{"id":"b738d55b-804b-4607-a399-6d830fa4c47f","arxiv_id":"1908.04463","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DL-PDE combines neural-network-based derivative estimation, synthetic meta-data generation, and sparse regression to recover PDEs from noisy, limited, discrete observations.","lead":"This paper introduces DL-PDE, a pipeline that trains a neural network on noisy measurements, generates synthetic data from that network, and uses sparse regression to recover the governing partial differential equation. It is a candidate tool for discovering physics from sparse field observations, such as well or satellite data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AD derivatives from a noisy-data-trained NN are never validated; meta-data resampling cannot correct derivative bias, so the robustness claim rests on an unproven mechanism.","rationale":"The reader's weakest assumption identifies the same core issue: the neural network trained on noisy and sparse observations must yield derivatives accurate enough for sparse regression, but no error bounds or direct validation are provided. My concern sharpens this by noting that meta-data generation, the paper's main novelty, cannot fix derivative bias because it resamples the same network; it only adds more rows to the regression. The paper's Table IX provides indirect evidence that NN derivatives degrade badly under noise when meta-data are not used, so the mechanism is not merely unproven but plausibly fragile. However, the final discovered equations in the main experiments are mostly correct, and the proposed test could resolve the concern in either direction. Therefore the reader's CONDITIONAL verdict is appropriate; the condition should be a direct derivative-accuracy check and full reporting of STRidge thresholds and training hyperparameters. This does not change the verdict, but it makes the condition explicit and testable.","tokens_in":13623,"tokens_out":5530,"duration_ms":63022,"concrete_test":"Reproduce the Burgers 5%-noise run (3000 training points, Table VI) and the groundwater 5%-noise run (2500 training points, Table II) using the paper's stated network architecture. After training, evaluate the AD derivatives (u_t, u_x, u_xx, and u_xxx where relevant) at the meta-data grid and compare them against the clean simulated true derivatives, reporting relative L2 errors. Then rerun STRidge with the true derivative columns substituted while keeping the same meta-data locations and thresholds. If the NN derivative errors exceed, say, 30% yet the discovered equation is unchanged, the claimed mechanism (accurate derivatives via AD) is not what drives the result; if the discovered equation changes materially, noise robustness fails under direct derivative scrutiny. Repeat over at least 10 random seeds and report the STRidge threshold values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the assumption that automatic differentiation of a neural network trained on noisy, sparse observations yields derivatives accurate enough for STRidge to recover the correct PDE. The paper never tests this directly: Eq. (13) compares solutions of the learned PDE with the true solution, not the derivatives used in Eq. (4). Universal approximation is cited for function values, not derivative approximation, and the training loss (Eq. 8) contains no derivative penalty or smoothing. A flexible deep network can fit noisy data by oscillating, and its AD derivatives can then be far from the true derivatives. Generating meta-data from the same trained network adds more rows of the same biased derivative estimates; this can reduce variance in STRidge but cannot remove systematic derivative bias. The paper's own Table IX shows the failure mode: for the groundwater equation without meta-data, the recovered coefficient collapses from 1.0 to 0.026 at 1% noise and 0.00024 at 5% noise, indicating the NN derivatives are unreliable under noise. Yet the paper concludes robustness without smoothing from final PDE-form recovery alone, which may reflect STRidge thresholding on a small candidate library rather than accurate derivatives. This is the most load-bearing weak point because if the derivatives are inaccurate, the claimed advantage over direct numerical differentiation is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DL-PDE, a data-driven method for discovering governing partial differential equations from discrete and noisy observations. The pipeline trains a feedforward deep neural network to fit the observed field u(x,t), uses automatic differentiation of the trained network to obtain the derivatives appearing in a candidate library, generates a large set of meta-data by querying the trained network on a denser grid, and finally applies sequential threshold ridge regression (STRidge) to identify a sparse PDE. The method is demonstrated on four model equations: groundwater flow, convection-diffusion, Burgers, and KdV. Experiments vary the data volume and noise level, include fixed-temporal and fixed-spatial monitoring settings, and compare against direct STRidge and against a no-meta-data ablation.","tokens_in":13887,"tokens_out":5473,"duration_ms":59875,"significance":"If the central claim holds, DL-PDE would be a useful addition to the PDE-discovery toolbox because it avoids direct numerical