{"id":"ba564acf-ec8b-43be-8c17-b96408111a54","arxiv_id":"2412.02924","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A loss decomposition that rewrites mean squared error into amplitude and phase terms is applied to a recurrent autoencoder for 1D linear advection, with claimed long-horizon accuracy gains shown on a single test case.","lead":"This paper proposes splitting the usual mean squared error training loss for a wave-predicting neural network into two parts, one tied to wave amplitude and one to wave phase, and reports improved long-horizon forecasts on a 1D linear advection test problem. The catch is that the decomposition is a standard algebraic identity and the reported improvement is shown only for a single test parameter value.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled single-case comparison cannot support the claimed phase-accuracy and horizon extension.","rationale":"I agree with the rejection's bottom line. The paper's mathematical identity is correct but is only a variance-correlation decomposition; the interesting claim is the empirical one that training with this loss beats MSE. That claim is load-bearing and is not yet demonstrated: one test case, no error bars, unspecified baseline tuning, and no report of α,β. If a properly controlled benchmark still shows a consistent advantage, the concern would be resolved. The additional issue about τDISP not being a pure phase error is real but secondary: the method could be useful even if the decomposition is not physically exact; what the paper needs is evidence. Therefore I would not change the REJECT verdict, but the rejection is for insufficient evidence rather than a refuted method.","tokens_in":4964,"tokens_out":6820,"duration_ms":72138,"concrete_test":"Run a controlled benchmark: train AB-CRAN with MSE and with the decomposed loss, using identical architecture, optimizer, data splits, and equal ASHA tuning budgets, over at least 10 random seeds; report per-time-horizon mean ± std error over all 19 test parameters and report the chosen α,β for every seed. If the error intervals overlap, the claimed phase-accuracy and horizon extension is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—'significantly improves phase accuracy ... and extends the prediction horizon' (§3)—is supported only by a single test parameter (μ=1.0125), one error-vs-time curve, and no error bars. The MSE baseline is not specified in terms of hyperparameter tuning, seeds, or per-test-parameter errors across the 19 test instances; because Eq. (6) is a reweighting of the identity τ = τDISS + τDISP, any advantage could arise from tuning α and β rather than from the decomposition. In addition, τDISP = 2(1−ρ)σ_a σ_d is a Pearson-correlation term, not a pure phase-error metric: it responds to any waveform decorrelation and saturates for large shifts, so the claimed clean separation of phase and amplitude errors is not established. The text's statement that β>0.5 'emphasizes dispersion' also conflicts with Eq. (6), where β multiplies τDISS, making the reported configuration unclear.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to improve long-horizon wave predictions made by an attention-based convolutional recurrent autoencoder by decomposing the mean squared error (MSE) into dissipation and dispersion components. In Section 2.3, the authors rewrite MSE as the sum of a variance-difference term, a mean-difference term, and a correlation term, calling the first two the dissipation error and the third the dispersion error. The training loss then reweights these two error terms (and the decoder versus propagator losses) with hyperparameters α and β. Section 3 evaluates the method on the linear advection equation for one test parameter (μ = 1.0125), with qualitative space-time plots and an error-versus-time curve comparing the decomposed loss against plain MSE. The authors conclude that the proposed loss significantly improves phase accuracy and extends the prediction horizon.","tokens_in":5106,"tokens_out":3104,"duration_ms":31407,"significance":"If the empirical claims were established, the proposed loss decomposition would be a simple and broadly applicable training objective for autoregressive neural network surrogates of wave-like PDEs. The algebraic derivation of the decomposition from MSE is correct and clearly presented, and the paper usefully connects numerical dissipation/dispersion concepts to loss design for deep learning. However, the current evidence is not sufficient for a journal-level claim: the central assertion of phase-accuracy improvement rests on a single test parameter, two qualitative figures, and no reported hyperparameter values. The contribution is therefore more of a promising idea with a proof-of-concept than a validated method.","major_comments":[{"comment":"The text states that setting α and β greater than 0.5 puts a stronger emphasis on reducing dispersion error, but in Eq. (6) β multiplies τDISS (the dissipation term), so β > 0.5 actually increases the weight on amplitude error, not phase error. The final tuned values of α and β are not reported anywhere, so