{"id":"da4e5f15-c38c-4bcf-ba62-0283f3964c3a","arxiv_id":"2411.14942","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using finite differences instead of automatic differentiation in the loss function makes a neural kink-soliton solver about 45 percent faster than PINN with comparable accuracy.","lead":"The paper compares a neural-network method that computes derivatives by finite differences (NNDE) with physics-informed neural networks (PINN) on two kink soliton equations. NNDE reaches similar accuracy in about 45 percent less training time, which could lower the cost of solving nonlinear field equations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline speed advantage is not yet method-level evidence: Section 2.3 and Table 2 report one unseeded run per configuration, a fixed 10k-epoch budget, and no learning rate or code, so the 45% time gap could be implementation or stochastic variation rather than a property of NNDE.","rationale":"The reader's conditional verdict identifies the weakest assumption correctly: the measured speed advantage is assumed to be a property of NNDE rather than an artifact of the particular TensorFlow implementation and hyperparameter choices. My stress-test agrees with that assessment. I considered whether there is a more fundamental mathematical flaw in the NNDE construction, such as the fixed dx=10^-2 finite-difference stencil limiting asymptotic accuracy when training data is refined; this is a real limitation, but the paper explicitly compares averaged accuracy over a range of step sizes and reports only a small MSE difference, so it does not by itself invalidate the 'comparable accuracy' part of the claim. The more load-bearing issue is the lack of repeated runs, seeds, code, and learning-rate reporting, combined with the fixed-epoch protocol that prevents a time-to-accuracy comparison. These are all empirical-evidence gaps, not logical contradictions. The paper does have genuine positive features: it tests on two problems with known exact solutions, so accuracy can be measured directly, and the NNDE/PINN distinction is clearly described. But without reproducibility information, a skeptical reader cannot determine whether the 45% speedup is method-level or implementation-level. I therefore would not change the reader's CONDITIONAL verdict. The requested concrete test—multi-seed reruns with time-to-threshold logging—would settle the concern directly. My agreement with the reader is 'agree' because the reader's weakest_assumption identifies the same core weakness, even though I add the fixed-epoch/time-to-accuracy angle as part of the same concern.","tokens_in":6231,"tokens_out":7365,"duration_ms":79835,"concrete_test":"Release the code and rerun the full 5x4 configuration grid with at least 10 independent random seeds for both NNDE and PINN on the same GPU, reporting mean +/- std for wall-clock time and MSE. In the same runs, log the training epoch at which each method first reaches the other method's final MSE (or a fixed threshold such as MSE=0.003) and compare time-to-threshold. If the roughly 45% time advantage does not persist in the mean, or if the error bars overlap substantially, or if PINN reaches the target accuracy in less time, the central efficiency claim should be downgraded to a single-implementation observation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NNDE 'achieves shorter computation times while maintaining the same level of accuracy' as PINN. For this claim to hold as a statement about the method, the measured times must reflect NNDE's finite-difference derivative scheme rather than accidental implementation details. That condition is not secured by the reported evidence. Section 2.3 states that for each combination of training step size (five values) and batch size (four percentages) the authors train once and then average MSE and time; Table 2 reports those averages without error bars, seeds, or any measure of run-to-run variation. No learning rate is reported, and no PINN architecture or initialization details are given beyond the shared optimizer=Adam, dx=10^-2, and epochs=10000. Because the two methods minimize different loss landscapes (finite-difference residuals versus exact-autodiff residuals), the same default hyperparameters can favor one method arbitrarily. Additionally, the comparison uses a fixed epoch budget rather than time-to-target-accuracy; the abstract's wording 'same level of accuracy' is evaluated at the end of 10,000 epochs, but the relevant speed comparison for practical use would be the time each method needs to reach a given MSE. If PINN reaches NNDE's accuracy in fewer epochs, or if NNDE's apparent speed advantage is due to a faster but less thoroughly optimized PINN implementation, the headline conclusion would not generalize. This is not an internal inconsistency in the paper; it is a load-bearing gap in the empirical support for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural Network for Difference Equation (NNDE), which uses finite-difference approximations of derivatives inside a fully connected network with shared weights, and compares it with Physics-Informed Neural Networks (PINNs) that use