{"id":"4b43a8d2-072d-4b00-a56c-a75391205ac4","arxiv_id":"2511.20687","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Overlapping Schwarz iteration can glue operator-inference ROMs to each other and to full-order finite-element models for 3D nonlinear solid dynamics, with online speedups up to 106x.","lead":"This paper couples machine-learned reduced-order models with full physics simulations inside an overlapping Schwarz domain-decomposition loop, and tests it on 3D solid-mechanics problems. It reports accurate hybrid models and online speedups up to 106x versus full-order Schwarz coupling.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tension-specimen accuracy claim rests on a FOM-FOM reference that hit the 32-iteration cap; reported O(1e-4) errors may be errors against an unconverged iterate, not the converged coupled solution.","rationale":"The reader's weakest-assumption analysis and this stress-test identify the same load-bearing point: the tension-specimen FOM-FOM reference is not a converged solution, yet it defines the errors used to support the central accuracy claim. This is explicitly visible in Tables 7–8 (mean/max 32.0/32 Schwarz iterations) and in the text's decision to cap iterations at 32. The 106x speedup inherits this baseline issue and also excludes offline training costs, so the efficiency claim is weaker than the abstract implies. My read does not change the conditional verdict: the method is plausibly correct and well-supported by the 1D exact-solution tests and the bolted-joint/torsion cases, but the flagship tension-specimen demonstration needs a converged reference and a transparent cost accounting before the strongest claims should be taken at face value. No fatal flaw is apparent; the concern is about validation and accounting in one test case, not about the internal consistency of the algorithm.","tokens_in":41045,"tokens_out":3351,"duration_ms":31062,"concrete_test":"Re-run the Section 5.4 tension-specimen FOM-FOM O-SAM with the iteration cap removed (or raised to at least 200) at the stated tolerances (δrel=1e-8, δabs=1e-6), and record the final-time solution and CPU time. Then recompute the Table 7/8 ROM relative errors and the speedup relative to this converged FOM-FOM reference. If the displacement/stress errors shift by more than the reported O(1e-4) margin, or if the converged FOM-FOM CPU time increases materially, the accuracy/efficiency claims as stated are not supported. Separately, include the offline FOM-FOM training runs and OpInf training time in an end-to-end wall-clock speedup for one prediction; if the speedup drops below 1, the 106x figure overstates practical efficiency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 1.2, supported by Section 5.4) is that hybrid OpInf/FOM O-SAM couplings are accurate and efficient, with speedups up to 106x. Tables 7 and 8 report errors computed via Eq. (28) against a FOM-FOM O-SAM reference. In that flagship tension-specimen case, the FOM-FOM run does not converge: it hits the 32-iteration cap with mean/max Schwarz iterations of 32.0/32, while the ROM-containing couplings converge in 7–10 iterations. If the reference itself is not converged to the stated tolerances (δrel=1e-8, δabs=1e-6), then the reported O(1e-4) displacement errors only show that the ROM reproduces the same unconverged iterate, not that either model is accurate with respect to the true PDE solution. The manuscript explicitly states that no monolithic reference is computed (Section 5), so there is no external accuracy check in this case. The 106x speedup is likewise computed against a cap-limited FOM-FOM run; a converged baseline could take materially longer, and the ROM errors could change if the converged FOM-FOM solution differs. In addition, the speedup excludes the offline FOM-FOM training runs and OpInf operator-regression time, so end-to-end efficiency for a single query is not established. The methodology may still be sound, but the headline quantitative claims in the tension case are not yet anchored to a converged ground truth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid coupling method for subdomain-local non-intrusive Operator Inference (OpInf) reduced order models (ROMs) and full order models (FOMs) using the overlapping Schwarz alternating method (O-SAM). The offline stage builds subdomain-local POD bases and infers cubic (or lower-order) polynomial ROM operators, including a new boundary-POD compression and a regularization-parameter selection procedure. The online stage couples these ROMs with each other or with FOMs using O-SAM, allowing disparate meshes, element types, and time integrators. Numerical experiments on 1D wave propagation, a 3D bolted joint, a 3D torsion specimen, and a 3D tension specimen demonstrate, in most cases, good accuracy relative to FOM-FOM O-SAM references and substantial online speedups, with the largest reported speedup being 106x in the tension specimen. The paper emphasizes a 'top-down' training strategy and