{"id":"fa7202e3-5664-4061-a9af-3c1f314d05a1","arxiv_id":"2501.04105","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DeepVIVONet reconstructs and forecasts riser vibrations under vortex shedding from three sparse sensors and optimizes where those sensors should sit.","lead":"The paper trains a neural network called DeepVIVONet to reconstruct and predict the bending motion of an offshore marine riser from just three strain sensors, and then uses the trained model to choose better sensor locations. It matters because better sensor placement with fewer instruments could cut the cost and risk of monitoring deep-sea risers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed sensor-placement superiority over POD rests on a single MSE pair with no error bars; the best POD placement is within 0.3%, so the central practical claim is not yet supported.","rationale":"I read the paper as claiming two things: (i) 3 sparse strain sensors plus DeepVIVONet can reconstruct/forecast riser strain fields, and (ii) the learned sensor locations beat POD-based placement. The reader's weakest assumption targeted (i) via low-dimensionality. I agree that this is a real assumption, but the POD analysis already provides partial support (96.91% variance in 3 modes), and the reconstruction figures are visually plausible. The more immediately load-bearing weakness is in (ii): the only quantitative comparison, Table 2, shows a learned placement whose MSE differs from a manually selected POD placement by only 0.3%, with no uncertainty quantification. Because the paper's title, abstract, and conclusion emphasize optimal sensor placement, an unsupported superiority claim undermines the central practical contribution. The lack of error bars and the unspecified 'reconstructed dataset' in Section 4.1 make the reported advantage non-evaluable. This does not overturn the conditional verdict; it strengthens it. The reader's rationale already noted that the POD baseline essentially matches, so my concern is a sharper version of that point rather than a disagreement. Hence verdict_should_be remains UNCHANGED (CONDITIONAL), with the condition being a statistically grounded and fair comparison to POD on raw sensor data.","tokens_in":12451,"tokens_out":8968,"duration_ms":90173,"concrete_test":"Retrain DeepVIVONet with at least 5 random seeds for each configuration in Table 2 (initial guesses, learned locations, POD (29,43,81), and an automated POD selection such as Q-DEIM or greedy maximization of the mode-observability determinant), keeping architecture and 500k-iteration schedule fixed; report mean and standard deviation of prediction-window MSE. Also state explicitly how the 500-point reconstructed dataset in Section 4.1 was generated, and if feasible recompute the same comparison on the raw 23-sensor readings plus two pinned boundaries. If the learned placement's mean MSE is not significantly lower than the best automated POD placement (e.g., overlapping 95% confidence intervals), the superiority conclusion should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central practical contribution is that DeepVIVONet's learned observer locations are more precise and cost-effective than POD-based placement (Abstract; Section 4.4). Table 2 is the only direct evidence: learned locations (4,67,92) give prediction-window MSE 1.290e-06, while POD placement (29,43,81) gives 1.294e-06. The gap is 0.3%, smaller than the run-to-run variation expected after 500,000 stochastic training iterations, and no seeds, error bars, or statistical comparison are reported. The POD baseline also used manually chosen extrema (Figure 12) rather than an automated placement criterion, so the comparison is not a fair protocol. Additionally, Section 4.1 states that optimization was performed on a 'reconstructed dataset' discretized at 500 equidistant points, but the method used to construct that dense field is never specified; if it relies on interpolation or on the same surrogate family, the Table 2 MSEs may not reflect performance on the original 23-sensor field data. Together these gaps mean the abstract's claim of 'more precise and cost-effective configurations' is not established by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DeepVIVONet, a DeepONet-based surrogate that maps time histories from m=3 observer strain sensors on a marine riser to the full CF/IL strain field, and uses the trained network in an outer-loop optimization to learn sensor locations. Experiments use NDP VIV field data for three shear-flow cases (U=0.50, 1.50, and 2.20 m/s) for reconstruction and forecasting, one transfer-learning case (U=1.5 to U=1.4 m/s), and a sensor-placement comparison against three manually selected POD-based location sets. The paper claims that DeepVIVONet provides accurate reconstruction and forecasting and that its learned sensor locations are more precise and cost-effective than POD-based placements.","tokens_in":12681,"tokens_out":4388,"duration_ms":41463,"significance":"If the claims held, the paper would offer a practical sparse-sensing framework for riser monitoring and a data-driven alternative to POD-based sensor placement. The use of real experimental NDP data, the direct comparison to POD, and the clear description of the DeepONet input/output structure are strengths. However, the central quantitative evidence is thin: the reconstruction/forecasting and transfer-learning claims are supported mainly by qualitative plots, and the sensor-placement superiority