{"id":"de11e365-eb7e-4ac9-aa87-562556dcf744","arxiv_id":"2601.16598","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Replay training on the DNN-MG's own perturbed trajectories removes the long-time instability of the hybrid Navier-Stokes solver, while Transformers or larger patches improve accuracy on unseen geometries.","lead":"An existing hybrid neural-network/finite-element flow solver is made stable for long simulations by retraining on trajectories of its own errors, and by giving the network more spatial context with Transformers or larger patches. The result is a 2D Navier-Stokes solver that corrects a cheap coarse solution to near-fine accuracy at roughly 2-2.6x lower computational cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy claims rest on self-consistent replay targets; Sec 7.3 admits no independent mesh-refinement validation, so low divergence/local error may certify stability on a wrong branch.","rationale":"The reader's weakest assumption is indeed the most load-bearing issue: the evaluation relies on fine-grid solutions as the gold standard, but those fine-grid solutions are generated from the same NN-perturbed state, making the metrics self-referential. The manuscript's own Sec 7.3 admits that mesh-refinement validation is not possible for sensitive cases, which directly undermines the title's 'robust and stable' and the abstract's accuracy claims. The paper is otherwise well-documented, reproducible (Zenodo code/weights), and the replay-training stability improvements are internally consistent; the concern is not that the method fails, but that the reported 'accuracy' gains may not reflect physical convergence. The reader already conditioned acceptance on amending claims and adding error bars, so no change to the verdict is needed. The proposed concrete test—running ro1 against a finer reference and checking chaotic cases against an ensemble of perturbed fine simulations—would settle whether the self-consistency concern actually lands.","tokens_in":30631,"tokens_out":3946,"duration_ms":41966,"concrete_test":"Re-run the replay-trained MLP and Transformer on ro1 (stable periodic flow) for at least 10 periods and compare mean drag, lift amplitude, and lift frequency against a reference computed on the next finer level (L+2) with halved time step. The paper already cites such finer-level convergence (Sec 6.3: ampJl→0.578, meanJd→0.493), so this is feasible. If hybrid statistics match the finer reference within the coarse-to-fine gap, the concern is resolved. For the sensitive sq9* case, run 10 fine-grid reference simulations with O(1e-6) random perturbations of the initial condition and check whether the hybrid trajectory's time-averaged velocity and lift remain within the ensemble spread; if it leaves the spread, low divergence is not evidence of physical accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that replay training 'prevents instabilities and also improves every metric' and that Transformers 'improve the accuracy' — is only as strong as the yardstick used to measure accuracy. Two metrics dominate the evaluation: divergence Jdiv (15) and local error tau (17). Both are computed relative to a fine-grid solution obtained from the same NN-perturbed state: tau assumes the previous step is correct and solves forward one step; replay training generates its reference by running the fine solver from the NN-perturbed state. These procedures measure how well the NN corrects the coarse input for the state it is given; they cannot detect a systematic bias that shifts the whole trajectory onto a different physical branch, because the reference is always conditioned on that branch. Section 7.3 explicitly concedes that 'it is currently not possible to validate a solution by increasing the mesh resolution and using a smaller time step' for sensitive cases. Figure 1 already shows that an unstable DNN-MG can settle into a stable but nonphysical equilibrium; after replay the divergence is low, but this does not by itself establish that the equilibrium is the physically correct one. The only cross-check offered is in Sec 6.3 for ro1, where drag/lift functionals approach finer-level reference values; no such independent check is provided for the chaotic generalization cases (sq9*, ro3*) or for the Re-doubling experiments. Hence the gains in Tables 1-6 may measure self-consistency rather than solution accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports improvements to the hybrid neural-network / finite-element solver DNN-MG for two-dimensional instationary Navier–Stokes equations. The method solves on a coarse level and uses a neural network to predict a defect correction on a finer level. The contributions are: input normalization; noise and rotation data augmentation; a replay-buffer retraining stage that generates reference targets from the hybrid simulation itself; and a comparison of MLP, RNN, Transformer, and larger-patch architectures. The central empirical claims are that replay training 'consistently prevents instabilities and also improves every metric' (Section 6.3, Table 1), that Transformers and enlarged receptive fields improve the accuracy of DNN-MG without a significant performance penalty (Section 7), and that the resulting hybrid simulation runs at roughly 2.2--2.6x the coarse-grid cost while approaching fine-grid accuracy. The paper provides an extensive experimental