{"id":"0887312f-f290-44c0-a966-ea90a198b3c2","arxiv_id":"2601.09623","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey classifying ROM–domain-decomposition coupling methods into intrusive and data-driven families, with engineering-oriented recommendations and open-challenge scoping.","lead":"This paper categorizes how reduced-order models — fast surrogates for expensive simulations — are combined with domain decomposition, which splits a large geometry into repeatable pieces. It organizes the methods into equation-based and data-driven families and names the most promising coupling strategies for engineering.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The binary intrusive/non-intrusive taxonomy is not sound because PINN is classified as non-intrusive despite not being 'purely data-driven' per the paper's own §1 definition.","rationale":"The reader's weakest assumption concerned the representativeness of the ~300 references and the lack of a documented search protocol. That is a valid concern about external coverage, but the more load-bearing issue is internal: the paper's classification criterion shifts when handling PINNs. The paper's central claim is that the literature can be adequately classified into exactly two categories. This requires a single, consistent, and exhaustive principle. The paper defines intrusive as projection-based and non-intrusive as purely data-driven, but then classifies PINNs as non-intrusive because they are not algebraic solvers—a different criterion. This is not a minor labeling quibble; PINN methods are a major branch of the review, and the paper itself stresses their ability to work without data. The taxonomy therefore does not satisfy its own definitions, which weakens the credibility of the 'most promising' comparisons. However, this inconsistency is repairable and does not necessarily invalidate the entire review, so the existing CONDITIONAL verdict remains appropriate. The concrete test—reclassifying all families under the paper's original definitions—would settle whether the binary partition is genuinely exhaustive or requires a third category or a unified definition.","tokens_in":51278,"tokens_out":3515,"duration_ms":36733,"concrete_test":"Apply the paper's §1 operational definitions to every family in Table 1, including PINN, PGD, and optimization-based methods, without invoking the §5.1.3 redefinition. Specifically, determine whether each PINN formulation cited in §5.5 (e.g., [142,141,157]) can be trained using only PDE residuals and no FOM snapshots; if at least one such 'data-free' PINN exists, the non-intrusive label fails the paper's own criterion. Also check whether any other family (e.g., PGD in §4.3.1) fits neither category without ad hoc adjustment. If any family cannot be uniquely assigned, the binary taxonomy is unsound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central taxonomic claim (§6) is that all methods can be adequately classified into two categories: intrusive and non-intrusive. However, the paper's own definitions in §1 are inconsistent with this partition. §1 states that intrusive methods 'manipulate the governing PDEs using Galerkin projection to derive a reduced system' and that non-intrusive methods are 'purely data-driven, which adopts a set of sampled data to train a surrogate model.' In §5.1.3, the paper admits that PINNs can solve PDEs without requiring data and that they incorporate the PDE residual into the loss function, yet it reclassifies PINNs as non-intrusive solely because they do not 'explicitly solve the PDE system through algebraic manipulation.' This is an ad hoc shift of the defining criterion: under the §1 definitions, PINNs are neither intrusive (no Galerkin projection) nor non-intrusive (not purely data-driven). The paper even acknowledges the tension but does not resolve it. Since PINN-based methods constitute a substantial part of the data-driven review (§5.5), the alleged dichotomy is not exhaustive. If the categories are not defined by a single consistent principle, the taxonomy is not a faithful organization of the field, and cross-category judgments such as 'DGRBEM is the most promising' are built on an unsound foundation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of reduced order modeling (ROM) combined with domain decomposition (DD). It proposes a hierarchical taxonomy: common preliminaries (domain decomposition, parameterization, dimensionality reduction), followed by a binary split into intrusive (projection-based) and non-intrusive (data-driven) coupling frameworks. Intrusive methods are further divided into monolithic and iterative schemes, and the review presents the underlying formulations for RBEM, RBHM, RDF, SCRBEM, DG-based methods, partition-of-unity