automatic differentiation. The method is applied to the time-independent phi^4 and Sine-Gordon kink equations on a compactified coordinate. The reported results, averaged over five training step sizes and four batch sizes, show that NNDE achieves MSE close to PINN (0.002333 vs 0.002192 for phi^4; 0.014344 vs 0.015208 for Sine-Gordon) while reducing training time by roughly 45%. The authors conclude that NNDE is a faster alternative with comparable accuracy.","tokens_in":6503,"tokens_out":5214,"duration_ms":48568,"significance":"If the efficiency comparison is robust, the paper offers a simple and potentially useful practical variation on PINNs for solving one-dimensional ODEs: replacing autodiff by shared-weight finite differences reduces computational cost with comparable accuracy against known exact kink solutions. The study is clearly presented, benchmarks against analytic solutions, and does not fit any parameter to the comparison metric. The main limitation is that the central speed claim is not yet supported by enough statistical and reproducibility evidence; the manuscript as it stands is more of a proof-of-concept benchmark than a definitive method-level comparison.","major_comments":[{"comment":"The central claim that NNDE achieves a 45% time reduction at nearly equal accuracy rests on a single training run per configuration, with results averaged over step-size and batch-size combinations but not over repeated runs. No random seeds, standard deviations, or learning rate are reported. The measured time gap could therefore reflect stochastic variation or implementation choices in the TensorFlow code rather than a property of the NNDE scheme. Please provide repeated runs with error bars, the full hyperparameter set (learning rate, initialization, optimizer settings), and code or detailed pseudocode so the comparison is reproducible.","section":"§2.3, Table 2"},{"comment":"The comparison uses a fixed epoch budget (10,000 epochs) rather than time-to-target accuracy. The abstract's claim of 'shorter computation times while maintaining the same level of accuracy' is evaluated only at the end of this budget; if PINN reaches NNDE's final MSE in fewer epochs, the practical conclusion would change. Please report MSE as a function of training time (or epochs) for both methods and compare the time needed to reach a given accuracy threshold.","section":"§2.3, Table 2"},{"comment":"NNDE and PINN minimize different objective functions: the NNDE residual uses finite-difference derivatives with dx=10^-2, while the PINN residual uses exact automatic differentiation. The paper should quantify the truncation error introduced by this discretization and verify that the reported accuracy comparison is not biased by the choice of dx. In particular, it would be useful to show how the NNDE MSE and computation time vary as dx is reduced, to establish that the speed advantage is not tied to solving a looser discretized problem.","section":"§2.1, Eqs. (3)-(5)"}],"minor_comments":[{"comment":"There is a typo: 'mass sqauared' should be 'mass squared'.","section":"§2.2.1"},{"comment":"In the sentence introducing the models, 'we will apply NN models' should be 'we apply NN models', and the phrase 'call the kink soliton' should be 'called the kink soliton'.","section":"§2.2"},{"comment":"The sentence 'so numerical methods has to be relied upon' contains a subject-verb agreement error; consider 'numerical methods have to be relied upon'.","section":"§1"},{"comment":"The relative time reductions are reported as 45.75% and 45.84% in the text while Table 2 lists 0.457523 and 0.458447; please unify the notation and the number of significant digits.","section":"§3.1-3.2"},{"comment":"The axis labels and legends in Figures 3 and 4 are too small to read after typesetting; larger fonts and a legend entry for each method would improve readability.","section":"Figures 3 and 4"},{"comment":"Reference [12] is an arXiv preprint and is missing its arXiv identifier and year; please provide the full citation and consider whether a peer-reviewed reference is available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is exactly what it says on the tin: a comparison of PINN against NNDE, which is just PINN with derivatives computed by central finite differences instead of autodiff. That is a routine modification, and the authors missed prior finite-difference neural PDE solvers, so the novelty is low. But within that narrow scope, the work is honest and clearly presented. The two test problems have exact kink solutions, the loss is standard, and the consistency of the time reduction across five step sizes and four batch sizes suggests the effect is real at the implementation level.\n\nWhat it does well: the benchmark is clean and reproducible in principle. They state the architecture (32-64-32-1, tanh), the optimizer (Adam), dx = 1e-2, 10k epochs, and the hardware. They also note the GPU memory tradeoff for higher-order derivatives, which is a fair practical point. The figures showing MSE and time versus step size and batch size are useful.