{"id":"26c4ecc7-4354-4a06-ba49-77c880679808","arxiv_id":"2411.14262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EED-ECSW speeds up non-intrusive identification of cubic stiffness tensors in geometric-nonlinearity ROMs by using an ECSW-reduced mesh to cheaply compute the tangent stiffness, with reported offline speedups of 3.18x and 13.86x.","lead":"This paper combines two existing model-reduction techniques, Enhanced Enforced Displacement and Energy Conserving Sampling and Weighting, to identify nonlinear stiffness tensors for reduced-order structural models faster. The result is a shorter offline construction phase for non-intrusive ROMs used in acoustic fatigue analysis, with 3.18x and 13.86x speedups on two test panels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ECSW cubature is trained on SQM force snapshots, but EED-ECSW tensor identification uses the same cubature for tangent stiffness at probe displacements that are not on the SQM, with no error bound connecting epsilon_ECSW to this extrapolation, so the claimed generality is not yet established…","rationale":"The paper's central claim has two parts: reduced offline cost and no degradation of ROM accuracy. The cost reduction is convincingly quantified (3.18x and 13.86x speedups), but the no-degradation part rests on the undocumented transfer of ECSW weights from SQM force-reproduction training to tangent-stiffness accuracy at EED probe displacements. The reader's weakest assumption identifies exactly this gap, and I agree. The numerical PSD comparisons are strong evidence for the two specific structures and loads considered, but they do not establish the general claim because PSD agreement can be insensitive to systematic errors in individual tensor coefficients. The proposed direct comparison of Eq. (26) against the exact reduced tangent stiffness, and of the identified tensors, is feasible using data the authors already possess from their standard-EED runs; it would either validate the transfer property or force a limitation statement. This does not change the CONDITIONAL verdict: the concern is real but testable, and the paper's demonstrations remain meaningful.","tokens_in":25967,"tokens_out":7144,"duration_ms":72792,"concrete_test":"Using the standard-EED tangent stiffness data already generated for Tables 1 and 2, compute the relative Frobenius norm error of the Eq. (26) approximation against the exact reduced nonlinear tangent stiffness V^T(K_t - K^(1))V at all EED probe displacements, separating probes along VM columns from probes along SMD columns; then compute the relative Frobenius norm error of the resulting EED-ECSW identified tensors versus the standard-EED tensors. If both errors are small (say below 5-10%) for SMD probes as well as VM probes, the extrapolation concern is resolved. If the tangent-stiffness errors are large but the PSDs still agree, the ROM accuracy claim is load-case-specific and the paper should state that limitation explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's central operation, Eq. (26), approximates the reduced nonlinear tangent stiffness with an ECSW cubature whose weights xi* are determined in Eq. (24) from reduced nonlinear force vectors on SQM samples q = Gamma(gamma) generated by Latin Hypercube sampling (Eq. (27)). The EED identification then uses Eq. (26) at probe displacements q = eta_r v_r and q = eta_s v_s + eta_r v_r, where the v_r are VM and SMD columns of the physical RB. These probes are outside the training distribution; in particular, a pure SMD displacement is not on the SQM (for gamma = 0, the SQM gives zero, and nonzero SMD components are coupled to VM amplitudes by theta_ij gamma_i gamma_j/2, so a displacement q = eta_r theta_ij with no VM component is never sampled). The validation error epsilon_ECSW in Eq. (35) measures force-reproduction error on SQM samples only and provides no bound on the error of the tangent-stiffness approximation at EED probes. Because Eqs. (20) and (21) are solved from the approximate tangent stiffness, any such error propagates directly into the identified tensors. The excellent PSD agreement in Section 4 is a weak probe of this: PSDs are aggregate statistics and may be insensitive to tensor errors in weakly excited or symmetric components, so the claim that EED-ECSW works without degrading ROM accuracy is not yet supported