{"id":"aa2740be-fe18-41ee-8a6d-8cc84f3fd5b3","arxiv_id":"2412.00636","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A finite-element-inspired neural network with locally focused basis blocks and an adaptive refinement loop solves PDEs with sharp features more accurately than standard PINNs in the reported experiments.","lead":"This paper introduces a neural network for solving PDEs whose building blocks mimic the local basis functions used in finite element methods, plus an adaptive loop that adds new blocks where the current solution is inaccurate. The authors report that the approach captures sharp peaks and singularities with errors one to two orders of magnitude below standard physics-informed networks in their experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Residual indicator (4.1) demonstrably mislocates high-error regions in the Burgers test (Sec 5.4); without a control showing adaptive placement beats random placement, the central 'automatic identification' claim is unsupported.","rationale":"I read the paper as making two central claims: fixed BI-DNN beats standard PINN at similar parameter counts, and ABI-DNN automatically finds a good architecture by placing new blocks where they are needed. The first claim is supported by consistent numerical trends across function fitting, Poisson problems, and Burgers, and the ReLU BI-block construction is exact, so I do not see an internal inconsistency there. The second claim, however, is exactly what the paper's own Sec. 5.4 undermines: the residual indicator does not identify the high pointwise-error region for Burgers. That is an internally acknowledged failure of the mechanism, not merely a disagreement with an external baseline. Because ABI-DNN's novelty is the targeted adaptive enhancement, this is the most load-bearing weakness. A random-placement control is the cleanest way to determine whether the indicator actually contributes to the reported gains or whether the gains come from retraining a larger network. The reader's strongest assumption already identified this as the weak point, and the conditional verdict is appropriate: the paper should be accepted only after the indicator's role is tested, not assumed. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":20069,"tokens_out":10137,"duration_ms":97898,"concrete_test":"Run the Burgers experiment (5.7) with matched total epochs and parameter counts, comparing three enhancement strategies at each ABI-DNN iteration: (i) the proposed eta_p-based DBSCAN placement; (ii) random centroid placement with the same number of new BI-blocks per iteration; (iii) oracle placement at the true high pointwise-error locations. Repeat with at least 10 random seeds and report mean +/- std of final relative L2 error. If strategy (i) is not statistically better than (ii), the residual indicator is not load-bearing for the reported gains and the adaptive-targeting claim is unsupported; if (i) beats (ii) despite the qualitative mismatch, the mechanism needs re-examination but the headline result may stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty of ABI-DNN is that the residual indicator eta_p = |L u_NN(x_p) - f(x_p)| (Eq. 4.1), combined with DBSCAN clustering, identifies where new BI-blocks should be inserted so that network growth is targeted. This mechanism is load-bearing: without it, ABI-DNN reduces to an architecture-growing heuristic, and the reported gains over fixed BI-DNN could be attributed to added capacity and warm-start retraining rather than intelligent placement. The paper itself provides a counterexample to the mechanism. In Sec. 5.4 (Burgers equation), Figs. 22(d)-(f) and 23(a)-(b) show that the marked clusters centered near x=0 do not overlap the regions of high pointwise error, and the text explicitly states that 'the error indicator (4.1) may not be sufficiently effective in complex problems like the Burgers equation.' The final ABI-DNN still attains 4.89e-3 versus 6.18e-3 for fixed BI-DNN, so the immediate result is not contradicted; the concern is that the claimed automatic targeted-enhancement mechanism is not established in the one nonlinear, time-dependent example, giving no reliability guarantee for the method's advertised advantage beyond the manufactured Poisson and function-fitting cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Basis-inspired Blocks (BI-blocks), which are small subnetworks designed to reproduce or approximate finite-element hat functions, and builds two architectures on them: BI-DNN, a fixed network with BI-blocks in the first layer per input dimension, and ABI-DNN, which repeatedly solves a PINN loss, marks