evaluates both reproductive and predictive scenarios.","tokens_in":41521,"tokens_out":3522,"duration_ms":32401,"significance":"If the central claims hold, the paper provides a practical, plug-and-play framework for coupling non-intrusive ROMs with FOMs in nonlinear solid dynamics, addressing a long-standing barrier in multiscale simulation. The methodology is well-motivated and the exposition is generally clear. Notable strengths include: (i) a reproducible implementation with commit-hashed code and input decks; (ii) predictive test cases that use parameter/initial-condition combinations not in the training set; (iii) a demonstration that quadratic/cubic polynomial ROMs can serve as local surrogates even for non-polynomial hyperelastic models; and (iv) a thorough comparison with related work. The main caveat is that the flagship tension-specimen accuracy and speedup claims rest on a FOM-FOM reference that did not meet the stated Schwarz convergence tolerances, which undermines the headline numbers.","major_comments":[{"comment":"The FOM-FOM reference run in the tension specimen hits the 32-iteration cap with mean/max Schwarz iterations of 32.0/32, i.e., it does not satisfy the stated tolerances (δrel=1e-8, δabs=1e-6). Since the reported hybrid errors are computed via Eq. (28) relative to this FOM-FOM solution, the O(1e-4)-O(1e-3) errors and the 106x speedup are not validated against a converged coupled solution. Please either provide a converged FOM-FOM reference (or a monolithic reference), or explicitly characterize the errors as being relative to an unconverged iterate and perform a sensitivity study with respect to the Schwarz iteration cap.","section":"Section 5.4, Tables 7-8"},{"comment":"The reported speedups are online-only. The 106x figure compares online CPU times (4m 42.1s vs. 8h 19m 29.5s) and excludes the offline FOM-FOM runs required for 'top-down' training as well as the OpInf operator-inference time. For a single query, the end-to-end speedup is therefore not established. Please state this limitation explicitly and, if appropriate, provide an amortized cost analysis for multi-query scenarios.","section":"Section 5 (above §5.1) and §5.4, Tables 7-8"}],"minor_comments":[{"comment":"Typo: 'denots' should be 'denotes'.","section":"Section 2 (text near Eq. (8))"},{"comment":"The notation for the convergence check is hard to parse; consider simplifying the subscript/superscript structure or adding a short explanation after the equation.","section":"Eq. (22)"},{"comment":"CPU time is listed as '1m 43.5' without a unit; add 's' for consistency.","section":"Table 7, QOpInf-QOpInf 1-iter row"},{"comment":"It would help the reader if the text explicitly states that QOpInf is chosen as a quadratic surrogate for the non-polynomial Neohookean model, rather than leaving the reader to infer this from the model assignment.","section":"§5.3-5.4"},{"comment":"Scenario 3 is predictive in initial condition but not in material parameters; the distinction between state-predictive and parameter-predictive should be stated more clearly to avoid overstating the predictive scope.","section":"§5.1.3"},{"comment":"The claim of being 'first' to apply SAM-based OpInf coupling to realistic 3D nonlinear solid dynamics could be softened given the closely related works [16] and [19]; consider phrasing as 'among the first'.","section":"§1.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its formulation and the numerical evidence is strong except for the tension-specimen case, where the reference solution is not converged. This is a load-bearing issue for the paper's headline accuracy/efficiency claims. The authors should be able to address it by either running the FOM-FOM baseline to convergence (possibly with relaxed tolerances) or by adding a clear disclaimer and a Schwarz-convergence study. The reproducibility practices (commit hashes, input decks) are exemplary and should be commended."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nWorth a look if you work on ROM/domain-decomposition coupling. The paper does something concrete and useful: it takes the overlapping Schwarz alternating method, couples subdomain-local OpInf ROMs (linear, quadratic, cubic) with FOMs and with each other, and demonstrates on 3D solid dynamics problems with different meshes, element types, and time integrators. Code is pinned by commit hashes, input decks are provided, and there are genuine predictive test cases (different initial conditions/parameters than training). The authors also show that a single Schwarz iteration—the approach in earlier work—gives wrong answers on the nonlinear test problems; iterating to convergence matters. That is a real, useful negative result.