{"id":"519d7d89-4bc3-4b20-bbc9-79ea9d5c47e0","arxiv_id":"2411.17433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LC-SVD-DLinear and LC-HOSVD-DLinear forecast high-resolution cylinder wake snapshots by applying DLinear to temporal coefficients of low-cost SVD/HOSVD modes.","lead":"This paper combines a low-cost singular value decomposition with the DLinear linear forecasting network to predict future high-resolution flow snapshots from sparse velocity measurements. It reports accurate one-step forecasts and long visual reconstructions for a laminar cylinder wake at Re=220 and an experimental turbulent wake at Re=2600.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported test errors may be inflated because the spatial SVD modes are computed from the full dataset, including the test period, so the test-set reconstruction is not out-of-sample.","rationale":"The reader identified the fixed-basis assumption as the weakest structural point: future snapshots must lie in the span of W_rec. My concern is more specific and more immediately decisive: the spatial basis is not merely fixed but is computed from data that includes the test period, so the reported test-set reconstruction errors are not a valid measure of forecasting performance. This is a concrete, checkable flaw in the evaluation protocol rather than a physical limitation of the method. If the test were re-run with a proper temporal split, the method might still work, but the current numbers cannot support the central claim. I therefore recommend UNVERDICTED rather than REJECT: the method may be salvageable, but its validity cannot be assessed from the evidence as presented. The reader's weakest_assumption captures a related but distinct issue, so my agreement is partial.","tokens_in":22158,"tokens_out":4248,"duration_ms":50279,"concrete_test":"Re-run the full pipeline for both datasets with a strict temporal split before any SVD: compute W_rec and T_rec using only the training snapshots (e.g., the first 70% of the temporal window), then apply the fixed W_rec to forecast and reconstruct the held-out test snapshots. Compare the resulting test RRMSE against the reported values. If the RRMSE increases materially (especially for the turbulent case), the reported errors are not out-of-sample and the central claim must be re-evaluated.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim that LC-SVD-DLinear and LC-HOSVD-DLinear forecast high-resolution snapshots with low cost rests on the test-set RRMSE values (1.384%, 0.571%, 11.554%, 10.571%). As described, these numbers are not out-of-sample evidence for forecasting. Section 2.4 states that the temporal coefficients are split into train/validation/test only after LC-SVD has reconstructed the spatial modes and temporal coefficients; eqs. (2.12)-(2.14) then reconstruct snapshots using W_rec, the spatial basis computed from the entire input dataset. Consequently, the spatial structure of the test snapshots has already informed W_rec, and only the temporal coefficients are genuinely out-of-sample. The reported RRMSE therefore measures reconstruction of test data with a basis that has seen that data, not the error of forecasting genuinely unseen future states. For the turbulent case, retaining only 6 modes makes this leakage especially consequential because the basis is low-rank and may encode test-period coherent structures. The paper also gives no quantitative error for the extended forecasts (Nsnap=1000 laminar, Nsnap=200 turbulent), only visual comparisons, so the advertised multi-step forecasting capability is not directly validated. This does not refute the method, but it undermines the headline evidence for the central claim as currently presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LC-SVD-DLinear and LC-HOSVD-DLinear, hybrid reduced-order models that combine low-cost singular value decomposition (or its higher-order variant) with the DLinear neural architecture. The pipeline decomposes an under-resolved dataset into spatial modes and temporal coefficients, uses DLinear to forecast the temporal coefficients autoregressively, and then reconstructs high-resolution snapshots by multiplying the forecast coefficients by reconstructed spatial modes. The method is demonstrated on a three-dimensional laminar cylinder wake at Re=220 and an experimental turbulent cylinder wake at Re=2600, with reported test reconstruction errors of RRMSE=1.384% and 0.571% for the laminar case and 11.554% and 10.571% for the turbulent case using