{"id":"c657c103-da9d-496d-a8c8-25eba2ba568f","arxiv_id":"2504.14971","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid neural surrogate combining Fourier neural operators, time-fused ConvLSTM, and an MLP wall-shear model predicts tandem-cylinder vortex-induced vibration flow fields and structural responses at Re=200.","lead":"Researchers built a neural network surrogate that predicts vortex-induced vibrations of two cylinders in tandem, replacing the fluid solver in a fluid-structure interaction loop with a deep learning model. The model runs about a thousand times faster than CFD for this configuration and matches the simulated vibration response reasonably well.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GIP displacement feedback (Sec. 4.4.2) is load-bearing but its training-time oracle selection mismatches the averaged inference procedure; an ablation is needed to support the claimed accuracy.","rationale":"The central claim is that the hybrid model accurately and reliably predicts VIV for two tandem cylinders. That claim is supported almost entirely by a single closed-loop roll-out on a held-out temporal segment of the same trajectory used for training. For such a roll-out, the only quantity that mediates all coupling is the displacement fed back through Eq. 26. The paper's own description of the GIP correction in Sec. 4.4.2 introduces an oracle-selection step during training (choose the candidate displacement closer to the labeled one) that is not used at inference (average instead). This is a train/test mismatch in the most load-bearing variable. It also makes the method under-specified: no GIP algorithm, thresholds, or validation is given, so the result cannot be reproduced or checked. Generalization to unseen configurations and missing code are serious but secondary; even for the tested configuration, the reported accuracy could be an artifact of the oracle. The concrete ablation proposed would settle this: retrain with the inference-time averaging rule and compare errors. If the model degrades, the claim needs to be revised. If it does not degrade, the concern is resolved and the paper's conditional acceptance is justified. The reader's conditionality is therefore unchanged.","tokens_in":30030,"tokens_out":8522,"duration_ms":79808,"concrete_test":"Retrain the full USFNO-based hybrid model on the same 800 training snapshots, but replace the 'better one' oracle selection with the prediction-phase rule: always use the average of the ODE displacement and the GIP displacement for feedback, applied from epoch 1 onward. Then evaluate on the same hold-out segment (tU0/D = 160-240) and compare displacement MAE/RMSE for upstream and downstream cylinders against Tables 4-5 and flow-field MAE/MRE against Table 6. If the displacement MAE rises above the claimed 0.0006 m, or the roll-out diverges within the 400-step horizon, then the reported accuracy depends on the oracle selection and the central claim is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The two-cylinder FSI claim rests on closed-loop displacement feedback into the flow-field network. In Sec. 4.4.2 the authors state that during training, after epoch 180, 'the better one between the computationally derived vibration displacement and that captured via gradient analysis is selected,' while in the predicting phase 'the average of the aforementioned two forms is adopted.' The 'better one' is evidently the candidate closer to the labeled displacement, so the feedback distribution used in training is oracle-selected and differs from the averaged procedure used in deployment. This is a train/test mismatch in the most load-bearing component: Eq. 26 injects y_{i-1} multiplicatively into the low-dimensional flow features, so the network is trained under more accurate boundary feedback than it receives at inference. The reported displacement MAE below 0.0006 m and the 400-step roll-out stability (Tables 4-6) could therefore depend on a correction mechanism that is under-specified (no algorithm, thresholds, or validation) and whose training procedure cannot be reproduced without labels. The GIP reads cylinder position from the predicted flow field, so it is not independent of the flow-field network's errors; averaging it with the ODE displacement may mask drift rather than correct it. If the GIP correction is inaccurate or the exponential coupling is not a faithful encoding of the moving boundary, the closed-loop prediction will drift. This threatens the central claim even for the tested tandem configuration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid neural-physics FSI solver for vortex-induced vibration of two elastically mounted cylinders in tandem and of an isolated cylinder at Re = 200. The fluid part is a novel USFNO-FConvLSTM network that predicts the next flow field from N historical flow fields and the current structural displacement; a separate MLP predicts wall shear stress; the structural response is advanced by a fourth-order Runge-Kutta solver. The two components exchange surface