\n\nThe main caveat is the one the stress-test flags, and it is not manufactured: in the flagship tension-specimen case (Section 5.4), the FOM-FOM O-SAM baseline used as reference never meets the Schwarz tolerance; it runs into the 32-iteration cap (mean/max 32/32) in both reproductive and predictive runs. The ROM couplings converge in 7–10 iterations, and errors are computed against that unconverged FOM-FOM solution. So the reported O(1e-4) errors and the 106x speedup are anchored to an iterate that is not a converged coupled solution, and there is no monolithic reference in the paper. The authors are transparent about this (they explicitly state they don't compare to a monolithic solve), but it means the headline accuracy claims are not validated against the true PDE solution. The tension case is also where the ROM looks best, which makes the gap more important.\n\nOther soft spots are minor. The speedup figures are online-only; offline training (including the FOM-FOM runs) is excluded, which is standard in this literature but worth stating. Some predictive multi-query cases (bolted joint COpInf-COpInf) have large errors—velocity errors around 10% and von Mises stress errors above 30%—but the authors report those honestly and the location of maximum stress is still right. Novelty is incremental: Farcas et al. and Gkimisis et al. did OpInf-Schwarz couplings, and the authors have prior SAM papers; the new pieces are cubic OpInf, boundary POD, full-iteration comparisons, and the 3D solid mechanics demonstrations. That is enough for a solid applied paper, not a breakthrough.\n\nOverall: the methodology is coherent, the code is real, the derivations look right. If you want to use the numbers, read Section 5.4 carefully before quoting them. The paper deserves a serious referee; the main revision request should be an anchored reference—a converged FOM-FOM or a monolithic solve on at least one 3D case—and a clear online/offline cost breakdown. I'd send it to review and cite it with the caveat.","headline":"Solid, honest engineering demonstration of Schwarz-coupled OpInf ROMs in 3D solid dynamics, but the headline accuracy and speedup claims rest on a non-converged FOM-FOM reference.","tokens_in":41932,"tokens_out":3278,"would_cite":true,"duration_ms":28974,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M55","65M60","74S05","74B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the overlapping Schwarz alternating method can couple subdomain-local data-driven reduced-order models, built by operator inference, with each other and with full-order finite element models, achieving accurate soluti","keywords":["operator inference","overlapping Schwarz alternating method","domain decomposition","reduced order model","solid dynamics","hyperelasticity","hybrid FOM-ROM coupling","non-intrusive model reduction"],"falsifier":"Re-run the tension-specimen case allowing the FOM-FOM O-SAM to iterate until the stated tolerances are met (or obtain a monolithic reference with a compatible mesh) and count the offline training cost; if the converged reference differs from the 32-iteration reference by more than the claimed hybrid-model errors, or if the end-to-end time including training exceeds the FOM-FOM time for a single query, the reported accuracy and 106× speedup are not validated.","tokens_in":40936,"feed_emoji":"🧩","tokens_out":6298,"duration_ms":52647,"temperature":0.7,"pith_summary":"The paper claims that the overlapping Schwarz alternating method (O-SAM) can stitch together subdomain-local reduced-order models built by operator inference — learned from data without modifying the full-order code — and couple them to each other or to full-order finite-element models, even when subdomains use different meshes, element types, and time integrators. The central demonstration is on fully nonlinear three-dimensional solid dynamics problems: bolted joints, torsion, and tension specimens, where the paper reports speedups up to about 106× relative to a pure FOM-FOM Schwarz coupling, with relative errors typically in the 0.1–5% range. The authors position this as the first application of Schwarz-iterated operator-inference coupling to realistic 3D nonlinear solid mechanics, and argue it enables workflows in which only a few physically motivated subdomains need to be meshed or modeled at high fidelity.","feed_headline":"Learned subdomain models cut solid-dynamics solve time 106x","feed_subtitle":"By gluing data-driven local ROMs to full-order models with the Schwarz method, errors stay near 0.1-1%.","key_machinery":"The central mechanism is the overlapping Schwarz alternating method (O-SAM) itself, which exchanges Dirichlet boundary data between subdomain solves and iterates until displacement/velocity convergence; the paper's novel twist is to let some (or all) subdomain solvers be OpInf ROMs. OpInf (operator inference) learns reduced linear/quadratic/cubic operators from snapshots of a FOM-FOM Schwarz simulation, using a least-squares fit with regularization. Two enabling components are introduced: boundary POD bases (Φ∂S, Φ∂φ) that compress the boundary forcing term so the learned boundary operator does not scale with interface DoF count, and a subdomain-local regularization sweep that chooses λ by a","core_discovery":"The paper's central