collocation points with large PDE residual via Eq. (4.1), clusters them with DBSCAN, and inserts new BI-blocks at cluster centroids. The authors report numerical experiments on singular and high-frequency function fitting, Poisson equations with peaks and a re-entrant corner, and the viscous Burgers equation, claiming that BI-DNN outperforms PINNs with comparable parameter counts and that ABI-DNN automatically grows an architecture with error comparable to or better than the fixed BI-DNN.","tokens_in":20385,"tokens_out":5426,"duration_ms":48088,"significance":"If the performance claims hold, the work offers a practical way to inject FEM-like locality into PINNs, with an interpretable analogue of h-refinement. The BI-block construction is transparent, and the adaptive insertion via residual marking and density clustering is a clear algorithmic contribution. The paper is honest about the Burgers indicator limitation. However, the significance is tempered by missing statistical rigor (single runs), a theoretical gap for the Tanh block, and the absence of a control that isolates targeted placement from capacity increase.","major_comments":[{"comment":"The paper's own Burgers experiment shows that the residual indicator (4.1) marks clusters near x=0 that do not overlap the high pointwise-error regions, and the text concedes that the indicator 'may not be sufficiently effective' in that problem. Since targeted placement is the central novelty of ABI-DNN, this counterexample undercuts the claim that the adaptive mechanism, rather than added capacity and warm-start retraining, is responsible for the gains in Table 8. A control experiment that adds the same number of BI-blocks at random or uniformly spaced locations and retrains with the same epochs is needed to support the adaptive-placement claim.","section":"§5.4, Figs. 22-23"},{"comment":"All numerical results are from a single run per configuration; no seeds, repeats, or error bars are reported. The statements 'significantly lower' and 'comparable' rest on point comparisons that could be within run-to-run variance, especially for the cases where ABI-DNN improves on fixed BI-DNN by factors of 2-4 (e.g., Tables 1 and 5). The authors should either provide repeated runs with statistics or clearly restrict their claims to the particular trajectories shown.","section":"Tables 1-8"},{"comment":"The Tanh BI-block is obtained by replacing the piecewise-linear saturation s(x) with tanh(x) after a first-order Taylor expansion near x=0, but no error bound is given for the discrepancy between the Tanh block and the intended FEM basis function, nor for the resulting residual in the PINN loss. Because all numerical experiments use the Tanh block, the exactness of the ReLU representation (3.2) does not justify the architecture actually deployed; a quantitative approximation statement is needed to ground the 'basis-inspired' interpretation.","section":"§3.1.2, Eq. (3.5)"},{"comment":"For Problem (5.2) with 5 initial blocks, Table 4 reports 39 adaptive iterations, while Section 5.1 states that the maximum number of adaptive iterations is J=10 in all experiments. This inconsistency must be resolved, and the stopping criterion (η ≤ ηtol vs. J cap) clarified, because it affects reproducibility and the claim that ABI-DNN stops once a prescribed tolerance is met.","section":"Table 4"}],"minor_comments":[{"comment":"Algorithm 1 line 3 uses 'while η > ϵ', but ϵ is not defined; the tolerance is called ηtol elsewhere in the paper.","section":"Algorithm 1"},{"comment":"The DBSCAN parameters ε=0.1 and MinPts=1 are fixed across all problem domains (e.g., [0,1], [−1,1]^2, and the space-time domain), and no sensitivity study is given; a brief discussion of how these choices scale with domain size would help.","section":"§5.1"},{"comment":"The d-dimensional BI-DNN (3.13) is a fully connected generalization of the KST-inspired form (3.12), but the text does not discuss whether (3.13) retains any approximation guarantees; the motivation should be stated more cautiously.","section":"§3.2.3"},{"comment":"In Eq. (3.3), W2^1 is a diagonal matrix while W1^1 is a vector; a brief note on the intended dimensions would improve readability.","section":"§3.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a promising architecture paper but the evidence base is thinner than the claims. The Burgers counterexample in the paper itself should be addressed head-on, perhaps by repositioning the contribution as a heuristic h-refinement that sometimes works, or by adding controlled experiments. I would not accept the paper in its current form. The fit to the journal's scope is appropriate; the KST link is decorative but harmless."