{"id":"38217579-7bba-4b53-aad1-772894084e62","arxiv_id":"2504.21328","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Switching between coarse and fine collocation datasets during PINN training, like a multigrid V-cycle, improves accuracy by up to 64 percent on lid-driven cavity benchmarks.","lead":"Training a PINN on multiple datasets of different collocation density, switched in a V-cycle pattern, improves accuracy on high-frequency and stiff PDE benchmarks like lid-driven cavity flow. The paper offers a simple training schedule that may help PINNs reach higher Reynolds numbers without redesigning the network.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 30–60% gain is not causally attributed to dataset switching: no epoch-matched or scheduler-matched control exists, so per-level schedules or longer training could explain it.","rationale":"I agree with the reader's conditional verdict. The empirical demonstrations are plausible: the ODE and Smith-Hutton results keep total epochs comparable and show improvement, and the Re=400 comparison includes several sequence variants. The load-bearing weakness is causal attribution: the method as described changes dataset composition, epochs, and per-level optimizer/scheduler settings simultaneously, and the frequency mechanism in Section II.A is asserted rather than measured. The paper's own Section IV defers numerical frequency analysis, and no code or multi-seed statistics are provided. A no-switch control with matched schedule and epoch budget would settle whether the active ingredient is the multi-level collocation data or simply the training schedule; until then, conditional acceptance is appropriate. The concern does not justify rejection because there is independent evidence of improvement in the same-epoch Smith-Hutton experiment, and no reported result directly contradicts the central claim; however, causal attribution and reproducibility remain unverified.","tokens_in":15797,"tokens_out":7300,"duration_ms":85415,"concrete_test":"Run the Re=400 Random V-cycle of Table VIII with all three levels using the same 6565-point D1 dataset (no dataset switch), keeping per-level durations and scheduler settings identical, for 5 random seeds; report mean±std MAE against the Ghia reference. Run the same control on the Smith-Hutton V-cycle (replace D2/D3 with D1) with 5 seeds. If the no-switch control matches the reported 4.25e-03/1.38e-02 MAE or the Smith-Hutton 2.10e-03 Rel. L2 within seed variance, the claimed benefit comes from the epoch/scheduler schedule rather than from multi-level collocation datasets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central '30–60% improvement' claim is not yet attributable to multi-level datasets. In the Re=400 headline comparison (Table IX), the random V-cycle uses 163,000 epochs while the single-level baseline uses 104,000, and every proposed variant also uses 162,000–184,000 epochs; no comparison is run at a fixed epoch budget. Appendix Tables XIII–XVII show per-level scheduler factors and learning rates, so the V-cycle could benefit from learning-rate cycling rather than from the asserted low-frequency content of smaller collocation sets. The ODE and Smith-Hutton sections keep total epochs equal, but they are single runs without seed variability, and the ODE still changes scheduler settings between D1 and D2 (Table XIII). Section II.A's claim that 'spectral bias of smaller collocation sets results in lower training frequencies' is never measured; Section IV explicitly defers numerical analyses of training frequencies to future work. The Re=5000 result (Table XII) is likewise compared only against other papers, not against the same architecture trained without the proposed schedule, so the multilevel contribution there is untested. Without an ablation that holds total epochs and optimizer/scheduler trajectory fixed while toggling the dataset sequence, the accuracy gain cannot be assigned to the proposed mechanism, and the method's portability to new problems remains uncertain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-level dataset training procedure for Physics-Informed Neural Networks (PINNs), inspired by multigrid ideas. Instead of training on a single collocation dataset, the method alternates between datasets of different collocation densities, each equipped with its own optimizer and learning-rate scheduler, using a transfer-learning-style handoff. The authors report accuracy gains on a 1D high-frequency ODE, the Smith-Hutton convection-diffusion problem, and steady lid-driven cavity flows at Re=400, 1000, and 5000, claiming 30--60% accuracy improvements and attributing the gains to mitigating spectral bias through exposure to different frequency scales. The paper also combines the approach with curriculum training for the higher-Reynolds-number cavity cases and compares against published PINN results and CFD references.","tokens_in":16122,"tokens_out":5794,"duration_ms":61516,"significance":"If the claimed gains are causal, the method is attractive because it is a training-level modification that leaves the network architecture and loss formulation unchanged, and it is compatible with transfer and curriculum learning. The paper's use of external benchmarks (Ghia et al., Smith-Hutton) and the detailed per-level training tables in the appendix are positive features. However, the current evidence does not establish the mechanism or the headline improvement magnitude: