{"id":"4e002694-9277-46ad-a23d-d1be448e7c90","arxiv_id":"2507.21716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A HODMD-based reduced order model reconstructs and predicts idealized left ventricle flow from a few training cycles with reported errors below 5-10% and a speed-up near 10^5.","lead":"Researchers used a data-driven method called HODMD to build a fast surrogate model of blood flow inside idealized left ventricle models. The model reportedly reconstructs and predicts flow with errors below 5-10% and runs about 100,000 times faster than full CFD simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported sub-5%/10% errors are normalized by the worst-case error and are never compared against a trivial periodic baseline, so the zero-growth ROM's apparent accuracy may reflect the inherent cycle-to-cycle variability of the CFD data rather than predictive skill.","rationale":"The paper is a competent application of HODMD to a challenging moving-domain hemodynamics problem, and the out-of-sample prediction setup is a genuine strength. However, the headline accuracy numbers are defined in Eq. (5) relative to the maximum absolute error, and the model's core validity rests on the zero-growth assumption in Section 2.3, which forces exact periodicity onto a flow that the authors themselves show is only approximately periodic in Section 3.2. Because no baseline model is compared, the reader cannot distinguish HODMD's predictive skill from simple cycle repetition. The manuscript's own limitation statement, which sets a 5-10% baseline uncertainty margin, is in-scope evidence that reinforces this concern. The issue is significant enough to keep the verdict conditional, but not so severe as to reject the work: the method may well be useful, but its quantitative claims need to be benchmarked against the data's intrinsic variability and a nonperiodic baseline. This aligns with the reader's verdict, so I recommend no change.","tokens_in":14924,"tokens_out":4740,"duration_ms":58291,"concrete_test":"Recompute the T-3 and T-10 ROM errors in the V-10 window using a relative L2 norm, ||v - v_pred||_2 / ||v||_2, and compare them against (a) the same metric applied between CFD cycle 11 and cycle 20 (and between consecutive validation cycles) to establish the data-variability baseline, and (b) a trivial periodic baseline that repeats the last training cycle (or the average of the last two training cycles) over V-10. If the HODMD relative error is comparable to the cycle-to-cycle variability, or if the periodic-repeat baseline achieves the same or better errors, then the central accuracy claim is not supported. Also report unnormalized mean and maximum absolute errors, since Eq. (5) can obscure them.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Eq. (5), which defines E(%) = |v - v_pred| / max|v - v_pred| * 100, i.e., each error is expressed as a percentage of the largest absolute error anywhere in the field, not of the flow magnitude. A histogram showing 99% of points below 5% only means those points are within 5% of the worst error; if the worst error is large relative to characteristic velocities, the absolute accuracy can be poor. This normalization is combined in Section 2.3 with the decision to force all retained growth rates to zero (δ_m = 0). The training window [0, pT] is explicitly transient, so nonzero negative growth rates in that window encode convergence to the later periodic regime; zeroing them removes that mechanism and replaces it with an assumed exact periodicity. Section 3.2 documents that the validation cycles are not exactly periodic: the RRMSE of total kinetic energy reaches about 4%, local discrepancies can be larger, and the authors state that predictions should be interpreted within a baseline uncertainty margin of approximately 5-10%. No comparison is made against a trivial baseline that simply repeats the last training cycle over the validation window. Consequently, the reported sub-5%/sub-10% errors may be no better than the inherent cycle-to-cycle variability of the CFD data, and the zero-growth periodic model may not be responsible for the apparent success.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a reduced-order modeling framework based on higher-order dynamic mode decomposition (HODMD) for reconstructing and extrapolating intraventricular flow in two idealized left-ventricle geometries. The method uses HOSVD-based tensor compression, DMD mode extraction with spectral filtering, and enforces zero growth rates for prediction. The authors train on 1.5, 3, or 10 cardiac cycles from transient CFD simulations and validate on cycles 10–20, reporting normalized errors below 5% (Ideal 1) and below 10% (Ideal 2) and speed-ups of order 10^5.","tokens_in":15301,"tokens_out":4189,"duration_ms":45294,"significance":"The application of HODMD to cardiac hemodynamics is novel and the out-of-sample validation design is a strength: the validation window is not used in training. The method provides interpretable modes tied to vortex-ring dynamics. If the quantitative claims are substantiated with appropriate error metrics and baselines, this could be a useful tool for fast surrogate modeling. The authors also provide open code and tutorials, which supports reproducibility.","major_comments":[{"comment":"The error metric in Eq. (5) normalizes the absolute error by the maximum absolute error in the prediction window. Consequently, the