{"id":"ae1fefc4-a8c8-4221-8ea6-8e0e6629416e","arxiv_id":"2603.26396","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Augmented-Lagrange domain decomposition enforces interface continuity between local neural nets and is reported to converge faster than the plain Lagrange version on 2D/3D elasticity surrogates.","lead":"The paper splits the input space of a neural-network surrogate into small subdomains, trains a separate network on each, and uses Lagrange-multiplier constraints to keep the pieces continuous at their boundaries. If it holds up, this would let engineers train large simulation surrogates in parallel, with better accuracy in hard-to-fit regions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scalability claim is unsupported: all tests use ≤10 subdomains on one node, wall-clock time grows nearly linearly with M, and the paper's own conclusion admits larger problems are future work; abstract overstates ALMA as preferred for large-scale problems.","rationale":"The paper proposes a coherent extension of domain decomposition to data-driven NN surrogates, and the visual/quantitative evidence on the 2D problem supports the qualitative claims of improved continuity and local accuracy. The algorithms are described in enough detail to be plausible, and the broad experimental trends are consistent with the narrative. However, the abstract's scalability claim is not supported by the experiments. The largest test uses 10 subdomains on one shared-memory node, far below the 'large-scale' regime claimed, and the measured scaling is actually linear in M rather than demonstrating parallel speedup. The paper's own conclusion explicitly defers large-scale assessment to future work, which is a self-identified limitation. The reader's weakest assumption identified the same issue: the practical recommendation of ALMA as the preferred large-scale method is an extrapolation. Thus I agree with the reader's conditional verdict. My concrete test would settle the scalability question directly by running a 32-subdomain case and comparing against a global baseline under matched conditions. I did not elevate the shifted-error-metric or train/test concerns to the primary objection because those primarily affect the accuracy-quantification claim, while the scalability claim is both more central to the headline and more clearly contradicted by the paper's own text.","tokens_in":19704,"tokens_out":6212,"duration_ms":65502,"concrete_test":"Run ALMA and a single global NN on the same 2D compression problem with M=32 subdomains (4x8) on the same 32-core node, using the same total training budget (e.g., 100 outer iterations) and reporting wall-clock time and maximum relative error on a held-out test set. If ALMA's total wall-clock time grows linearly or worse with M relative to the M=10 result, or if its test error does not beat the global NN at matched total parameter count, the scalability claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ALMA is 'the preferred choice for large-scale problems' rests on wall-clock times from a 2D cylinder with at most 10 subdomains and a 3D cylinder with 2 subdomains, all on a single 32-core node. No strong- or weak-scaling study is presented, and Fig. 3b shows ALMA total time growing roughly linearly with the number of subdomains, not sublinearly. Multi-directional LMA runs exceeding 168 h and ALMA configurations with >30% error are excluded from the headline comparisons, so the favorable LMA-vs-ALMA contrast is partly a selection artifact. More importantly, the Conclusion states: 'In the future, ALMA will be applied to larger-scale problems to further assess its scalability... larger problems would highlight also the method's computational advantages.' This is an explicit admission that the computational/scalability advantage is not demonstrated in the present study, directly contradicting the abstract's scalability claim. Because 'preferred choice for large-scale problems' is the key differentiator of ALMA, the unsupported scalability extrapolation is load-bearing: if it fails, the contribution reduces to a small-problem DDM with modest accuracy gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two domain-decomposition strategies for training data-driven neural-network surrogate models in solid mechanics. The spatial domain is partitioned into non-overlapping subdomains, each served by a local feedforward NN, and interface continuity (C0 and C1) is enforced through either a Lagrange-multiplier formulation with iteratively linearized constraints (LMA) or an augmented-Lagrange formulation with a penalty term (ALMA). Both methods are applied to synthetic FEM data from a 2D linear-elastic compression problem and a 3D cylinder with stochastic material parameters. The reported results indicate that the constrained local models reduce interface discontinuities and improve accuracy in high-gradient regions compared with an unconstrained global NN, and that ALMA converges faster than LMA in the tested configurations. The abstract and conclusion further claim that ALMA's scalability makes it the preferred choice for large-scale problems.","tokens_in":20023,"tokens_out":4529,"duration_ms":48896,"significance":"If the central claims are sustained, the paper would be a useful contribution to the literature on domain decomposition for neural-network surrogate models, extending techniques previously developed for PINNs to purely data-driven regression. The paper contains several strengths: clear algorithmic descriptions (Algs. 1-4), a systematic sweep over the number of subdomains, a sensitivity study of ALMA to interface gaps and training-point density, and