{"id":"88849961-20b1-426c-87d4-eb895c0a8679","arxiv_id":"2508.10908","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"KIST-Ocean is a proposed deep-learning 3D global ocean model that claims to capture coupled ocean-atmosphere responses such as equatorial waves and wind-driven vertical motion.","lead":"The abstract describes KIST-Ocean, a deep-learning global ocean model said to reproduce ocean-atmosphere coupling such as equatorial Kelvin and Rossby waves and wind-driven vertical motion. The submitted full text is an unrelated condensed-matter paper about flat bands in non-Hermitian lattices, so this report can only judge the abstract.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No held-out evaluation or physics-consistency test is shown to separate learned wave dynamics from memorized reanalysis training fields; the abstract's central skill claim is therefore not yet evidenced.","rationale":"The reader's weakest assumption correctly identifies that reproducing Kelvin and Rossby waves is treated as evidence of learned physics, but no protocol separates generalization from memorization. My concern matches that premise. The reader's UNVERDICTED verdict remains appropriate: the submission provides an abstract with unsupported positive claims and a mismatched full text, so there is neither enough evidence to reject nor to accept the central claim. I do not propose a verdict change because the issue is evidentiary insufficiency, not a demonstrated internal inconsistency. A strict temporal holdout test would settle whether the wave-capture claim reflects genuine predictive skill or training-data retrieval.","tokens_in":1866,"tokens_out":1258,"duration_ms":17515,"concrete_test":"Obtain the actual KIST-Ocean manuscript and data, then run a strict temporal holdout: train through December 2019 and evaluate NINO3.4 SST anomaly, tropical Pacific sea-surface-height variability, and vertical-velocity response to wind-stress curl for 2020-2023. Report root-mean-square errors relative to climatology and persistence baselines, with confidence intervals. Additionally, check that the diagnosed Kelvin and Rossby wave phase speeds in these held-out forecasts match theoretical values within uncertainty, rather than appearing only in reconstructed training states.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that KIST-Ocean accurately captures Kelvin and Rossby wave propagation and wind-stress-driven vertical motion, demonstrating learned ocean-atmosphere coupling. The load-bearing premise is that this behavior reflects generalization rather than retrieval from the training distribution. Kelvin and Rossby waves and cyclonic/anticyclonic wind-stress responses are prominent, high-amplitude features in the reanalysis fields such a model would be trained on. An autoregressive deep-learning model can reproduce these patterns by memorization or interpolation if the evaluation uses the same periods or fields as the training data. The abstract provides no held-out periods, no error bars, no comparison to baselines, and no physics-consistency or perturbation test (e.g., generating a forecast from an initial condition not in the training set and checking wave phase speeds). This is not evidence that the model is wrong; it is evidence that the central conclusion is unsupported. Compounding this, the submitted full text is an unrelated condensed-matter paper on non-Hermitian flat bands, so the described evaluation protocol cannot be verified from this submission at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript, arXiv:2508.10908, is submitted under the title 'Data-driven global ocean model resolving ocean-atmosphere coupling dynamics' and presents, in its abstract, a deep-learning model called KIST-Ocean. The abstract claims that KIST-Ocean is a global three-dimensional ocean general circulation model based on a U-shaped visual attention adversarial network, that it achieves robust predictive skill, and that it accurately captures Kelvin and Rossby wave propagation in the tropical Pacific and vertical motions induced by cyclonic and anticyclonic wind stress. The submitted full text, however, is an unrelated condensed-matter paper titled 'Exceptional flat bands in bipartite non-Hermitian lattices' (arXiv:2508.10901), which contains no mention of KIST-Ocean, oceans, or atmospheric coupling. No methods, equations, training data description, evaluation metrics, baselines, or code are provided for the claimed ocean model.","tokens_in":2053,"tokens_out":1572,"duration_ms":16671,"significance":"If the claims in the abstract were substantiated, the result would be significant: a purely data-driven global ocean model that reproduces key coupled ocean-atmosphere phenomena such as Kelvin and Rossby waves would be an important step toward DL-based climate prediction. However, the significance cannot be assessed because the manuscript contains no evidence for these claims beyond a short abstract, and the body text is a different paper entirely. The lack of metrics, held-out evaluation, baselines, and physics-consistency tests means that even the abstract's assertions are not verifiable from the submitted material. No credit can be given for machine-checked proofs, reproducible code, or falsifiable predictions because none are present.","major_comments":[{"comment":"The submitted full text is the paper 'Exceptional flat bands in bipartite non-Hermitian lattices' (ArXiv:2508.10901), which is a condensed-matter study with no relation to the abstract's KIST-Ocean model. The