{"id":"ee08e95f-6c9c-43b0-aef6-6383ec3034bc","arxiv_id":"2412.21132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DeepF-fNet uses two convolutional networks plus a shift-and-scale routine to estimate locally resonant metamaterial parameters that place a vibration bandgap on a user-specified frequency.","lead":"This paper trains a two-part neural network that turns a target vibration frequency into the geometry of a metamaterial designed to block it. The online lookup is about 500,000 times faster than a genetic algorithm, but that comparison excludes the expensive offline training and data generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SICE4's correction step feeds the IEPS shifted/rescaled spectra that may lie outside the training distribution; only a single 200 Hz target is FE-verified, so the broad 'real-time optimal parameters' claim is not yet supported.","rationale":"The reader's conditional verdict is appropriately cautious. The method has real strengths: a plausible dual-network architecture, an inverse map from dispersion spectra to parameters, and one end-to-end FE validation at 200 Hz where the SICE4 result produces a bandgap matching the GA solution and the target. The 500000x figure is an online-inference comparison and is not by itself misleading, provided the intended use is after one-off offline dataset construction. However, the core algorithmic innovation is the SICE4 prediction-correction that manufactures a spectrum by shifting and rescaling. The validity of this operation is precisely where the paper is thinnest. The IEPS is trained on physical spectra; the synthetic spectra are not physical; the 200 Hz test is a single realization and may be an easy target because the closest dataset sample already has a nearby bandgap. The time-series demonstration stops at predicted parameters, without FE confirmation, so it does not close the gap. The physics-informed terms are numerically irrelevant, removing the only mechanism that might regularize out-of-distribution inputs. The proposed frequency sweep is cheap relative to the dataset already generated (2000 FE runs) and would either shore up or falsify the central claim. I therefore agree with the reader's weakest_assumption and keep the conditional verdict, requiring such a sweep or equivalent multi-target FE validation before acceptance.","tokens_in":14759,"tokens_out":6141,"duration_ms":62241,"concrete_test":"Sweep target frequencies across the operating range (e.g., 50-750 Hz in 25 Hz steps). For each target, run SICE4, use the predicted (r,L,s) to rebuild the unit cell, and recompute the dispersion curves with the same COMSOL eigenfrequency study as used for the dataset. Record whether the obtained first bandgap contains the target frequency and the absolute error of its center. If, for any nontrivial fraction of targets (e.g., >5%) the gap misses the target by more than the spectral resolution used in training, the generalization claim is falsified. As a secondary check, compare the FE spectrum of the predicted parameters against the synthetic spectrum that was fed to IEPS to see whether the inversion is actually consistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 1 (lines 6-10), the synthetic spectrum 'new DC' is created by shifting all branches n>1 by Δ and rescaling branch n=1 by Γ = min(new DC_{n=2})/max(DC_{n=1}[ib]). This is a rigid frequency shift plus a branch-wise scaling; it does not correspond to any physical (r,L,s) geometry, because eigenfrequencies do not transform this way under geometric changes. The IEPS was trained exclusively on FE-generated spectra from valid parameter triples, so the manufactured input lies on a different manifold. The paper's only external FE check of the full SICE4 pipeline is a single 200 Hz target compared with a single GA run (Table 7/Fig. 10). The time-series simulation (Fig. 9) outputs parameters for 300 targets between 250 and 750 Hz but never re-runs the FE solver to confirm that the resulting bandgaps actually bracket those targets. The claimed speed advantage is about inference, but the correctness of the parameters is only demonstrated at one frequency. The near-zero PDE/BC weights (Table 5) mean no physical constraint is active to pull off-manifold inputs back to realistic solutions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeepF-fNet, a dual-CNN neural network that maps dispersion curves of a locally resonant metamaterial unit cell to the geometric parameters (r, L, s), and SICE4, a prediction-correction algorithm that transforms a stored dispersion curve to approximate a user-specified