a 3D uncertainty-propagation demonstration. The convergence curves in Figs. 2 and 11 and the error maps in Figs. 7, 12 and 13 provide qualitative evidence that the interface constraints behave as intended. However, the evidence base is narrower than the claims: the 2D accuracy numbers are computed on training data, the headline comparisons exclude several poor-performing runs, a central optimizer statement is contradictory, and the scalability claim is explicitly deferred to future work in the Conclusion. These gaps concern load-bearing parts of the paper's contribution, so the manuscript needs substantial revision before the advertised claims are supported.","major_comments":[{"comment":"The claim that ALMA's scalability makes it 'the preferred choice for large-scale problems' is not supported by the experiments. All runs are on a single 32-core node with at most M=10 subdomains in 2D and M=2 in 3D. Fig. 3b shows total wall-clock time growing roughly linearly with M, not sublinearly; no strong- or weak-scaling study is provided. Multi-directional LMA runs did not finish within 168 h and are omitted from Figs. 3a/6, while multi-directional ALMA runs are omitted from Fig. 6 because their errors exceeded 30%. The Conclusion itself states that larger problems are needed 'to further assess its scalability' and that 'larger problems would highlight also the method's computational advantages.' Because the scalability preference is a key differentiator of ALMA, this extrapolation is load-bearing. Please either add scaling experiments (e.g., weak scaling with fixed work per subdo","section":"Abstract; §4.1, Figs. 3 and 6; §5"},{"comment":"The accuracy assessment uses the modified relative error e_rel = |u_hat - u|/(|u|+1). Since the outputs are normalized to [-1,1], the denominator ranges only from 1 to 2, so the reported 'maximum relative error' is not a standard relative error and can differ from the conventional definition by up to a factor of 2, especially where u is small. More importantly, the 2D errors are computed from the same FEM dataset used for training; no held-out test split is described for the 2D problem. The 3D section does mention test samples, but the 2D accuracy claims in Figs. 6 and 7 are therefore vulnerable to overfitting. Please recompute the 2D accuracy on an independent test set, report the standard relative error with a small floor in the denominator, and include the excluded multi-directional ALMA configurations (or state explicitly that the headline comparison covers only converged, low-error","section":"§4.1, Eq. (45), Figs. 6-7"},{"comment":"There is a direct contradiction about the optimizer used for LMA. The text in §4.1 states that 'LMA utilizes the Nadam optimizer' [30], while §3.1, Algorithms 1 and 3, and Table 2 specify L-BFGS for the primal updates. This is not a cosmetic inconsistency: the convergence behavior in Fig. 2, the interpretation of 'LBFGS restarts' in §4.2, and the reproducibility of the method all depend on which optimizer is actually used. Please correct the text/algorithms so that the stated optimizer matches the implementation, and report the hyperparameters actually used for LMA.","section":"§4.1, Fig. 2; §3.1; Table 2"},{"comment":"The algorithmic description is not sufficiently specified to reproduce the reported timings and convergence. The main loop uses tolerances epsilon_pr, epsilon_lambda, epsilon_ij, epsilon_l, and the inner loops use D_m and P_m, but no numerical values are given for any of these quantities. Fig. 3 is described as 'total computational time over 100 training iterations,' yet Algorithms 1 and 2 terminate by convergence checks, not by a fixed 100-iteration count. It is unclear how the 100 iterations relate to the outer loop k, the dual loop d, and the primal loop p. In addition, Algorithm 2 line 7 sets D_m = 0 inside the loop when lambda = 0, which appears to disable the dual loop under that condition and needs clarification; Algorithm 3 lists rho as a parameter but never uses it, while Algorithm 4 lists epsilon_l but never uses it. Please provide the hyperparameter values and correct the pseu","section":"Algorithms 1-4; §4.1"}],"minor_comments":[{"comment":"There are numerous typographical and formatting issues, e.g., 'G¨ odde', 'T able 1', 'B. Rosi´ c', 'domain decomposition methods (DDM)' used inconsistently, and 'the third update' versus 'the fourth dual update' in §4.2. A careful proofread is needed.","section":"Throughout"},{"comment":"The notation for the symmetric gradient is inconsistent: Eq. (10) writes ∇s u(ω) while Eq. (11) defines a mapping ∇s : u1(x)u2(ω) -> (∇s u1(x))u2(ω). Please define the symbol once and use it consistently.","section":"§2, Eqs. (10)-(11)"},{"comment":"The index in the local objective is written 'i,...,M'; it should be 'i=1,...,M'.","section":"§3, Eq. (21)"},{"comment":"The caption says 'iteration of the loop outside the given loops,' which is confusing. Please specify which outer iteration is shown and why that iteration was chosen.","section":"§4.1, Fig. 2 caption"},{"comment":"The text states that 'the fourth dual update is used to show θ_i convergence,' while Fig. 11b caption says 'dual update ALMA iteration 2.' Please align the text and captions.","section":"§4.2, Table 4 and Fig. 11"}],"recommendation":"major_revision","confidential_remarks":"The paper has a credible algorithmic core and the experiments are directionally encouraging, but the advertised scalability and accuracy claims outrun the evidence. The most serious issue is the scalability extrapolation: the authors themselves defer the relevant test to future work, yet the abstract and conclusion present