{"id":"8f46c89d-e99c-43b3-86b8-09e29cc72f7b","arxiv_id":"2409.01949","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"ELM-FBPINNs replace subdomain networks in FBPINNs with extreme learning machines, reducing PDE training to structured linear least-squares while maintaining competitive accuracy on benchmarks.","lead":"The paper proposes ELM-FBPINNs, a hybrid method combining multilevel domain decomposition with random feature models to solve PDEs by turning training into linear least-squares. This could speed up and stabilize physics-informed neural network solvers for scientific simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Linear least-squares solve may become ill-conditioned for finer decompositions or stiffer PDEs, undermining robustness claims","rationale":"The reader's weakest assumption directly isolates the same point: whether the ELM substitution preserves both expressivity and numerical stability. Because the full manuscript supplies only a numerical study without reported condition numbers or conditioning analysis, that assumption remains the single load-bearing uncertainty; the reader's UNVERDICTED verdict is therefore unchanged.","tokens_in":1674,"tokens_out":326,"duration_ms":14115,"concrete_test":"For each benchmark in the numerical study, extract the normal matrix of the final least-squares problem (or its largest singular value ratio) at the reported optimal feature count and subdomain level; if the 2-norm condition number exceeds 1e8 or grows by more than an order of magnitude when subdomain count is doubled, the stability/robustness claim requires qualification or additional regularization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction replaces nonlinear subdomain optimization with a single structured linear least-squares problem whose matrix is assembled from random-feature residuals and partition-of-unity weights. For the claim of improved robustness and competitive accuracy to hold, this matrix must remain sufficiently well-conditioned across the reported benchmarks and parameter ranges. No a-priori bound or regularization strategy is indicated in the abstract, and the numerical study would need to demonstrate that condition numbers stay moderate when subdomain count or feature count is increased; otherwise the observed stability may be an artifact of the specific test problems rather than a general property of the ELM-FBPINN formulation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes ELM-FBPINNs, a hybrid method that replaces the trainable subdomain networks in finite-basis PINNs (FBPINNs) with extreme learning machines (random feature models). This yields an explicit structured linear least-squares problem that eliminates backpropagation. The multilevel extension combines this with domain decomposition and partition-of-unity weighting. A systematic numerical study on benchmark PDEs is claimed to show that both ELM-FBPINNs and multilevel ELM-FBPINNs attain competitive accuracy while accelerating convergence and improving robustness to architectural and optimization parameters.","tokens_in":1828,"tokens_out":362,"duration_ms":11657,"significance":"If the performance and robustness claims are substantiated, the approach would supply a computationally lighter alternative to standard PINNs and FBPINNs for localized PDE approximation, removing iterative nonlinear optimization while retaining the benefits of domain decomposition.","major_comments":[{"comment":"Abstract: the central claim of 'competitive accuracy' and 'significantly accelerating convergence' is asserted on the basis of a systematic numerical study, yet the abstract (and the information supplied) contains no quantitative error values, convergence rates, error bars, or specific benchmark comparisons, preventing verification of the performance assertions.","section":"Abstract"},{"comment":"Numerical study / method formulation: the linear least-squares matrix assembled from random-feature residuals and partition-of-unity weights is asserted to remain well-conditioned and robust across the reported parameter ranges, but no condition-number measurements, regularization strategy, or a-priori bounds are indicated; without such evidence the robustness claim cannot be separated from the specific test problems chosen.","section":"Numerical study"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major point below.","responses":[{"response":"We agree that the abstract would be strengthened by including quantitative indicators. We will revise the abstract to report representative error values, convergence improvements, and benchmark comparisons from the numerical study.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim of 'competitive accuracy' and 'significantly accelerating convergence' is asserted on the basis of a systematic numerical study, yet the abstract (and the information supplied) contains no quantitative error values, convergence rates, error bars, or specific benchmark comparisons, preventing verification of the performance assertions."},{"response":"The robustness is evidenced by consistent performance across the reported parameter sweeps in the numerical study. We acknowledge that explicit condition-number measurements would provide additional support. We will add condition-number results for representative matrices and clarify any regularization in the revised manuscript.","revision_made":"yes","referee_comment":"[Numerical study] Numerical study / method formulation: the linear least-squares matrix assembled from random-feature residuals and partition-of-unity weights is asserted to remain well-conditioned and robust across the reported parameter ranges, but no condition-number measurements, regularization strategy, or a-priori bounds are indicated; without such evidence the robustness claim cannot be separated from the specific test problems chosen."}],"tokens_in":1323,"tokens_out":305,"duration_ms":13047,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The central contribution is replacing the nonlinear optimization inside each subdomain of FBPINNs with random-feature extreme learning machines. This produces an explicit linear least-squares formulation that eliminates backpropagation while keeping the multilevel partition-of-unity structure. The authors run a systematic numerical comparison against plain PINNs and standard FBPINNs on several benchmark PDEs, plus ablation experiments that isolate the effects of domain decomposition versus random-feature enrichment. Those experiments are the part that actually adds value: they give concrete evidence that accuracy remains competitive while training becomes faster and less sensitive to hyperparameter choices such as feature count or subdomain layout. The formulation itself is straightforward and the linear-algebra step is reproducible in principle. The main soft spot is conditioning of the assembled least-squares matrix. The