beyond the two reported load cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hyperreduction-accelerated variant of the Enhanced Enforced Displacement (EED) method for non-intrusive identification of cubic stiffness tensors in Galerkin ROMs of geometrically nonlinear structures. The reduced nonlinear tangent stiffness needed in EED is evaluated with an ECSW cubature whose reduced mesh and weights are trained on force snapshots generated from a static quadratic manifold, avoiding HFM time-integration simulations for training. Tensors are identified in the physical VM/SMD basis and subsequently transformed to an orthogonalized basis. The method is tested on a curved panel (RB of 35) and a nine-bay fuselage panel (RB of 50), reporting speed-ups of 3.18x and 13.86x in ROM construction while producing displacement PSDs in close agreement with standard EED and Abaqus HFM.","tokens_in":26342,"tokens_out":19004,"duration_ms":173240,"significance":"If confirmed, the contribution is practically significant because EED tensor identification is the dominant offline cost for large ROMs in acoustic-loading applications, and the method preserves full non-intrusiveness with respect to commercial FE codes. The numerical study is carefully designed: ECSW training is simulation-free, the comparison against both standard EED and the HFM is appropriate, and the tolerance study in Section 4.2.4 gives useful practical guidance. The main caveat is that the transfer of ECSW weights from force training to tangent-stiffness evaluation at EED probe displacements is heuristic; the paper does not provide an error bound, and its validation relies on aggregate PSD comparisons. I therefore view the empirical claims as well supported for the two tested structures, but the generality claim needs either a diagnostic or an explicit limitation.","major_comments":[{"comment":"The number of EED enforced displacements is stated inconsistently. Section 2.5 and Algorithm 2 give (m^2+5m)/2, but the reported 665 displacements for m=35 (Section 4.1.3) and 1325 for m=50 (Section 4.2.1) equal (m^2+3m)/2. The sentence in Section 2.5.2 saying 'additional m(m−2)/2 displacements' is non-integer for the m values used and is presumably a typo for m(m−1)/2. Please correct the formulas so the complexity count and the implementation match; this is necessary for reproducibility.","section":"§2.5, §2.5.2, Algorithm 2, §4.1.3, §4.2.1"},{"comment":"The ECSW weights in Eq. (24) are trained to reproduce reduced nonlinear forces on SQM samples q=Γ(γ), but Eq. (26) uses the same weights to approximate the reduced nonlinear tangent stiffness at EED probe displacements q=η_r v_r and q=η_s v_s+η_r v_r. These probes are not in the SQM training distribution: a pure SMD displacement has no VM component and cannot be represented on the SQM of Eq. (10). The validation error in Eq. (35) only measures force reproduction on held-out SQM samples and provides no bound on the tangent-stiffness error at the probes. Since Eqs. (20) and (21) are solved from the approximate tangent stiffness, this is a genuine extrapolation step. The excellent PSD agreement in Section 4 is reassuring, but PSDs are aggregate statistics and may be insensitive to systematic tensor errors in weakly excited components. Please either add a direct numerical check of the ECSW tangent-stiffness approximation at representative EED probes, or state explicitly as a limitation that no transfer guarantee is provided.","section":"§3.2, §3.3.2, Eq. (26), Eq. (35)"}],"minor_comments":[{"comment":"In Eq. (21), the last term should read K~(3)_issj η_s^2, not η_r^2, for the case r<s<j.","section":"Eq. (21)"},{"comment":"In Table 3 the 'accuracy' column is only quantified by ϵ_ECSW; the dynamic accuracy is shown in Fig. 10. Please clarify in the caption that ϵ_ECSW is the SQM force-reproduction error and not a direct measure of ROM accuracy.","section":"Section 4.2.4, Table 3"},{"comment":"The text uses both 'EED-ECSW' and 'ECSW-EED' to refer to the same method; please standardize the terminology.","section":"Section 4.1.4"},{"comment":"The introduction contains a typo 'pyhsics-based strategies' (should be 'physics-based strategies').","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core empirical demonstration is solid. The main issues are internal consistency of the EED count and the unquantified