differentiation of noisy data, handles irregularly spaced and limited observations, and does not require explicit smoothing. The paper's experimental design has genuine strengths: it validates the learned PDE by solving it and comparing with the true solution using a relative-error metric, and it includes a direct comparison with STRidge on a common regular grid as well as an ablation that isolates the effect of meta-data generation. These elements make the empirical claims substantially more credible than a bare table of recovered coefficients. The main gaps are that the accuracy of the automatic-differentiation derivatives is never directly assessed, the role of meta-data is asserted rather than mechanistically demonstrated, and no uncertainty quantification or sensitivity analysis is provided for the hand-chosen STRidge threshold. These are fixable with additional experiments and analysis, so the paper is a solid candidate after major revision.","major_comments":[{"comment":"The central mechanism of the method—that a neural network trained on noisy, sparse observations yields, through automatic differentiation, derivatives accurate enough for the sparse regression in Eq. (4)—is never directly tested. The training loss in Eq. (8) contains only pointwise function-value matching, with no derivative penalty or smoothing, and Eq. (13) validates the solution of the learned PDE against the true solution, not the derivatives actually fed into STRidge. A flexible deep network can fit noisy data by oscillating, in which case its AD derivatives are biased estimates of the true derivatives. Generating meta-data from the same trained network adds more rows with the same systematic bias and cannot remove it. I request a direct assessment of derivative accuracy, for example error norms between NN-AD derivatives and the true derivatives as a function of noise level and data volume, or a PDE-residual check on the derivatives, to support the claimed robustness without smoothing.","section":"II.B, II.C, and Eq. (13)"},{"comment":"The meta-data ablation in Table IX is informative but also exposes a weakness in the interpretation. Without meta-data, the groundwater coefficient collapses from 1.110 (clean data) to 0.026 at 1% noise and 0.00024 at 5% noise, while with meta-data it remains 0.971 and 0.906. This dramatic difference shows that the raw AD derivatives from the noisy-data-trained network are highly unreliable. The paper attributes the improvement to meta-data enhancing the stability of sparse regression, but no mechanism is given for why averaging over the same biased derivative estimates removes systematic bias. The authors should either provide a variance-reduction argument with supporting evidence, or acknowledge that the recovered coefficients at higher noise levels (e.g., 0.906 vs. 1.0 for groundwater at 5% noise) remain nontrivial offsets that need explanation and uncertainty quantification.","section":"III.A.1 and Table IX"},{"comment":"The paper reports a single run for each configuration and does not report the STRidge threshold tol, the neural-network training hyperparameters (learning rate, epochs, batch size, initialization, normalization, early stopping), or any sensitivity analysis. Because the method's output is a single discovered equation whose form and coefficients depend on the threshold and on the trained network, the reader cannot determine whether coefficient deviations such as 0.940 vs. 1.0 for the KdV equation at 10% noise are typical or artifacts of one seed and one threshold. I ask for repeated independent runs (different NN initializations and noise draws), reported tol values, and a small sensitivity sweep over tol; without these, the 'satisfactory' and 'robust' conclusions are not yet fully supported.","section":"III.A.1, III.A.4, and II.C"}],"minor_comments":[{"comment":"The statement 'According to the universal approximation theorem of neural network, functions of any complexity can be approximated by a neural network with any precision' cites Ref. [24], which is a paper on numerical differentiation and regularization, not the universal approximation theorem. Moreover, universal approximation results apply to function values, not to derivatives of arbitrary order. Please correct the citation and qualify the statement.","section":"II.C and Refs. [24]"},{"comment":"The network architecture is described per case (e.g., nine-layer tanh for most cases, five-layer sin for KdV and engineering settings), but training details such as learning rate, number of epochs, batch size, data normalization, and early stopping are omitted. These details are needed for reproducibility.","section":"II.B and III.A"},{"comment":"The text says DL-PDE is 'robust to 10% noise' for the KdV equation, but Table VII shows the coefficient of u_xxx is 0.940 at 10% noise, a 6% deviation. Please define 'robust' quantitatively or soften the wording to match the reported accuracy.","section":"III.A.4"},{"comment":"The relative error definition divides by the range of the true solution, which can be small or zero for solutions that change sign, such as Burgers or KdV waves. Please specify the evaluation domains and confirm that the denominator does not make the metric misleading.","section":"Eq. (13)"},{"comment":"The fixed-temporal and fixed-spatial monitoring