{"id":"68ac71c2-e2f0-4af8-97da-1fcf9188db1a","arxiv_id":"2607.15171","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A domain-decomposition solver with three cascade-trained neural surrogates replaces all fine-scale local operations, reaching stable ~3-6% L2 accuracy on unseen synthetic vascular networks without global assembly.","lead":"This paper learns neural surrogates for the local pieces of a domain-decomposition solver and reuses them to simulate large multiscale 3D-1D problems without assembling the global fine-scale model. If it holds up, it offers a path to cheaper repeated simulations in vascularized tissue, fractured media, and composites.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform-accuracy premise is not verified in the norm required by Prop. 1: training losses and reported errors are L2, while the error bound needs an energy/interface norm over all of bP.","rationale":"The reader's weakest assumption — uniform representability/accuracy over the localized parameter space — is the same load-bearing concern I identify, but it can be sharpened. The chain of the central claim requires: (i) exact DD contraction (Prop. 2), (ii) uniform surrogate error ε in the norm used by Assumption 3, and (iii) stability constant CDD. The paper provides no construction or verification of (ii) beyond average/relative L2 test errors on unseen geometries from the same generative model. In fact, the training losses are L2 and Euclidean, so the ε appearing in Prop. 1 is not even the quantity being controlled; the broken-energy and interface norms of Prop. 2 are more demanding. The weak-scaling experiment supports this concern: the plateau grows from 2.7% to 5.8% L2 and 7.8% to 15.8% broken energy as Nsub goes from 27 to 125, showing that the localized states encountered online move away from the training distribution. I also noted a secondary issue: the appendix proof of Prop. 2 derives an iteration-dependent contraction factor and does not explicitly establish a uniform q<1, though this may be repairable with trace inequalities. That makes me keep the CONDITIONAL verdict, not reject: the numerical evidence is positive but the theoretical guarantee as stated is not fully established. I agree with the reader's emphasis on Assumptions 1–2; the norm mismatch strengthens it.","tokens_in":31381,"tokens_out":14793,"duration_ms":123179,"concrete_test":"Using the trained surrogates, compute the local error e⋆(bµ) on a large held-out set (e.g., 50 new vascular geometries and DD states from the fully surrogate iteration) in the energy norm induced by the local Robin bilinear form and the interface trace norm of Prop. 2, instead of L2. Estimate the supremum over this set. If the energy-norm sup is substantially larger than the L2 test error (or grows with Nsub), the ε in Prop. 1 is not controlled by the training losses; if it is bounded and close to the reported L2 values, the norm-mismatch concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central bound (Prop. 1) requires a uniform surrogate error bound ε = sup_{bµ∈bP} ||e⋆(bµ)||_{V(Ω⋆)}, with the norm used in Assumption 3, and the discrete Robin–Robin contraction of Prop. 2 is stated in a broken-energy plus interface norm. However, every training loss in §4.2.1 (eqs. (15), (17), (18)) is a discrete L2(Ω⋆) or Euclidean residual, and Table 4 reports the same L2 measures. No inequality is provided linking these L2 errors to the energy/interface norm in which the perturbation δ(k) is bounded. For the microstructure-informed closure in particular, an L2-small error can be large in H1/energy near the embedded network. Hence the ε that would enter Prop. 1 is not the quantity being minimized or measured, so the reported test errors (e.g., 5.6% for Srom) do not substantiate the theoretical error floor. Moreover, Assumption 2 is a supremum over the whole localized parameter space bP, while training uses 65 geometries and 5 DD iterations; the weak-scaling results (Table 6) show the plateau growing with Nsub, indicating states less represented in the offline distribution. This is precisely the premise whose failure would collapse the bound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MS-DD-ROM, a non-intrusive reduced-order framework for multiscale mixed-dimensional problems. Domain decomposition (DD) is used to localize the solution operator; three neural surrogates replace the local fine-scale operations: the transmission map, the local solution map, and the coarse-residual contribution. The paper formalizes local representability, analyzes the surrogate-perturbed DD iteration as an inexact contraction, and proves that the asymptotic global error is bounded by CDD/(1-q) times the uniform surrogate error (Proposition 1). A two-level Robin-Robin method for 3D-1D coupled problems is instantiated, with cascaded training designed to match online distributions. Numerical experiments on a 3D-1D oxygen-perfusion model show a stable error plateau on unseen microstructures