claim rests on a 0.3% MSE difference with no statistical characterization. The contribution is primarily empirical, and the conclusions need stronger quantitative support before they can be relied upon.","major_comments":[{"comment":"The claim that DeepVIVONet's learned observer locations are \"more precise and cost-effective\" than POD-based placement rests on a single MSE pair: learned locations (4,67,92) give 1.290e-06, while the best POD placement (29,43,81) gives 1.294e-06, a difference of about 0.3%. No seeds, error bars, or statistical comparison are reported, and the POD placements are chosen manually from Figure 12 rather than by an automated criterion. The observed gap is smaller than the run-to-run variation expected after 500,000 stochastic training iterations, so the practical superiority claim is not established by the evidence presented.","section":"Section 4.4, Table 2"},{"comment":"The reconstruction and forecasting results for the CF cases and the IL case are reported only as qualitative time-series and FFT plots; no quantitative error metric (MSE, relative L2, or similar) is given for the test window, and no confidence intervals are provided. This makes it impossible to assess the abstract's claims of \"accurately reconstructing\" and \"accurate predictions\" across the tested velocity regimes. The only quantitative numbers in the paper appear in Table 2 for the sensor-placement study.","section":"Section 3.1, Figures 5-8"},{"comment":"The transfer-learning demonstration uses a single near-neighbor case (U=1.5 m/s to U=1.4 m/s) and reports no quantitative metric or comparison with a model trained directly on the target case, so the abstract's claim of generalization to other flow conditions via transfer learning is not supported. At minimum, the authors should report the prediction-window error for the transferred model and compare it with a model trained directly on the target case.","section":"Section 3.2"},{"comment":"The sensor-placement optimization is performed on a \"reconstructed dataset\" discretized at 500 equidistant points, but the construction of this dense field is never described. If it relies on interpolation or on the same surrogate family as DeepVIVONet, the MSE values in Table 2 may not reflect performance on the original 23-sensor field data. The construction should be specified, and the final evaluation should be reported on the original sensor locations as well.","section":"Section 4.1"}],"minor_comments":[{"comment":"The definition of the matrix E is inconsistent: the text says \"with 100 spatial elements, and 100 time samples,\" but the displayed matrix has 250 rows and the preceding sentence mentions 250 time samples. Please clarify the number of time samples used in the POD analysis.","section":"Section 4.2"},{"comment":"The sentence \"The observer locations optimized by DeepVIVONet consistently delivered superior results\" is misleading given that POD placement (29,43,81) achieves an MSE of 1.294e-06, nearly identical to the learned placement's 1.290e-06; a more precise wording would acknowledge that the two are statistically indistinguishable on the reported evidence.","section":"Section 4.4"},{"comment":"The text says \"we use our DeepVIVONet model as a surrogate\" and describes sampling r realizations from the location distributions, but the number r and the method for obtaining strain values at non-instrumented sampled locations are not specified. Please state r and explain how the strain at those sampled points is obtained from the discrete sensor data.","section":"Section 4.1"},{"comment":"The choice m=3 is not justified beyond the POD variance argument; no experiments with other sensor counts are shown, so it is unclear how sensitive the method is to the number of observers. A brief sensitivity study or at least a discussion of the choice would strengthen the paper.","section":"Section 3.1"},{"comment":"There are several typographical and grammatical errors, e.g., \"explaind\" in Section 4, \"the the motion\" in the Introduction, and \"We find that that\" in the Abstract. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece is treating observer locations as learnable Gaussian distributions (means and stds) and alternating their optimization with the network weights. That is a legitimate extension of DeepONet and of existing POD-based placement ideas, and it is applied to real NDP riser data, which is a plus. The reconstruction and forecasting results across three flow speeds look plausible, and the transfer-learning case (U=1.5 to U=1.4) is at least a sensible near-neighbor test.\n\nThe soft spots are real but not fatal. The main practical claim—that DeepVIVONet yields more precise and cost-effective sensor configurations than POD—rests entirely on Table 2, where the learned locations (4,67,92) give MSE 1.290e-06 and the best POD placement (29,43,81) gives 1.294e-06. That is a 0.3% gap, far smaller than the run-to-run variation you would expect after 500k stochastic training iterations, and there are no seeds or error bars. The POD baseline also used manually picked mode extrema rather than an automated criterion, so the comparison is not a fair protocol. The stress-test note about the 'reconstructed dataset' at 500 equidistant points is worth taking seriously: Section 4.1 never says how that dense field was built, and if it came from the same surrogate family, the Table 2 numbers may not reflect performance on the original 23-sensor data. That should be clarified.