experiments are only presented for the convection-diffusion and Burgers equations; the main text should state this scope explicitly rather than implying that all four test cases were evaluated in engineering settings.","section":"III.B and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The contribution relative to Berg and Nystrom [22] and to the authors' earlier work [7,8] is somewhat incremental, with the main new ingredient being the generation of meta-data from the trained network. The authors should be asked to sharpen the novelty statement. The citation error for the universal approximation theorem is a warning sign that the reference list should be carefully re-checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper's actual new ingredient is generating synthetic meta-data from a trained neural network before doing sparse regression. That is a real, if modest, step beyond Berg and Nystrom, and the end-to-end experiments on four PDEs look plausible. Second, the central robustness claim is not backed by what is reported. The paper says automatic differentiation of a noise-trained network gives reliable derivatives, but it never checks that directly. The relative error in Eq. (13) compares solutions of the learned PDE with truth, not the derivatives fed into Eq. (4). Universal approximation speaks to function values, not derivative accuracy, and the training loss in Eq. (8) has no derivative or smoothing term.\n\nThe paper does several things well. The meta-data idea is clear and the ablation in Table IX shows it helps a lot in some cases. The engineering-setting tests with fixed spatial or temporal observation points are a good practical addition, and the comparison with direct STRidge in Appendix C is honest even if the smoothing details are sparse. The manuscript is written in a straightforward way and does not oversell to the point of hysteria.\n\nWhere it is soft, in proportion: the load-bearing assumption that AD derivatives are accurate under noise is unproven, and Table IX actually undermines it. Without meta-data, the groundwater coefficient collapses from 0.992 clean to 0.026 at 1% noise and 0.00024 at 5% noise, which tells me the network derivatives are unreliable once noise is present. Generating more meta-data from the same network can reduce variance in the regression, but it will not remove a systematic derivative bias. So the claimed mechanism—robustness via AD—is not established; the empirical success is real on these benchmarks, but the explanation is not.\n\nReproducibility is also thin. There is no code, no STRidge threshold values, no network training hyperparameters, and no repeated runs or error bars. The tables show single runs with no indication of variance. For a method paper, that is a significant omission.\n\nCircularity is not a real issue here. The true PDE is used only to generate training/validation data; the regression has many possible outcomes. Self-citations [7] and [8] are used for comparison, not as inputs.\n\nBottom line: this is a paper for people working on data-driven PDE discovery who want a quick script to try on noisy field data. It deserves a serious referee, because the idea is sensible and the experiments are suggestive. But it needs a major revision: direct validation of derivative accuracy, error bars or repeated runs, and full hyperparameter disclosure. I would send it to review rather than desk-reject, and ask for those changes before publication.","headline":"A plausible pipeline for PDE discovery that generates meta-data from a trained network, but the paper never validates the derivative accuracy that carries the robustness claim.","tokens_in":14388,"tokens_out":2911,"would_cite":false,"duration_ms":33015,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M32","35R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep neural network trained on sparse, noisy observations, combined with automatic differentiation and sparse regression, recovers the governing partial differential equation.","keywords":["data-driven discovery","partial differential equations","deep neural network","sparse regression","automatic differentiation","noisy data","meta-data generation"],"falsifier":"On a clean synthetic Burgers solution with known coefficients, add 5% uniform noise to a sparse training subset and run DL-PDE; if the automatic-differentiation derivatives from the noisy-trained network differ from a high-order finite-difference reference on a clean grid by more than the finite-difference error, the recovered equation will pick up spurious terms, falsifying the noise-robustness claim.","tokens_in":13427,"feed_emoji":"🧠","tokens_out":8102,"duration_ms":72096,"temperature":0.7,"pith_summary":"The paper proposes DL-PDE, a pipeline that discovers the partial differential equation governing a physical process from discrete, noisy observations. It trains a fully connected neural network to fit the data, uses automatic differentiation to obtain time and space derivatives from that network, generates a large synthetic meta-data set of solution points and derivatives, and then applies sparse regression (STRidge) to select the few terms in a candidate library that form the equation. The authors test it on groundwater flow, convection-diffusion, Burgers, and KdV equations, and report correct