and algorithmic weak scalability up to 125 subdomains.","tokens_in":31784,"tokens_out":5754,"duration_ms":52783,"significance":"The framework is timely and useful: it combines domain decomposition with learned local operators, avoids fine-scale assembly and local high-fidelity solves online, and proposes a cascaded training strategy to reduce offline-online distribution shift. The abstract perturbation analysis (Proposition 1) is correct under its assumptions, and the numerical experiments are extensive, including generalization to unseen geometries and weak scaling. The main value of the paper is conditional on verifying the stated uniform-accuracy assumption in the norm required by the theory; as it stands, the numerical evidence does not substantiate that assumption. The paper is a solid contribution to DD-ROMs and would be suitable for publication after the norm-mismatch and uniformity issues are addressed.","major_comments":[{"comment":"Assumption 2 requires sup_{bµ∈bP} ||e⋆(bµ)||_{V(Ω⋆)} ≤ ε, and Assumption 3 bounds the perturbation δ(k) in the broken-energy/interface norm ||·||⋆ of Proposition 2. However, all training losses in eqs. (15), (17), (18) and the reported errors in Table 4 are discrete L2(Ω⋆) or Euclidean for the coarse residual. No inequality is given linking these quantities to V(Ω⋆) or ||·||⋆. For Srom, whose closure term is microstructure-dependent, an L2-small error can be large in H1/energy near the embedded network. Therefore the reported 5.58% test error does not substantiate the ε entering Proposition 1. Please train/report errors in the energy/interface norm or provide and verify norm-equivalence constants over bP.","section":"§2.3, §4.2.1, Table 4"},{"comment":"Assumption 2 is a supremum over the full localized parameter space bP, but the experimental coverage is limited to 65 training geometries, 5 DD iterations, and test geometries drawn from the same generative model. The weak-scaling study shows the error plateau growing with Nsub (relative L2 error from 2.67% for 27 subdomains to 5.79% for 125 subdomains; broken-energy error from 7.76% to 15.8%). This degradation indicates that larger decompositions produce local DD states progressively less represented in the offline distribution, which is precisely the premise whose failure would collapse the Proposition 1 bound. The paper should either provide evidence that the localized parameter space is covered in the norm of Assumption 2, or soften the claim and discuss adaptive enrichment as a necessary component.","section":"§5.2.1, Table 6"},{"comment":"The proof concludes q = sqrt(1 - θ_{k-1}/2) with θ_{k-1} = A/(A+B) ∈ (0,1]. As stated, this q depends on the iterate and can approach 1 as θ_{k-1} → 0, so no uniform constant q < 1 independent of k is established. Proposition 1 requires a fixed contraction constant q < 1. A uniform bound needs an additional argument, e.g., using the inverse/trace inequality to bound B from above by a multiple of A, or a different contraction estimate. As written, the proof of Proposition 2 is incomplete for the purpose of supporting Proposition 1.","section":"Appendix, Proposition 2, eqs. (24)-(25)"}],"minor_comments":[{"comment":"The phrase 'the bilinear formal(·,·)' should read 'the bilinear form a_l(·,·)'.","section":"Appendix, Prop. 2"},{"comment":"Reference [14] contains a typo: 'heterognous' should be 'heterogeneous'.","section":"References"},{"comment":"The notation 'MINN 0.3' is defined in Section 4.1, but a one-line reminder in the table caption would improve readability.","section":"Table 3"},{"comment":"The weak formulation of the local Robin problem is given for a pair of subdomains; it may help to state explicitly that the case of multiple interfaces follows by summing the corresponding interface terms.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is original and within the scope of the journal. The main theoretical result is sound under its assumptions, but the numerical validation does not currently verify the uniform-accuracy assumption in the correct norm, and the proof of uniform contraction needs completion. These issues are addressable with additional analysis and experiments, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine step forward for operator-level DD-ROM. The new thing is that transmission, local solve, and coarse residual are all replaced by three cascade-trained neural surrogates, so the online iteration needs no fine-scale operator assembly or local high-fidelity solves. The local-representability framing is useful, the cascaded training distribution is the right idea, and the experiments are honestly reported: ten unseen geometries, weak scaling from 27 to 125 subdomains, and no attempt to hide the error plateau or its growth.