ALMA as preferred for large-scale problems. The accuracy comparisons are also weakened by the modified error metric and the exclusion of non-converged or high-error runs. I would support a revised version that either provides scaling evidence and a clean held-out accuracy comparison or substantially qualifies the claims. The contradictory optimizer statement must be fixed before any decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper proposes something genuinely useful—data-driven, non-overlapping domain decomposition for neural network surrogates, with separate interface networks and C1 constraints enforced by Lagrange (LMA) and augmented Lagrange (ALMA) multipliers. The FEM-based experiments are the right kind of testbed, and the authors look at things the field often skips: missing data near interfaces, varying point density, and a 3D uncertainty propagation problem. The two algorithms are clearly specified in pseudo-code, the 3D example includes test statistics at the interface, and the conclusion is honest enough to admit that scalability to large problems is future work.\n\nThe soft spots are exactly where the abstract and conclusion stop agreeing. The headline claim that ALMA's scalability makes it the preferred choice for large-scale problems is not supported by the evidence in the paper. All runs use at most 10 subdomains on a single 32-core node; Fig. 3b shows near-linear growth in ALMA compute time, not sublinear scaling; multi-directional LMA runs over 168 h and ALMA runs with >30% errors are excluded from the headline comparisons; and there is no strong- or weak-scaling experiment anywhere. That makes the scalability paragraph in the abstract a stretch, and the reader's stress-test note lands. The future-work admission at the end says it plainly.\n\nThere are also smaller correctness and reporting issues. The 2D accuracy numbers appear to be computed on training data, with the modified error metric erel with denominator |ui|+1, which flatters the numbers. The paper describes LMA as using L-BFGS in the method section and then Nadam in the results—those cannot both be true. No code or data are released, so the numbers are not independently checkable. And the closest prior work, reference [19] with Lagrange multipliers for partitioned PINNs, is cited but never benchmarked, making the novelty claim harder to evaluate.\n\nNone of this kills the core idea. The setup is coherent, the interface constraint formulation is reasonable, and on the small problems tested the constrained methods do improve interface continuity and reduce errors near boundary layers compared to a single global NN. The paper deserves a serious referee and probably a major-revision request: align the abstract with the actual evidence, add a held-out test split, fix the optimizer contradiction, release at least the code and data, and either add a real scaling study or drop the scalability language. Right now it is a promising small-problem method with large-problem aspirations.","headline":"A useful small-scale demonstration of Lagrange-multiplier domain decomposition for data-driven NN surrogates, with the scalability claim out ahead of the evidence.","tokens_in":20518,"tokens_out":2148,"would_cite":true,"duration_ms":23758,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting the input space into subdomains makes neural-network surrogates more accurate and faster to train when interface continuity is enforced with augmented Lagrange multipliers.","keywords":["domain decomposition","neural network surrogate","augmented Lagrange multipliers","interface continuity","L-BFGS","finite element","uncertainty propagation","surrogate modeling"],"falsifier":"Run ALMA on a large problem—hundreds of subdomains or high-dimensional stochastic inputs—and measure wall-clock time and interface error: if per-subdomain cost grows faster than linearly or interface errors exceed the tolerances used here, the scalability claim fails. Alternatively, find a problem where LMA with a feasible linearization converges in fewer outer iterations than ALMA with a tuned penalty, contradicting the claim that ALMA always converges faster.","tokens_in":19563,"feed_emoji":"🧩","tokens_out":5147,"duration_ms":50073,"temperature":0.7,"pith_summary":"The paper tries to show that a single large neural network is not the best way to build surrogates for fields with local nonlinearities: splitting the input space into non-overlapping subdomains, training a small network on each, and forcing the networks to agree on interfaces recovers or beats global-network accuracy while reducing training time and parameters. Two enforcement schemes are compared: classic Lagrange multipliers with iteratively linearized constraints (LMA) and augmented Lagrange multipliers with a quadratic penalty (ALMA). On 2D and 3D elastic compression problems the interface constraints cut discontinuities dramatically, DDM approximations in nonlinear regions are more accurate than a global NN, and ALMA converges faster than LMA at a small accuracy cost. The paper concludes that ALMA is the better choice for large-scale engineering surrogates, with three vertically stacked subdomains in the 2D example giving the best accuracy-to-time balance, and a two-subdomain split reproducing 3D displacement statistics within a few percent.","feed_headline":"Split the domain, halve the error: local networks beat one big NN","feed_subtitle":"With Lagrange interface constraints, small local surrogates cut seams and train faster; augmented Lagrange is the scalable pick.","key_machinery":"The central object is a substructured (non-overlapping) domain decomposition of the