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The BI-block idea is the real contribution here. Writing a hat-shaped basis function as a small ReLU network with the weights tied to nodal positions is exact and clean, and the two-layer square-root reparameterization to tame large weights is sensible. The Tanh block is a reasonable smooth stand-in; the paper is honest that it is an approximation, though it gives no error bound for that replacement. The r-adaptivity observation, that training moves the effective nodes, is a nice way to connect the architecture to FEM intuition. On the experiments, I credit the consistent order-of-magnitude gains over standard PINN across the Poisson and function-fitting tests. That is a real pattern, not a single lucky run, and the cluster-guided insertion figures make the mechanism visible.\n\nThe soft spots are where the advertisement outruns the evidence. The central ABI-DNN claim is that the residual indicator (4.1) finds where new blocks are needed. The paper itself shows that in the Burgers test the marked clusters around x=0 do not overlap the high pointwise-error regions, and the text concedes the indicator 'may not be sufficiently effective in complex problems.' That is a load-bearing admission. Without a control that compares adaptive placement against random placement of new blocks, the reported ABI-DNN gain over fixed BI-DNN could just be extra capacity plus warm-start retraining. The stress-test note is right about this, and I see nothing in the paper that refutes it. Second, all accuracy tables are single runs with no repeats or error bars. For a method whose whole pitch is reliability, I need at least a few seeds. Third, the KST motivation for the d-dimensional assembly is heuristic: equation (3.13) is a generalization of the KST-inspired form, not a consequence of it. That is fine, but it should be stated as a design choice rather than a theorem-driven architecture. Missing baseline: the ANE method [17] is cited but never compared, and it is the closest architectural cousin.\n\nWho should read this: people working on adaptive PINN architectures and on embedding FEM structure into networks will get value from the BI-block construction. It deserves a serious referee, but the referee should push for repeated runs, a random-placement control, and either a better indicator or a more modest claim for the Burgers class. My own verdict is conditional.","headline":"A genuinely new FEM-inspired block architecture with real gains in the reported tests, but the adaptive placement mechanism is not yet supported in the one nonlinear case and the experiments need repeats; worth refereeing, not ready to crown.","tokens_in":20865,"tokens_out":1023,"would_cite":true,"duration_ms":12841,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M50","68T99","35Q68","35J75"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a neural network whose first layer is made of finite-element-style \"basis-inspired blocks\"—each initialized to reproduce a hat function—can capture sharp gradients and singularities in PDEs far better than a…","keywords":["Adaptive basis-inspired deep neural network","physics-informed neural networks","partial differential equations","singularity","finite element basis function","Kolmogorov Superposition Theorem","localized features"],"falsifier":"In the Burgers setting, compute the overlap between the clusters marked by equation (4.1) and the top-decile set of pointwise testing errors after each adaptive iteration; the paper already displays non-overlap in its Figures 22 and 23. A run where the final ABI-DNN, trained with the same epochs and initial model as a fixed BI-DNN of the final size, fails to beat that fixed BI-DNN—or where moving the clusters to the true high-error regions changes the outcome—would falsify the claim that residual-guided block insertion is what produces the gain.","tokens_in":19877,"feed_emoji":"📐","tokens_out":7882,"duration_ms":68510,"temperature":0.7,"pith_summary":"The paper is trying to establish that the weakness of physics-informed neural networks on localized features—cusp singularities, steep peaks, shocks—is fixable at the architecture level. It introduces the BI-block, a small subnetwork whose weights are initialized from a nodal position and its neighboring interval lengths so that the block behaves like a linear finite element basis function; a network assembled from these blocks is BI-DNN. Building on the Kolmogorov Superposition Theorem, the paper extends the construction from one dimension to several without tensor-product