{"id":"0585de1c-ea31-40d8-a8df-35b2048c028a","arxiv_id":"2505.13042","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A teaching paper demonstrates perceptron training, PINNs, autoencoders, and SINDy on pendulum examples, with code notebooks provided online.","lead":"This tutorial walks physics students through four neural-network techniques, each applied to a simple pendulum: fitting the gravitational constant, solving the pendulum ODE with a physics-informed network, estimating latent dimensionality with autoencoders, and rediscovering the equation of motion with SINDy. It is a teaching resource with notebooks on GitHub, not a new research result.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SINDy example's reported 8/10 success is internally inconsistent, undermining the model-discovery claim.","rationale":"The reader's weakest assumption is the misstated Theorem II.1, which is a genuine correctness problem in the pedagogical narrative. However, the central claim is that the four applications demonstrate NN capabilities and are reproducible. The capstone SINDy example's own text cannot be squared internally: 8/10 recovered equation (23) versus 'some initializations recovered equation (24)' and mixed sin+linear forms. This is not a matter of external consensus but a direct internal inconsistency, and it affects the headline demonstration rather than just background theory. The theorem, by contrast, is wrong but does not feed into the numerical constructions. A conditional accept remains appropriate because the notebooks exist and could settle the counts; the recommended revision is to re-run the SINDy experiments, report exact counts per outcome, and correct the text. The concrete test above would settle the matter.","tokens_in":30960,"tokens_out":5769,"duration_ms":54988,"concrete_test":"Run the SINDy notebook from [73] ten times with distinct random seeds (documenting the environment/commit), and classify each run's discovered model: (a) exactly ẍ = -sin z (normalized eq. 23), (b) exactly ẍ = -z (eq. 24), (c) mixed ẍ = ξ0 sin z + ξ1 z with both coefficients nonzero, or (d) all coefficients pruned to zero. Compare the counts with the text's '8 correct, 2 no model.' Any run in category (b) or (c) invalidates the sentence as written and requires a corrected, quantified report. Also check whether the notebook itself records per-run outcomes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III D states: 'Among the 10 initializations we tested, 8 recovered the correct model, i.e., equation (23), while the other 2 did not recover any model.' The immediately following sentence says 'in some initializations the network recovered the model from equation (24), which is the linear approximation of the original equation, while at other times it recovered equations of the form ẍ = ξ0 sin z + ξ1 z.' Equation (24) is the small-angle linear ODE, not equation (23); a model with both sin z and z terms is also not equation (23) unless ξ1 = 0 and ξ0 matches the normalized coefficient. If some successful runs produced these alternatives, the '8 recovered the correct model' count cannot be right. If those alternatives occurred among the 2 failures, then 'did not recover any model' is wrong. The text is therefore self-contradictory. Example 4 is the capstone demonstration that NNs can discover governing equations, so this inconsistency strikes at the central claim. The notebooks [73] could resolve it, but no commit hash or random-seed details are provided, so a reader cannot tell which description is accurate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a tutorial aimed at physics undergraduates and engineering students. It introduces perceptrons, feed-forward neural networks, backpropagation, convolutional networks, and autoencoders, and then presents four Jupyter-notebook applications to a single physical system, the pendulum: supervised estimation of g from a known small-angle solution; a physics-informed neural network (PINN) solving the small-angle ODE; an autoencoder that compresses pendulum images into a low-dimensional latent space and denoises them; and a SINDy-autoencoder that attempts to discover the nonlinear pendulum equation from image data. The paper argues that this sequence demonstrates how neural networks can perform parameter fitting, ODE solving, latent-dimensionality estimation, and model discovery, and it directs readers to a GitHub repository for the notebooks.","tokens_in":31199,"tokens_out":13813,"duration_ms":132140,"significance":"If the technical content were accurate, this would be a valuable educational resource. The pedagogical structure, with four increasing-complexity tasks built around one familiar physical