projection","chaotic dynamics","Kuramoto-Sivashinsky equation","Kolmogorov flow","energy spectrum"],"falsifier":"Hold the number of modes $r$ fixed on a chaotic test case, sweep the number of clusters $K$ from 1 to well past the BIC elbow, and measure the prediction horizon and cluster-affiliation statistics. If the ql-ROM's advantage over the g-ROM does not grow up to the elbow and then degrade, cluster geometry is not the operative mechanism; if the trajectory begins to chatter at cluster boundaries for large $K$, the switching map is the weak point.","tokens_in":23648,"feed_emoji":"🌀","tokens_out":11073,"duration_ms":101198,"temperature":0.7,"pith_summary":"This paper sets out to establish that a chaotic system's attractor can be modeled more reliably by partitioning it into clusters and building a separate reduced-order model in each cluster than by using one global model for the whole manifold. The proposed quantized local reduced-order model, or ql-ROM, first applies K-means clustering to snapshots, then constructs a POD-Galerkin model around each cluster centroid, and switches between local models according to which centroid is nearest. With the same number of retained modes as a global POD-Galerkin model, the local models are claimed to be numerically stable where the global model diverges, to extend the short-term prediction horizon, and to reproduce long-term statistics such as energy spectra and probability distributions. The online cost stays nearly the same because only one local model is integrated at any time. A reader should care because this offers a simple, interpretable route to stable reduced-order models for chaotic and turbulent partial differential equations without increasing degrees of freedom.","feed_headline":"Per-cluster models beat a global ROM at the same cost","feed_subtitle":"Splitting the attractor into clusters gives stable local reduced-order models with longer prediction horizons and accurate statistics.","key_machinery":"The load-bearing construction is the quantized local reduced-order model itself: K-means partitions the snapshots into $K$ clusters, each with a centroid $\\boldsymbol{c}_k$; the fluctuations around the centroid feed a per-cluster POD basis; and Galerkin projection yields $K$ low-dimensional systems $\\frac{d\\boldsymbol{a}_k}{dt} + \\boldsymbol{B}_k \\boldsymbol{a}_k + \\boldsymbol{N}_k(\\boldsymbol{a}_k, \\boldsymbol{c}_k) + \\boldsymbol{f}_k = \\mathbf{0}$. The piece that stitches the local models together is the change-of-basis map $\\boldsymbol{a}_j = \\boldsymbol{U}_j^{H} \\boldsymbol{U}_i \\boldsymbol{a}_i + \\boldsymbol{U}_j^{H}(\\boldsymbol{c}_i - \\boldsymbol{c}_j)$, applied when the nearest-centroid assignment switches from cluster $i$ to cluster $j$. This map converts the reduced coordinates from the outgoing cluster's basis to the incoming cluster's basis, and every matrix in it is precomputed offline. The number of clusters $K$ is selected with a BIC elbow and the number of modes $r$ by test-set reconstruction error, so the method remains a standard intrusive POD-Galerkin workflow with clustering inserted as a preprocessing step.","core_discovery":"The central discovery is that the heterogeneity of the attractor, not just its dimension, is what defeats global intrusive reduced-order models, and that local modeling around cluster centroids removes much of that difficulty. The authors demonstrate this on the Kuramoto-Sivashinsky equation and on two-dimensional Kolmogorov flow, covering bursting, chaotic, quasiperiodic, and turbulent regimes. In the KS bursting case, a global model with nine modes is unstable while the quantized local model with nine modes is stable and accurate, and in the KS chaotic case the local model extends the prediction horizon by roughly a factor of two while avoiding the high-wavenumber aliasing that contaminates the global model's energy spectrum. For Kolmogorov flow, the local model captures the spatial energy spectrum and the cluster-visit statistics that the global model misses. The claim is that these advantages come with essentially no extra online cost, because at each time step only the local model of the currently active cluster is integrated.","pith_inferences":["Beyond the paper: the quantization idea should transfer to non-intrusive learners such as reservoir computers or neural-ODE models, whose global latent models often struggle with multi-modal attractors; the clustering and switching machinery does not depend on the local model being a Galerkin projection.","Beyond the paper: the hard nearest-centroid switch is the most fragile piece, and a natural extension is a soft or probabilistic mixture of local models near