statement that 'in at least 99% of spatial locations the error is below 5%' (Section 4.1) refers to 5% of the worst error, not 5% of a characteristic flow velocity. This metric cannot support the abstract's claim that reconstruction and prediction errors remain below 5%/10%. Please recompute errors using a flow-physics-based normalization (e.g., normalized RMSE by mean inlet velocity or TKE) and report the implied absolute errors; this is essential for assessing the method's accuracy.","section":"Section 2.3, Eq. (5)"},{"comment":"The zero-growth enforcement is not validated against a trivial baseline. Since the validation cycles exhibit cycle-to-cycle variability (Section 3.2 reports TKE RRMSE up to 4% and a baseline uncertainty of 5–10%), a model that simply repeats the last training cycle may achieve similar or better performance. Please add a comparison with a periodic-repeat baseline and, if possible, with a version of the ROM that retains estimated growth rates, over the same validation window. Without this, the reported accuracy cannot be attributed to the HODMD predictive mechanism.","section":"Section 2.3 and Section 3.2"},{"comment":"The tunable growth-rate threshold delta_tune is chosen separately for each case (1.0, 0.5, 0.01, with the text mentioning 0.05 in Section 3.3), and no sensitivity study is reported. Because this parameter directly controls which modes are retained, the claimed robustness of the spectral filtering is not established. Please provide a systematic variation of delta_tune (or a data-driven selection criterion) and report the resulting error distributions for at least the T-3 case.","section":"Table 2 and Section 2.3"},{"comment":"The speed-up factor compares the CFD cost on 40 CPUs with the ROM time, but the text does not state whether t_ROM includes the offline HODMD/HOSVD training cost or only the evaluation of Eq. (2). Since the abstract highlights the 10^5 speed-up, please clarify the cost accounting and, if the offline cost is excluded, report the total (offline + online) speed-up.","section":"Eq. (7) and Section 4.1"}],"minor_comments":[{"comment":"The threshold in Fig. 4 is delta_tune = 5e-2, but Table 2 lists delta_tune = 1.0, 0.5, 0.01. Clarify the relationship between these values and the selection criterion.","section":"Section 3.3"},{"comment":"The notation \\hat{bR}_1... in Eq. (4) is undefined; define the dimensions of these matrices or state that they are the blocks of the modified Koopman matrix.","section":"Section 2.2"},{"comment":"The sentence 'A tunable tolerance epsilon determines...' should specify that epsilon_1 = epsilon_2 = epsilon, as stated later, and define how epsilon maps to P_i and N.","section":"Section 2.1"},{"comment":"The expression 'p \\in R' should be 'p \\in {1.5,3,10}' or a defined set; using real numbers is imprecise.","section":"Section 3.1"},{"comment":"The histograms are based on the normalized error of Eq. (5); please state this in the captions to avoid over-interpretation.","section":"Figure 8 and 11"},{"comment":"The statement 'the ROM prediction over the same time span was completed in just 36 seconds' should specify the hardware and whether this includes data loading and mode selection.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of physics.flu-dyn and the method is presented clearly. The principal concern is that the central quantitative claims rest on an error metric that is normalized by the worst-case error; this is a fixable but substantive issue. I recommend major revision. Author self-citations are extensive but mostly appropriate for a methods paper; no concerns about novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the application, not the method: HODMD/md-HODMD is established, but this is the first time it has been pointed at intraventricular flow with moving meshes and transient cardiac cycles. The out-of-sample setup is a real strength. They train on the first p cycles (p = 1.5, 3, 10) and predict cycles 10–20, which are not used to fit the modes. The reported speed-up of about 10^5 is credible if you accept the ROM runtime as the cost of prediction after the offline training. The two geometries bracket different vortex breakdown behaviors, which gives the test some range. I also appreciate that the authors explicitly acknowledge a 5–10% baseline uncertainty from cycle-to-cycle variability; not everyone would include that.\n\nThe soft spots are real and mostly quantitative. The error metric in Eq. (5) is normalized by the maximum absolute error in the field, not by a characteristic flow velocity. So a histogram with 99% of points below 5% means those points are within 5% of the worst error; it does not mean the prediction is within 5% of the true velocity. That makes the abstract's 'below 5%' and 'below 10%' claims weaker than they read. My second concern is the absence of a trivial baseline: the validation cycles are not exactly periodic (their own TKE RRMSE shows about 4% cycle-to-cycle variation, with local discrepancies larger), so a model that simply repeats the last training cycle might achieve similar errors. The zero-growth enforcement, which is central to the long-term prediction, is an ad hoc fix, and the per-case delta_tune is chosen with knowledge of the outcomes. None of this is fatal, but it means the paper demonstrates feasibility, not robust predictive skill.