\n\nWhat it does well: previous DD-ROM work typically replaces only the local solve or uses intrusive projection. Here the authors go all the way to a fully non-intrusive two-level iteration with a deterministic coarse solve. Proposition 1 is elementary perturbation theory, but it is correctly applied. Proposition 2 is a credible discrete contraction proof for the Robin-Robin iteration under lifting-stability assumptions, with the contraction factor explicitly bounded. The numerical study is not cherry-picked, and the paper is transparent about the plateau behavior.\n\nThe soft spots are real. First, the stress-test note is right: the training losses and Table 4 measure L2(Omega*) errors (and Euclidean for the coarse residual), while Proposition 1 bounds the global error in the energy/interface norm. An L2-small error can be large in H1 near embedded 1D features. The paper never supplies the norm-consistency inequality, so the reported test errors do not substantiate the uniform error floor. Second, Assumption 2, the supremum over the whole localized parameter space, is not established; 65 synthetic vascular geometries from one generative model is a thin basis for a uniform guarantee over an infinite family. The weak-scaling results make this visible: the plateau rises from 2.7% to 5.8% relative L2 error as Nsub grows, which is consistent with states drifting outside the offline distribution. To the paper's credit, Section 6 admits that a detailed error characterization is still missing. Third, there is no code or data release, and the timing comparison is per-subdomain rather than end-to-end parallel. These are minor relative to the norm issue, but they matter for reproducibility.\n\nWho this is for: anyone working on DD-ROM, non-intrusive surrogates for multiscale PDEs, or mixed-dimensional 3D-1D problems. It deserves a serious referee: the framework is coherent, the experiments are careful within their scope, and the main theoretical claim is conditional on an assumption the paper does not yet verify. I would accept it for peer review and push for the missing norm analysis and some evidence of uniform accuracy over a broader parameter space. I would cite it if I worked in this area. For a reading group it is a useful example of how to combine DD and neural surrogates, but the norm gap should be a discussion point.","headline":"A solid, subfield-level contribution that replaces all fine-scale local operations in a two-level Robin-Robin DD iteration with cascade-trained neural surrogates; the theory is standard and the main caveat is that the uniform-accuracy premise is asserted rather than verified in the norm the bound actually needs.","tokens_in":32222,"tokens_out":2731,"would_cite":true,"duration_ms":26867,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65N30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing the exact local operators of a two-level Robin-Robin domain-decomposition iteration with neural surrogates preserves the contraction mechanism, and proves the asymptotic global error stays bounded by CDD/(1-q","keywords":["domain decomposition","model order reduction","neural surrogate models","Robin-Robin method","mixed-dimensional problems","local representability","multiscale simulation","3D-1D coupling"],"falsifier":"Take a local Robin subproblem whose localized parameter lies outside the training distribution — e.g., a vascular graph with vessel radius or branching density substantially different from the 75 synthetic networks — and measure the surrogate's error supremum. If the measured global error plateau exceeds CDD/(1-q) ε (or the local surrogate errors exceed ε), the representability or uniform-accuracy premise is falsified.","tokens_in":31312,"feed_emoji":"🫀","tokens_out":4439,"duration_ms":40735,"temperature":0.7,"pith_summary":"This paper proposes a way to solve multiscale problems without ever assembling or solving the global fine-scale system. It decomposes the domain into subdomains, maps every subdomain to a common reference shape, and learns three local maps — the local solution, the Robin transmission data, and the coarse residual contribution — with neural networks. These surrogates are inserted into a two-level Robin-Robin iteration, and the paper proves a stability result: if each local surrogate has uniform error epsilon, the global iteration converges to a neighborhood of the true solution of size at most CDD/(1-q) epsilon. The method is tested on a 3D-1D blood-perfusion model with unseen vascular geometries, where it reaches a stable error plateau without losing interface stability, and it weak-scales to over a million degrees of freedom.","feed_headline":"Surrogates replace fine-scale solves inside domain-decomposition loops","feed_subtitle":"A proof and a 3D-1D perfusion demo show the iteration stays stable, hitting a bounded error floor.","key_machinery":"The load-bearing object is the local representability property: after mapping each subdomain to a common reference domain, all local solution maps that appear during the DD iteration are assumed to belong to one parametrized family s*(bµ) on a reference problem. This justifies training a single reusable surrogate. The iteration itself is a