input feature space: the domain is split into disjoint subdomains, each handled by a small feedforward network, with an extra network on every interface used to fix the shared values and normal derivatives. The C0/C1 interface conditions enter as constraints in each local loss. LMA enforces them with Lagrange multipliers after linearizing the network predictions around the current weights, requiring feasibility checks; ALMA replaces this with an augmented Lagrange term (λᵀQ + ρ‖Q‖²) that needs no linearization and uses the penalty ρ as multiplier step size. Local networks are trained with a quasi-Newton optim","core_discovery":"On its own terms, the paper establishes that non-overlapping domain decomposition is a viable and advantageous training strategy for data-driven neural network surrogates. By assigning each subdomain its own small network and an interface network, and by enforcing C0 and C1 continuity through Lagrange multiplier terms in the local loss, both proposed algorithms (LMA and ALMA) produce approximations with substantially reduced interface discontinuities compared to training without constraints. In the 2D compression benchmark the best DDM results halve the maximum relative error of a single-domain network while also reducing errors near the boundary where the global network struggles. ALMA is t","pith_inferences":["A fairer head-to-head would compare a global network with the same total parameter budget as all local and interface networks combined; the paper's global baseline is a single 80×2 network, so part of the accuracy gain may come from parameter distribution rather than decomposition per se.","For genuinely large runs, the number of Lagrange multipliers scales with interface points times stochastic samples (the paper notes this cost); the method's practical ceiling likely depends on compressing interface representations, e.g., by modeling interface gaps as random fields—an extension that would also replace the stochastic, run-dependent gap solutions with a single well-defined approximat","The robustness to missing data near interfaces suggests the same ALMA machinery could be applied to experimental data, such as digital image correlation fields with occluded zones, where the interface network bridges unsensed regions; the paper notes the application area but does not test real measurements.","The scalability conclusion would be strengthened by a test on more than 10 subdomains or a higher-dimensional stochastic input; as reported, the fastest ALMA configuration used only 2–3 subdomains, so 'large-scale' preference is extrapolated."],"forward_implications":["If the central claim holds, surrogate models for design optimization can be built from ensembles of small networks rather than one large one, with training parallelized over subdomains.","Interface continuity enforcement is what makes the decomposition usable: unconstrained local networks leave visible seams, and both Lagrange formulations remove them in the tested cases.","The augmented Lagrange variant offers a practical default: it needs no constraint linearization and its wall-clock time grows roughly linearly with subdomain count for vertical splits, while the classic Lagrange variant does not.","For stochastic problems, the method splits only the spatial domain while keeping the random parameter space global, which the 3D cylinder test shows is enough to reproduce mean and variance of displacement fields.","A practical recipe emerges: three vertically stacked subdomains with ALMA gave about 4% maximum relative error in about 10 hours on the 2D benchmark, compared to 3.2 minutes for a single network but with roughly double the error."],"fun_headline_variants":["Divide and conquer: domain decomposition sharpens neural surrogates","Local networks, Lagrange-stitched, outperform one big NN","Augmented Lagrange: the scalable split for neural surrogates","Domain decomposition: local NNs beat a global network on accuracy","Split the feature space, improve the fit: domain-decomposed NNs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that the augmented Lagrange method is the preferred choice for large problems rests on wall-clock measurements taken on a 32-core cluster with at most 10 subdomains in 2D and 2 subdomains in 3D, with the slowest LMA runs and least accurate ALMA runs left out of the headline comparisons.","fun_headline_variants_meta":{"raw":{"variants":["Divide and conquer: domain decomposition sharpens neural surrogates","Local networks, Lagrange-stitched, outperform one big NN","Augmented Lagrange: the scalable split for neural surrogates","Domain decomposition: local NNs beat a global network on accuracy","Split the feature space, improve the fit: domain-decomposed NNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001381,"raw_usage":{"total_tokens":5454,"prompt_tokens":791,"completion_tokens":4663,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":4574}},"tokens_in":535,"tokens_out":4663,"duration_ms":28776,"temperature":1.0,"reasoning_tokens":4574,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:16:02.361494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ALMA on a large problem—hundreds of subdomains or high-dimensional stochastic inputs—and measure wall-clock time and interface error: if per-subdomain cost grows faster than linearly or interface errors exceed the tolerances used here, the scalability claim fails. Alternatively, find a problem where LMA with a feasible linearization converges in fewer outer iterations than ALMA with a tuned penalty, contradicting the claim that ALMA always converges faster.","supporting_citations":[],"review_version":1}