target bandgap frequency before querying the network. The intended application is real-time semi-active vibration isolation. The case study trains on 2000 finite-element-computed dispersion spectra, validates the inverse solver (IEPS) by feeding predicted parameters back to the FE solver and comparing the resulting dispersion curves, and compares SICE4 against a genetic algorithm (GA) for a single 200 Hz target. The authors report a speed advantage of up to 5e5 times and comparable first-bandgap accuracy.","tokens_in":15044,"tokens_out":6003,"duration_ms":57245,"significance":"The paper contains a credible validation of the first-bandgap prediction: predicted parameters are re-run through the FE solver and shown to reproduce the target bandgap, which rules out simple overfitting to training labels. The SICE4 heuristic of selecting the nearest stored spectrum and then applying a shift/scale correction is a practical idea for inverse eigenvalue problems in this setting. If the off-manifold generalization of SICE4 were demonstrated across a range of target frequencies, the framework would be a valuable tool for fast structural optimization. However, the current evidence supports only a narrow, single-frequency claim; the broad real-time claims in the abstract and conclusions are not yet backed by the experiments.","major_comments":[{"comment":"The correction step constructs an input spectrum by shifting all branches n>1 by Δ and rescaling branch 1 by Γ = min(new DC_{n=2})/max(DC_{n=1}[ib]). This synthetic spectrum is not a physically realizable dispersion curve of any (r,L,s) triple, because eigenfrequency branches do not transform in this way under geometric changes. The IEPS was trained exclusively on FE-generated spectra from valid parameter triples (Section 3.2), so the corrected input lies on a different manifold. The paper's only FE verification of the full SICE4 pipeline is a single 200 Hz target compared with a single GA run (Table 7, Fig. 10). The time-series simulation in Fig. 9 outputs parameters for 300 targets between 250 and 750 Hz but does not re-run the FE solver to confirm that the resulting bandgaps actually bracket those targets. To support the central claim of real-time optimal parameter estimation over a user-defined frequency range, the authors should FE-verify a representative sample of target frequencies, or otherwise quantify the network's error on artificial SICE4 inputs versus genuine FE spectra.","section":"Section 2.1, Algorithm 1 lines 6-10"},{"comment":"The speed comparison is asymmetric. The SICE4 inference time of 0.0157 s excludes the offline cost of generating the 2000-sample FE dataset and training both networks, whereas the reported GA time of 7690 s is the full optimization time. Offline costs can legitimately be amortized in real-time deployment, but the 'up to 500,000 times faster' statement in the abstract and Section 4 should be presented as an inference-time comparison. Moreover, the GA settings (population size, number of generations, objective function definition) are not specified, so the 7690 s figure cannot be assessed or reproduced.","section":"Section 3.3, Table 7"},{"comment":"The PDE loss weight is 10^-14 and the BC loss weight is 10^-12 for both networks, while the data loss weights are 10^8 and 10^12. With these settings the physics terms contribute negligibly to the total loss, so the framework is effectively a purely data-driven surrogate. The paper's claim that 'incorporating physical laws into training reduces data requirements and improves accuracy' (Section 4) is therefore not supported by the experiments. The authors should either show an ablation with meaningful physics weights, or revise the characterization of the method as physics-informed.","section":"Section 2.2, Table 5"}],"minor_comments":[{"comment":"'Its width (along the y-axis) was optimized' appears to be a typo; the width is along the x-axis according to Eq. (4c) and Fig. 3a.","section":"Section 3.1"},{"comment":"The Floquet-Bloch condition is written for a vector w0 but the surrounding text refers to a scalar displacement; please make the notation consistent.","section":"Section 3.1, Eq. (6)"},{"comment":"The paper describes DeepF-fNet as based on DeepONets, but the architecture in Table 3 is a standard CNN without a branch/trunk structure. Clarify the connection to DeepONets or avoid the term.","section":"Section 2.2"},{"comment":"The compliance conditions are said to be 'violated for some combinations of parameters,' but it is not specified how these violations are treated during training or what 'empirically observed' means