{"id":"f9b6650f-e2ab-4703-bca2-8b630440b13a","arxiv_id":"2505.06014","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Convolutional-autoencoder and neural-network surrogates trained on 150 CFD sea states can reproduce time-averaged wave-induced pressure fields and ship-fixed added resistance, with mean absolute drag error 1.77e-5 on held-out cases.","lead":"The paper trains a neural network surrogate to predict pressure fields around a container ship in regular waves, then integrates those fields to estimate the added drag from seaway resistance. The goal is a fast, data-driven way to support ship routing and fuel-saving decisions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The routing claim is unsupported because the surrogate is evaluated only in ship-fixed coordinates; Section 5.1 shows time-averaging raises the global-frame drag error to 3.42e-5 versus 0.94e-5, so the headline 1.77e-5 does not apply to inertial-frame added resistance.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the surrogate and its evaluation are restricted to ship-fixed coordinates, while the stated routing application requires global-frame resistance. The paper is transparent about this limitation, but the abstract and conclusion nevertheless frame the method as a route-guidance tool, so the gap is central rather than incidental. The paper's own numbers make the concern concrete: Section 5.1 shows the time-averaged transfer pipeline has a global-frame drag error of 3.42e-5 versus 0.94e-5 for instantaneous integration, and Section 5.2 deliberately switches to ship-fixed evaluation. No global-frame evaluation of the full ML surrogate is reported, so the headline 1.77e-5 error cannot be used to support the routing claim. This is a correctness risk for the stated application, not merely a disagreement with external consensus. The mirror-split and data-availability concerns are real but secondary; the frame mismatch alone is sufficient to require either a global-frame evaluation or a substantially qualified claim. Because the reader already assigned CONDITIONAL based on this weakness, the stress-test does not change the verdict; it sharpens the concrete condition that would resolve it.","tokens_in":21649,"tokens_out":5315,"duration_ms":51326,"concrete_test":"Compute the global-frame added resistance error of the trained surrogate on the 18 held-out test cases: for each case, take the surrogate-predicted time-averaged pressure field on meta-grid B with the Int-Mean strategy, apply the same integration routine used in Section 5.1 for the time-averaged strategy (projecting onto the CFD hull surface, adding hydrostatic pressure, and rotating by the time-averaged pitch angle from the corresponding CFD run), and compare the resulting Delta C_D with the CFD values in Table 7. Report the mean absolute error E_D and per-case values. If the mean E_D is close to 0.94e-5, the ship-fixed evaluation suffices; if it is near 3.42e-5 or larger, the routing claim is not supported by the surrogate as built.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the surrogate predicts added resistance for routing. Routing requires the added resistance in the global/inertial frame, i.e., the x1-component of the wave-induced force in Eq. (1). Section 5.1 quantifies that transferring time-averaged pressure fields to the meta-grid and integrating with an averaged pitch angle yields a mean global-frame drag error of 3.42e-5 on grid B, versus 0.94e-5 when instantaneous fields are integrated. Section 5.2 then evaluates the ML surrogate only in the ship-fixed frame, reporting E'_D = 1.77e-5 (Table 5, grid B, Int-Mean). The paper explicitly states that 'only the ship-fixed coordinate system will be used to determine the quality of the entire surrogate process' and acknowledges in Section 2.3 that reconstructing global forces from time-averaged ship-fixed pressures 'would require a matching pitch angle'. Because the surrogate is trained on time-averaged data, it cannot recover the instantaneous pitch motion on which global-frame accuracy depends. Therefore the headline accuracy does not transfer to the routing quantity; the full pipeline's global-frame error is at least the time-averaging error of 3.42e-5, roughly twice the reported ship-fixed value, unless the ML contribution compensates, for which no evidence is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an offline/online surrogate-modeling pipeline for predicting wave-induced added resistance of a free-floating KRISO container ship. A convolutional autoencoder reduces time-averaged, ship-fixed CFD pressure fields to a low-dimensional latent space, and a fully connected neural network maps the three sea-state parameters (wave height, encounter angle, wave length) to that latent space; the decoded pressure fields are then projected onto the hull and integrated to obtain added-resistance coefficients. The manuscript evaluates the pipeline for 150 CFD-simulated regular sea states (augmented to 285 cases by mirroring) on two structured meta-grids, reports data-transfer errors separately from machine-learning errors, and presents per-case drag and pressure-field errors on 18 held-out test cases. The stated headline result is a mean ship-fixed added-drag error of 1.77e-5 for the best configuration on meta-grid B with interior cells filled by iterative averaging (Section 5.2, Table 5).","tokens_in":22010,"tokens_out":3362,"duration_ms":37879,"significance":"If the claimed accuracy carried over to the routing-relevant quantity, the work would be a useful engineering contribution: it demonstrates a complete