\n\nThere is also the inherent low-dimensionality assumption: three instantaneous strain point measurements mapping to the full riser field. The paper's own POD analysis (first three modes, 96.91% variance) supports this for one case, but it is not tested for other sensor counts or for the transfer-learning scenario, and no quantitative errors are reported anywhere—just qualitative time traces and FFTs. So the evidence is suggestive, not conclusive.\n\nThe citation patterns are fine; DeepONet and the NDP dataset are both properly credited. No code or data is shipped, which hurts reproducibility, but the paper is not circular: the held-out prediction window means the reconstruction result is not just memorization.\n\nWho gets value: someone working on sparse sensing for offshore structures or neural-operator-based sensor placement will find a useful demonstration and a clear framework. It deserves a serious referee—the methodological idea is worth engaging with—but the authors should be pushed to add error bars, multiple seeds, a fair POD baseline, and a description of how the dense dataset was constructed. My verdict: conditional accept with major revision.","headline":"A reasonable DeepONet application to VIV reconstruction and sensor placement, but the headline claim about beating POD rests on a single 0.3% MSE gap with no error bars.","tokens_in":13228,"tokens_out":642,"would_cite":false,"duration_ms":7589,"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":"DeepVIVONet reconstructs and forecasts the full cross-flow and in-line strain fields of a VIV-excited marine riser from only three observer sensors, and learns the sensor locations that minimize prediction error.","keywords":["deep operator network","vortex-induced vibrations","sensor placement optimization","transfer learning","marine riser","sparse measurements","proper orthogonal decomposition","strain reconstruction"],"falsifier":"Train DeepVIVONet on the test2430 shear case with m=2 observers instead of m=3, keep the same optimization procedure, and measure prediction-window MSE; if the error rises by more than an order of magnitude relative to the reported 1.290e-06, the 'three sensors suffice' claim is specific to that sensor count rather than a general low-dimensional property. A second check is to train on test2430 and test on test2500 (1.50 m/s vs 2.20 m/s) without transfer learning; the transfer claim predicts graceful degradation, while a collapse would bound the generalization to near-neighbor velocities.","tokens_in":12239,"feed_emoji":"📡","tokens_out":5440,"duration_ms":48354,"temperature":0.7,"pith_summary":"This paper claims that the full vibration field of an offshore marine riser can be recovered from just three streamed strain measurements, using a neural operator it calls DeepVIVONet. The trained model both reconstructs the strain along the whole riser in the training window and forecasts it in a future window from the same three observer signals. The paper further claims that making the observer locations into learnable parameters and optimizing them alongside the network gives a sensor layout (sensors at 4, 67, and 92) whose prediction-window error is the lowest among the tested choices, including three manually selected POD-based placements. If right, this would mean offshore riser monitoring needs far fewer strain gauges than current practice, and sensor placement can be designed by the same model that does the monitoring.","feed_headline":"Three sensors reconstruct a riser's full strain field","feed_subtitle":"A neural operator learns the sensor locations that give the lowest forecast error on vortex-induced vibrations.","key_machinery":"The load-bearing object is the DeepONet decomposition $G_\\theta(u)(z^*) = \\sum_{k=1}^P B_k(u) T_k(z^*) + B_0$, where the branch net $B_k$ consumes the $m$ observer strain values at a time step (or a look-back stack for forecasting) and the trunk net $T_k$ consumes the query location $(t_j, z^*)$. For sensor placement, observer locations are turned into learnable Gaussian distributions $(\\mu_i, \\sigma_i)$; at each training step $r$ realizations are sampled from each distribution, and the location parameters are optimized alternately with the network weights so the $\\sigma_i$ shrink toward the informative positions. The POD baseline supplies three manually selected three-sensor combinations from the first three spatial modes, which the paper both matches and compares against.","core_discovery":"On the paper's terms, the central discovery is that a DeepONet-style operator can learn a map from the m=3 observer strain histories to the entire strain field ε(z,t), and that this map is accurate enough that the prediction window (data never used in training) is matched to the level shown in the time-domain and FFT comparisons. The supporting quantitative comparison is Table 2, where the learned observer locations (4, 67, 92) give prediction-window MSE 1.290e-06, versus 3.745e-05 for the initial guesses and 9.838e-06, 1.294e-06, and 5.410e-06 for three POD-based choices. The learned locations are not uniformly better than every POD choice, but the paper argues POD selection is sensitive to manual choices while the learned locations are produced automatically. The paper also claims transfer learning: a network trained on shear-flow case test2430 (U=1.50 m/s) predicts the neighboring case test2420 (U=1.40 m/s) without retraining.","pith_inferences":["Editorial inference: the same