matrix: four networks per configuration, multiple channel-flow geometries, hold-out validation cases, perturbations of the geometry, and Reynolds-number doubling, with code and trained weights promised on Zenodo [82].","tokens_in":31048,"tokens_out":3599,"duration_ms":39529,"significance":"If the claims hold, DNN-MG would be a practical and reproducible hybrid strategy for long-time 2D Navier–Stokes simulation at substantially reduced cost, with useful guidance on architecture choice and on stabilizing autoregressive NN/FEM coupling. The strengths of the study are its breadth: 4 seeds per architecture, separate train/validation/replay stages, multiple accuracy metrics, explicit runtime profiling, and the public code/weights deposit. The authors also honestly report negative results (fine-velocity feedback in Appendix C, limitations in Section 7.3). However, the central accuracy claims rest on metrics that condition the reference on the NN-perturbed trajectory (Eq. (17) and the replay generation in Section 6.3), and the paper itself states in Section 7.3 that for sensitive cases validation by mesh refinement is currently not possible. This limits the strength of the generalization claims, especially for the chaotic sq9* and ro3* cases.","major_comments":[{"comment":"The two load-bearing accuracy metrics, the local error τ in Eq. (17) and the replay-training targets in Section 6.3, are computed from a fine-grid reference that is started from the NN-perturbed state. This measures how well the NN corrects the coarse input for the state it is given, but it cannot detect a systematic bias that pushes the whole trajectory onto a different physical branch. Section 7.3 explicitly concedes: 'it is currently not possible to validate a solution by increasing the mesh resolution and using a smaller time step' for sensitive cases. The only independent check provided is the ro1 drag/lift comparison against higher-level references in Section 6.3 and Table 2. No such check is offered for sq9*, ro3*, or the Re-doubling experiments in Section 7.1. Since the title-claim 'generalizes to different geometries' is substantiated mainly by Tables 4–6, this is a load-bearing","section":"§7.3 and §6.3 / Eq. (17)"},{"comment":"All central comparison tables report only 'mean' and 'best' over 4 seeds, without standard deviations, individual values, or a measure of spread. With N=4, the claims that replay training 'consistently prevents instabilities' and that 'Transformer > MLP > RNN' are statistically fragile: the authors themselves report that 2 of 4 Transformers become unstable at Re=200 (Fig. 9), 1 of 4 baseline MLPs fails on sq9*, and 2 of 4 MLP-M1 networks deviate on sq9* (Section 7.1). The absence of spread makes it impossible for the reader to assess whether the reported metric differences are meaningful. Please provide per-seed values in an appendix, or at least standard deviations/ranges, and state the number of seeds used for each table.","section":"Tables 1–6"},{"comment":"The summary claim of Section 7 that Transformers improve accuracy 'without a significant impact on performance' is overstated as stated. Table 7 shows Transformer training time of 39.3 h vs. 24.0 h for MLP, and Fig. 10 shows roughly 12x slower inference, though the simulation runtime is dominated by the solver. More importantly, the accuracy improvement is not uniform: on sq9* (Table 5) the Transformer's mean divergence (5.2428) is worse than both MLP (5.1221) and RNN (5.0278), and only the best value is comparable; on ro3* (Table 6) the Transformer is best on eJd and e_v but worst on mean divergence. The claim should be qualified to the metrics and cases where the improvement actually holds.","section":"§7 and Table 7 / Fig. 10"}],"minor_comments":[{"comment":"The abstract uses 'instationary' while the title and Introduction use 'nonstationary'. Please make the terminology consistent throughout.","section":"Abstract / Title"},{"comment":"Typo: 'AdamW turned out to be essential essential for stability' should be 'essential for stability'.","section":"§4.4.1"},{"comment":"Typo: 'generaly therms' should be 'generally these terms'.","section":"§2"},{"comment":"The definition of τ has a double comma in the argument list and the notation is otherwise dense. Please rewrite the argument list for clarity, e.g. τ(\\tilde v_{n-1}, \\tilde v_n, f_{n-1}, f_n).","section":"Eq. (17)"},{"comment":"In the text after Table 5, 'the training data where J_l ≈0.s0' contains an apparent typo ('0.s0'). Also, the figure captions in Appendix B (Figs. B.14, B.15) refer to 'ro6' where the text says 'sq9*' and 'ro3*'; please correct the captions.","section":"Table 5 / §7.1"},{"comment":"Several figure captions in Appendix B repeat the same description and do not always match the listed table numbers (e.g., Fig. B.11 is called for Table 1, B.12 for Table 2, etc.). Please align captions with the tables they illustrate.","section":"§5 / B.11–B.15"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid empirical study with reproducible artifacts and an unusually honest limitations section. The main technical concern is that the accuracy metric used for the generalization cases is self-referential: it measures local correction ability on the trajectory actually followed, not closeness to the physical solution. This is a fixable issue within the manuscript's scope by adding one independent validation run for