extrapolation from ECSW force training to tangent-stiffness evaluation; both are fixable without changing the method. I did not find evidence of circularity or unsupported claims beyond the need for an explicit limitation statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers what it says: EED-ECSW cuts offline ROM construction time by 3.18x and 13.86x on two Abaqus models while keeping PSDs visually on top of standard EED and the high-fidelity model. The combination is new relative to the closest prior work, Kim et al. 2023: ECSW is used inside an indirect identification scheme (EED) rather than a direct one, and that is what makes the method compatible with commercial FE packages. That is a real, practically useful step. The two test cases are well chosen. The nine-bay panel is substantial (200k dof, 50-vector RB, 1325 tangent stiffness evaluations), and the authors compare against standard EED and the HFM, include a convergence study over VM count, and run a tolerance sweep on the ECSW training. The LHS-based SQM training is thoughtful, and the physical-to-orthogonal RB tensor transformation is a sensible way to avoid imposing nonphysical displacement shapes during identification. The timing breakdown is honest and useful. Citation pattern looks solid; it builds on the right literature and positions itself clearly against Kim et al. The main soft spot is exactly the one in the stress-test note: ECSW weights are trained on nonlinear force snapshots on the SQM, then used in Eq. (26) to approximate reduced tangent stiffness for EED probe displacements, pure SMD directions and VM+SMD combinations, that are not on the SQM. There is no error bound linking epsilon_ECSW to the tensor identification error, and PSD agreement is aggregate evidence. I do not think this invalidates the paper: the two test cases show the method works, the authors are transparent about the empirical nature of the validation, and the central claim is about speedup plus preserved accuracy on these cases, not a universal theorem. But it does limit the generality claim. A second soft spot is the absence of code and data, so exact reproduction is not possible. Third, the stochastic response comparisons have no error bars; with one 10s realization per model that is acceptable for a methods paper, but worth noting. For peer review: send it. A serious referee should ask for a numerical probe of tangent-stiffness accuracy at the EED probe displacements, or an error analysis, plus code/data if feasible, but this is a solid methods paper with practical impact and deserves referee time.","headline":"A solid, practically useful methods paper that combines EED and ECSW to cut offline ROM construction time by 3-14x on two Abaqus models, with a real but non-fatal gap: no error bound for the cubature's extrapolation to tangent stiffness at off-manifold probe displacements.","tokens_in":783,"tokens_out":782,"would_cite":true,"duration_ms":36950,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74S05","74H45"],"pacs":[],"model":"deepseek-v4-flash","headline":"A hyperreduction scheme speeds nonlinear structural ROM construction up to 13.9x by replacing exact stiffness evaluations with a weighted mesh subset, matching full-model power spectra.","keywords":["Geometric Nonlinearity","Reduced Order Modeling","Galerkin-ROM","Modal Derivatives","Hyperreduction","ECSW","EED","Acoustic Loading"],"falsifier":"Evaluate the ECSW approximation in Eq. (26) at EED's probe displacements, comparing the weighted reduced-mesh stiffness against a full-mesh assembly; if the relative error there is far larger than the ECSW validation error $\\epsilon_{\\rm ECSW}$, the identified tensors will be biased even when the training looks accurate.","tokens_in":25773,"feed_emoji":"⚙️","tokens_out":10094,"duration_ms":91612,"temperature":0.7,"pith_summary":"The paper tries to establish that the most expensive step in building non-intrusive reduced order models for geometrically nonlinear structures—identifying the cubic stiffness tensors from a commercial finite element code—can be accelerated with hyperreduction. The proposed EED-ECSW method replaces each exact evaluation of the reduced nonlinear tangent stiffness in the Enhanced Enforced Displacement scheme by an Energy Conserving Sampling and Weighting approximation over a small, weighted subset of