LC-SVD-DLinear and LC-HOSVD-DLinear, respectively. The paper also presents extended forecasts of 1000 (laminar) and 200 (turbulent) snapshots and states that the approach provides uncertainty quantification and low computational cost.","tokens_in":22394,"tokens_out":6150,"duration_ms":60444,"significance":"If the evidence were fully out-of-sample, the proposed hybrid approach would be a useful contribution to sparse-sensor flow forecasting: the combination of LC-SVD/LC-HOSVD with a lightweight linear forecaster is sensible, the DLinear temporal extrapolation is a genuine held-out forecast of the temporal coefficients, and the LC-HOSVD variant appears to be new. The reported low reconstruction errors on the test sets would support the central claim of forecasting high-resolution snapshots from sparse measurements. However, as presented, the headline test errors are weakened by a basis-leakage issue, the abstract's promise of uncertainty quantification is not fulfilled, the extended forecasts are only visually assessed, and no runtime or baseline comparisons support the 'low-cost' claim. These issues are fixable within the scope of the manuscript, so the underlying idea remains viable after revision.","major_comments":[{"comment":"The reported test-set RRMSE values are not out-of-sample evidence for forecasting because the spatial modes W_rec are computed from the entire input dataset before the train/validation/test split. Section 2.4 states that the temporal coefficients are split only after LC-SVD has reconstructed the spatial modes and temporal coefficients, and Eq. (2.12) constructs W_rec from the full snapshot matrix. Consequently, the spatial structure of the test-period snapshots has already informed W_rec, and the test reconstruction is a projection onto a basis that has seen those snapshots. This is especially consequential for the turbulent case, where only 6 modes are retained. The authors should recompute W_rec and T_rec using only the training snapshots, freeze those quantities, and then reconstruct the held-out test snapshots; only then would the reported RRMSEs (1.384%, 0.571%, 11.554%, 10.571%) validate the forecasting claim.","section":"§2.4 and Eq. (2.12)"},{"comment":"The abstract promises that the forecasting and reconstruction results are evaluated 'including uncertainty quantification,' but no uncertainty intervals, confidence bands, ensemble statistics, or any other UQ tool appears in the error-analysis section or in the results. All reported errors are single point estimates (e.g., MAE=0.454, RRMSE=1.384%, Wasserstein distance at a single snapshot). If uncertainty quantification is part of the contribution, the manuscript must define how intervals are constructed and report them for the test-set reconstructions; otherwise the abstract should be revised to not promise UQ.","section":"Abstract and §2.7/§4"},{"comment":"The multi-step forecasting capability is not quantitatively validated. The extended forecasts (Nsnap=1000 for the laminar case, Nsnap=200 for the turbulent case) are supported only by visual inspection of selected snapshots (Figs. 12-13, 19-20, 26-27, 33-34) and qualitative statements that velocity ranges remain stable. No error metric is reported for these extrapolated snapshots against any ground truth. Because the central claim includes forecasting many steps ahead, the authors should provide quantitative error measures on a held-out portion of the data, or clearly state that these extended forecasts are illustrative and not error-assessed.","section":"§4.1 and §4.2, extended forecasts"},{"comment":"The 'low computational cost' part of the central claim is asserted but not measured for the hybrid pipeline. The paper cites speedups for LC-SVD from Ref. [56] but reports no runtime, memory, or FLOP comparisons for LC-SVD-DLinear or LC-HOSVD-DLinear against standard SVD/HOSVD, against DLinear applied directly to full-resolution data, or against a POD-DLinear baseline. Without such measurements, the reader cannot verify the claimed cost advantage. Adding a small benchmark table with training time, inference time, and memory for the full pipelines would directly support the paper's central claim.","section":"§1 and §4, computational-cost claim"}],"minor_comments":[{"comment":"The notation in Eqs. (2.12) and (2.13) is ambiguous: the superscripts on the snapshot matrix are not clearly defined, and it is not specified whether the reconstruction uses the low-resolution snapshot matrix or a paired high-resolution snapshot