quantitatively; please clarify.","section":"Section 3.3, Fig. 5(e)"},{"comment":"The notation DC, f_bg, and ib is not fully defined; in particular, the index ib runs over dataset samples while n indexes branches. Please define all variables explicitly.","section":"Algorithm 1"},{"comment":"The GA comparison uses a single run and a single target frequency; please report the GA's population size, generation count, and the objective function, and ideally repeat with several random seeds to provide error bars.","section":"Section 3.3, Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper does not provide code, data, or trained models, and the term 'proprietary algorithms' is used despite the algorithm being described in the text; this limits reproducibility. The main technical risk is the off-manifold generalization of SICE4, which I believe can be addressed by adding FE verification on multiple target frequencies and by qualifying the speed claims. The topic is appropriate for physics.comp-ph."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something genuinely useful: it extends Molinaro's neural inverse operator idea from a fixed eigenfrequency to a variable target frequency, and couples it with a heuristic correction scheme (SICE4) that shifts and rescales stored dispersion curves. The first-bandgap validation is the strongest part—when predicted parameters are fed back into the FE solver, the dispersion curves land close to the target, and the comparison against a GA at 200 Hz shows comparable parameters. That is real evidence the network works on at least one point in the design space.\n\nThe central soft spot is exactly what the stress test flags. SICE4 manufactures input spectra by rigidly shifting branches and rescaling the first branch with a single factor Γ. Nothing guarantees those spectra correspond to any physical (r,L,s) triple, and the network was trained only on FE-generated spectra. The fact that the physics losses have weights like 1e-14 and BC 1e-12 means no physical constraint is active enough to pull off-manifold inputs back into a realistic regime. The one FE-verified case at 200 Hz does not establish that the time-series outputs in Fig. 9 are correct—those 300 targets were never re-checked against FE. That is the load-bearing gap.\n\nThe speed comparison is also slanted. Reporting 5e5x faster by comparing SICE4 inference time (0.0157 s) against GA optimization time (7690 s) ignores the offline cost of generating 2000 FE samples and training. That may be acceptable if the claim is 'online retuning,' but 'up to 500,000 times faster than traditional methods' is the abstract's headline and it is not a fair comparison. The GA is also a single run with no variance reported.\n\nMinor but worth noting: the PDE/BC losses are numerically negligible, so calling this a physics-informed network is a stretch. The paper also does not release code or data, which limits reproducibility. The self-citation concern from the reader is overblown—citing Molinaro and then extending it is fine.\n\nWho benefits? The NVH and metamaterial design community. It is an incremental but actionable contribution, provided the SICE4 off-manifold issue is addressed or at least empirically shown to be benign across multiple targets. I would accept this for peer review with the expectation that the authors add multi-frequency FE verification, a fairer cost accounting, and either code/data or a clearer physical justification for the SICE4 transform.\n\nNet: worth engaging, but only with a strong request for additional validation.","headline":"A plausible incremental extension of neural inverse operators to variable-frequency vibration isolation, but the headline speed claim and the SICE4 correction step need serious scrutiny before the paper's conclusions can be trusted.","tokens_in":15521,"tokens_out":1194,"would_cite":false,"duration_ms":14679,"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":"DeepF-fNet, a dual-network physics-informed surrogate, estimates locally resonant metamaterial parameters that place a vibration bandgap at a user-specified target frequency in real time.","keywords":["physics-informed neural network","DeepONet","vibration isolation","locally resonant metamaterial","inverse eigenvalue problem","bandgap tuning","semi-active vibration control","genetic algorithm"],"falsifier":"A direct check: take the parameters predicted by SICE4 for a target frequency, run a fresh finite-element simulation of that unit cell, and compare the simulated dispersion curves with the corrected spectrum that was fed to the network. If the first bandgap