methods, optimization-based methods, and Schwarz-type iterations. Non-intrusive methods are organized into Schwarz-based, interpolation-based, optimization-based, and physics-informed neural network (PINN) groups. The paper concludes that the field can be adequately classified in this way and recommends DGRBEM as the most promising method for engineering problems, with localized training and archetype blocks as the most cost-effective offline strategy.","tokens_in":51484,"tokens_out":3305,"duration_ms":37568,"significance":"If the taxonomy and the engineering recommendations are accepted, the paper provides a useful map of a fragmented literature and could help newcomers choose coupling strategies. Its strengths are the large collection of references, the worked algebraic descriptions for the intrusive families, and the up-to-date treatment of PINN-based domain decomposition, including cPINN, XPINN, DPINN, FBPINN, and related variants. The paper is less useful as a critical guide, however, because the central classification claim and the ‘most promising’ judgments rest on definitional choices and a corpus whose representativeness is not audited.","major_comments":[{"comment":"The central dichotomy is internally inconsistent. In §1, non-intrusive methods are defined as ‘purely data-driven’ methods that ‘adopt a set of sampled data to train a surrogate model.’ In §5.1.3, the paper states that PINNs can solve PDEs ‘without requiring data’ and that the PINN loss includes the PDE residual, yet it classifies PINNs as non-intrusive because they do not ‘explicitly solve the PDE system through algebraic manipulation.’ This changes the defining criterion mid-paper. Under the original definition, PINNs are neither intrusive (no Galerkin projection) nor non-intrusive (not purely data-driven). Because PINN methods are reviewed extensively in §5.5 and are folded into the two-category conclusion in §6, the claimed dichotomy is not exhaustive. A consistent definition—or a third category such as ‘physics-constrained’ or ‘PDE-informed’—is needed, and the summary conclusions sh","section":"§1 vs. §5.1.3"},{"comment":"The paper makes global knowledge claims—‘the available methods can be adequately classified into two categories’ and ‘DGRBEM is the most promising for engineering problems’—but it gives no literature-search protocol, inclusion/exclusion criteria, or coverage statistics. The set of roughly 300 references is therefore not auditable as a representative corpus. The scope note in §1 explicitly excludes clustering-based local ROMs, yet the concluding claim is unqualified. In addition, several of the recommended methods and illustrative examples come from the authors’ own research circle (e.g., refs. [16,74,107,108,135,136,138,177,178,205,217,246,261,291]); without explicit evaluation criteria (offline cost, online speedup, accuracy, implementation maturity, generality), the ‘most promising’ judgment is not transparent. I recommend adding a short methods subsection on corpus construction and ev","section":"§6 and §1"}],"minor_comments":[{"comment":"The sentence following Eq. (10) is incomplete: ‘because w_{m,i}|’ is cut off. Please complete the explanation of why the interface term vanishes for the test space.","section":"Eq. (10)"},{"comment":"The opening of §5 describes non-intrusive methods as operating ‘without requiring any modification of the underlying governing equations,’ which is a different criterion from the ‘purely data-driven’ definition in §1. This is exactly the ambiguity that leads to the PINN classification problem; the terminology should be harmonized.","section":"§5 opening"},{"comment":"The notation in Eq. (7) is typeset ambiguously: ‘F(w_i) v_j ∈ V, w_i ∈ W’ should be separated into two clauses. Also, the use of n for both the trial-space dimension and the outer normal is confusing in the same section.","section":"Eq. (7) and §4.1"},{"comment":"Heading capitalization is inconsistent (e.g., ‘parameterization techniques’ vs. ‘Physical Informed Neural Network’). Please standardize.","section":"Headings"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful survey and the technical material it presents is mostly standard. The main issue is that the headline taxonomy is not defined consistently, and the concluding recommendations would be more credible if the corpus and evaluation criteria were stated. I would not reject the paper, but I would ask for a focused revision of the definitions and the conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a genuinely useful review for newcomers to ROM+DD, mostly because it organizes a diffuse literature into a clear hierarchy and makes explicit engineering judgments. It is not a research result, and the central taxonomy has a real inconsistency that the authors should fix.