\n\nThe soft spots are real and load-bearing. Table 2 reports averages over 20 configurations, but each configuration was trained once. No seeds, no standard deviations, and no learning rate are reported. The 45% speed advantage could shift under repeated runs or a different learning rate, especially since the two methods minimize different loss landscapes and the same default hyperparameters can favor one arbitrarily. A bigger problem is that the comparison uses a fixed epoch budget rather than time-to-target-accuracy. If PINN reaches NNDE's final MSE in fewer epochs, the headline speedup reverses. The reported MSEs are also not small: 0.002 for phi4 and 0.014 for Sine-Gordon. Both methods are mediocre at these kinks, so \"comparable accuracy\" mostly means they fail in similar ways. The abstract and discussion oversell the result with phrases like \"novel method\" and \"transformative impact\"; the actual contribution is a modest benchmark on two static 1D equations.\n\nWho is this for? Someone working on neural solvers who wants a quick data point on finite-difference residuals versus autodiff. It is a useful data point, not a definitive study. The paper deserves a serious referee because the central claim is falsifiable and a competent referee can check it, but it needs major revision: seeds, error bars, the learning rate, a time-to-accuracy comparison, and ideally code. I would not cite it in my own work, though it is worth a reading-group discussion about benchmarking practices.","headline":"A clean but thin empirical benchmark: a finite-difference variant of PINN runs about 45% faster on two 1D kink equations with comparable accuracy, but the single-run evidence and fixed epoch budget do not yet support a method-level speed claim.","tokens_in":7066,"tokens_out":1815,"would_cite":false,"duration_ms":20251,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65L10","35Q51"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network that approximates derivatives by finite differences solves kink soliton equations as accurately as physics-informed neural networks while taking about 45 percent less training time.","keywords":["topological solitons","kink solitons","physics-informed neural networks","finite difference discretization","neural network differential equation solver","phi4 theory","Sine-Gordon equation","training time comparison"],"falsifier":"Rerun the comparison in Table 2 with the same equations, architecture, and hyperparameters across at least ten random seeds, recording wall-clock time and MSE for each run; if the mean training-time gap is not near 45 percent, or if NNDE's mean MSE no longer falls within a few percent of PINN's, the central claim fails.","tokens_in":6011,"feed_emoji":"⚛️","tokens_out":6067,"duration_ms":52701,"temperature":0.7,"pith_summary":"This paper tries to establish that a neural network which computes derivatives by finite differences, called NNDE, can replace the standard physics-informed neural network approach for finding topological kink solitons. On the phi4 and Sine-Gordon equations, NNDE matches PINN accuracy to within a few percent while cutting training time by roughly 45 percent. The authors argue the saving comes from avoiding automatic differentiation for the derivatives inside the equation, which forces PINN to do extra backpropagation passes. If true, the method lowers the computational barrier for studying solitons and other nonlinear systems.","feed_headline":"Kink solitons solved 45 percent faster by finite-difference neural net","feed_subtitle":"Discretized derivatives match PINN accuracy on kink equations while cutting training time by 45 percent.","key_machinery":"The load-bearing object is the NNDE architecture: an input $x$ is expanded in a second layer into three points $x-dx$, $x$, $x+dx$; each point passes through dense layers that share weights, producing approximate solution values at the three points. The loss is the squared residual of the differential equation plus squared boundary conditions, with derivatives replaced by centered finite differences, $f'(x)=\\frac{f(x+dx)-f(x-dx)}{2dx}$ and $f''(x)=\\frac{f(x+dx)+f(x-dx)-2f(x)}{dx^2}$. Because these differences are computed with a fixed stencil rather than by automatic differentiation, the network avoids PINN's extra backpropagation for derivatives. A compactifying coordinate $\\tilde{x}=\\tanh(x)$ maps the real line to $[-1,1]$ and turns the kink boundary conditions into three pinned points, $f(-1)$, $f(0)$, and $f(1)$, which are also enforced in the loss. This combination is what makes the speed comparison concrete.","core_discovery":"The central discovery is a direct comparison on two soliton-bearing equations: for phi4 theory, NNDE achieves average MSE 0.002333 versus PINN's 0.002192, with training time 1786.6 seconds versus 3293.3 seconds; for Sine-Gordon, NNDE achieves MSE 0.014344 versus 0.015208, with time 1702.8 seconds versus 3144.2 seconds. The relative time reductions are about 45.8 percent in both cases, while the accuracy differences are around 6 percent. The paper attributes the gap to the derivative mechanism: NNDE discretizes derivatives with neighboring points, so backpropagation is needed only for the loss itself, whereas PINN must backpropagate through automatic differentiation