manuscript therefore contains no methods, architecture details, training procedure, or evaluation protocol for the claimed data-driven ocean model. The central claim of the abstract cannot be checked against any accompanying evidence.","section":"Full text (entire)"},{"comment":"The abstract asserts 'robust ocean predictive skill' and 'accurate' reproduction of Kelvin and Rossby wave propagation and wind-stress-driven vertical motions, yet it provides no quantitative metrics, no comparison to established baselines (e.g., numerical ocean models or other DL models), no error bars, and no description of data splits or held-out periods. These assertions are therefore unsupported.","section":"Abstract"},{"comment":"The abstract's key inference is that reproducing Kelvin and Rossby waves demonstrates the model's ability to represent ocean-atmosphere coupling. However, these phenomena are prominent features of the reanalysis fields such a model would be trained on, so their reproduction could result from memorization or interpolation rather than learned physics. Without a test of generalization—such as forecasting from initial conditions outside the training set or checking wave phase speeds—the claim that the model 'captures' coupling dynamics is not established.","section":"Abstract (generalization claim)"}],"minor_comments":[{"comment":"The phrase 'U-shaped visual attention adversarial network' is not defined; the reader cannot determine what architecture components are involved (e.g., attention layers, discriminator, loss function) or how they are combined.","section":"Abstract"},{"comment":"The abstract does not specify the input and output variables of the model (e.g., sea surface height, temperature, salinity, wind stress), the training data source (e.g., which reanalysis product), or the temporal and spatial resolution.","section":"Abstract"},{"comment":"Given the mismatch between the abstract and the full text, the manuscript appears to have been assembled incorrectly; the authors should ensure that the submitted PDF corresponds to the claimed research.","section":"Full text"}],"recommendation":"reject","confidential_remarks":"The manuscript presents a serious integrity or submission error: the abstract describes a deep-learning ocean model, but the full text is an unrelated condensed-matter paper on non-Hermitian flat bands. This is not a case of incremental revision; the central claim of the abstract is entirely unsupported by the submitted body. Even if the correct full text were provided, the abstract's claims would still need substantial quantitative evidence and held-out generalization tests to be publishable. In the current state, the paper cannot be evaluated on its merits and does not meet the minimum standards for review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jim,\n\nThis one is not in a reviewable state. The metadata and abstract describe KIST-Ocean, a deep-learning global ocean model with claims about Kelvin/Rossby wave propagation and ENSO-relevant coupling. The full text attached is a condensed-matter paper about exceptional flat bands in non-Hermitian lattices. There is no overlap between the two. So as submitted, there is no manuscript to evaluate.\n\nThat said, the abstract alone points at a real and timely problem. Extending AI weather models to the ocean and to subseasonal-to-seasonal timescales is an active gap, and the architecture outlined—partial convolution, visual attention, adversarial training, transfer learning—is a plausible toolkit for it. If the actual full text matched the abstract, this would be a candidate for serious refereeing because the research direction matters and the authors are clearly working in a relevant area.\n\nThe soft spots are substantial. The abstract's central claim—\"accurately captures realistic ocean response\"—is supported by no numbers: no RMSE, no anomaly correlation, no baselines, no held-out periods, no error bars. The stress-test note is right that Kelvin and Rossby waves are prominent features in the reanalysis fields such a model would train on, so demonstrating them without a held-out or perturbation test does not separate learned physics from retrieval. That critique lands. But the proximate problem is the full-text mismatch, which makes even this abstract-level evaluation moot.\n\nOn citation patterns: the abstract cites no prior DL ocean models, which is a gap, but that is minor next to the mismatch.\n\nMy recommendation: do not send this to peer review. Return it to the authors to fix the submission. If the ocean-model paper is real and the full text was swapped by mistake, then a complete manuscript with evaluation details would deserve to be reviewed. As it stands, there is nothing coherent to referee.