matrix. Please clarify the dimensions and state explicitly which data matrix is used in each equation.","section":"§2.2, Eqs. (2.12)-(2.13)"},{"comment":"The conclusions state that 'the effects of the number of retained modes on the models performance accuracy has also been tested,' but no such sensitivity experiment appears in Section 4. Either add the experiment or remove the claim from the conclusions.","section":"§5 and §4"},{"comment":"There are several cross-reference errors and typos: the text refers to 'test data predictions in fig. 12 and fig. 19' where it likely means Figs. 9 and 16; the turbulent HOSVD subsection refers to 'fig. 26' and '27' where it means Figs. 33 and 34; 'Hofp bifurcation' should be 'Hopf bifurcation'; 'sings' in Section 2.2 should be 'signs'; and 'M SE' should be 'MSE.'","section":"Various"},{"comment":"The abstract says DLinear enables the model to 'capture the non-linear dynamics of the temporal data,' but DLinear is a linear model. Please clarify that the nonlinearity comes from the decomposition and the autoregressive rolling-window loop, not from the linear layers themselves.","section":"Abstract and §2.5"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the train/test protocol: the spatial basis is constructed from the full dataset, so the reported test RRMSEs are not out-of-sample. This is likely an oversight rather than intentional, because the temporal-coefficient forecast itself is held out. The fix is straightforward—rebuild W_rec on the training split only—but it changes the headline numbers. The missing uncertainty quantification and the absence of any computational-cost benchmark are also important for the journal's readers. I do not see a fundamental flaw in the method, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read of arXiv:2411.17433. The genuinely new piece is LC-HOSVD, a tensor extension of their earlier LC-SVD, and the specific pairing of LC-SVD/HOSVD with DLinear for forecasting from sparse sensors. The pipeline is internally coherent and the two cylinder benchmarks are standard. They report RRMSE numbers for test-set reconstructions (1.384%, 0.571%, 11.554%, 10.571%).\n\nThe main problem is that those numbers are not clean out-of-sample evidence. The spatial modes W_rec are reconstructed from the entire snapshot matrix (eq. 2.12) before the temporal coefficients are split into train/validation/test (Sec. 2.4). So the test snapshots' spatial structure has already informed the basis. Only the DLinear temporal step is genuinely extrapolating. The reported RRMSE is therefore optimistic — especially in the turbulent case, where only 6 modes are retained and the basis could be encoding test-period coherent structures. This is fixable: recompute the basis from training snapshots only and project the test snapshots onto that fixed basis. But as it stands, it undermines the central claim.\n\nOther soft spots: no baselines (POD + linear regression, DLinear-only, etc.), no error bars or repeated runs despite the abstract promising uncertainty quantification (what's given is a Wasserstein distance and worst-snapshot histograms — diagnostics, not UQ), inconsistent forecast-horizon statements (intro says 100x input, laminar run is 1000 steps from L=15, turbulent is 200 from L=100, conclusion says 30x), and no code or data. Hyperparameters are tuned per dataset with no sensitivity analysis. The long-horizon forecasts (1000/200 snapshots) are only shown visually, never quantified.\n\nCredit where due: the writing is clear, the method is reproducible in principle from the equations, the literature coverage is adequate, and LC-HOSVD follows naturally from LC-SVD.\n\nThis is for people working on sparse-sensor reconstruction and ROM forecasting. They'll get a useful pipeline description and a cautionary example of validation leakage.\n\nRecommendation: send it to peer review, but tell the authors the out-of-sample issue must be fixed before any claim about forecasting unseen states is accepted.","headline":"The pipeline is coherent and LC-HOSVD is a real extension, but the headline test errors aren't clean out-of-sample because the spatial basis is fit to the full dataset, so the forecasting claim is over-supported.","tokens_in":22949,"tokens_out":2317,"would_cite":false,"duration_ms":24896,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid model combining low-cost singular value decomposition with the DLinear network can forecast high-resolution fluid flow snapshots from under-resolved or sparse