a chaotic case or by substantially tempering the accuracy claims. The missing variance information across seeds is also a blocking issue for the 'robust' and 'consistently' wording in the title and abstract. I do not see grounds for rejection, but the paper needs a revision that addresses these two points before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jendersie et al. report a solid, incremental improvement of their DNN-MG hybrid solver. The new pieces are real: replay-buffer retraining to close the train/inference gap, noise and rotation augmentation, a Transformer variant with a probabilistic tail, and a systematic comparison across MLP, RNN, and Transformer on several geometries and Reynolds numbers. The experiments are extensive — four seeds per architecture, multiple metrics, held-out cases, and code/weights promised on Zenodo. The replay results are convincing: divergence and local error drop substantially on training and validation cases, and the method remains stable when the Reynolds number is doubled. The limitations section is honest, including the point that mesh-refinement validation is currently impossible.\n\nThe main weakness is that the accuracy metrics are mostly self-consistency measures. The local error tau in (17) assumes the previous step is correct and solves one forward step from the NN-perturbed state; replay targets are generated the same way. Neither can detect a systematic bias that moves the whole trajectory to a different physical branch. The paper even shows such a branch shift in Fig. 1. The only independent check is the drag/lift comparison on ro1, which looks good but covers only one periodic case. For the chaotic generalization cases and the Re-doubling runs, there is no independent reference beyond the fine-grid solution computed from the same perturbed states. So the numbers in Tables 1-6 may overstate accuracy, and the title's 'robust and stable' is stronger than what is certified. On top of that, the main tables report only mean and best over four seeds, no standard deviations, which makes the architecture ranking (Transformer > MLP > RNN) less certain than it looks.\n\nThese are addressable. The authors should soften the claims, add variance information, and ideally provide at least one sensitive case where a physically meaningful functional (like drag/lift, or a time-averaged field) is compared against a reference that is not conditioned on the NN-corrected trajectory. The self-citation is heavy but justified by the method's lineage.\n\nWho is this for: researchers working on hybrid numerical/ML solvers, especially for CFD. It deserves a serious referee, not a desk reject. I would send it to review and expect revisions along the lines above.","headline":"Replay training genuinely stabilizes DNN-MG and the generalization tests are useful, but the accuracy claims rest on self-consistent metrics; 'robust and stable' is stronger than the evidence certifies.","tokens_in":31465,"tokens_out":2712,"would_cite":true,"duration_ms":32532,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65N55","76D05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Replay training stabilizes hybrid neural-network flow solver","keywords":["DNN-MG","Navier-Stokes equations","finite element method","deep learning","replay buffer","stability","generalization","computational fluid dynamics"],"falsifier":"Run a replay-trained DNN-MG simulation on a case with a sensitive symmetric solution, such as the perturbed nine-square-obstacle geometry, and compare its time-averaged velocity and lift against reference solutions computed on two successively refined meshes with smaller time steps; if the hybrid run keeps low divergence and low local error yet its mean velocity or force functionals drift away from the refined references beyond the coarse baseline, the central accuracy claim would be falsified.","tokens_in":30552,"feed_emoji":"🌊","tokens_out":5468,"duration_ms":64256,"temperature":0.7,"pith_summary":"The paper claims that the instabilities seen in the hybrid DNN-MG solver—a coarse finite element solve corrected by a neural network—are not inherent to the method but come from a distributional shift between training data and the solver's own evolving state. Retraining on data generated by the hybrid simulation itself, using replay buffers, is claimed to consistently prevent these instabilities and improve every reported accuracy metric. The paper further claims that enlarging the neural network's spatial receptive field, most effectively with a Transformer over local patch neighborhoods, improves accuracy on unseen geometries and at doubled Reynolds numbers without a meaningful performance penalty. If these claims hold, a neural-network-corrected coarse solver can sustain stable long-time Navier-Stokes simulations at roughly 2.2–2.6 times the coarse-grid cost while approaching fine-grid accuracy on a class of 2D flows.","feed_headline":"Replay training stabilizes hybrid neural-network flow solver","feed_subtitle":"Learned corrections let a coarse-grid solver match fine-grid accuracy on 2D flows at over 2x speed.","key_machinery":"The load-bearing mechanism is the defect-correction loop of Algorithm 1: after a coarse solve, the interpolated fine solution's residual is computed, and a neural network predicts a fine-level defect that is added to the interpolant; the corrected fine state then defines