\n\nBest,\n[You]","headline":"The abstract and the full text are two different papers, so this submission cannot be reviewed as an ocean-modeling contribution.","tokens_in":2585,"tokens_out":2605,"would_cite":false,"duration_ms":27476,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A data-driven ocean model reproduces the Kelvin and Rossby waves that couple ocean and atmosphere.","keywords":["deep learning","ocean general circulation model","El Niño–Southern Oscillation","Kelvin waves","Rossby waves","ocean-atmosphere coupling","adversarial network","partial convolution"],"falsifier":"Run the model on an equatorial wind stress perturbation that was held out from training and check whether the ocean response contains an eastward-moving Kelvin wave and westward-moving Rossby wave with realistic speeds and vertical structure; failure on such an out-of-distribution forcing would indicate the waves are memorized rather than generated by learned physics.","tokens_in":980,"feed_emoji":"🌊","tokens_out":1214,"duration_ms":67455,"temperature":0.7,"pith_summary":"The paper introduces KIST-Ocean, a deep-learning global ocean model built on a U-shaped visual attention adversarial network. It claims that purely data-driven simulation can realistically reproduce ocean responses to atmospheric forcing, including equatorial Kelvin and Rossby wave propagation and wind-driven vertical motions. That matters because these are the physical mechanisms by which the ocean and atmosphere couple in phenomena like the El Niño–Southern Oscillation. If the claim holds, data-driven models could extend from short-range weather prediction into coupled climate prediction, which currently relies on expensive numerical ocean-atmosphere models.","feed_headline":"Deep-learning ocean model reproduces El Niño's wave machinery","feed_subtitle":"A data-driven global ocean model claims to capture the waves that couple ocean and atmosphere.","key_machinery":"The central object is the U-shaped visual attention adversarial network architecture, a generator trained against a discriminator in an adversarial setup. Partial convolution layers handle irregular land-ocean boundaries, adversarial training produces physically realistic fields rather than smoothed averages, and transfer learning adapts the model to distribution shifts encountered during auto-regressive rollouts. Together these components allow the model to propagate ocean anomalies realistically rather than merely regress toward climatology.","core_discovery":"KIST-Ocean is a DL-based global three-dimensional ocean general circulation model that combines partial convolution, adversarial training, and transfer learning to handle coastal complexity and auto-regressive distribution drift. The paper's evaluations report robust predictive skill and efficiency, and the model captures realistic ocean response, specifically Kelvin and Rossby wave propagation in the tropical Pacific and vertical motions induced by cyclonic and anticyclonic wind stress. The authors interpret these results as evidence that the model can represent key ocean-atmosphere coupling mechanisms underlying climate phenomena including ENSO, reinforcing confidence in extending DL-based approaches to broader Earth system modeling.","pith_inferences":["A decisive test would be to force the model with wind stress from a historical El Niño not used in training and check whether it generates the correct sequence of equatorial Kelvin and Rossby waves; the abstract reports no such held-out validation, so the generalization claim remains open.","The full text supplied with this entry is a different manuscript (on exceptional flat bands in non-Hermitian lattices); if that text is genuinely part of the paper, the paper's internal coherence is in question. This extraction follows the abstract and reader notes.","One can also probe the 'learned physics' hypothesis by comparing the model's wave phase speeds and dispersion relations against linear equatorial wave theory; a data-driven model that reproduces the correct Kelvin wave speed would be strong evidence of emergent physics."],"forward_implications":["If correct, DL-based ocean models could replace or augment dynamical ocean components in coupled prediction systems, cutting the computational cost of seasonal-to-decadal forecasts.","Reproducing Kelvin and Rossby waves suggests data-driven models can capture propagating signals, not just spatial patterns, which is essential for forecasting ENSO evolution.","Success would extend the DL revolution beyond weather forecasting into coupled Earth system prediction, including ocean circulation and biogeochemistry.","The approach offers a template for building data-driven submodels of other Earth system components that must exchange fluxes with an atmosphere model."],"supporting_citations":[],"fun_headline_variants":["Deep-learning ocean model nails El Niño's wave patterns","Neural network ocean model sees ENSO's wave physics","AI ocean model reproduces tropical Pacific waves","Data-driven ocean model captures coupling dynamics"],"cache_read_input_tokens":4864,"weakest_assumption_plain":"The load-bearing premise is that the model's ability to reproduce Kelvin and Rossby waves and vertical motion reflects genuine predictive skill rather than recall of patterns present in its training data.","fun_headline_variants_meta":{"raw":{"variants":["Deep-learning ocean model nails El Niño's wave patterns","Neural network ocean model sees ENSO's wave physics","AI ocean model reproduces tropical Pacific waves","Data-driven ocean model captures coupling dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1410,"prompt_tokens":843,"completion_tokens":567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":459,"tokens_out":567,"duration_ms":6930,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:58:05.104161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the model on an equatorial wind stress perturbation that was held out from training and check whether the ocean response contains an eastward-moving Kelvin wave and westward-moving Rossby wave with realistic speeds and vertical structure; failure on such an out-of-distribution forcing would indicate the waves are memorized rather than generated by learned physics.","supporting_citations":[],"review_version":1}