measurements, with reconstruction errors as low as 0.571% on a laminar…","keywords":["low-cost singular value decomposition","low-cost HOSVD","DLinear","hybrid reduced order model","sparse sensor measurements","cylinder wake forecasting","data-driven fluid dynamics"],"falsifier":"Run the pipeline on a flow that undergoes a controlled change of regime after the training window, such as a cylinder wake with a sudden Reynolds-number change or an oncoming gust, and measure the RRMSE of the forecast snapshots against a full simulation. If the error grows sharply when the new coherent structure appears, that confirms the fixed-basis assumption is the limiting factor; if the error stays near the reported 0.5-12% levels, the basis is more transferable than assumed.","tokens_in":21936,"feed_emoji":"🌊","tokens_out":10194,"duration_ms":81745,"temperature":0.7,"pith_summary":"The paper proposes a hybrid forecasting pipeline that takes under-resolved or sparse flow measurements, reconstructs clean high-resolution spatial modes and temporal coefficients using low-cost singular value decomposition, feeds the temporal coefficients into the shallow DLinear network to predict future values, and multiplies the forecast by the reconstructed modes to produce new high-resolution snapshots. The claim is that this works at low computational cost because the expensive operations are done on the reduced, low-dimensional data. The authors validate the idea on a three-dimensional laminar cylinder wake and on an experimental turbulent cylinder wake, reporting test reconstruction relative root-mean-square error (RRMSE) of 1.384% and 0.571% for the laminar case and 11.554% and 10.571% for the turbulent case, respectively for LC-SVD-DLinear and LC-HOSVD-DLinear. A reader should care because the method promises high-resolution flow forecasting from a handful of sensors, which would cut the cost of experiments and simulations.","feed_headline":"Hybrid SVD network forecasts high-res flow from sparse measurements","feed_subtitle":"Testing on cylinder wakes: reconstruction errors as low as 0.571% laminar, 10.571% turbulent.","key_machinery":"The load-bearing object is the reconstructed spatial modes matrix $W^{\\mathrm{rec}}$ together with the reconstructed temporal coefficients $T^{\\mathrm{rec}}$, obtained from an under-resolved snapshot matrix by the LC-SVD reconstruction formulas (Eqs. 2.12-2.14). The temporal coefficients are then decomposed by DLinear into trend and seasonality using an average-pooling layer, forecast autoregressively with one linear layer per component, and recombined as $\\hat{V} = W^{\\mathrm{rec}} \\bar{\\Sigma} (\\hat{T}^{\\mathrm{rec}})^\\top$ to form new snapshots. LC-HOSVD applies a component-wise tensor decomposition before the same LC-SVD reconstruction, allowing mode filtering per spatial component.","core_discovery":"The central discovery is that the temporal coefficients produced by a low-cost SVD or HOSVD reconstruction form a smooth, low-dimensional time series that a simple linear decomposition network can forecast accurately, and that this forecast can be lifted back to full spatial resolution using the reconstructed spatial modes. LC-SVD reconstructs the spatial modes $W^{\\mathrm{rec}}$ and temporal coefficients $T^{\\mathrm{rec}}$ from under-resolved data using Eqs. (2.12)-(2.14); DLinear splits $T^{\\mathrm{rec}}$ into trend and seasonality via an average-pooling decomposition, forecasts each component with one linear layer per mode, and the autoregressive predictions are recombined with $W^{\\mathrm{rec}}$ and the singular values to form future snapshots. LC-HOSVD performs a component-wise tensor decomposition before the same reconstruction step, which the paper finds gives cleaner mode filtering and slightly lower reconstruction errors.","pith_inferences":["Editorial extension: because the fixed spatial basis is built from paired training data, the method should be most trustworthy while the flow's coherent structures remain unchanged; a flow that changes topology after the training window would likely produce growing forecast error.","Editorial extension: the same pipeline could be tested on other quasi-periodic flows, such as flapping wings or bluff bodies at different Reynolds numbers, where DLinear's trend-seasonality decomposition should be able to exploit the periodic temporal coefficients.","Editorial extension: an explicit comparison against training DLinear directly on full-resolution data, or against a recurrent network on the same