forces and displacement feedback (Eqs. 26-28). The authors report displacement MAE below 0.0006 m for the tandem case, pressure-field MAE/MRE of 0.0159 and 14.46% in the prediction phase, and prediction speeds over 1000x faster than CFD, with additional comparisons against CNN- and Unet-based baselines.","tokens_in":1904,"tokens_out":2260,"duration_ms":63023,"significance":"If the reported accuracy and speed hold under a properly specified training and evaluation protocol, the paper would make a useful contribution: a Fourier-operator-based recurrent surrogate for multi-body FSI and a modular coupling framework that goes beyond single-cylinder studies. The paper's strengths are its systematic study of fusion time steps (N = 16), its controlled architecture comparisons against CNN-FConvLSTM and Unet-FConvLSTM baselines, and the inclusion of both an isolated- and a tandem-cylinder case. These relative comparisons support the claim that the proposed flow-field architecture outperforms the authors' baselines. However, the absolute accuracy claims are not yet fully established because the closed-loop displacement feedback is trained with an oracle-selected correction scheme and the evaluation is limited to later time segments of the same simulation trajectories; no code or data are released, which further limits reproducibility.","major_comments":[{"comment":"The GIP displacement feedback is load-bearing, but the training procedure is not reproducible and is mismatched with inference. The text states that during training, after epoch 180, 'the better one between the computationally derived vibration displacement and that captured via gradient analysis is selected,' whereas in the predicting phase 'the average of the aforementioned two forms is adopted.' Since 'better' is evidently defined relative to the labeled displacement, the training feedback distribution is oracle-selected and differs from the averaged inference procedure. Because Eq. (26) injects y at time i-1 multiplicatively into the low-dimensional fluid features, the network is trained under more accurate boundary feedback than it receives in deployment, so the displacement MAE values in Tables 4-5 and the long roll-out accuracy shown in Figures 24-27 cannot be attributed to the published training procedure without an ablation. Please provide a full specification of the GIP algorithm (gradient computation, threshold, validation) and add an ablation comparing (a) ODE-only feedback, (b) average feedback from epoch 1, and (c) the oracle-selection schedule, reporting displacement and flow-field errors for each.","section":"Sec. 4.4.2, Eq. (26)"},{"comment":"The recurrent training procedure is under-specified, which matters because the model is autoregressive over the FSI loop. The manuscript does not state whether the USFNO-based FSI recurrent unit is trained with teacher forcing or closed-loop rollouts, how the loss is accumulated over the sequence, or how the epoch-180 GIP selection threshold was chosen. These details determine the distribution of the structural displacement seen by the flow-field network during training and directly influence the error-accumulation behavior reported in Figures 24-27. Please specify the exact training loop and report a sensitivity study over the switch epoch and over the choice of teacher forcing versus closed-loop training.","section":"Sec. 4.1 and Sec. 4.4.2"},{"comment":"The evaluation does not currently support the abstract's claim that the model can 'accurately and reliably predict' VIV evolution generically. For both the isolated-cylinder and tandem-cylinder cases, the prediction region is a later time segment of the same simulation trajectory used for training, so the test is in-sample temporal extrapolation rather than prediction for a new configuration. No experiment with different initial conditions, reduced velocity, or spacing ratio is reported, and the tandem case has only one trajectory. Please either add at least one unseen trajectory per case (e.g., different reduced velocity or initial disturbance) or explicitly restrict the claims to interpolation of the trained trajectory, which would be a more modest but defensible statement.","section":"Sec. 4.4.1 and Sec. 4.4.2"},{"comment":"The wall shear stress module is described as a function of the previous wall shear stress alone, with no dependence on the predicted pressure field, the structural displacement, or the flow-field state. Yet the wall shear force from Eq. (28) is a component of the coupling force that drives the structural solver. As written, this module is an open-loop autoregressive model of the shear stress and does not receive the current boundary feedback used by the flow-field network. Please clarify whether the MLP is trained jointly with the flow-field model and whether the structural displacement or the predicted flow field enters its inputs; if the MLP is open-loop, the wall-shear force predictions in Figure 28 need to be re-evaluated in closed loop, and the 'fluid-structure interaction' characterization