the method as described is internally inconsistent and cannot be reproduced or interpreted. Please correct the description and report the actual tuned values, or reconcile the text with the equation.","section":"Section 2.3, Eq. (6)"},{"comment":"The paper's central claim that the decomposed loss 'significantly improves phase accuracy' and 'extends the prediction horizon' is supported only by a single randomly selected test parameter (μ = 1.0125), one error-versus-time curve, and no quantitative error metrics, confidence intervals, or results across the other 18 test parameters. A journal-level claim requires aggregate statistics over the full test set, together with a description of the MSE baseline's hyperparameter tuning budget so that the comparison is fair. Without this, the observed improvement could be due to tuning the reweighting coefficients rather than to the decomposition itself.","section":"Section 3, Figure 4"},{"comment":"The dispersion term τDISP = 2(1 − ρ)σ(ua)σ(ud) is a Pearson-correlation-based measure of overall waveform decorrelation, not a pure phase-error metric: it also responds to amplitude mismatch or any shape change, and it saturates for large phase shifts. The claim that the decomposition cleanly separates phase and amplitude errors is therefore not established. A concrete test would be to evaluate τDISS and τDISP for synthetic perturbations with known pure phase shifts and pure amplitude changes, and to show that they behave as claimed; alternatively, compare against a direct phase-lag metric such as cross-correlation lag.","section":"Section 2.3, Eq. (5)"},{"comment":"The hyperparameters α and β are tuned using Ray Tune ASHA, but the search ranges, the number of trials, and the final selected values are not reported. Because Eq. (4) is an algebraic identity, any choice of α and β is just a reweighted MSE; the paper's claim of a principled advantage requires demonstrating that the selected weights are robust across the test set and that plain MSE trained with the same tuning effort does not match the performance. Please provide these details.","section":"Section 2.3 and Section 2.4"}],"minor_comments":[{"comment":"The phrase 'novel theoretical framework' overstates what Eq. (4) provides, since the decomposition is an algebraic identity; consider rephrasing to 'loss reweighting strategy' or 'decomposition-based loss formulation'.","section":"Abstract and Section 2.3"},{"comment":"The denoising decoder loss is mentioned but no ablation or description of its interaction with the decomposed propagator loss is given; clarify whether the reported improvement is attributed to the decomposition alone or to the combination with denoising.","section":"Section 2.4"},{"comment":"The caption says 'with a time-horizon of 10 time-steps', while the paper discusses long-horizon predictions; please clarify the relationship between this horizon, the 200 simulated time steps, and the non-dimensional times t* = 0.036 and 0.392.","section":"Figure 4 caption"},{"comment":"The phrase 'We randomly selected a parameter from the test set' is vague; specify the selection procedure and, ideally, report results for all 19 test parameters rather than only one.","section":"Section 3"},{"comment":"There are several presentation issues: the typo 'Coodinate' in Figure 3, the inconsistent use of 'AB-CRAN' before it is defined, and some incomplete reference formatting (e.g., [15]). These should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a short workshop-style contribution, and its current empirical support is far too thin for a journal publication. The algebraic identity is correct, but the central claim of phase-accuracy improvement is not quantified over the test set. I recommend major revision rather than outright rejection because the idea is plausible and the deficiencies (multi-test evaluation, hyperparameter reporting, and correction of the β inconsistency) are addressable in a revision. If the authors cannot provide the missing statistics or if the β inconsistency turns out to reflect the actual tuned configuration, the paper may need to be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The loss decomposition in Eqs. 3-5 is the textbook variance-covariance identity for mean squared error. The paper is honest that Le Guen and Thome already applied this idea to time-series forecasting, so the novelty here is limited to the application domain (linear advection) and the AB-CRAN architecture. The algebra is correct, and the authors do report some training details (285 epochs, Ray Tune for alpha and beta). That is the extent of the credit I can give. The empirical claim is the problem. 