of the simulated cell does not center near the target frequency for a substantial fraction of the tested targets, the shift-and-rescale assumption fails.","tokens_in":14549,"feed_emoji":"🔧","tokens_out":5811,"duration_ms":55763,"temperature":0.7,"pith_summary":"The paper claims that a physics-informed neural network, trained once on finite-element simulations, can invert the bandgap-placement problem for a locally resonant metamaterial: given a target frequency to isolate, it outputs the geometric parameters that put a bandgap there. The authors pair this network, DeepF-fNet, with a prediction-correction scheme called SICE4 that shifts and rescales a stored dispersion curve to approximate the target spectrum before asking the network for parameters. They test the idea on a trichiral honeycomb resonator attached to a steel plate and report parameters close to those found by a genetic algorithm, while running roughly 500,000 times faster. The practical stake is semi-active vibration isolation: a device that can retune itself at control-loop rates instead of requiring minutes or hours of optimization.","feed_headline":"Physics-informed net tunes vibration filters 500,000x faster","feed_subtitle":"DeepF-fNet maps a target frequency to metamaterial geometry in 0.016 seconds, enabling semi-active vibration isolation in vehicles.","key_machinery":"The load-bearing object is the SICE4 prediction-correction update on dispersion curves: a shift by $\\Delta$ and a rescaling by Gamma that convert a stored spectrum into an approximate target spectrum without solving the direct eigenvalue problem. This is paired with DeepF-fNet's dual CNN architecture, where the IEPS inverts spectra into structural parameters and the WES enforces physical plausibility by evaluating the wave-equation residual and boundary-condition residual on the predicted mode shapes, making the inverse mapping respect the underlying physics.","core_discovery":"DeepF-fNet is a dual convolutional network: the inverse eigenvalue problem solver (IEPS) maps a set of dispersion curves to the three unit-cell parameters r, L, and s, while the wave equation solver (WES) takes the discretized geometry defined by those parameters and predicts mode shapes. Training combines a data loss with two physics-informed terms, the harmonic wave-equation residual and the boundary-condition residual, so the inverse mapping is constrained by the governing elastodynamics. The SICE4 deployment algorithm selects from the training dataset the spectrum whose first bandgap is closest to the user's target frequency, shifts all curves except the first by the difference $\\Delta$, rescales the first curve by Gamma = min(second curve)/max(first curve), and feeds the corrected spectrum to the IEPS. The paper reports that for a 200 Hz target this pipeline returns parameters essentially equivalent to those found by a genetic algorithm, with a first bandgap centered near the target, while consuming about 0.0157 s of CPU time versus 7690 s for the GA.","pith_inferences":["The shift-and-rescale correction in SICE4 is effectively an affine augmentation of the training manifold; a natural testable extension is to check whether corrected spectra lie close to genuine finite-element spectra in the IEPS latent space, and if not, to add a small number of physically realizable spectra near the target-bandgap manifold to close the distribution shift.","Because the method succeeds on the first bandgap, it likely transfers to other periodic structures, such as phononic plates or beams with attached resonators, whose bandgap formation is dominated by the lowest dispersion branches; the same SICE4 logic could be reused if the input spectra are normalized consistently.","The reported 500,000x speedup is a single-case comparison against one genetic algorithm implementation without parallelization or warm starting; the practical gain in deployment will be smaller once sensor readout, actuation, and safety margins are included, though the core advantage of offline training with online feed-forward inference remains.","The spectral bias at higher eigenfrequencies suggests that multi-bandgap targets will require a higher-fidelity operator, such as a Fourier neural operator, as the authors propose; a focused study on bandgaps above the first would test whether the correction step preserves accuracy there."],"forward_implications":["A trained DeepF-fNet under SICE4 can estimate optimal unit-cell parameters for a time-varying target frequency at control-loop rates; the simulated 2 s, 150 Hz test signal tracks a 250-750-500 Hz sequence.","For a single 