should be qualified.","section":"Sec. 2.1.1 and Eq. (1)"}],"minor_comments":[{"comment":"The MRE formula is ambiguous as typeset: it should be written with explicit parentheses as the sum over i of absolute errors divided by the sum over i of absolute truth values, multiplied by 100 percent. Please rewrite the equation to avoid confusion between the summation over the error and the summation over the denominator.","section":"Eq. (40)"},{"comment":"The discrete Fourier transform formulas contain garbled notation, including unclear exponential arguments and summation bounds. These should be rewritten with proper parentheses and explicit bounds for x1, x2, k1, and k2 so that the modes are unambiguous.","section":"Eqs. (8)-(9)"},{"comment":"The speedup claim of '4020 times faster' is reported without describing how the comparison was timed (GPU inference batching, number of CFD cores, warm-up, or memory transfer). Please provide the timing methodology and report both training time and inference time per rollout.","section":"Sec. 4.3"},{"comment":"The data availability statement says data are available from the corresponding author on request, but no code or pretrained models are mentioned. Given the many structural hyperparameters and the GIP threshold, releasing the implementation would substantially improve reproducibility.","section":"Data availability"},{"comment":"The manuscript contains several typographical and notation inconsistencies, such as 'physis field features' in Section 2.1.2, 'FIS neural model' in Section 5, and the inconsistent use of the flow-field state notation; a careful proofreading pass is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's main claim rests on a closed-loop feedback mechanism that is evaluated with an oracle-selected training signal, and the evaluation is limited to in-sample temporal extrapolation. These issues are fixable with an ablation and additional experiments, so I do not recommend rejection, but the revision needs to be substantive. I would also encourage the editor to ask for a clearer novelty statement relative to Han et al. (2022) and Fan and Wang (2024), since the current framing is largely architectural and the FSI-coupling mechanism resembles those earlier hybrids."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a genuine architecture paper for FSI surrogates, but the displacement-feedback mechanism (the GIP correction in Sec. 4.4.2) has a train/test mismatch that makes the headline accuracy numbers unreliable until it is fixed or ablated. The paper deserves a serious referee, but I would not cite it as evidence of a working multi-cylinder VIV surrogate yet.\n\nWhat is new: the USFNO-FConvLSTM flow model (U-shaped FNO with time-fused ConvLSTM) plus MLP wall-shear submodel coupled to a fourth-order Runge-Kutta structural solver is a combination I have not seen in the cited literature. The application to two tandem cylinders at Re=200 is also new. The paper does solid homework: it validates the CFD data against Griffith et al. and Zhao, checks the nearest-neighbor interpolation by comparing lift from the processed fields, and reports systematic ablations of the fusion time step and comparisons against its own CNN- and Unet-based baselines. The wall-shear term is a nice, physically motivated addition at this Reynolds number.\n\nThe soft spot is in Sec. 4.4.2. The GIP 'gradient analysis' correction is load-bearing: the displacement feedback enters the flow network multiplicatively via Eq. 26. Yet the authors state that after epoch 180 they select 'the better one' between the ODE displacement and the GIP position during training, while at inference they average the two. 'Better one' evidently means closer to the label, so the network is trained with boundary feedback that is more accurate than what it receives at deployment. That is exactly the kind of oracle selection that can produce single-trajectory roll-out stability without real prediction skill. The text gives no algorithm, thresholds, or validation for GIP, and there is no ablation running the model with and without it. The GIP reads the cylinder position from the predicted flow field, so averaging it with the ODE solution may be hiding drift rather than correcting it.\n\nThe other reservations are more minor. The demonstration is one trajectory at one Re and spacing ratio; the conclusions generalize beyond the evidence. Training details (teacher forcing? loss weighting?) are not specified. No code or data release, just 'available upon request.' None of this is disqualifying by itself, but combined with the GIP issue, the accuracy claims are not yet established.