'Significantly improves phase accuracy' and 'extends the prediction horizon' are supported by one randomly selected test parameter, mu=1.0125, two qualitative figures, and one error-vs-time curve. No aggregate errors across the 19 test instances, no confidence intervals, no ablations, no stated values for alpha and beta. The MSE baseline's hyperparameters and seeds are not reported. Since the loss is a reweighting of an identity, any observed improvement could simply come from tuning alpha and beta on the test case. The single-case comparison cannot carry the generalization claim. There is also an internal inconsistency in the method description. The text says that alpha and beta greater than 0.5 give 'a stronger emphasis on reducing dispersion error,' but Eq. 6 defines Lpropagator = (1-beta)*tau_DISP + beta*tau_DISS. Larger beta puts more weight on tau_DISS, i.e., dissipation, not dispersion. Either the equation or the text is wrong, and without the actual tuned values we cannot tell which was used. This is load-bearing because the whole story is about prioritizing dispersion. A subtler issue: tau_DISP = 2(1-rho)*sigma_a*sigma_d is a Pearson-correlation term, not a pure phase-error metric. It responds to any waveform decorrelation and saturates for large phase shifts. The claim that the decomposition cleanly separates phase and amplitude errors is therefore not established by the math alone. The paper is a workshop note, and that is the level of evidence it provides. The idea is not worthless; there may be a useful empirical study hiding here, but as written the data do not support the conclusions. I would not cite it, and I would not send it to a serious journal as is. My recommendation to the editor: desk reject, with an invitation to resubmit if the authors provide per-test-parameter error statistics, report alpha and beta, fix the inconsistency, and add ablations against plain MSE with matched tuning.","headline":"A standard MSE decomposition applied to wave forecasting, but the central claim rests on one uncontrolled test case and an inconsistent hyperparameter description.","tokens_in":661,"tokens_out":811,"would_cite":false,"duration_ms":26571,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M06"],"pacs":[],"model":"deepseek-v4-flash","headline":"A loss function that tracks phase and amplitude separately keeps neural wave forecasts accurate for longer.","keywords":["loss decomposition","phase error","amplitude error","dispersion","dissipation","long-horizon prediction","wave propagation","deep neural networks"],"falsifier":"Compare the decomposed loss against MSE on two controlled data sets, one where the only error is a time shift (pure phase error, no amplitude change) and one where the only error is amplitude scaling (pure dissipation, no phase shift); if the decomposed loss does not reduce error on the pure phase-shift case, or if it fails to beat MSE on more than the single reported test parameter, the claimed mechanism is not supported.","tokens_in":4755,"feed_emoji":"🌊","tokens_out":6235,"duration_ms":99988,"temperature":0.7,"pith_summary":"This paper tries to establish that the error accumulation that limits neural network forecasts of wave propagation can be reduced by splitting the training loss into phase and amplitude parts instead of using plain mean squared error. The key identity rewrites MSE as a dissipation term (differences in variance and mean) plus a dispersion term (a correlation deficit), and the authors train their autoencoder with a reweighted combination that emphasizes the propagator and the dispersion error. On a linear advection problem with random wave speeds, the decomposed loss reportedly keeps predictions phase-aligned over the full horizon and lowers error growth relative to an MSE-trained model. If true, this gives a cheap, architecture-independent way to inject numerical concepts of dissipation and dispersion into deep learning for physical forecasting.","feed_headline":"Phase-aware loss extends neural wave forecast horizon","feed_subtitle":"Separating dispersion from dissipation in the loss cuts error growth and keeps wave predictions phase-aligned.","key_machinery":"The load-bearing object is a statistical rewriting of MSE into two named error types borrowed from numerical analysis: dissipation (amplitude error) and dispersion (phase error). Concretely, if $u_a$ is the ground truth and $u_d$ the prediction, Eq. 4 reads $$\\tau = [\\$\\sigma$(u_a)-\\$\\sigma$(u_d)]^2 + (\\bar{u}_a - \\bar{u}_d)^2 + 2(1-\\rho)\\$\\sigma$(u_a)\\$\\sigma$(u_d),$$ and the paper defines $\\tau_{\\mathrm{DISS}}$ as the first two terms and $\\tau_{\\mathrm{DISP}}$ as the third. The machinery works by giving the optimizer a target that can penalize a wave shape that is right in magnitude but shifted in time differently from a wave that has the right phase but wrong height, which plain MSE cannot distinguish.","core_discovery":"The central claim is that decomposing the mean squared error into separate dissipation and dispersion components, and weighting those components differently during training, substantially improves long-horizon predictions for wave propagation. For the linear advection equation, the authors use the identity $\\tau = [\\sigma(u_a)-\\sigma(u_d)]^2 + (\\bar{u}_a - \\bar{u}_d)^2 + 2(1-\\rho)\\sigma(u_a)\\sigma(u_d)$, call the first two terms dissipation error and the third term dispersion error, and train the AB-CRAN model with $L_{\\mathrm{propagator}} = (1-\\beta)\\tau_{\\mathrm{DISP}} + \\beta\\tau_{\\mathrm{DISS}}$, with hyperparameters