system, is effective, and the examples are checked against the known analytical pendulum solution rather than only against their own predictions. The availability of accompanying notebooks is a genuine strength, and the SINDy capstone is an ambitious and appropriate topic for the intended audience. The main problems are localized: a misstated mathematical theorem, a numerically inconsistent worked example, and a self-contradictory account of the SINDy results. All three are reparable, but until they are fixed the manuscript cannot serve as a dependable introduction.","major_comments":[{"comment":"Theorem II.1 is not a correct statement of the Kolmogorov-Arnold theorem, and it is not the universal approximation theorem that the surrounding discussion needs. The statement says a continuous function g 'can be approximated by' f(x) = sum_q phi_q(sum_p phi_{q,p}(x_p)) and then writes '= g(x)'; as written it is an exact equality, not an approximation. It also treats phi_q and phi_{q,p} as arbitrary given nonlinear functions, whereas the theorem asserts the existence of such univariate functions for a particular target g. Moreover, the Kolmogorov-Arnold representation theorem alone does not imply that a standard two-layer MLP with a fixed activation function and trainable weights can approximate every continuous function; that is the content of the universal approximation theorem (Cybenko; Hornik et al.). Because this theorem is the paper's only justification for the claim that a chain of weighted sums and activations can approximate the desired function, the mathematical foundation of the tutorial is currently incorrect. Please replace this theorem with an accurate statement and keep the two theorems clearly separated.","section":"Section II.B, Theorem II.1 and Eq. (15)"},{"comment":"The reported success rate for model discovery is self-contradictory. The text reads 'Among the 10 initializations we tested, 8 recovered the correct model, i.e., equation (23), while the other 2 did not recover any model' and immediately adds that in some initializations the network recovered the model of equation (24) (the small-angle linear equation) and in others recovered equations of the form z'' = xi0 sin z + xi1 z. Equation (24) is not equation (23), and a retained z term alongside sin z is also not equation (23) unless xi1 = 0 and xi0 equals the normalized coefficient. If the linear or mixed forms were successful runs, the '8 recovered the correct model' statement is wrong; if they were failures, 'did not recover any model' is wrong. This matters because Example 4 is the capstone demonstration that neural networks can discover governing equations. Please give an exact breakdown of the ten runs, define the success criterion, and provide random seeds or a pinned version of the notebook so the claim is checkable.","section":"Section III.D, Example 4 (SINDy results)"},{"comment":"The step-by-step regression example is numerically inconsistent. Equation (11) initializes b = -0.138, but the batch computation uses b = -0.13; the first entry of wt+b is written as -0.138, which agrees with the former, not the latter, and for t = 0.25 the two choices give -0.01375 and -0.00575, respectively. After the first update the text reports w_new = 0.4906 and b_new = -0.1464, but using those updated weights the ReLU outputs are not the stated [0, 0.036, 0.178, 0.32, 0.46]; they are [0, 0, 0.0989, 0.2216, 0.3442], and the reported post-update loss 0.0643 does not follow from either set of numbers. Since the paper explicitly promises a step-by-step guide, these arithmetic errors need to be corrected and the whole walkthrough rechecked.","section":"Section II.A.2, linear regression walkthrough"}],"minor_comments":[{"comment":"Grammar and spelling need a pass; examples include 'neural network and their basic concepts', 'Austrália' in the affiliation, and inconsistent capitalization in the enumerated list of applications.","section":"Abstract and Introduction"},{"comment":"The code uses variables x and y that are not defined in the snippet (the text defines x_j and y_j); as printed the code will raise a NameError. Listing 2 defines these variables correctly; the two listings should be made consistent.","section":"Section II.A.1, Listing 1"},{"comment":"The sentence 'On Table I there are some of the main cost functions used in ML' refers to Table I, which actually lists activation functions; Table II is the cost-function table.","section":"Appendix A, near Table I"},{"comment":"'Lorentz attractors' should be 'Lorenz attractors'.","section":"Section III.D"},{"comment":"The caption lists predictions at epochs 0, 1, 2, and 100 but refers only to red, green, and blue lines; please specify the