claim is that converged O-SAM iterations with subdomain-local OpInf ROMs — linear, quadratic, or cubic — reproduce the accuracy of FOM-FOM Schwarz solutions while costing far less online CPU time. The claim is supported by four test problems: a 1D elastic wave with mixed time integrators, a 3D bolted joint with cubic Saint Venant–Kirchhoff nonlinearities, a 3D torsion bar with generic hyperelastic nonlinearities modeled by quadratic OpInf, and a 3D tension specimen where quadratic OpInf models coupled to each other are reported to match the FOM-FOM reference to roughly 0.01–5% error while running up to about 106× faster. A key reported observation is that the hybrid coupl","pith_inferences":["The 106× speedup is an online-only figure; it excludes the offline FOM-FOM training runs and OpInf operator learning. In a one-off analysis the method would be slower end-to-end, so the practical benefit is in multi-query settings where training costs amortize.","In the tension-specimen benchmark the FOM-FOM reference run consumed all 32 allowed Schwarz iterations without meeting the tolerance (mean/max 32.0/32), so the reported errors and speedup are measured against a reference that may not be fully converged; an independently converged reference could change those numbers.","The observation that ROM-involved couplings converge in fewer Schwarz iterations suggests a testable extension: using an intentionally coarsened or smoothed model in selected subdomains may accelerate Schwarz convergence even without formal ROM accuracy, as long as the final iterate is checked against tolerances.","Because the coupling is formulated at the algebraic level and the paper frames O-SAM as PDE-agnostic, the same recipe should transfer to other nonlinear PDEs (e.g., thermo-mechanics or plasticity) provided the polynomial OpInf form is replaced or augmented appropriately."],"forward_implications":["Multi-query analyses (design, uncertainty quantification) on complex geometries can replace one subdomain's high-fidelity solver with a learned surrogate, cutting online solve time while preserving Schwarz-level accuracy.","Different meshes, element types, and time integrators within one simulation become routine: ROM subdomains can use implicit time stepping with coarse meshes while FOM subdomains run explicit, as demonstrated in the torsion and wave cases.","Because hybrid couplings converge in fewer Schwarz iterations than FOM-FOM, the speedup is not solely from reduced DoFs but also from faster coupling convergence, making ROM placement in 'simpler' subdomains an attractive accuracy-preserving acceleration strategy.","Higher-order polynomial OpInf (quadratic and cubic) is sufficient to reproduce the global response of fully non-polynomial hyperelastic models once localized via Schwarz, suggesting the polynomial structural assumption is not an obstacle in the tested regime.","A single Schwarz iteration per time step (an additive-Schwarz-like shortcut) can converge to the wrong solution in nonlinear problems; iterating to the tolerance is necessary — a direct caution for cheaper Schwarz variants."],"fun_headline_variants":["Hybrid Schwarz coupling speeds 3D solid dynamics 106x","Operator inference ROMs + Schwarz: 106x faster solid dynamics","Data-driven subdomain models glue via Schwarz for 106x speedup","Schwarz-coupled ROMs speed solid dynamics by 106x","Operator inference meets Schwarz: 106x speedup in solid dynamics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The accuracy and speedup numbers rest on treating the FOM-FOM Schwarz solution as the ground truth, but in the headline tension-specimen case that reference never reached the Schwarz convergence tolerances within the 32-iteration cap, so reported errors are relative to an unconverged baseline.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid Schwarz coupling speeds 3D solid dynamics 106x","Operator inference ROMs + Schwarz: 106x faster solid dynamics","Data-driven subdomain models glue via Schwarz for 106x speedup","Schwarz-coupled ROMs speed solid dynamics by 106x","Operator inference meets Schwarz: 106x speedup in solid dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3664,"prompt_tokens":717,"completion_tokens":2947,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":2870}},"tokens_in":461,"tokens_out":2947,"duration_ms":17765,"temperature":1.0,"reasoning_tokens":2870,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:03:17.556265+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the tension-specimen case allowing the FOM-FOM O-SAM to iterate until the stated tolerances are met (or obtain a monolithic reference with a compatible mesh) and count the offline training cost; if the converged reference differs from the 32-iteration reference by more than the claimed hybrid-model errors, or if the end-to-end time including training exceeds the FOM-FOM time for a single query, the reported accuracy and 106× speedup are not validated.","supporting_citations":[],"review_version":1}