chain from CFD data through projection, dimension reduction, regression, and force integration, with per-case error tables, a careful separation of data-transfer versus machine-learning errors, openly available training data, and honest reporting of outlier cases such as the beam-sea case e1-3m. The data-transfer verification in Section 5.1 is a particular strength because it isolates the interpolation error before any ML is introduced. However, the central routing claim in the abstract and introduction requires added resistance in the inertial/global frame, and the manuscript deliberately evaluates the surrogate only in the ship-fixed frame; Section 5.1 shows that time-averaging degrades the global-frame drag error substantially. That gap, together with unresolved issues around test-set selection and mirrored-case leakage, means the paper currently supports a narrower claim than the one it advertises.","major_comments":[{"comment":"The headline accuracy of E'_D = 1.77e-5 (Table 5) is computed in the ship-fixed coordinate system, while the routing motivation requires the inertial-frame added resistance, i.e., the x1-component in Eq. (1). Section 5.1 reports that transferring time-averaged pressure fields to meta-grid B and integrating with an averaged pitch angle yields a mean global-frame drag error of 3.42e-5, compared with 0.94e-5 when instantaneous fields are integrated; the paper explicitly states that only the ship-fixed system will be used to judge the surrogate. Because the ML surrogate is trained on time-averaged fields and no model of the instantaneous pitch angle is described, the route-relevant global-frame error is at least the time-averaging error of 3.42e-5 unless the ML contribution compensates, for which no evidence is provided. The authors should either evaluate the full surrogate in the inertial frame or clearly restrict the claims to ship-fixed added resistance.","section":"§5.2, Table 5 vs. §5.1"},{"comment":"The test set in Table 5 contains nine mirrored cases marked 'm' (e.g., b2-3m, d1-2m, e1-3m), which are exact reflections of simulated cases at the complementary encounter angle. The paper does not state whether the non-mirrored counterpart of each mirrored test case was excluded from the training and validation sets. If any such counterpart appears in training, the corresponding test errors are not independent because the mirrored pressure field is a deterministic transformation of the original. The authors should describe the random split explicitly and either exclude counterpart cases or report the test errors separately for mirrored and non-mirrored cases.","section":"§4.1, Dataset split and mirrored test cases"},{"comment":"The reported mean error of 1.77e-5 corresponds to the configuration (meta-grid B, Int.-Mean, simultaneous training) that performs best on the same 18-case test table used for evaluation. Selecting the best of four configurations on the test set introduces selection bias, so the reported accuracy is optimistic relative to what would be obtained on a truly unseen configuration choice. The authors should either use a separate hold-out set for configuration selection or present the selection rule and its uncertainty transparently.","section":"§5.2, Table 5, configuration selection"}],"minor_comments":[{"comment":"The sentence 'Thanks to the the now widely available hardware and software' contains a duplicated article; it should read 'Thanks to the now widely available hardware and software'.","section":"§1, Introduction"},{"comment":"The symbol E'_D is introduced in the Table 5 caption but the text in Section 5.2 defines E_D and E_p; the notation should be unified so the ship-fixed drag error has a single name.","section":"§5.2, Table 5 caption and text"},{"comment":"Some numeric entries contain stray characters, e.g., '7 .765' in the c1-2 row, '4 .160' in the f3-2 row, and '2 .467' in the j4-1 row; these should be corrected.","section":"Appendix A, Table 6"},{"comment":"The citation in Section 4.3 contains a stray question mark ('(?Wu et al. [2021], Schwarz et al. [2025b])'), and the reference to Muzaferija et al. contains a typo ('Syposium' instead of 'Symposium').","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an engineering feasibility study rather than a methodological novelty paper. The main obstacle is not the ML machinery but the mismatch between the routing-oriented claim and the ship-fixed evaluation target; this can plausibly be fixed by a focused global-frame validation or by a clear reframing of the contribution. I would not reject on the basis of the current gaps, but the reported numbers should not be presented as route-relevant added-resistance errors without that additional work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper to know about if you work on ML surrogates for ship hydrodynamics, but the abstract promises more than the validation delivers. The surrogate is evaluated only in ship-fixed coordinates, and the paper's own Section 5.1 shows that the time-averaging strategy used by the ML pipeline has a mean global-frame drag error of 3.42e-5 on the fine grid, versus 0.94e-5 for instantaneous integration. The reported 1.77e-5 applies to ship-fixed drag, which is not the quantity a route optimizer needs. The authors acknowledge the coordinate-frame issue in Section 2.3, then proceed anyway.\n\nWhat is genuinely useful: the systematic data-transfer verification, the per-case error table with outliers flagged, and a held-out test set built on parameter combinations not seen in training. The application of this CAE/NN architecture to a free-floating container ship in regular waves is new, and the data availability statement is a plus. The paper is honest about the worst case (e1-3m) and about the reasons for it.