explosion. It then wraps BI-DNN in an adaptive \"solve, estimate, mark, enhance\" loop, ABI-DNN, in which new BI-blocks are added at clusters of high PDE residual. The reported experiments support the claim that BI-DNN reaches relative errors roughly one to two orders of magnitude below PINN with comparable trainable parameters, and that ABI-DNN reaches a prescribed tolerance with an architecture whose error matches or beats a fixed BI-DNN of the same size.","feed_headline":"FEM-style neural blocks beat PINNs on sharp singularities","feed_subtitle":"Basis-inspired blocks cut relative error one to two orders, and the network grows blocks where residuals cluster.","key_machinery":"The carrying object is the Basis-inspired Block (BI-block): a short subnetwork defined by equations (3.2) (ReLU) and (3.5) (Tanh), whose weights and biases are initialized from a nodal coordinate $x_j$ and adjacent interval lengths $h_{j-1}, h_j$ so that its output is a finite element hat function $\\varphi_j(x)$—exactly for ReLU, approximately for Tanh, which is used for smoothness in PINN training. Because each block's attention region is set by its initialization, the network designer knows which part of the domain a block is responsible for. BI-DNN stacks these blocks and feeds them into fully connected layers; the multi-dimensional version follows the Kolmogorov Superposition Theorem to avoid tensor products. ABI-DNN repeats solve-estimate-mark-enhance: it evaluates the residual indicator $\\eta_p = |\\mathcal{L} u_{NN}(x_p) - f(x_p)|$ on a fixed sample set, marks points above a fraction $\\gamma$ of the maximum, clusters them with DBSCAN, and inserts new BI-blocks at cluster centroids with radius-scaled supports. This machinery is what turns \"add neurons\" from an empirical gamble into a localized enrichment operation.","core_discovery":"On the paper's own terms, the central discovery is that translating the finite element hat function into a trainable neural block gives the network a built-in notion of locality: the block's attention region is known from its initialization, so adding a block at a specified location is a controlled act of local enrichment rather than a blind widening of a hidden layer. In one dimension the block is exact for ReLU and approximate for Tanh; in higher dimensions the blocks are assembled per coordinate under the Kolmogorov Superposition Theorem, and the remaining layers mix their outputs. The same trainable node positions give a form of r-adaptivity, and the adaptive loop adds blocks where the residual marks trouble, a form of h-adaptivity. The paper's numerical section claims this combination yields one-to-two-order-of-magnitude relative $L^2$ error reductions over standard PINN on cusped and high-frequency function fitting, Poisson problems with one and two steep peaks, a re-entrant corner problem, and the viscous Burgers equation, with the adaptive version typically matching or improving on the fixed BI-DNN at the same final parameter count.","pith_inferences":["Editorial inference: the BI-block front end can be read as a learnable mesh with known support, which suggests a natural diagnostic—monitor the trained nodal positions and block supports to visualize where the network actually concentrated its approximation power, a tool the paper does not build.","Editorial inference: the paper's own Burgers experiment shows the residual indicator's marked clusters can miss the regions of largest pointwise error, so replacing equation (4.1) with a gradient- or feature-aware indicator (or a dual-weighted residual) is the most direct testable upgrade; the paper names this as future work.","Editorial inference: to isolate whether the gain is architectural rather than due to extra training or parameter tuning, one could compare ABI-DNN against PINN with the same adaptive-sampling markers but no BI-blocks; the paper compares against fixed PINN only.","Editorial inference: the method's philosophy suggests a bridge to existing adaptive-sampling PINNs—use their residual-based sampling to place BI-blocks instead of just collocation points—which would combine both lines of work."],"forward_implications":["For localized-feature problems—cusp singularities, steep Gaussian peaks, re-entrant corners, shock-like Burgers profiles—BI-DNN with a similar parameter count to PINN is claimed to reduce relative $L^2$ error by one to two orders of magnitude.","ABI-DNN can start from a small network and grow it until a tolerance is met, so the user no longer has to guess the width in advance; the final network's error is comparable to or better than a fixed BI-DNN with the same