the central cavity comparison is confounded with total epoch count and schedule changes, the ODE benchmark contains an incorrect derivative that invalidates the experiment as written, run-to-run variability is not reported, and the spectral-bias mechanism is asserted rather than measured. The Smith-Hutton experiment provides some controlled evidence (equal total epochs, equal per-level optimizers and schedulers), but it is a single run and shows a modest improvement. With corrected benchmarks and controlled ablation experiments, the contribution would be of genuine interest to the PINN community.","major_comments":[{"comment":"The derivative in Eq. (9) is incorrect: d/dx sin(10x) = 10 cos(10x), not 10 sin(10x). As written, the residual equation is inconsistent with the stated exact solution f(x)=sin(10x), and the boundary conditions f(0)=f(2π)=0 do not determine the stated solution. The 64% test-error improvement in Table IV is therefore not a valid validation of the method until Eq. (9), the corresponding loss definition, and the reported experiment are corrected and rerun.","section":"§III.B, Eq. (9)"},{"comment":"The headline Re=400 comparison is not controlled. The 'Random, 1 level' baseline runs 104,000 epochs, while 'Random, V-cycle' runs 163,000 epochs and the other variants run 162,000--184,000 epochs; additionally, the per-level schedulers in Table XV differ across levels. The claimed 64.1% error decrease could therefore be due to extra training, learning-rate cycling, or the combined schedule rather than to dataset switching. An epoch-matched baseline with the same optimizer/scheduler trajectory, or at least a matched total number of parameter updates, is required to attribute the gain to the proposed mechanism.","section":"§III.D, Table IX"},{"comment":"The mechanistic premise that smaller collocation sets expose the network to lower-frequency components is never tested. Section II.A asserts this spectral relationship, but Section IV explicitly defers 'numerical analyses on training frequencies' to future work. Without frequency-content measurements or a mechanism-based ablation, the analogy to multigrid remains a hypothesis rather than a demonstrated explanation of the observed accuracy gains.","section":"§II.A and §IV"},{"comment":"The abstract claims the method avoids time spent 'fine-tuning of neural network structures, loss weights as well as hyperparameters,' but the experiments hand-select per-level optimizers, learning rates, scheduler factors, stopping criteria, and network widths (Tables XIII--XVII), and different architectures are used for different problems. This overstates the turnkey character of the proposed approach.","section":"Abstract and §II.A"},{"comment":"The Re=5000 result is not evidence for the multi-dataset procedure: it compares the proposed pipeline against other published methods, but it does not include a same-architecture, same-epoch ablation trained without the V-cycle/curriculum schedule. The contribution of dataset switching at high Reynolds number is therefore untested.","section":"§III.D, Table XII"}],"minor_comments":[{"comment":"Eq. (11a) appears to contain a typo in the diffusion term: '∂²φ/∂xy' should be '∂²φ/∂x² + ∂²φ/∂y²'. Also, 'Pelect number' should be 'Péclet number'.","section":"§III.C, Eq. (11)"},{"comment":"The text says the proposed method improves accuracy by 20%, but the relative L2 reduction in Table VI is (2.54e-3 - 2.10e-3)/2.54e-3 ≈ 17%; the reported percentage should be reconciled.","section":"§III.C, Table VI"},{"comment":"The column 'Error decrease' is reported with negative percentages, which is confusing because a decrease should be a positive number; the formula also averages the relative reductions of u and v without justification. Please clarify the sign convention and the aggregation.","section":"§III.D, Table IX"},{"comment":"All experiments appear to be single runs without seed variability. Reporting mean and standard deviation over at least three seeds, particularly for the ODE and Smith-Hutton cases, would calibrate the magnitude of the reported gains.","section":"§III.B--D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a machine-learning journal, and the central idea is worth pursuing, but the evidence as presented is not yet sufficient. The incorrect derivative in Eq. (9) and the confounded epoch counts in Table IX will require new experiments, not just textual revision. The DWroot weighting is cited to a same-group master's thesis (ref. 54); this is not a reason to reject, but the editors may wish to verify public accessibility of the thesis. The paper does not release code; given the large number of hand-set per-level hyperparameters, code availability would substantially improve reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look for the training schedule alone: single network, sequential switching between collocation datasets of different densities, with per-dataset optimizers and schedulers in V- or W-cycle patterns. That is an adoptable idea that targets spectral bias without changing network architecture. The single network also distinguishes it from MPINNs, which use multiple sub-networks.