favoring the propagator and dispersion. On a held-out wave speed, the model trained this way preserves phase across the prediction horizon, while the MSE-trained model shows phase lag; error-versus-time comparisons show reduced error accumulation. The authors regard this as a framework that can extend to nonlinear wave propagation and other physical forecasting tasks such as fluid dynamics and climate modeling.","pith_inferences":["Because the dispersion term is a correlation-based statistic, it could serve as a standalone diagnostic for phase drift in any autoregressive model, independent of training.","The behavior of the decomposition at small batch sizes may differ from full-horizon statistics, so its effectiveness could depend on batch size in ways plain MSE does not exhibit.","A testable extension is a training schedule that varies the relative weight of dissipation and dispersion over time, for instance emphasizing amplitude early and phase later, which the paper does not explore."],"forward_implications":["Networks trained with the decomposed loss should stay phase-aligned over the full rollout, whereas MSE-trained networks develop a growing phase lag.","The usable forecast horizon, meaning the number of time steps before error exceeds a fixed threshold, increases for the linear advection problem.","The improvement transfers to nonlinear wave propagation, since the authors note the approach has been applied to the inviscid Burgers equation.","The same weighting strategy, emphasizing the propagator and the dispersion term, can be carried over to other autoregressive spatiotemporal forecasters without changing their architecture."],"supporting_citations":[{"why":"supplies the dissipation and dispersion error distinction from numerical analysis that the loss decomposition is built on.","marker":"[14]"},{"why":"prior work on decomposed loss functions for time-series forecasting that this method extends to wave propagation.","marker":"[15]"},{"why":"provides the AB-CRAN architecture, data generation, and training setup used in the experiments.","marker":"[5]"},{"why":"defines dissipation and dispersion in numerical schemes, grounding the terminology of the two loss terms.","marker":"[12]"},{"why":"introduces the deep convolutional recurrent autoencoder structure adapted for the propagator.","marker":"[10]"},{"why":"supplies the hyperparameter tuning method used to set the weighting coefficients alpha and beta.","marker":"[18]"}],"fun_headline_variants":["Splitting loss into phase and amplitude extends wave forecasts","Phase-aware loss decomposition reduces long-horizon wave errors","Dissipation and dispersion split improves neural wave forecasting","Decomposed loss keeps wave predictions phase-aligned longer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the claim that the variance, mean, and correlation statistics in Eq. 5 isolate physical amplitude error from physical phase error; if those statistics mix the two error types, the weighted loss is just arbitrary reweighting of MSE with no guaranteed advantage.","fun_headline_variants_meta":{"raw":{"variants":["Splitting loss into phase and amplitude extends wave forecasts","Phase-aware loss decomposition reduces long-horizon wave errors","Dissipation and dispersion split improves neural wave forecasting","Decomposed loss keeps wave predictions phase-aligned longer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00063,"raw_usage":{"total_tokens":2856,"prompt_tokens":836,"completion_tokens":2020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1955}},"tokens_in":452,"tokens_out":2020,"duration_ms":17800,"temperature":1.0,"reasoning_tokens":1955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:56:26.095302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the decomposed loss against MSE on two controlled data sets, one where the only error is a time shift (pure phase error, no amplitude change) and one where the only error is amplitude scaling (pure dissipation, no phase shift); if the decomposed loss does not reduce error on the pure phase-shift case, or if it fails to beat MSE on more than the single reported test parameter, the claimed mechanism is not supported.","supporting_citations":[{"cited_title":"Comparison of accurate methods for the integration of hyperbolic equations","cited_arxiv_id":null,"evidence_quote":"supplies the dissipation and dispersion error distinction from numerical analysis that the loss decomposition is built on."},{"cited_title":"Predicting waves in fluids with deep neural network.Physics of Fluids, 34(6), 2022","cited_arxiv_id":null,"evidence_quote":"provides the AB-CRAN architecture, data generation, and training setup used in the experiments."},{"cited_title":"Griffiths","cited_arxiv_id":null,"evidence_quote":"defines dissipation and dispersion in numerical schemes, grounding the terminology of the two loss terms."},{"cited_title":"A system for massively parallel hyperparameter tuning","cited_arxiv_id":null,"evidence_quote":"supplies the hyperparameter tuning method used to set the weighting coefficients alpha and beta."}],"review_version":1}