color-to-epoch correspondence.","section":"Figure 6 caption"},{"comment":"The experiment estimates the dimensionality of the image manifold for a single simulated trajectory, but the text repeatedly says it estimates 'the dimensionality of the parameter space.' Please adjust the wording to avoid suggesting that the autoencoder is measuring the number of physical parameters of the pendulum.","section":"Section III.C and the abstract"},{"comment":"The repository is cited without a commit hash or DOI; for a tutorial whose numerical claims readers are expected to reproduce, please pin the exact version.","section":"Supplementary material entry [73]"}],"recommendation":"major_revision","confidential_remarks":"The paper is best evaluated as a teaching resource rather than a primary research contribution, and as such it is appropriate for a physics education venue. The reported defects are real and are not matters of taste: the theorem statement is mathematically false, the regression walkthrough does not add up, and the SINDy outcomes are described inconsistently. I do not see grounds for rejection, because the pedagogical architecture and the four pendulum applications are sound in conception and can be repaired. I would ask the authors to correct the theorem, redo all by-hand arithmetic, give an exact SINDy outcome table, and pin the notebook version before the paper is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a teaching tutorial, not a research contribution. The four-example arc around a single physical system is a good idea, and the GitHub notebooks are real assets. But the paper needs a careful correction pass before students are pointed at it; Theorem II.1 is misstated, and the SINDy success-rate paragraph contradicts itself.\n\nWhat's actually new: nothing for the ML or physics community. The methods are all established, and applying them to a pendulum is a textbook exercise. That's fine for physics.ed-ph. The paper's value is pedagogical: one notebook per example, code snippets in the text, and a progression from fitting a parameter to discovering an ODE. The SINDy setup is explained in more detail than most intros for a second-order system, which I appreciate.\n\nThe soft spots are real but fixable. The statement of the Kolmogorov-Arnold theorem (Theorem II.1) is wrong as written: it says approximation with equality, drops the dependence of the outer functions on f, and calls the construction a polynomial-size universal approximator. That conflates K-A with the standard universal approximation theorem and will confuse the exact audience the paper targets. The linear regression walkthrough also has a bias inconsistency (-0.138 initial but -0.13 used in the first batch computation), and the reported weight update doesn't match the preceding arithmetic. These are small in the big picture, but in a tutorial, arithmetic slips undermine trust.\n\nThe more serious issue is in the SINDy example. The paper says 8 of 10 initializations recovered equation (23) and the other 2 \"did not recover any model,\" then immediately says some initializations recovered equation (24) or mixed sin z + z forms. Those statements can't both be true. If the linear/mixed forms count as successes, the 8/10 number is wrong; if they were failures, \"did not recover any model\" is wrong. Without a commit hash or seeds, the notebooks can't settle it from the paper. The authors need to correct this in plain language.\n\nI'd send this to review, not desk reject, because the structure and material are worth having after revision. But I'd make the theorem correction and the SINDy count mandatory, and I'd ask for the notebook commit hash.","headline":"A useful teaching tutorial with real notebook assets, held back by a misstated theorem and an internally inconsistent SINDy success count; worth publishing after a careful correction pass.","tokens_in":31780,"tokens_out":2474,"would_cite":false,"duration_ms":25720,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One pendulum can teach a physics student the four major neural-network uses: fitting, solving, compressing, and discovering.","keywords":["Neural networks","Physics education","Tutorial","Simple pendulum","Physics-informed neural networks","Autoencoders","SINDy","Equation discovery"],"falsifier":"Compare the paper's equation (15) with the statement of the Kolmogorov-Arnold theorem in the two references cited for it. The standard theorem states an exact representation on a compact cube with inner functions built for the specific target function, while the paper's version claims an approximate-and-exact equality for any continuous function on $\\mathbb{R}^N$ and counts $O(N^2)$ univariate functions. A reader can settle