architecture.","Because BI-block weights encode nodal positions, training the network is also moving the mesh (r-adaptivity), while adding blocks is mesh refinement (h-adaptivity); both mechanisms are available inside one optimization.","The KST-based assembly gives a coordinate-wise construction for multivariate problems, so the local-enrichment idea does not require tensor-product basis grids that grow exponentially with dimension."],"supporting_citations":[{"why":"Defines the standard PINN loss and training setup that serves as the baseline throughout the paper.","marker":"[2]"},{"why":"Introduces the adaptive network enhancement idea whose physical-partition cost motivates the new BI-block construction.","marker":"[17]"},{"why":"Supplies the Kolmogorov Superposition Theorem used to assemble multi-dimensional BI-DNN without tensor products.","marker":"[22]"},{"why":"Provides the simplified KST formulation with a fixed number of univariate inner functions that the paper quotes as Theorem 1.","marker":"[23]"},{"why":"Provides the DBSCAN density clustering routine used to turn marked high-residual points into placements for new BI-blocks.","marker":"[31]"},{"why":"Supplies the reference solution and test-grid conventions for the Burgers equation benchmark.","marker":"[32]"}],"fun_headline_variants":["Neural net borrows FEM hat functions for sharp PDE features","Adaptive basis blocks carve out PDE singularities with less error","BI-DNN: FEM-inspired blocks cut PINN errors on cusps and shocks","Local enrichment via basis blocks tames PDE singularities","Basis-inspired blocks add neurons where PDE residuals spike"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the residual indicator $\\eta_p = |\\mathcal{L} u_{NN}(x_p) - f(x_p)|$ evaluated on the fixed sample set identifies the regions where the approximation truly needs new capacity; the paper's own Burgers experiment shows the marked clusters can lie away from the high pointwise-error regions, so if the indicator misfires the added blocks cannot deliver the claimed improvement.","fun_headline_variants_meta":{"raw":{"variants":["Neural net borrows FEM hat functions for sharp PDE features","Adaptive basis blocks carve out PDE singularities with less error","BI-DNN: FEM-inspired blocks cut PINN errors on cusps and shocks","Local enrichment via basis blocks tames PDE singularities","Basis-inspired blocks add neurons where PDE residuals spike"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1472,"prompt_tokens":1004,"completion_tokens":468,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":620,"tokens_out":468,"duration_ms":5478,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:09:05.115029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the Burgers setting, compute the overlap between the clusters marked by equation (4.1) and the top-decile set of pointwise testing errors after each adaptive iteration; the paper already displays non-overlap in its Figures 22 and 23. A run where the final ABI-DNN, trained with the same epochs and initial model as a fixed BI-DNN of the final size, fails to beat that fixed BI-DNN—or where moving the clusters to the true high-error regions changes the outcome—would falsify the claim that residual-guided block insertion is what produces the gain.","supporting_citations":[{"cited_title":"Karniadakis","cited_arxiv_id":null,"evidence_quote":"Defines the standard PINN loss and training setup that serves as the baseline throughout the paper."},{"cited_title":"Self-adaptive deep neural network: Numerical approx- imation to functions and PDEs","cited_arxiv_id":null,"evidence_quote":"Introduces the adaptive network enhancement idea whose physical-partition cost motivates the new BI-block construction."},{"cited_title":"On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition","cited_arxiv_id":null,"evidence_quote":"Supplies the Kolmogorov Superposition Theorem used to assemble multi-dimensional BI-DNN without tensor products."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the simplified KST formulation with a fixed number of univariate inner functions that the paper quotes as Theorem 1."},{"cited_title":"DBSCAN: Density-based spatial clustering of appli- cations with noise","cited_arxiv_id":null,"evidence_quote":"Provides the DBSCAN density clustering routine used to turn marked high-residual points into placements for new BI-blocks."},{"cited_title":"Moving sampling physics-informed neural networks induced by moving mesh PDE","cited_arxiv_id":null,"evidence_quote":"Supplies the reference solution and test-grid conventions for the Burgers equation benchmark."}],"review_version":1}