\n\nThe strongest evidence is epoch-matched. On the 1D ODE, Table IV shows Rel. L2 dropping from 1.15e-01 to 4.16e-02 over the same 30,000 epochs. On Smith-Hutton, the V-cycle uses 50,000 total epochs just like the uniform baseline, and Table VI shows a 17% improvement (2.54e-03 to 2.10e-03). These are real, though more modest than the abstract suggests. The Re=400 cavity results are impressive on the surface—up to 64% error decrease—but those runs use 163,000–184,000 epochs versus 104,000 for the baseline. No fixed-epoch control is reported, so the gain could come from longer training or from the per-level learning-rate schedules rather than from dataset switching itself. That is the paper's main soft spot.\n\nThe mechanism claimed in Section II.A—that smaller collocation sets expose the network to lower frequencies—is asserted but not measured. The authors acknowledge this in Section IV, which is honest, but it means portability to new problems is uncertain. There are also smaller issues: Eq. 9 gives the wrong derivative for sin(10x), all results are single runs with no error bars, and no code is provided. The citation pattern is acceptable; the DWroot weighting is self-cited (ref 54) but it is an input, not fitted to the benchmark outputs.\n\nOverall, the central idea is plausible and the epoch-matched results do support a real effect. The paper deserves serious peer review, and a competent referee should ask for a clean ablation that toggles the dataset sequence while holding total epochs and the scheduler trajectory fixed. That would settle whether the multilevel schedule is the cause or just a proxy for more training. I would bring it to a reading group that works on PINN training methods, and I would not mind citing it as related work if I were writing on spectral bias in PINNs.","headline":"A simple, plausible dataset-switching schedule for PINNs with real but modest epoch-matched gains; the larger 30–60% claims are not yet causally attributed to the method.","tokens_in":16624,"tokens_out":1776,"would_cite":false,"duration_ms":21572,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M99","76D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Multi-level dataset cycling cuts PINN errors by 30–60 percent and reaches Re=5000.","keywords":["physics-informed neural networks","multi-level dataset training","spectral bias","multigrid-inspired training","lid-driven cavity flow","curriculum training","transfer learning"],"falsifier":"Hold total epochs and the sequence of learning-rate schedules fixed but replace the multi-dataset V-cycle with a single collocation set; if accuracy matches the multi-dataset run, dataset switching is not the cause. Then check the network's prediction spectrum after each level: the mechanism predicts low frequencies dominate on the smallest set and high frequencies appear on the largest.","tokens_in":15593,"feed_emoji":"🧮","tokens_out":21649,"duration_ms":185783,"temperature":0.7,"pith_summary":"The paper tries to establish that Physics-Informed Neural Networks can be made substantially more accurate on stiff and high-frequency PDEs by a training-level trick: instead of feeding one fixed set of collocation points (locations where the PDE residual is evaluated), the network is trained successively on several datasets of different point densities, with the order and pace controlled by multigrid-inspired V- and W-cycles. The authors' numerical evidence covers a 1D high-frequency ODE, a 2D convection-diffusion benchmark, and the steady lid-driven cavity flow, where the switching method reduces errors by 30–60 percent relative to a single-dataset baseline and, combined with curriculum training over increasing Reynolds numbers, produces a usable solution at $Re=5000$. The payoff of the claim is practical: it offers a way to improve PINN accuracy without restructuring the network, tuning loss weights, or hand-crafting features, using only the ordering and density of the training collocations.","feed_headline":"Switching collocation sets reduces PINN error by 30–60%","feed_subtitle":"V-cycle training inspired by multigrid predicts lid-driven cavity flow at Re=5000.","key_machinery":"The central object is a multi-level dataset training schedule: a fixed set of collocation datasets $D_1, D_2, D_3$ with decreasing numbers of points (e.g., 6670, 1770, 435 for the convection-diffusion case; 6565, 1685, 445 for the cavity at $Re=400$), cycled in sequences such as V-cycle ($D_1\\to D_2\\to D_3\\to D_2\\to D_1$) or W-cycle. The mechanism is that each level is trained with its own Adam optimizer and learning-rate scheduler and then connected to the next level through transfer learning, so that the network sees the same architecture but different collocation densities. For the cavity problem, the composite loss weights are fixed by the dimensional-analysis DWroot rule rather than by adaptive weighting. The intended function of the schedule is multigrid-like frequency separation: small datasets are supposed to teach low-frequency error components and large datasets high-frequency ones, so that alternation removes both kinds of error.","core_discovery":"The paper's central claim is that spectral bias in PINN training can be counteracted by cycling through collocation datasets of different sizes, treating each switch as a transfer-learning step with its own optimizer and learning-rate scheduler. On the 1D equation $f(x)=\\sin(10x)$ in $[0,2\\pi]$, a two-level schedule (64 points, then 32, then 64) lowers the relative $L^2$ test error from $1.15\\times10^{-1}$ to $4.16\\times10^{-2}$ in the same 30,000 epochs. On the 2D convection-diffusion benchmark, a three-level V-cycle (6670, 1770, 435, 1770, 6670 points) improves the relative $L^2$ error from $2.54\\times10^{-3}$ to $2.10\\times10^{-3}$ at fixed 50,000-epoch cost. For the steady lid-driven cavity at $Re=400$, random V-cycling over three collocation sets reduces the best mean absolute error by 64 percent versus a single random dataset, and with curriculum training through $Re=400,1000,2000,3200,5000$ the model reaches about $7\\times10^{-2}$ mean absolute error at $Re=5000$. All runs keep the same MLP with a sine first layer and Swish hidden layers; only the dataset sequence and per-level schedules change.","pith_inferences":["A control that replays the same per-level optimizer and scheduler steps on a single dataset would determine whether dataset switching, rather than the added optimization structure, causes the gains; the paper does not report such a control.","If the multigrid analogy is exact, the optimal number of levels and the collocation ratios could be derived from the problem's frequency content or from the network's NTK spectrum, turning the schedule into an automatic design.","The approach should combine with residual-based adaptive sampling—using the coarse levels to locate regions needing refinement—though the paper does not test this combination.","The reported improvement is relative to single-dataset baselines with hand-picked schedulers, so in other applications the practical gain may depend on how well the baseline is already tuned."],"forward_implications":["Because the method only changes the collocation schedule, it can be added to any existing PINN architecture without new layers, loss terms, or feature engineering.","Layered onto curriculum training, it extends PINN solvability to Reynolds numbers like $Re=5000$ where single-dataset training does not converge, and with fewer parameter updates than the published comparison methods.","The reported 30–60 percent error reductions suggest the scheme can help on any PDE with multiple frequency scales, including sharp-front transport problems and oscillatory solutions.","Per-level learning-rate schedulers become the main tuning knob, giving practitioners a concrete substitute for hand-tuning a single global schedule."],"supporting_citations":[{"why":"It establishes the low-frequency-first learning bias of neural networks that the proposed method targets.","marker":"[31]"},{"why":"It introduces the multi-scale DNN idea of learning across frequency scales that the dataset-cycling design draws on.","marker":"[39]"},{"why":"It supplies the multigrid frequency-smoothing perspective that motivates the V- and W-cycle sequences.","marker":"[40]"},{"why":"It provides the curriculum-learning strategy used in the high-Reynolds-number cavity runs.","marker":"[28]"},{"why":"It defines the transfer-learning procedure used to connect consecutive dataset levels.","marker":"[46]"},{"why":"It gives the dimensional-analysis loss weighting (DWroot) applied in all lid-driven cavity experiments.","marker":"[54]"},{"why":"It supplies the benchmark reference solution for evaluating lid-driven cavity accuracy.","marker":"[56]"}],"fun_headline_variants":["Multigrid-style data cycling cuts PINN error by 30–60%","Fix PINN spectral bias by cycling collocation sets","V-cycle training grids improve PINN accuracy up to 60%","Switch collocation sets to reduce PINN error by 30–60%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that training on a smaller collocation set exposes the network to lower-frequency components and that switching dataset sizes—rather than the extra training time, the per-level schedulers, or transfer-learning warm starts—is what produces the 30–60 percent gains.","fun_headline_variants_meta":{"raw":{"variants":["Multigrid-style data cycling cuts PINN error by 30–60%","Fix PINN spectral bias by cycling collocation sets","V-cycle training grids improve PINN accuracy up to 60%","Switch collocation sets to reduce PINN error by 30–60%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2889,"prompt_tokens":1145,"completion_tokens":1744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":761,"completion_tokens_details":{"reasoning_tokens":1667}},"tokens_in":761,"tokens_out":1744,"duration_ms":12742,"temperature":1.0,"reasoning_tokens":1667,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:06:16.234061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold total epochs and the sequence of learning-rate schedules fixed but replace the multi-dataset V-cycle with a single collocation set; if accuracy matches the multi-dataset run, dataset switching is not the cause. Then check the network's prediction spectrum after each level: the mechanism predicts low frequencies dominate on the smallest set and high frequencies appear on the largest.","supporting_citations":[{"cited_title":"Rahaman , author A","cited_arxiv_id":null,"evidence_quote":"It establishes the low-frequency-first learning bias of neural networks that the proposed method targets."},{"cited_title":"Wienands \\ and\\ author W","cited_arxiv_id":null,"evidence_quote":"It supplies the multigrid frequency-smoothing perspective that motivates the V- and W-cycle sequences."},{"cited_title":"Krishnapriyan , author A","cited_arxiv_id":null,"evidence_quote":"It provides the curriculum-learning strategy used in the high-Reynolds-number cavity runs."},{"cited_title":"\\ Chou \\ and\\ author C.-A","cited_arxiv_id":null,"evidence_quote":"It gives the dimensional-analysis loss weighting (DWroot) applied in all lid-driven cavity experiments."}],"review_version":1}