the matter by reading the cited sources and checking whether they contain the equality and domain printed in the paper; they do not, which shows the stated theorem is not the cited theorem and the tutorial's foundation, as printed, is not sound.","tokens_in":30727,"feed_emoji":"🧠","tokens_out":15075,"duration_ms":144561,"temperature":0.7,"pith_summary":"This tutorial paper sets out to show that a physicist who knows only undergraduate mechanics can learn the core ideas of neural networks through a single familiar system, the pendulum. The authors walk from the perceptron and backpropagation to deep architectures, then present four pendulum applications of increasing complexity: a one-neuron network that estimates gravitational acceleration, a physics-informed network that solves the small-angle equation, an autoencoder that measures the system's one-dimensional latent space and filters image noise, and a SINDy autoencoder that rediscovers the nonlinear pendulum equation from images. The paper's practical claim is that each of these steps is reproducible with the companion notebooks, so the reader learns both the concepts and the implementation. If the demonstrations work as described, the tutorial offers a low-cost route into physics-informed machine learning for students and researchers outside the field.","feed_headline":"One pendulum teaches physicists four neural-network tasks","feed_subtitle":"Fitting, solving, compressing, discovering: reproducible notebooks make each pendulum step concrete.","key_machinery":"The carrying object is the loss function: each application works by building the physics, or the physics to be discovered, into the cost rather than leaving the network to free-form pattern match. Example 1 uses $L = \\frac{1}{M}\\sum_i (\\mathrm{data}_i - \\theta_0 \\cos(\\sqrt{g'/\\ell}\\, t_i))^2$; Example 2 adds the ODE residual and the initial condition to the data term; Example 3 minimizes reconstruction error $\\lVert x - \\psi(\\varphi(x))\\rVert^2$ through an encoder-decoder pair; and Example 4 minimizes $L_{\\mathrm{SINDy}} = L_{\\mathrm{recon}} + \\lambda_1 L_{\\dot{x}} + \\lambda_2 L_{\\dot{z}} + \\lambda_3 \\lVert \\Xi \\rVert_1$, where $\\Theta(z)$ is a library of candidate functions and $\\Xi$ is thresholded to force sparsity. The theoretical warrant offered for why such networks can represent the needed functions is Theorem II.1, stated as a two-layer universal approximation result; the paper calls it the Kolmogorov-Arnold theorem.","core_discovery":"The core demonstration is that the four standard ways neural networks meet physics — supervised parameter estimation, physics-informed solution of ordinary differential equations, unsupervised latent-dimensionality estimation, and sparse model discovery — can each be taught on the same pendulum. Example 1 fits $g$ by composing the network output with $\\theta_0 \\cos(\\sqrt{g'/\\ell}\\, t)$ and minimizing the squared error against simulated angle data. Example 2 adds the harmonic-oscillator residual $d^2\\theta/dt^2 + (g/\\ell)\\theta$ and the initial condition to the loss. Example 3 trains a convolutional autoencoder to reconstruct pendulum images and finds that a latent space of dimension one is enough, with dimension two tolerating up to 40 percent added noise. Example 4 constrains the latent dynamics to lie in a library $\\Theta(z)$ with sparse coefficients $\\Xi$ and reports that eight of ten initializations recover an equation of the form $d^2 z/dt^2 + \\sin z = 0$, with the other two recovering the linearization. The intended moral is that neural networks are not black-box substitutes for physics but tools that can be constrained by, and can recover, physical laws.","pith_inferences":["A natural test of the tutorial's pedagogical bet is to run the companion notebooks on a different but equally familiar system — the harmonic oscillator or a falling body — and see whether the four exercises transfer without extra instruction; the paper does not run that test.","The reported 8-in-10 SINDy success rate suggests that reporting a seed sweep and the coefficient-pruning schedule would strengthen reproducibility, since initialization alone changes whether the exact or the linearized equation is found.","The tutorial's practical demonstrations would survive a corrected foundation: replacing the printed Theorem II.1 with the standard universal approximation statement (density of multilayer perceptron outputs on compact sets) would preserve every one of the four exercises, since the theorem supplies the warrant, not the code.","The denoising result at 20-40 percent noise could be turned into a quantitative benchmark: measure reconstruction error versus noise level and compare