the right-hand side for the next coarse solve. Stability comes from replay buffers: a meta-loop that runs the hybrid simulation, records fine-grid reference corrections for the network's own perturbed states, and finetunes the network on these replayed samples plus the original data, without needing a differentiable solver. The architectural comparison centers on the receptive field: patch size M, temporal stacking, and, for","core_discovery":"On the paper's own terms: DNN-MG computes each time step by solving on a coarse level, prolongating to a fine level, evaluating the fine residual, and letting a neural network predict a defect that is added to the solution; this corrected solution feeds back into the next right-hand side. The paper's central discovery is that this loop can be made stable and accurate by closing the gap between training and inference. The mechanism is replay training: run the hybrid simulation, compute a fine-grid reference for the neural-network-perturbed state at each step, and finetune on the resulting data. In their experiments this replay training consistently prevents instabilities and also improves eve","pith_inferences":["Editorial inference: if the claims hold, the replay meta-loop is architecture-agnostic and could transfer to other learned-correction PDE solvers; the main open cost is that replay generation is per-network and per-case, so the economics depend on how often the retrained network is reused.","Editorial inference: the paper's evidence for fine-level accuracy is mostly divergence, local error, and force functionals over moderate horizons; a more demanding test would compare long trajectories against a reference on two successively refined meshes, a check the paper says is currently impossible for sensitive cases.","Editorial inference: the failure of fine-state feedback suggests the next bottleneck is not single-step accuracy but a stable differentiable coupling between solver and network; implementing true backpropagation-through-time in a differentiable solver could unlock the temporal receptive field that the paper found unhelpful as input.","Editorial inference: a testable extension would be to train one replay-stabilized network on a family of geometries and measure how far it extrapolates in Reynolds number or obstacle shape before the replay loop must be rerun; the paper's mixed-geometry cases are a start but leave the boundary of applicability unspecified."],"forward_implications":["Long-time hybrid simulations can remain stable without backpropagation through time: replay training alone closes the distributional shift, at the price of one extra training stage.","Coarse-grid runs corrected this way run about 2.2–2.6 times faster than the fine reference and add only 20–30% over the coarse solve in the tested cases; neural network inference is a minor part of the runtime.","Replay-trained networks generalize to unseen obstacle arrangements, perturbed meshes, and doubled Reynolds numbers; local error drops by up to a factor of 16 and mean-velocity error by up to a factor of 5 relative to the coarse solution.","Larger receptive fields, whether from a Transformer or a larger patch, improve accuracy on non-local flows; Transformers provide this on unstructured meshes while preserving locality, without significant performance impact.","Direct feedback of the network's own corrected fine solution as input destabilizes the run, so the stable configuration deliberately keeps the coarse solve in the loop."],"fun_headline_variants":["Replay training stops error drift in hybrid flow solver","Neural-FEM hybrid flow solver stabilized by replay retraining","Replay buffer stabilizes hybrid neural network flow solver","Hybrid solver for 2D flows made robust with replay training","Neural-FEM solver generalizes better with replay training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that low divergence and low one-step local error against a fine-grid reference certify a physically correct long-time trajectory; the paper itself notes in its limitations section that for sensitive cases this cannot currently be validated by increasing mesh resolution and reducing the time step.","fun_headline_variants_meta":{"raw":{"variants":["Replay training stops error drift in hybrid flow solver","Neural-FEM hybrid flow solver stabilized by replay retraining","Replay buffer stabilizes hybrid neural network flow solver","Hybrid solver for 2D flows made robust with replay training","Neural-FEM solver generalizes better with replay training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1189,"prompt_tokens":760,"completion_tokens":429,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":347}},"tokens_in":504,"tokens_out":429,"duration_ms":4919,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:29:45.886785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a replay-trained DNN-MG simulation on a case with a sensitive symmetric solution, such as the perturbed nine-square-obstacle geometry, and compare its time-averaged velocity and lift against reference solutions computed on two successively refined meshes with smaller time steps; if the hybrid run keeps low divergence and low local error yet its mean velocity or force functionals drift away from the refined references beyond the coarse baseline, the central accuracy claim would be falsified.","supporting_citations":[],"review_version":1}