two-level non-overlapping Robin-Robin method; the paper interprets the surrogate-embedded version as a perturbed fixed-point iteration DD + δ, and Proposition 1 bounds the propagation of δ via the contraction factor q and a stability constant CDD. A cascaded training procedure (transmission first, then solution, then coarse residual) makes each surrogate's","core_discovery":"The central claim is that domain decomposition turns a globally intractable solution map into a family of locally representable operators, and that these local operators can be replaced by neural surrogates without destroying the convergence of the iterative solver. Concretely, for the mixed-dimensional 3D-1D problem, the paper builds three surrogates — the local solution map Srom, the Robin transmission map Trom, and the coarse-residual map Crom — deployed inside a two-level Robin-Robin iteration. Proposition 1 states that if the exact DD operator is a contraction with factor q and if the local surrogate error is uniformly bounded by epsilon, then the surrogate iteration satisfies limsup ||","pith_inferences":["A direct corollary the paper does not develop: the error floor CDD/(1-q) ε can be used to allocate training effort — local samples should target the subdomains and DD states contributing the largest residual or interface errors, which the weak-scaling results already hint at.","The abstract formulation is not tied to the 3D-1D setting; if representability holds, the same operator-level replacement could apply to other parametrized multiscale PDEs with reusable local decompositions, but the assumption would have to be re-verified for each new problem class.","Because Srom does not enforce interface continuity in its outputs, the observed interface jumps (order 1e-3, with pointwise errors near 5%) suggest that adding an interface-consistency penalty to the training loss could lower the broken-energy plateau without changing the non-intrusive structure.","The reported online speedup of about 2.3x per subdomain excludes batched GPU inference; a parallel implementation that evaluates all subdomain surrogates as one batch would likely raise the speedup considerably."],"forward_implications":["Uniform local surrogate accuracy (error ≤ ε over the localized parameter space) is sufficient to keep the global iteration within a bounded error floor of size CDD/(1-q) ε.","The online stage needs no fine-scale operator assembly and no local high-fidelity solves: only neural evaluations and a small deterministic coarse solve remain.","The same reference surrogates can be reused across unseen microstructures and across larger decompositions without retraining, as demonstrated on 27, 64, and 125 subdomains.","The two-level coarse correction preserves the contraction behavior as the number of subdomains grows, so the surrogate iteration inherits weak scalability from the underlying DD method.","The cascaded training strategy reduces the offline-online distribution shift, so each surrogate is accurate on the perturbed states it actually encounters when deployed."],"fun_headline_variants":["Neural surrogates replace fine-scale solves in domain decomposition","Bounded error in neural surrogate DD: scaling without assembly","Mixed-dimensional DD-ROM: stable iteration with learned local operators","Domain decomposition goes neural: surrogates keep the iteration stable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire error bound collapses if Assumptions 1 and 2 fail: every local subproblem that arises during the iteration must be exactly representable by one reference family on a common domain, and the finite training set must make the surrogate uniformly accurate (error ≤ ε) over the whole localized parameter space — neither premise is verified beyond the specific generative model used for training.","fun_headline_variants_meta":{"raw":{"variants":["Neural surrogates replace fine-scale solves in domain decomposition","Bounded error in neural surrogate DD: scaling without assembly","Mixed-dimensional DD-ROM: stable iteration with learned local operators","Domain decomposition goes neural: surrogates keep the iteration stable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2702,"prompt_tokens":820,"completion_tokens":1882,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1813}},"tokens_in":564,"tokens_out":1882,"duration_ms":11875,"temperature":1.0,"reasoning_tokens":1813,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T23:54:53.594008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a local Robin subproblem whose localized parameter lies outside the training distribution — e.g., a vascular graph with vessel radius or branching density substantially different from the 75 synthetic networks — and measure the surrogate's error supremum. If the measured global error plateau exceeds CDD/(1-q) ε (or the local surrogate errors exceed ε), the representability or uniform-accuracy premise is falsified.","supporting_citations":[],"review_version":1}