against principal component analysis at the same latent dimension, since the paper notes PCA would be forced to linearize."],"forward_implications":["A student who follows the four notebooks can, without prior machine-learning coursework, recover a physical constant, solve an ODE, estimate a system's intrinsic dimensionality, and discover a governing equation from image data.","Embedding physical constraints in the loss reduces the amount of training data and epochs needed: the PINN reaches a comparable loss in about half the epochs of the pure data-fit example, and the SINDy model needs roughly 15 percent of the data other architectures would.","The autoencoder's latent space can act as a dimensionality detector: one latent coordinate reconstructs the pendulum's state, and enlarging the latent space to two coordinates raises the noise tolerance from 20 to 40 percent.","SINDy-type loss terms can yield an explicit, interpretable equation rather than a black box, since the sparse coefficient vector $\\Xi$ identifies which terms in the library are active.","The same SINDy formulation extends to multi-variable systems such as the Lorenz attractor when cross terms are added to the library $\\Theta$."],"supporting_citations":[{"why":"supplies the companion notebooks that let the reader reproduce all four pendulum applications.","marker":"[73]"},{"why":"defines the physics-informed neural network method whose loss terms Example 2 uses to solve the pendulum ODE.","marker":"[28]"},{"why":"introduces the SINDy autoencoder architecture and sparse-regression loss that Example 4 uses for equation discovery.","marker":"[17]"},{"why":"provides the autoencoder and backpropagation background that Examples 3 and 4 build on.","marker":"[15]"},{"why":"cited as the source of Theorem II.1, the universal approximation statement the tutorial's conceptual promise rests on.","marker":"[83]"},{"why":"the other citation for Theorem II.1, supplying the approximation-theory reference for why neural networks can fit continuous functions.","marker":"[64]"},{"why":"a prior SINDy pendulum implementation that the notebook in Example 4 adapts.","marker":"[106]"}],"fun_headline_variants":["One pendulum, four neural-net lessons for physicists","Four neural-network tools, one simple pendulum","Neural networks for physicists, taught by a pendulum","One pendulum, four ways to teach NN to physicists","Pendulum-based crash course in neural nets for physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that neural networks can approximate any continuous function well enough for these exercises to work, and the paper grounds that premise in Theorem II.1; as stated the theorem is not correct, so even if the notebook examples run, the tutorial's conceptual foundation is unreliable.","fun_headline_variants_meta":{"raw":{"variants":["One pendulum, four neural-net lessons for physicists","Four neural-network tools, one simple pendulum","Neural networks for physicists, taught by a pendulum","One pendulum, four ways to teach NN to physicists","Pendulum-based crash course in neural nets for physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3416,"prompt_tokens":957,"completion_tokens":2459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2385}},"tokens_in":573,"tokens_out":2459,"duration_ms":17014,"temperature":1.0,"reasoning_tokens":2385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:21:05.743518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the paper's equation (15) with the statement of the Kolmogorov-Arnold theorem in the two references cited for it. The standard theorem states an exact representation on a compact cube with inner functions built for the specific target function, while the paper's version claims an approximate-and-exact equality for any continuous function on $\\mathbb{R}^N$ and counts $O(N^2)$ univariate functions. A reader can settle the matter by reading the cited sources and checking whether they contain the equality and domain printed in the paper; they do not, which shows the stated theorem is not the cited theorem and the tutorial's foundation, as printed, is not sound.","supporting_citations":[{"cited_title":"Survey of optimization algorithms in modern neural net- works","cited_arxiv_id":null,"evidence_quote":"supplies the companion notebooks that let the reader reproduce all four pendulum applications."},{"cited_title":"A survey of transformers","cited_arxiv_id":null,"evidence_quote":"cited as the source of Theorem II.1, the universal approximation statement the tutorial's conceptual promise rests on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"a prior SINDy pendulum implementation that the notebook in Example 4 adapts."}],"review_version":1}