elements. The reduced mesh and weights are trained simulation-free on force snapshots from a static quadratic manifold, so no full high-fidelity dynamic runs are needed. On a curved panel and a nine-bay aeronautical panel under random acoustic loading, the method matches the power spectral densities of standard EED and of the high-fidelity model while cutting ROM construction time by factors of 3.18 and 13.86. If correct, it makes large tensorial ROMs practical for industrial acoustic-fatigue studies.","feed_headline":"Hyperreduction speeds nonlinear ROM construction up to 13.9x","feed_subtitle":"Offline build time drops from hours to under an hour, while predicted spectra match the full model.","key_machinery":"The load-bearing object is the ECSW approximation of the reduced nonlinear tangent stiffness: $\\tilde{\\mathbf K}^{(nl)} \\approx \\sum_{e\\in \\tilde E}\\xi_e\\mathbf V_e^{T}(\\mathbf K^t_e-\\mathbf K_e)\\mathbf V_e$. ECSW is a hyperreduction method that replaces the full element sum by a sparse non-negative weighted subset of elements, determined here with the sNNLS algorithm. The weights are trained on element forces from static quadratic manifold displacements $\\mathbf q=\\Gamma(\\boldsymbol\\gamma)$ with Latin-hypercube amplitudes, after subtracting each element's linear force. This same weighted subset is then used inside EED to evaluate the left-hand side of the identification equations for probes along individual basis vectors and pairs, providing all $(m^2+5m)/2$ tangent-stiffness evaluations at reduced cost. The transformation of the identified tensors from the physical to the orthogonalized basis, via $\\mathbf U=(\\mathbf V^T\\mathbf V)^{-1}\\mathbf V^T\\mathbf W$, completes the construction.","core_discovery":"The central discovery is that the ECSW reduced mesh trained to reproduce reduced nonlinear internal forces on a static quadratic manifold also reproduces the reduced nonlinear tangent stiffness needed for EED tensor identification. Specifically, the paper approximates $\\tilde{\\mathbf K}^{(nl)} = \\mathbf V^{T}(\\mathbf K_t-\\mathbf K^{(1)})\\mathbf V$ by $\\sum_{e\\in\\tilde E}\\xi_e \\mathbf V_e^{T}(\\mathbf K^t_e-\\mathbf K_e)\\mathbf V_e$, computes the weights from nonlinear force snapshots with linear parts removed, and uses this cheap tangent stiffness in the identification equations to determine the tensors. Identified on the physical basis $\\mathbf V$ and then transformed to an orthogonalized basis $\\mathbf W$, the tensors produce ROMs whose displacement PSDs overlay the standard EED ROM and the finite element model, while the tensor-construction time drops from 1908.5 s to 599.9 s on the curved panel and from 13.759 h to 0.993 h on the nine-bay panel.","pith_inferences":["Editorial extension: since the ECSW reduced mesh is trained only from vibration modes and static modal derivatives, the same reduced mesh could plausibly be reused across load cases or small geometry perturbations sharing a reduction basis, which the paper motivates but does not test.","Editorial extension: the paper introduces a Modal Derivative Participation Factor for ranking static modal derivatives, but does not isolate its effect on accuracy; a dedicated comparison against frequency-based or random selection would test whether this ranking is the right one.","Editorial extension: the absence of an error bound connecting $\\tau$ or $\\epsilon_{\\rm ECSW}$ to tensor identification error suggests a practical safeguard the paper does not explore: re-evaluate a few EED probe equations with the full mesh and stop refining the reduced mesh only when the identified tensors stop changing.","Editorial extension: the same EED-ECSW construction could be applied to updated-Lagrangian or co-rotational element formulations where the polynomial force model is approximate; the physical-to-orthogonalized basis transformation may reduce the resulting bias, but the paper leaves that setting untested."],"forward_implications":["Offline tensor identification time drops by factors of 3.18 and 13.86 on the two test panels, with the largest gains coming from faster finite element runs and faster reading of tangent-stiffness matrices.","ROM accuracy is preserved: power spectral densities from EED-ECSW tensors overlay those from standard EED and from the high-fidelity model, including nonlinear peak