\n\nFor you: if you need a surrogate for tandem VIV, this paper is worth reading for the architecture, not for the numbers. I would send it to peer review because the idea is useful and the comparisons are honest, but I would ask for a GIP ablation, a precise training protocol, and at least one unseen configuration before accepting the central claim.","headline":"Useful FSI surrogate architecture, but the GIP displacement-feedback correction has a training/inference mismatch that undermines the headline accuracy; deserves peer review, not citation yet.","tokens_in":30893,"tokens_out":3050,"would_cite":false,"duration_ms":28471,"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 hybrid neural network predicts vortex-induced vibration of two tandem cylinders over 1000 times faster than CFD while tracking flow fields and structural motion.","keywords":["vortex-induced vibration","fluid-structure interaction","two tandem cylinders","Fourier neural operator","convolutional long short-term memory","moving boundary","surrogate modeling","flow field prediction"],"falsifier":"Feed the trained network the true CFD displacement at each step instead of the network's own GIP-averaged displacement and compare flow-field errors. If the errors do not drop substantially, the displacement feedback is not carrying the boundary information; if they do, the closed-loop scheme depends on the readout being accurate. A second check: roll the network out for more than one vibration period and track the predicted cylinder boundary against the CFD mesh position; phase error growing without bound would falsify long-horizon reliability.","tokens_in":29834,"feed_emoji":"🌊","tokens_out":9091,"duration_ms":78106,"temperature":0.7,"pith_summary":"The paper sets out to show that a hybrid neural solver can replace the fluid part of a fluid–structure interaction (FSI) simulation for vortex-induced vibration (VIV) of two elastically mounted cylinders in tandem at Reynolds number 200. Its fluid model, a U-shaped Fourier neural operator joined to a time-fused convolutional LSTM, takes sixteen historical flow-field snapshots and the current cylinder displacement and outputs the next flow field; a structural solver then integrates the surface pressure and wall-shear forces to update displacement and velocity. On the tandem case the reported prediction errors are MAE 0.0159 for pressure, MRE 14.46%, and displacement MAE below 0.0006 m, with a speedup over numerical simulation of more than 1000 times. If these numbers stand, closed-loop multi-body FSI prediction becomes feasible without re-meshing, which matters for simulating riser bundles and other vibrating structures where the moving boundary is the expensive part.","feed_headline":"Neural solver predicts twin-cylinder vibration 1000x faster","feed_subtitle":"Tracks pressure fields and cylinder motion without re-meshing, closing the fluid–structure loop for two tandem cylinders.","key_machinery":"The load-bearing object is the USFNO-FConvLSTM flow-field model inside a recurrent FSI unit. The FNO part is a Fourier neural operator: a learnable integral operator that acts in frequency space using only the first $k_{\\max}$ Fourier modes, which is what lets the network resolve the sharp zero-padded cylinder boundary as high-frequency content. The FConvLSTM part is the time-fusion module: it swaps the time and channel axes, runs a ConvLSTM over the $N$ historical steps, then linearly projects the fused features to one future state. The U-shaped tensor additions carry encoder features into the decoder to restore information lost in down-sampling. Around that core sit the MLP wall-shear model, which advances wall shear stress one step, the structural dynamics integrator, and the two coupling transfers: $F_b = F_{b-P} + F_{b-wall}$ from fluid to structure, and $U_F' = U_F \\exp(y_{i-1})$ from structure to fluid. Together these form the USFNO-based FSI recurrent unit, and the whole network is a sequence of such units sharing hidden states.","core_discovery":"The central claim is that the proposed USFNO-based hybrid neural network can accurately and reliably predict the VIV evolution of two cylinders in tandem, including instantaneous flow-field states, wall shear stress, lift force, and structural displacement and velocity, at a fraction of the cost of CFD. The paper argues that this works because the flow-field model combines a Fourier neural operator, which captures the high-frequency gradient information at the moving cylinder edges, with a U-shaped encoder–decoder and a ConvLSTM variant that fuses $N$ historical time steps before projecting to one future state. The structural response is not learned: it is computed by fourth-order Runge–Kutta integration of the mass-damping-spring equation driven by forces integrated from the predicted pressure and wall-shear fields. Coupling is closed through force transfer from fluid to structure and through multiplicative displacement feedback $U_F' = U_F \\exp(y_{i-1})$ from structure to fluid. The claim is comparative as well: with similar parameters, USFNO-FConvLSTM beats CNN- and Unet-based alternatives, and the multi-time-step FSI model beats its single-step CNN counterpart, especially for the interacting downstream cylinder.","pith_inferences":["A testable extension would remove the exponential feedback and instead feed displacement as an extra input channel or additive term; if accuracy holds, the exponential form is