to obtain the equation's derivatives. The authors also note the tradeoff that NNDE consumes more GPU memory as higher-order derivatives require more neighboring points.","pith_inferences":["A testable extension is to push both methods to higher-order equations, for instance a fourth-order kink or a coupled system, and check whether NNDE's time advantage widens as the derivative-mechanism explanation would predict.","The reported times are single runs with no error bars; rerunning the same table with multiple random seeds would reveal whether the 45 percent gap is larger than run-to-run variance.","For very small $dx$, finite differences can suffer cancellation and roundoff; one could search for a critical $dx$ below which NNDE's accuracy degrades while PINN's does not.","The compactified coordinate trick is not restricted to kinks; the same network should apply to other localized solutions such as lumps or domain walls, provided boundary conditions can be pinned in the compactified interval."],"forward_implications":["For static one-dimensional kink problems, NNDE can be used in place of PINN at about half the training cost, making it practical to scan parameters, batch sizes, and initial guesses.","The paper's explanation implies the speed gap comes from derivative evaluation, so the advantage should be most visible when the equation contains high-order derivatives, since PINN needs deeper automatic differentiation while NNDE only adds neighboring points.","Because boundary conditions enter only as squared penalty terms, the same network handles local and non-local boundary conditions, including periodic ones, without architectural changes.","The accuracy of NNDE inherits the error of finite differences, so the discretization step $dx$ is a genuine hyperparameter that trades truncation error against memory, as the paper notes.","The paper reports that NNDE consumes more GPU memory than PINN; the method trades memory for training time."],"supporting_citations":[{"why":"Supplies the kink soliton solutions of phi4 theory and the Sine-Gordon equation that are used as exact test-data ground truth for accuracy measurements.","marker":"[1]"},{"why":"Provides the universal approximation theorem that underlies the premise that a neural network can represent the solution of the differential equation.","marker":"[4]"},{"why":"Defines the physics-informed neural network baseline whose automatic-differentiation derivative computation is the comparison point for accuracy and training time.","marker":"[11]"}],"fun_headline_variants":["Neural net with discretized derivatives cuts soliton training time by 45%","Finite-difference neural net matches PINN accuracy, trains 45% faster","Soliton solving: discretized derivatives slash training time without losing accuracy","NNDE beats PINN on speed for topological solitons","phi4 and Sine-Gordon solitons solved 45% faster with discretized derivatives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that would sink the headline is that the measured 45 percent time saving is a property of the NNDE method and not an artifact of the specific TensorFlow implementation, hyperparameters, or a single training run per configuration, since no seeds or error bars are reported.","fun_headline_variants_meta":{"raw":{"variants":["Neural net with discretized derivatives cuts soliton training time by 45%","Finite-difference neural net matches PINN accuracy, trains 45% faster","Soliton solving: discretized derivatives slash training time without losing accuracy","NNDE beats PINN on speed for topological solitons","phi4 and Sine-Gordon solitons solved 45% faster with discretized derivatives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3014,"prompt_tokens":860,"completion_tokens":2154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":2051}},"tokens_in":476,"tokens_out":2154,"duration_ms":15243,"temperature":1.0,"reasoning_tokens":2051,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:41:17.100173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the comparison in Table 2 with the same equations, architecture, and hyperparameters across at least ten random seeds, recording wall-clock time and MSE for each run; if the mean training-time gap is not near 45 percent, or if NNDE's mean MSE no longer falls within a few percent of PINN's, the central claim fails.","supporting_citations":[{"cited_title":"Comparative Study of Neural Network Methods for Solving Topological Solitons","cited_arxiv_id":"2411.14942","evidence_quote":"Supplies the kink soliton solutions of phi4 theory and the Sine-Gordon equation that are used as exact test-data ground truth for accuracy measurements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the universal approximation theorem that underlies the premise that a neural network can represent the solution of the differential equation."},{"cited_title":"Ab initio solution of the many-electron schr¨ odinger equation with deep neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the physics-informed neural network baseline whose automatic-differentiation derivative computation is the comparison point for accuracy and training time."}],"review_version":1}