\n\nWhat is new: the hierarchical classification — three common preliminaries plus coupling strategies, with monolithic vs iterative on the intrusive side and Schwarz/interpolation/optimization/PINN on the data-driven side — is a helpful organizing device. I don't know of another single text that puts RBEM, RDF, SCRBEM, DGRBEM, Schwarz variants, Gappy-POD, and PINN-based couplings side by side. The equations presented for the main intrusive methods are standard and mostly internally consistent. The paper is also honest about its own limits: Section 6 explicitly says the treatment may be too shallow for some readers and that a posteriori error analysis is omitted.\n\nSoft spots: first, the binary intrusive/non-intrusive classification is not as clean as Section 6 claims. Section 1 defines non-intrusive as purely data-driven, but Section 5.1.3 admits PINNs can solve PDEs without data and include the PDE residual in the loss. The paper then classifies PINNs as non-intrusive because they don't algebraically manipulate a reduced system. That is an ad hoc shift of the defining criterion. It doesn't sink the review, but the claim that all methods can be 'adequately classified' into two categories is too strong. Better to present the data-driven grouping as a pragmatic division rather than a logical dichotomy. Second, the reader's concern about corpus audibility is fair. There is no documented search protocol, no inclusion/exclusion criteria, no coverage statistics. Many of the 'most promising' recommendations come from the authors' own research circles, and clustering-based local ROMs are excluded in one sentence. That makes the engineering conclusions hard to audit. Both are addressable in revision.\n\nThis is a paper for practitioners and new graduate students who want a map of the field and a starting bibliography. I would read it, and I would send it to peer review: the taxonomy issues are fixable, and the scope and clarity justify referee time.","headline":"Useful engineering-oriented map of ROM+DD, but the binary intrusive/non-intrusive taxonomy is over-claimed and the corpus is not auditable.","tokens_in":52059,"tokens_out":1475,"would_cite":true,"duration_ms":17538,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"All reduced-order domain-decomposition methods fit two families","keywords":["reduced order models","domain decomposition","subdomain coupling","intrusive methods","non-intrusive methods","local reduced bases","Schwarz methods","physics-informed neural networks"],"falsifier":"A systematic literature search with explicit inclusion/exclusion criteria that finds a substantial family of coupling methods missing from the two-category map would falsify the completeness claim. Alternatively, a head-to-head benchmark on a repeating-geometry engineering problem (e.g., a thermal fin or blood-vessel flow) using the same snapshot budget — comparing DGRBEM against RBEM/SCRBEM, Dirichlet-Neumann Schwarz, and a data-driven interpolation method — would settle the 'most promising' claim: if DGRBEM is not the accuracy-per-cost winner, the recommendation fails.","tokens_in":51083,"feed_emoji":"🧩","tokens_out":6827,"duration_ms":71354,"temperature":0.7,"pith_summary":"This review tries to bring order to the many ways researchers combine reduced-order models with domain decomposition, a technique that splits a geometry into subdomains and glues local fast surrogates into a global solution. It claims that, despite dozens of individually adapted techniques, all coupling methods rest on a few concepts and can be sorted into intrusive (projection-based) and non-intrusive (data-driven) families. Within these, it maps monolithic versus iterative coupling and Schwarz-based, interpolation, optimization, and physics-informed neural network approaches. It further argues that training local models on small repeated archetype blocks is the most cost-effective offline strategy, and that a discontinuous-Galerkin-based reduced-basis method is the most promising route for engineering problems. If the review is right, a newcomer can choose a coupling method from the map, and research effort can concentrate on a few recommended families.","feed_headline":"All reduced-order domain-decomposition methods fit two families","feed_subtitle":"The survey sorts intrusive and data-driven coupling schemes and predicts which one reaches engineering first.","key_machinery":"The central machinery is the local-ROM assembly pipeline and its two-family classification. Three common preliminaries — domain decomposition, parameterization, and local reduced-basis construction — precede every method. The coupling principle is uniform: minimize discontinuities at interfaces while satisfying the PDEs and boundary