200 Hz target, SICE4 produces essentially the same first bandgap as a genetic algorithm while using about 500,000 times less CPU time, because training is performed offline.","The physics-informed losses allow the inverse mapping to be trained from a relatively small dataset of 2000 finite-element samples and to remain accurate within the parameter range covered by that dataset.","The first, lowest bandgap is predicted accurately even though higher eigenfrequencies exhibit spectral bias, so the method is directly usable for low-frequency single-bandgap isolation, while multi-bandgap targets need further development.","The framework positions semi-active vibration isolators as a practical replacement for fully active control systems in applications like automotive noise, vibration, and harshness, where rapid retuning is essential."],"supporting_citations":[{"why":"Supplies the DeepONet operator-learning framework used to learn the mapping from dispersion spectra to structural parameters.","marker":"[17]"},{"why":"Provides the neural inverse operator approach that handles inverse PDE problems, which SICE4 extends to variable target frequencies.","marker":"[19]"},{"why":"Establishes the physics-informed neural network methodology that underpins the combined data, PDE, and boundary-condition losses.","marker":"[13]"},{"why":"Defines the locally resonant metamaterial case study, including the host structure and attachable local resonator geometry used for validation.","marker":"[26]"},{"why":"Supplies the genetic algorithm implementation used as the baseline for accuracy and computational speed comparison.","marker":"[32]"},{"why":"Explains the low- and high-frequency bandgap formation mechanisms in chiral lattices, used to interpret parameter sensitivity.","marker":"[28]"},{"why":"Provides the finite-element solver (COMSOL) used to generate the dispersion curve and mode-shape dataset.","marker":"[21]"}],"fun_headline_variants":["Neural net slashes vibration isolation design time 500,000x","DeepF-fNet: physics-informed AI designs metamaterials in 0.016s","Real-time vibration isolation with neural net replaces active systems","This neural net solves inverse vibration problems 500,000x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The SICE4 correction step assumes that shifting a stored dispersion curve by $\\Delta$ and rescaling its first branch by Gamma produces an input spectrum that behaves like a real metamaterial spectrum for the network; if that artificial spectrum lies outside the training distribution, the predicted parameters can be wrong even though the network itself is accurate on genuine spectra.","fun_headline_variants_meta":{"raw":{"variants":["Neural net slashes vibration isolation design time 500,000x","DeepF-fNet: physics-informed AI designs metamaterials in 0.016s","Real-time vibration isolation with neural net replaces active systems","This neural net solves inverse vibration problems 500,000x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2760,"prompt_tokens":1012,"completion_tokens":1748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1671}},"tokens_in":628,"tokens_out":1748,"duration_ms":13496,"temperature":1.0,"reasoning_tokens":1671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:01:50.266961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: take the parameters predicted by SICE4 for a target frequency, run a fresh finite-element simulation of that unit cell, and compare the simulated dispersion curves with the corrected spectrum that was fed to the network. If the first bandgap of the simulated cell does not center near the target frequency for a substantial fraction of the tested targets, the shift-and-rescale assumption fails.","supporting_citations":[{"cited_title":"Jung, H.-G","cited_arxiv_id":null,"evidence_quote":"Defines the locally resonant metamaterial case study, including the host structure and attachable local resonator geometry used for validation."},{"cited_title":"Sayin, genetic-algorithm-matlab (2012)","cited_arxiv_id":null,"evidence_quote":"Supplies the genetic algorithm implementation used as the baseline for accuracy and computational speed comparison."},{"cited_title":"Baravelli, M","cited_arxiv_id":null,"evidence_quote":"Explains the low- and high-frequency bandgap formation mechanisms in chiral lattices, used to interpret parameter sensitivity."},{"cited_title":"URL http://www.comsol.com/products/multiphysics/ 26","cited_arxiv_id":null,"evidence_quote":"Provides the finite-element solver (COMSOL) used to generate the dispersion curve and mode-shape dataset."}],"review_version":1}