not load-bearing, while if it fails, the moving-boundary encoding is doing more work than the flow-field network alone.","Because errors accumulate with prediction horizon, re-initializing the flow-field state from a CFD snapshot or from a detected boundary position at regular intervals could stabilize very long rollouts; the paper does not test this.","The gradient-based position readout (GIP) suggests the network localizes boundaries from image gradients; this could be extended to flexible structures by tracking a parameterized deformed boundary instead of a rigid translation.","The same closed-loop architecture might apply to aeroelastic or marine-riser problems with multiple interacting bodies, but only if the training data cover the relevant reduced-velocity and spacing-ratio range; the paper explicitly lists multi-condition generalization as future work."],"forward_implications":["Multi-object VIV can be modeled without re-meshing: the next flow-field image is produced directly by the network, so the costly boundary-mesh update is bypassed.","Wall shear stress is a necessary part of the force transfer at Re = 200; pressure-only integration visibly mis-fits the lift force, so neural FSI surrogates for low-Re flows should include it.","Fusing about one vibration period of history (N = 16 of roughly 25 snapshots per cycle) improves accuracy; both shorter and longer histories degrade it, giving a design rule for sequence length.","The same architecture transfers to an isolated cylinder, sustaining phase-coherent displacement and velocity prediction over long horizons, which supports generalization beyond the tandem configuration.","For flow-field evolution the reported speedup is about 4020 times over the numerical solver, and for the full FSI loop more than 1000 times, which makes parameter sweeps and control studies practical."],"supporting_citations":[{"why":"Supplies the Fourier neural operator formulation and the truncated spectral integral operator used in the Conv-Fourier layers.","marker":"Li et al. (2021)"},{"why":"Provides the ConvLSTM cell that the paper modifies by time-fusion into FConvLSTM.","marker":"Shi et al. (2015)"},{"why":"Supplies the U-Net encoder–decoder idea behind the U-shaped USFNO architecture.","marker":"Ronneberger et al. (2015)"},{"why":"Establishes the CNN-plus-structural-dynamics FSI baseline that the paper extends and outperforms for multi-cylinder moving boundaries.","marker":"Han et al. (2022)"},{"why":"Provides the published VIV results for tandem and staggered cylinders used to validate the simulation dataset.","marker":"Griffith et al. (2017)"},{"why":"Provides published isolated-cylinder VIV results used to validate the in-house CFD dataset.","marker":"Zhao, (2022)"},{"why":"Documents the proximity-wake interference of two tandem VIV cylinders that motivates the multi-body test case.","marker":"Borazjani & Sotiropoulos (2009)"},{"why":"States the high-Reynolds wall-shear omission that the paper contrasts with its Re = 200 case, justifying the MLP wall-shear model.","marker":"Yin et al. (2022)"}],"fun_headline_variants":["Hybrid neural net predicts twin-cylinder VIV 1000x faster","USFNO-FConvLSTM predicts tandem cylinder FSI at 1000x speed","Neural solver for tandem cylinders: 1000x faster, high accuracy","Twin-cylinder vibration prediction: 1000x faster with deep learning","Hybrid deep learning predicts fluid-structure interaction for two cylinders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that multiplying the low-dimensional flow features by $\\exp(y_{i-1})$ and using a displacement that is itself partly read back from the predicted flow field faithfully encodes the moving boundary condition; if that feedback drifts, the closed-loop prediction will drift.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid neural net predicts twin-cylinder VIV 1000x faster","USFNO-FConvLSTM predicts tandem cylinder FSI at 1000x speed","Neural solver for tandem cylinders: 1000x faster, high accuracy","Twin-cylinder vibration prediction: 1000x faster with deep learning","Hybrid deep learning predicts fluid-structure interaction for two cylinders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1803,"prompt_tokens":1046,"completion_tokens":757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":662,"tokens_out":757,"duration_ms":6356,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:35:54.476679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the trained network the true CFD displacement at each step instead of the network's own GIP-averaged displacement and compare flow-field errors. If the errors do not drop substantially, the displacement feedback is not carrying the boundary information; if they do, the closed-loop scheme depends on the readout being accurate. A second check: roll the network out for more than one vibration period and track the predicted cylinder boundary against the CFD mesh position; phase error growing without bound would falsify long-horizon reliability.","supporting_citations":[],"review_version":1}