conditions. Named workhorses include monolithic RBEM (Lagrange multipliers), RDF (interface finite-element basis), SCRBEM (port/bubble static condensation), DGRBEM (penalized jumps), partition-of-unity weights, optimization-based functionals, and iterative Schwarz, Dirichlet-Neumann, and Robin-Robin schemes. The key organizational identity is the taxonomy itself,","core_discovery":"On its own terms, the review establishes that ROM+DD coupling methods, despite their diversity, rest on a few concepts and sort into two families: intrusive (projection-based) and non-intrusive (data-driven). Intrusive methods split into monolithic schemes (RBEM, RDF, SCRBEM, DGRBEM, partition of unity, optimization-based) and iterative Schwarz-type schemes; non-intrusive methods split into Schwarz-based, interpolation, optimization, and PINN groups. The review further argues that localized training on archetype blocks is the most cost-effective offline strategy, and that DGRBEM — a discontinuous-Galerkin reduced-basis element method whose jump-penalty interface terms glue local bases direct","pith_inferences":["Editorial extension: the 'most promising' ranking is only as strong as the representativeness of the cited corpus; a different selection of papers could shift the ranking, so the recommendation should be read as a hypothesis about the literature rather than a measured fact.","Editorial extension: the review's localized-training premise — small networks represent large-system behavior — is directly testable by training on a small assembly and predicting on progressively larger assemblies to map the accuracy decay with scale.","Editorial extension: the taxonomy invites a shared benchmark suite that reports accuracy, offline cost, and online cost for the same repeating-geometry problem across all mapped families; such a benchmark would be the natural next step and would turn the map into a quantitative decision tool."],"forward_implications":["A researcher entering ROM+DD can choose a coupling method by locating it on the two-family map instead of surveying the full literature.","If localized training on archetype blocks is as efficient as claimed, offline costs for large repeating geometries — heat exchangers, nuclear fuel assemblies, vascular networks — can be cut by training on small representative systems and transforming bases to each instance.","If DGRBEM is the most promising monolithic method, engineering-oriented ROM implementations will likely standardize on discontinuous-Galerkin assembly with jump penalties for interface continuity.","Iterative Schwarz-type couplings are predicted to reach commercial codes before monolithic methods because of their simpler formulation and natural fit with many subdomains and multiphysics.","The review implies research attention should move toward coupling many subdomains and toward robust FOM-ROM and ROM-ROM communication schemes."],"fun_headline_variants":["ROM-DD methods fall into two families: intrusive and non-intrusive","All ROM-DD coupling methods fit into two frameworks","Intrusive vs non-intrusive: The two families of ROM-DD methods","Survey: Reduced-order model methods for domain decomposition split two ways"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the cited literature is a representative, faithfully categorized sample of the field; without an auditable search protocol, a skewed corpus would change both the taxonomy's completeness and the 'most promising' ranking.","fun_headline_variants_meta":{"raw":{"variants":["ROM-DD methods fall into two families: intrusive and non-intrusive","All ROM-DD coupling methods fit into two frameworks","Intrusive vs non-intrusive: The two families of ROM-DD methods","Survey: Reduced-order model methods for domain decomposition split two ways"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3445,"prompt_tokens":739,"completion_tokens":2706,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2629}},"tokens_in":483,"tokens_out":2706,"duration_ms":16710,"temperature":1.0,"reasoning_tokens":2629,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:32:13.036283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic literature search with explicit inclusion/exclusion criteria that finds a substantial family of coupling methods missing from the two-category map would falsify the completeness claim. Alternatively, a head-to-head benchmark on a repeating-geometry engineering problem (e.g., a thermal fin or blood-vessel flow) using the same snapshot budget — comparing DGRBEM against RBEM/SCRBEM, Dirichlet-Neumann Schwarz, and a data-driven interpolation method — would settle the 'most promising' claim: if DGRBEM is not the accuracy-per-cost winner, the recommendation fails.","supporting_citations":[],"review_version":1}