cluster boundaries to remove the non-differentiability the authors mention in passing.","Beyond the paper: because the change-of-basis map is linear and precomputed, the method is naturally compatible with ensemble data assimilation; cluster affiliation could act as a categorical observation with local updates per cluster.","Beyond the paper: the claim of consistent superiority could be probed by comparing a ql-ROM with a g-ROM that is given the same total number of parameters, including the centroid terms, to isolate whether the gain comes from local linearity or from having more parameters overall."],"forward_implications":["A ql-ROM with the same number of retained modes as a g-ROM can remain stable and accurate in regimes where the g-ROM diverges, as in the bursting KS case with nine modes.","Short-term prediction horizons increase by roughly a factor of three in the bursting KS regime and roughly a factor of two in the chaotic KS regime relative to the global model.","Long-term statistics improve: the ql-ROM reproduces cluster-affiliation probabilities, kinetic-energy probability distributions, and spatial energy spectra, whereas the g-ROM shows spectral aliasing at high wavenumbers.","The online computational cost stays nearly identical to a global ROM, since only the active cluster's $r$-dimensional system is integrated and each switch costs one matrix-vector product plus a shift.","The method remains intrusive and interpretable, because each local model is a standard Galerkin projection of the governing equations with clustering changing only the reference mean and basis."],"supporting_citations":[{"why":"Supplies the POD snapshot method used to construct each cluster's local basis.","marker":"[85]"},{"why":"Establishes the intrusive POD-Galerkin projection framework that each local ROM inherits.","marker":"[6]"},{"why":"Predecessor that clusters snapshots and models local dynamics around centroids, which ql-ROM extends to per-cluster POD-Galerkin models.","marker":"[65]"},{"why":"Provides the data-driven manifold perspective and the bursting Kuramoto-Sivashinsky regime used to test the local-model hypothesis.","marker":"[73]"},{"why":"Provides the K-means/BIC formulation for clustering turbulent-flow data used in phase-space quantization.","marker":"[74]"},{"why":"Supplies the k-means++ seeding used to initialize the clustering optimization.","marker":"[79]"},{"why":"Supplies the Bayesian information criterion used to select the number of clusters.","marker":"[89]"},{"why":"Supports the cluster-affiliation probability diagnostic used to compare how models populate the phase space.","marker":"[99]"}],"fun_headline_variants":["Attractor clustering stabilizes chaotic ROMs","Local ROMs extend prediction horizon for chaos","Quantized local modeling beats global ROM","Cluster-based ROMs: stable, accurate, low cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands or falls on the assumption that K-means with Euclidean distance splits the high-dimensional attractor into regions where a low-dimensional linear model around the cluster center is sufficient, and that switching abruptly between these models at cluster boundaries does not corrupt the trajectory.","fun_headline_variants_meta":{"raw":{"variants":["Attractor clustering stabilizes chaotic ROMs","Local ROMs extend prediction horizon for chaos","Quantized local modeling beats global ROM","Cluster-based ROMs: stable, accurate, low cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4127,"prompt_tokens":1044,"completion_tokens":3083,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":3026}},"tokens_in":660,"tokens_out":3083,"duration_ms":26344,"temperature":1.0,"reasoning_tokens":3026,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:58:02.119737+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold the number of modes $r$ fixed on a chaotic test case, sweep the number of clusters $K$ from 1 to well past the BIC elbow, and measure the prediction horizon and cluster-affiliation statistics. If the ql-ROM's advantage over the g-ROM does not grow up to the elbow and then degrade, cluster geometry is not the operative mechanism; if the trajectory begins to chatter at cluster boundaries for large $K$, the switching map is the weak point.","supporting_citations":[{"cited_title":"The proper orthogonal decomposition in the analysis of turbulent flows","cited_arxiv_id":null,"evidence_quote":"Supplies the POD snapshot method used to construct each cluster's local basis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the data-driven manifold perspective and the bursting Kuramoto-Sivashinsky regime used to test the local-model hypothesis."}],"review_version":2}