\n\nThe soft spots, in rough order of seriousness. First, the coordinate-frame gap undermines the routing claim. Second, the mean relative RMS error of the predicted pressure field is 0.56, meaning about 56% error; that is high and undercuts the premise of predicting pressure fields rather than just tuning a drag coefficient. The authors do not discuss why that error is acceptable. Third, the best configuration is selected from the same Table 5 that reports the test errors; there is no nested validation. Fourth, nine of the eighteen test cases are mirrored, and the split of mirrored cases across training, validation, and test is not described, so statistical independence is in question. Fifth, there is an unresolved citation placeholder in Section 4.3. These are all addressable in revision.\n\nWho should read this: anyone building ML surrogates for sea-keeping loads. It is a clear engineering recipe with reproducible steps. It deserves a serious referee, but the referee should push on the inertial-frame error and the pressure-field accuracy. My recommendation: send it out, with major revision expected on the routing claim.","headline":"Solid engineering study of a CAE/NN surrogate for ship wave pressures, but the routing claim is unsupported by the ship-fixed-only validation and the time-averaging penalty in inertial-frame drag.","tokens_in":22535,"tokens_out":3306,"would_cite":false,"duration_ms":33935,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A convolutional-autoencoder/regression surrogate trained on time-averaged CFD pressure fields predicts the full wave-pressure field around a free-floating container ship from three sea-state parameters, yielding a held-out mean absolute…","keywords":["surrogate modeling","multiphase flows","pressure fields","seaway","ship routing","machine learning","convolutional autoencoder","added resistance"],"falsifier":"Run the surrogate in the inertial frame by combining its predicted ship-fixed pressure field with the CFD pitch-angle history (or a separately predicted pitch angle), and compare the global-coordinate mean absolute drag error against the 0.94e-5 benchmark for instantaneous integration; if the error stays at the time-averaged level of 3.42e-5 rather than dropping, the surrogate does not deliver routing-grade resistance in earth-fixed coordinates.","tokens_in":21405,"feed_emoji":"🚢","tokens_out":8352,"duration_ms":81135,"temperature":0.7,"pith_summary":"The paper aims to show that a data-driven surrogate can replace slow CFD when a routing system needs the added resistance a ship feels in waves. Instead of predicting a single resistance number, it reconstructs the entire three-dimensional, time-averaged pressure field around a free-floating container ship from just three sea-state parameters: wave height, wave length and encounter angle. A convolutional autoencoder compresses those fields into a ten-dimensional latent space and a small neural network maps the three parameters into that space, so a full pressure field can be generated almost instantly. On held-out sea states the surrogate reproduces the ship-fixed drag coefficient with a mean absolute error of 1.77e-5, about twice the error of the interpolation and integration step alone. The practical point is that if the method generalizes, existing CFD data sets become reusable fuel-saving tools for onboard routing.","feed_headline":"Sea-state-to-pressure surrogate predicts ship added resistance","feed_subtitle":"Trained on 150 CFD runs, it maps wave height, length and angle to a full 3D pressure field, holding drag error to about 1.8e-5.","key_machinery":"The load-bearing mechanism is the two-block CAE/NN surrogate. The convolutional autoencoder performs nonlinear order reduction: it squeezes a 128x64x96 pressure field down to ten latent numbers and reconstructs the field from them, capturing spatial dependencies with relatively few parameters. The fully connected regression network (200-200-10) maps the three sea-state control parameters into that same latent space, closing the loop from weather data to pressure field. Training uses a simultaneous loss that combines reconstruction error and prediction error on the full field, with the latent-matching term set to zero; interior ship cells are filled by iterated boundary means (Int.-Mean) so the convolutional filters near the hull are not corrupted by zeros.","core_discovery":"The central claim is that time-averaged, ship-fixed pressure deviations from calm water lie on a low-dimensional manifold parameterized by wave height, wave length and encounter angle, and that a CAE/NN pair can learn that manifold from 285 CFD-derived fields (150 simulated, 135 mirrored). The encoder reduces each field to ten latent variables, the regression network predicts the latent code from the three sea-state inputs, and the decoder outputs the full pressure field, which is then projected onto the hull and integrated to give added drag. The paper reports a held-out mean absolute error of 1.77e-5 for the ship-fixed drag coefficient with the finer meta-grid and iterative interior averaging, with head-sea cases reproduced well and the beam-sea short-wave case e1-3m as a notable outlier attributed to sparse data coverage. The authors evaluate only in the ship-fixed coordinate system, because for the inertial frame their own transfer analysis shows time-averaging introduces a larger error (3.42e-5) than