stress-test concern is reasonable: nothing in the abstract indicates a priori bounds or regularization, and if the reported runs do not track condition numbers as subdomain count or feature count grows, the robustness claim is limited to the specific test problems rather than a general property. The paper does not appear to contain internal contradictions or circular claims; the method is defined cleanly and the numerical claims are falsifiable. This is useful reading for anyone already working on domain-decomposed PINN variants who wants a faster training route without leaving the PINN framework. It is not paradigm-shifting, but the combination is new enough and the numerics are presented systematically enough that it merits a serious referee. I would send it to review and ask specifically for condition-number diagnostics and scaling tests on stiffer problems.","headline":"The paper swaps trainable subdomain networks for extreme learning machines inside multilevel FBPINNs, turning the problem into a single linear least-squares solve and reporting faster convergence on benchmarks.","tokens_in":2306,"tokens_out":384,"would_cite":false,"duration_ms":14714,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"ELM-FBPINN linear least-squares PDE solver is orthogonal to RS forcing chain","alignment":"orthogonal","rationale":"Paper's core construction (random-feature ELMs + partition-of-unity domain decomposition reducing FBPINN training to a single structured linear least-squares problem) lies entirely in numerical analysis for PDEs. It contains no J-cost, cosh identities, golden-ratio ladder, 8-tick periodicity, or parameter-free derivation of constants. No RS theorem (e.g., reality_from_one_distinction, J-uniqueness via Aczél, Alexander-duality D=3 forcing, or any Cost/Constants module) is invoked or paralleled.","tokens_in":44922,"confidence":"high","tokens_out":155,"duration_ms":10008,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"ELM-FBPINNs replace trainable subdomain networks in finite basis PINNs with extreme learning machines, reducing PDE training to a structured linear least-squares problem.","keywords":["extreme learning machines","physics-informed neural networks","domain decomposition","random features","PINNs","FBPINNs","multilevel methods","PDE solvers"],"falsifier":"On a standard benchmark such as the Poisson equation or a nonlinear wave problem, the ELM-FBPINN either produces solution errors substantially larger than those of FBPINNs or yields an ill-conditioned least-squares matrix whose solution deviates markedly from the true PDE solution.","tokens_in":2585,"feed_emoji":"🧮","tokens_out":672,"duration_ms":18849,"temperature":0.7,"pith_summary":"The paper establishes a hybrid method that combines multilevel domain decomposition and partition-of-unity constructions with random feature models. This replaces the iterative nonlinear optimization inside each subdomain of FBPINNs with extreme learning machines. Training then becomes a linear least-squares problem with no backpropagation required. Systematic comparisons on benchmark PDEs show that the resulting ELM-FBPINNs and multilevel variants reach competitive accuracy while accelerating convergence and increasing robustness to choices of architecture and optimizer. Ablation experiments separate the contributions of domain decomposition from those of the random-feature enrichment.","feed_headline":"Random features turn subdomain PINNs into linear least-squares solves","feed_subtitle":"ELM-FBPINNs reach accuracy comparable to nonlinear FBPINNs while accelerating convergence and reducing sensitivity to hyperparameters on PDE","key_machinery":"Extreme learning machines placed inside each subdomain of the multilevel finite-basis PINN framework, which reformulates the entire training task as a linear least-squares problem.","core_discovery":"By integrating multilevel domain decomposition and partition-of-unity constructions with random feature models, the multilevel ELM-FBPINN replaces trainable subdomain networks with extreme learning machines. This eliminates backpropagation entirely and reduces training to a structured linear least-squares problem, while achieving competitive accuracy on representative benchmark problems compared to standard PINNs and FBPINNs.","pith_inferences":["The linear least-squares structure may allow direct use of sparse or iterative linear solvers that scale to larger subdomain counts than nonlinear optimizers permit.","Random-feature enrichment could be inserted into other localized PINN variants to reduce per-subdomain training cost without altering the overall domain-decomposition layout.","Multilevel extensions of the same construction may further improve scalability for PDEs on complex geometries where single-level decompositions remain expensive."],"forward_implications":["ELM-FBPINNs achieve competitive accuracy while significantly accelerating convergence on benchmark problems.","Robustness to architectural and optimisation parameters improves relative to standard PINNs and FBPINNs.","Domain decomposition and random feature enrichment separately control expressivity, conditioning, and scalability.","Backpropagation is eliminated entirely, removing the need for iterative nonlinear solvers inside subdomains."],"fun_headline_variants":["ELM-FBPINNs reduce subdomain PINNs to least-squares","Random features eliminate backprop in FBPINNs","Linear solves replace nonlinear optimization in PINNs","ELM-FBPINNs achieve PINN accuracy with linear training"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Extreme learning machines retain enough approximation power inside each subdomain that the resulting linear least-squares problem stays well-conditioned and does not lose expressivity relative to nonlinearly trained networks.","fun_headline_variants_meta":{"raw":{"variants":["ELM-FBPINNs reduce subdomain PINNs to least-squares","Random features eliminate backprop in FBPINNs","Linear solves replace nonlinear optimization in PINNs","ELM-FBPINNs achieve PINN accuracy with linear training"]},"model":"grok-4.3","cost_usd":0.007155,"raw_usage":{"total_tokens":3286,"prompt_tokens":633,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":71549500,"prompt_tokens_details":{"text_tokens":633,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2588,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":633,"tokens_out":65,"duration_ms":18506,"temperature":1.0,"reasoning_tokens":2588,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-23T21:24:14.364276+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a standard benchmark such as the Poisson equation or a nonlinear wave problem, the ELM-FBPINN either produces solution errors substantially larger than those of FBPINNs or yields an ill-conditioned least-squares matrix whose solution deviates markedly from the true PDE solution.","supporting_citations":[],"review_version":1}