differentiable-location optimization could be applied to other sparse-sensing problems, such as temperature or pressure field reconstruction in structures, provided the underlying field is comparably low-dimensional.","Editorial inference: because the POD analysis in Section 4.2 shows 96.91% of variance in three modes for one shear case, the method's success likely tracks that low-dimensionality; a natural test is to run the same pipeline on a multi-frequency VIV case with richer modal content and check whether three observers still suffice.","Editorial inference: the paper's comparison does not include robustness of the learned locations to sensor failure; a testable extension is to retrain with one observer randomly dropped at test time and measure error degradation."],"forward_implications":["A three-sensor strain array on a riser can replace denser instrumentation for reconstruction and short-horizon forecasting of both cross-flow and in-line strain.","Sensor placement can be included in neural-operator training, making the placement problem differentiable rather than a separate heuristic.","A model trained at one maximum flow speed transfers to a neighboring speed without full retraining, reducing calibration cost when operating conditions drift.","The trained network acts as a fast surrogate, so monitoring can run in near-real time instead of relying on CFD or dense instrumentation.","POD-based placement remains a useful initialization but its manual mode-selection step causes large variation in error; the learned locations automate that selection."],"supporting_citations":[{"why":"Introduces the DeepONet architecture and the operator-learning formulation that DeepVIVONet builds on.","marker":"[31]"},{"why":"States the universal approximation theorem for nonlinear operators that motivates the DeepONet design.","marker":"[33]"},{"why":"Supplies the NDP experimental strain data for the riser cases (test2330, test2430, test2500) used in all experiments.","marker":"[34]"},{"why":"Provides the POD-based sensor-placement method that the paper adopts as the comparison baseline.","marker":"[38]"},{"why":"Demonstrates neural operators for predicting real-time response of floating offshore structures, the nearest prior application that DeepVIVONet extends.","marker":"[27]"},{"why":"Uses transformers for forecasting VIV from sparse vibration signals, providing the forecasting baseline context.","marker":"[30]"}],"fun_headline_variants":["Neural operator learns optimal sensor locations for VIV","Three sensors, full riser strain: deep operator picks optimal sites","Deep operator outperforms POD in sensor placement for VIV","Transfer learning predicts VIV from network trained on other flow","Sparse measurements, deep operator: optimal riser sensor placement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that three point strain measurements determine the entire riser strain field at that instant, meaning the VIV response is low-dimensional enough for a learned map from three observer values to all spatial locations to generalize across time windows and nearby flow speeds.","fun_headline_variants_meta":{"raw":{"variants":["Neural operator learns optimal sensor locations for VIV","Three sensors, full riser strain: deep operator picks optimal sites","Deep operator outperforms POD in sensor placement for VIV","Transfer learning predicts VIV from network trained on other flow","Sparse measurements, deep operator: optimal riser sensor placement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000995,"raw_usage":{"total_tokens":4211,"prompt_tokens":938,"completion_tokens":3273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":3190}},"tokens_in":554,"tokens_out":3273,"duration_ms":20511,"temperature":1.0,"reasoning_tokens":3190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:40:48.236184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DeepVIVONet on the test2430 shear case with m=2 observers instead of m=3, keep the same optimization procedure, and measure prediction-window MSE; if the error rises by more than an order of magnitude relative to the reported 1.290e-06, the 'three sensors suffice' claim is specific to that sensor count rather than a general low-dimensional property. A second check is to train on test2430 and test on test2500 (1.50 m/s vs 2.20 m/s) without transfer learning; the transfer claim predicts graceful degradation, while a collapse would bound the generalization to near-neighbor velocities.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the DeepONet architecture and the operator-learning formulation that DeepVIVONet builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"States the universal approximation theorem for nonlinear operators that motivates the DeepONet design."},{"cited_title":"Braaten, H","cited_arxiv_id":null,"evidence_quote":"Supplies the NDP experimental strain data for the riser cases (test2330, test2430, test2500) used in all experiments."},{"cited_title":"Yildirim, C","cited_arxiv_id":null,"evidence_quote":"Provides the POD-based sensor-placement method that the paper adopts as the comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates neural operators for predicting real-time response of floating offshore structures, the nearest prior application that DeepVIVONet extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Uses transformers for forecasting VIV from sparse vibration signals, providing the forecasting baseline context."}],"review_version":1}