integrating instantaneous fields (0.94e-5).","pith_inferences":["The strongest unstated consequence is negative: if routing requires resistance in the earth-fixed frame, the surrogate in its current form is incomplete, and the missing piece is quantified by the paper's own data (3.42e-5 for time-averaged versus 0.94e-5 for instantaneous integration).","A natural test of the method's ceiling is whether adding training cases near beam seas or explicitly encoding the mirror symmetry between beta and 360 degrees minus beta shrinks the outlier error of case e1-3m; the paper's data-coverage argument predicts it would.","The same CAE/NN pipeline should transfer to other hulls and to other field quantities such as air concentration or local wave elevation, provided a similarly smooth latent manifold exists; the paper's broader claim about complex multiphase flow fields is extrapolation rather than demonstration.","For an operational routing system, coupling this pressure-field surrogate with a cheap pitch-angle regression could recover most of the global-frame accuracy gap, since pitch enters the inertial-coordinate force integration in a nearly linear way."],"forward_implications":["A routing system could query the surrogate for many wave scenarios in the time it takes one CFD run to advance a few time steps, making added-resistance-aware route optimization feasible in practice.","Because the machine-learning stage adds only about twice the error of the pure interpolation/integration pipeline, the accuracy ceiling is currently set by data transfer to the structured meta-grid rather than by the autoencoder itself.","Head-sea conditions (encounter angles near 180 degrees) are predicted with the lowest errors, so initial deployment is most credible for head-sea routing.","The time-averaged surrogate is validated only for ship-fixed drag; a global-coordinate routing target would need a pitch-angle estimate or an instantaneous-field variant.","A latent dimension of ten already sits at the accuracy/cost sweet spot compared with dimensions five and twenty, so the approach does not need a large latent space to capture the wave-response manifold."],"supporting_citations":[{"why":"Supplies the autoencoder-plus-regression surrogate architecture previously applied to container-vessel aerodynamic forces, the direct template for the CAE/NN pipeline.","marker":"Pache and Rung [2022]"},{"why":"Provides the projection-based physics-ML formulation underlying parameter-to-reduced-space regression.","marker":"Swischuk et al. [2019]"},{"why":"Demonstrates autoencoder-based exploration and prediction for fluid systems, justifying the CAE reduction step.","marker":"Agostini [2020]"},{"why":"Shows that simultaneous training of the reduction and regression networks improves surrogate accuracy, the scheme adopted here.","marker":"Wu et al. [2021]"},{"why":"Earlier ML surrogate for ditching loads from CFD data, cited as evidence for simultaneous training and for the broader CFD-to-ML pipeline.","marker":"Schwarz et al. [2025b]"},{"why":"Defines the KRISO container ship geometry used in all simulations and validation.","marker":"SIMMAN [2008]"},{"why":"Describes the finite-volume CFD solver and pressure-correction scheme that generated the training data.","marker":"Rung et al. [2009]"},{"why":"Supplies the Hilbert-transform procedure used to select the whole-wave-period averaging time window.","marker":"Loft et al. [2023]"},{"why":"Establishes added resistance in waves as a key quantity for ship routing, motivating the target variable.","marker":"Liu et al. [2011]"}],"fun_headline_variants":["AI surrogate maps wave conditions to full ship pressure field","Wave-to-drag surrogate: CAE/NN predicts pressure field in seconds","Learned manifold maps sea state to hull pressure and drag","Pressure field surrogate: from wave inputs to added resistance","Neural net predicts full 3D pressure field for ship drag from waves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a time-averaged pressure field written in the ship's own moving frame contains enough information to predict the added resistance a routing system would act on; the paper only tests that premise in the ship-fixed frame, and its own data-transfer check shows that averaging raises the global-coordinate drag error from 0.94e-5 to 3.42e-5.","fun_headline_variants_meta":{"raw":{"variants":["AI surrogate maps wave conditions to full ship pressure field","Wave-to-drag surrogate: CAE/NN predicts pressure field in seconds","Learned manifold maps sea state to hull pressure and drag","Pressure field surrogate: from wave inputs to added resistance","Neural net predicts full 3D pressure field for ship drag from waves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000778,"raw_usage":{"total_tokens":3508,"prompt_tokens":1085,"completion_tokens":2423,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2336}},"tokens_in":701,"tokens_out":2423,"duration_ms":17801,"temperature":1.0,"reasoning_tokens":2336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:51:10.301434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the surrogate in the inertial frame by combining its predicted ship-fixed pressure field with the CFD pitch-angle history (or a separately predicted pitch angle), and compare the global-coordinate mean absolute drag error against the 0.94e-5 benchmark for instantaneous integration; if the error stays at the time-averaged level of 3.42e-5 rather than dropping, the surrogate does not deliver routing-grade resistance in earth-fixed coordinates.","supporting_citations":[],"review_version":1}