\n\nWho is this for? Someone working on reduced-order models for cardiovascular flows, or applying HODMD to moving-boundary problems, will find it useful. It is a competent engineering application, not a methodological breakthrough.\n\nMy recommendation: send it to peer review. The core result—that HODMD can produce reasonable predictions of idealized LV flow from as few as three transient cycles—is probably correct, and the flaws are addressable. A serious referee should ask for error normalized by flow magnitude, a comparison against a periodic baseline, and a sensitivity study on delta_tune. Without those, the quantitative claims should be softened.","headline":"A legitimate new application of HODMD to idealized left-ventricle flow with genuine out-of-sample prediction, but the headline error numbers are normalized in a way that flatters the method and no trivial baseline is tested.","tokens_in":15759,"tokens_out":1992,"would_cite":false,"duration_ms":23260,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["47.63.Cb"],"model":"deepseek-v4-flash","headline":"A HODMD-based reduced-order model reconstructs and predicts left-ventricle flow from as few as three transient cycles, keeping errors below 5-10% and speed-ups around 10^5.","keywords":["higher order dynamic mode decomposition","reduced order model","left ventricle hemodynamics","intraventricular flow","vortex ring","computational fluid dynamics","spectral filtering","flow prediction"],"falsifier":"Compute the per-cycle error maps for the T-3 model on each of the validation cycles 10–20; if any single cycle shows more than 10% error over a substantial region, the claimed below-10% long-term prediction would not hold for that cycle. A complementary test would be to train on the first three cycles of one CFD run and predict cycles 10–20 of a second run with a small perturbation to the peak inlet velocity, checking whether the error stays below the paper's thresholds.","tokens_in":14746,"feed_emoji":"🫀","tokens_out":9704,"duration_ms":97417,"temperature":0.7,"pith_summary":"The paper claims that a reduced-order model built on higher-order dynamic mode decomposition (HODMD) can reconstruct the full three-dimensional velocity field of idealized left-ventricle flows over complete cardiac cycles and predict it for up to ten cycles beyond the training window. The model uses spectral filtering and forces all retained growth rates to zero, and with only the first three transient cycles of CFD data it keeps reconstruction and prediction errors below 5% for one ventricular geometry and below 10% for another. If correct, this would turn expensive cardiac CFD simulations into fast surrogate models—roughly $10^5$ times cheaper—while keeping a physically interpretable modal description of the flow, centered on the vortex ring. The paper presents this as the first application of HODMD to cardiac flows and investigates how many training cycles are needed.","feed_headline":"Three heart cycles of data predict ventricle flow 100,000x faster","feed_subtitle":"Trained on the first three transient cycles, it reconstructs and forecasts ventricle flow with errors below 5 and 10%.","key_machinery":"The central object is the md-HODMD decomposition of the five-dimensional velocity tensor $\\boldsymbol{\\mathcal{V}} \\in \\mathbb{R}^{3 \\times J_2 \\times J_3 \\times J_4 \\times K}$: HOSVD compresses the tensor, DMD applied to the temporal coefficients extracts modes with amplitude, growth rate $\\delta_m$, and frequency $\\omega_m$, and predictions are obtained by evaluating the modal expansion at times beyond the training window. Two controls carry the argument: spectral filtering retains only modes whose growth rates fall below a tunable threshold $\\delta_{\\mathrm{tune}}$, and the zero-growth enforcement sets $\\delta_m = 0$ for the retained modes, turning the model into a purely periodic expansion. The paper identifies the intraventricular vortex ring as the dominant structure these modes must capture, and the two idealized geometries are chosen to present different vortex-ring behavior.","core_discovery":"The central claim, stated on the paper's own terms, is that the dominant intraventricular flow physics is sufficiently periodic that a spectrally filtered HODMD model with enforced zero growth rates ($\\delta_m = 0$) can serve as a reliable long-term reduced-order model. The paper shows that the dominant frequency and its harmonics are captured accurately even from 1.5 cycles, and that enforcing zero growth removes artificial damping and improves predictions. Using the first three cycles for training, the model reconstructs and predicts the flow in validation cycles 10–20 with errors below 5% for the semi-ellipsoidal Ideal 1 geometry and below 10% for the rounded Ideal 2 geometry, with speed-up factors of $1.5 \\times 10^5$ and $1.7 \\times 10^5$, respectively, relative to the full-order CFD runs.","pith_inferences":["The zero-growth assumption implies the ROM will degrade as beat-to-beat variability increases; a natural next test is patient-specific geometries with physiological cycle-to-cycle variation, where the retained mode set may need to grow or allow small nonzero growth rates.","Because the reported errors are global relative errors, clinically relevant local quantities such as wall shear stress or vortex-breakdown timing could be less accurate; a careful validation on those quantities would be needed before use in diagnosis.","A stronger