smearing, frequency shifts, and out-of-band response.","The method stays fully non-intrusive, so tensors can be built from the outputs of a commercial finite element code without accessing the element formulation.","The ECSW training tolerance $\\tau$ provides a tunable trade-off: smaller $\\tau$ gives larger reduced meshes and higher accuracy, while larger $\\tau$ gives faster construction, as demonstrated by the nine-bay panel tolerance study.","Because the identified ROM remains in tensorial form, online integration stays independent of the finite element code, preserving online speedups of roughly 400x and 267x over the high-fidelity model on the two test cases."],"supporting_citations":[{"why":"Introduces the Enhanced Enforced Displacement (EED) tensor identification scheme that this paper accelerates.","marker":"[31]"},{"why":"Introduces ECSW hyperreduction with non-negative weights and the sNNLS training algorithm used to build the reduced mesh.","marker":"[20]"},{"why":"Establishes the structure-preserving, stability, and accuracy properties of ECSW that justify using its tangent-stiffness approximation.","marker":"[36]"},{"why":"Provides the simulation-free static quadratic manifold lifting strategy that EED-ECSW adapts for training force snapshots.","marker":"[21]"},{"why":"Supplies an earlier simulation-free snapshot generation technique for ECSW training, which the paper contrasts with static quadratic manifold sampling.","marker":"[18]"},{"why":"Proposes the original enforced displacement tensor identification method whose cost EED was designed to reduce.","marker":"[28]"},{"why":"Establishes the exact cubic polynomial form of internal forces used to define the tensors being identified.","marker":"[24]"},{"why":"Uses an ECSW reduced mesh for tensor construction in a direct parametric identification setting, the closest prior use of this idea.","marker":"[38]"},{"why":"Provides the static modal derivative selection strategy on which the paper's new participation-factor ranking is based.","marker":"[44]"}],"fun_headline_variants":["Hyperreduction speeds nonlinear ROM build by 13.9x","ECSW hyperreduction cuts ROM tensor build time","Accelerating ROM construction: ECSW hyperreduction","13.9x faster non-intrusive ROM via hyperreduction","Hyperreduction accelerates nonlinear ROM tensor identification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small set of weighted finite elements chosen to reproduce nonlinear forces on static manifold shapes will also reproduce the tangent stiffness at EED's probe displacement shapes, which are outside the training set, and the paper provides no error bound connecting those two errors.","fun_headline_variants_meta":{"raw":{"variants":["Hyperreduction speeds nonlinear ROM build by 13.9x","ECSW hyperreduction cuts ROM tensor build time","Accelerating ROM construction: ECSW hyperreduction","13.9x faster non-intrusive ROM via hyperreduction","Hyperreduction accelerates nonlinear ROM tensor identification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1461,"prompt_tokens":1026,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":352}},"tokens_in":642,"tokens_out":435,"duration_ms":4328,"temperature":1.0,"reasoning_tokens":352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:21:19.499901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the ECSW approximation in Eq. (26) at EED's probe displacements, comparing the weighted reduced-mesh stiffness against a full-mesh assembly; if the relative error there is far larger than the ECSW validation error $\\epsilon_{\\rm ECSW}$, the identified tensors will be biased even when the training looks accurate.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the simulation-free static quadratic manifold lifting strategy that EED-ECSW adapts for training force snapshots."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies an earlier simulation-free snapshot generation technique for ECSW training, which the paper contrasts with static quadratic manifold sampling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes the original enforced displacement tensor identification method whose cost EED was designed to reduce."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the static modal derivative selection strategy on which the paper's new participation-factor ranking is based."}],"review_version":1}