temporal coefficients, would isolate how much of the accuracy comes from the SVD compression and denoising and how much from the linear forecaster.","Editorial extension: the reported compression ratios (17,066 and 835) suggest that the forecast stage's memory cost is dominated by the number of retained modes, so the method should scale to larger three-dimensional meshes as long as the mode count stays small."],"forward_implications":["Forecasts of 1000 snapshots for the laminar wake and 200 for the turbulent wake are generated from input sequences of length 15 and 100, respectively, with no visible drift in the velocity range.","Because the computation happens on 45 or 40 sensor points instead of the full mesh, the forecast step is cheap regardless of the output resolution.","The method works on both numerical and experimental data, and on both laminar and turbulent flows.","LC-HOSVD-DLinear gives lower reconstruction error than LC-SVD-DLinear in both test cases: 0.571% versus 1.384% for the laminar wake and 10.571% versus 11.554% for the turbulent wake.","For turbulent data, the paper reports that accuracy drops as more modes are retained, so the pipeline relies on keeping only a few robust modes."],"supporting_citations":[{"why":"supplies the LC-SVD reconstruction method that the entire pipeline is built on.","marker":"[56]"},{"why":"introduces DLinear, the trend-seasonality linear network used to forecast the temporal coefficients.","marker":"[57]"},{"why":"provides the optimal sensor placement used to generate under-resolved versions of high-resolution inputs.","marker":"[23]"},{"why":"is the source of the three-dimensional laminar cylinder wake dataset used as the numerical test case.","marker":"[38]"},{"why":"is the source of the experimental turbulent cylinder wake dataset used as the real-data test case.","marker":"[50]"},{"why":"establishes the multilinear singular value decomposition that LC-HOSVD extends to the low-cost setting.","marker":"[66, 67]"}],"fun_headline_variants":["SVD-DLinear: cheap sparse-data flow forecasting","Low-cost hybrid model forecasts fluid flow from sparse data","Sparse data to full flow via SVD and a shallow net","Physics-based SVD-DLinear cuts cost of flow forecasting","Under-resolved flow data? SVD-DLinear forecasts full field"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the high-resolution spatial modes $W^{\\mathrm{rec}}$ learned from paired training data remain a valid basis for the future snapshots being forecast; if the flow's coherent structures change after the training window, those modes cannot represent the new snapshots and the forecast error will grow regardless of how well the temporal coefficients are predicted.","fun_headline_variants_meta":{"raw":{"variants":["SVD-DLinear: cheap sparse-data flow forecasting","Low-cost hybrid model forecasts fluid flow from sparse data","Sparse data to full flow via SVD and a shallow net","Physics-based SVD-DLinear cuts cost of flow forecasting","Under-resolved flow data? SVD-DLinear forecasts full field"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2138,"prompt_tokens":1000,"completion_tokens":1138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1055}},"tokens_in":616,"tokens_out":1138,"duration_ms":11506,"temperature":1.0,"reasoning_tokens":1055,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:06:34.621278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a flow that undergoes a controlled change of regime after the training window, such as a cylinder wake with a sudden Reynolds-number change or an oncoming gust, and measure the RRMSE of the forecast snapshots against a full simulation. If the error grows sharply when the new coherent structure appears, that confirms the fixed-basis assumption is the limiting factor; if the error stays near the reported 0.5-12% levels, the basis is more transferable than assumed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces DLinear, the trend-seasonality linear network used to forecast the temporal coefficients."},{"cited_title":"de Silva, K","cited_arxiv_id":null,"evidence_quote":"provides the optimal sensor placement used to generate under-resolved versions of high-resolution inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the source of the three-dimensional laminar cylinder wake dataset used as the numerical test case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the source of the experimental turbulent cylinder wake dataset used as the real-data test case."}],"review_version":1}