falsification would be to train on the first three cycles of one simulation and predict the later cycles of a second simulation with a slightly perturbed inflow waveform, testing whether the extracted modes are stable features of the flow rather than artifacts of one CFD run.","The same spectral-filtering plus zero-growth recipe could be transferred to other periodic physiological flows, such as aortic or pulmonary flows, where a dominant fundamental frequency and harmonics are expected."],"forward_implications":["Training on only three transient cardiac cycles is enough to build a ROM that predicts cycles 10–20, so the long CFD simulations used for validation can be replaced by seconds-long ROM predictions in surrogate-database generation.","With speed-ups of at least $10^5$, parametric studies and real-time hemodynamic evaluation become feasible on ordinary computing hardware rather than HPC clusters.","The method works on two geometries with different vortex-ring life cycles, early breakdown versus a stable ring traveling to the apex, suggesting the approach is not tied to one flow regime.","The modal representation gives an interpretable physical picture: dominant modes sit at the cardiac frequency and its harmonics, so the ROM's output can be understood in terms of coherent flow structures rather than as a black box.","The comparison between 1.5, 3, and 10 training cycles establishes a practical rule of thumb: at least three cycles are needed, while one and a half is not reliable for long-term prediction."],"supporting_citations":[{"why":"Supplies the higher-order dynamic mode decomposition algorithm and its HOSVD-based dimensionality reduction, which the ROM is built on.","marker":"[20]"},{"why":"Provides the tensor-based multi-dimensional implementation of HODMD used to process the five-dimensional velocity data.","marker":"[22]"},{"why":"Defines standard DMD, the modal decomposition framework that HODMD extends for extracting coherent structures and their temporal evolution.","marker":"[25]"},{"why":"Provides the Ideal 1 semi-ellipsoidal left-ventricle geometry used as the first test case.","marker":"[28]"},{"why":"Provides the Ideal 2 rounded ventricular geometry used as the second, slower-flow test case.","marker":"[29]"},{"why":"Documents the vortex-ring formation and breakdown physics in these two geometries that the ROM is designed to capture and predict.","marker":"[15]"},{"why":"Describes the CFD modeling setup and validation for left-ventricle flow simulations, supporting the reliability of the reference data.","marker":"[35]"}],"fun_headline_variants":["Just 3 heart cycles yield 100,000x faster ventricle flow","From 3 cycles, HODMD forecasts ventricle flow 100,000x faster","Ventricle flow model uses 3 cycles for 100,000x speed-up","HODMD spins 3 cycles into 100,000x faster flow forecasts","3 cardiac cycles forecast flow 100,000x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the flow during validation cycles 10–20 is governed by exactly the same periodic modes extracted from the first training cycles, so forcing all retained growth rates to zero is a valid model; the paper itself reports cycle-to-cycle variability in total kinetic energy up to 4%, so this periodicity is only approximate.","fun_headline_variants_meta":{"raw":{"variants":["Just 3 heart cycles yield 100,000x faster ventricle flow","From 3 cycles, HODMD forecasts ventricle flow 100,000x faster","Ventricle flow model uses 3 cycles for 100,000x speed-up","HODMD spins 3 cycles into 100,000x faster flow forecasts","3 cardiac cycles forecast flow 100,000x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001573,"raw_usage":{"total_tokens":6300,"prompt_tokens":991,"completion_tokens":5309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":5208}},"tokens_in":607,"tokens_out":5309,"duration_ms":45509,"temperature":1.0,"reasoning_tokens":5208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:26:53.809956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the per-cycle error maps for the T-3 model on each of the validation cycles 10–20; if any single cycle shows more than 10% error over a substantial region, the claimed below-10% long-term prediction would not hold for that cycle. A complementary test would be to train on the first three cycles of one CFD run and predict cycles 10–20 of a second run with a small perturbation to the peak inlet velocity, checking whether the error stays below the paper's thresholds.","supporting_citations":[{"cited_title":"Hetherington, A","cited_arxiv_id":null,"evidence_quote":"Provides the tensor-based multi-dimensional implementation of HODMD used to process the five-dimensional velocity data."},{"cited_title":"Zheng, J","cited_arxiv_id":null,"evidence_quote":"Provides the Ideal 1 semi-ellipsoidal left-ventricle geometry used as the first test case."},{"cited_title":"Vedula, S","cited_arxiv_id":null,"evidence_quote":"Provides the Ideal 2 rounded ventricular geometry used as the second, slower-flow test case."},{"cited_title":"Lazpita, M","cited_arxiv_id":null,"evidence_quote":"Documents the vortex-ring formation and breakdown physics in these two geometries that the ROM is designed to capture and predict."},{"cited_title":"Lazpita, A","cited_arxiv_id":null,"evidence_quote":"Describes the CFD modeling setup and validation for left-ventricle flow simulations, supporting the reliability of the reference data."}],"review_version":1}