equation forms with coefficients close to the true values under noise levels up to 5–10% and with data volumes as low as 1–25% of a full grid. They also show it works when observations come only from fixed temporal or spatial monitoring points, which direct finite-difference sparse regression cannot handle.","feed_headline":"DL-PDE recovers governing equations from noisy, sparse data.","feed_subtitle":"Feeding automatic derivatives from a trained network into sparse regression recovers PDEs even from noisy, sparse data.","key_machinery":"The mechanism is a trained feed-forward neural network that approximates the solution as a function of $(x,t)$, with automatic differentiation supplying the derivatives $u_t$, $u_x$, $u_{xx}$, and so on at a large set of generated meta-data points; STRidge then selects a sparse coefficient vector $\\xi$ from the candidate library $\\Theta(u)$ to satisfy $U_t = \\Theta(U)\\xi$. The meta-data generation step is the load-bearing addition: it multiplies the number of equation-fitting points well beyond the original observations, which is what makes the sparse regression stable under noise and small data.","core_discovery":"The central claim is that a trained deep neural network can stand in for the unknown solution of a PDE well enough that its automatic differentiation yields accurate derivatives, and that sparse regression on those derivatives recovers the correct governing equation even when the original data are noisy and sparse. The novelty is the meta-data step: instead of evaluating derivatives only at the observed points, DL-PDE uses the trained network to generate a much larger set of points with their derivatives, which stabilizes the sparse regression. The paper demonstrates this on four classical equations and reports that the method is robust to uniform noise without any smoothing, outperforming direct STRidge and a version of DL-PDE without meta-data.","pith_inferences":["The same pipeline should extend to PDEs with smoothly varying coefficients in space or time, because the neural network can represent the coefficient fields as additional outputs and automatic differentiation can still supply the needed derivatives.","A practical sanity check suggested by the method's reliance on the network's derivatives is to compare them against finite-difference derivatives on a small clean holdout set before trusting the sparse regression result.","Running the discovery step across multiple network initializations and keeping only terms that survive every run would give a rough, data-driven measure of confidence in the recovered equation."],"forward_implications":["For groundwater flow, convection-diffusion, Burgers, and KdV equations, the correct PDE form is recovered with coefficients within a few percent of the true values, even under 5–10% uniform noise and with as little as 1–25% of the original data.","The method works in engineering settings where observations are taken only at fixed temporal or spatial monitoring points, a regime where finite-difference-based sparse regression cannot compute derivatives.","Generating meta-data from the trained network is essential: the ablation shows that without it, recovered coefficients degrade sharply under noise (for example, the groundwater coefficient falls to about 0.026 at 1% noise instead of 0.971).","On identical gridded data, DL-PDE maintains higher accuracy than direct STRidge as data volume shrinks and stays stable under noise, whereas direct STRidge selects an incorrect equation form at 5% noise."],"supporting_citations":[{"why":"Introduces the sparse identification approach that frames equation discovery as finding a sparse coefficient vector in a candidate library.","marker":"[1]"},{"why":"Supplies the STRidge algorithm and the PDE-discovery formulation that the paper adopts for its final sparse regression step.","marker":"[2]"},{"why":"Provides an alternative sparse-optimization route for learning PDE terms, cited as one of the methods the DL-PDE step can use.","marker":"[3]"},{"why":"Underpins the claim that automatic differentiation gives stable, noise-robust derivatives compared with numerical differentiation.","marker":"[17]"},{"why":"Previous neural-network automatic-differentiation PDE discovery that DL-PDE extends by generating meta-data and testing on limited, noisy data.","marker":"[22]"}],"fun_headline_variants":["Neural nets and sparse regression reveal PDEs from noisy data","DL-PDE: AI discovers equations from sparse noisy data","Neural network derivatives enable PDE discovery from sparse data","From noisy data to PDEs: DL-PDE's meta-data trick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a neural network trained on noisy and sparse observations yields, through automatic differentiation, derivatives of the true underlying solution that are accurate enough for sparse regression to recover the correct equation; no error bounds or convergence guarantees are provided.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets and sparse regression reveal PDEs from noisy data","DL-PDE: AI discovers equations from sparse noisy data","Neural network derivatives enable PDE discovery from sparse data","From noisy data to PDEs: DL-PDE's meta-data trick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2343,"prompt_tokens":880,"completion_tokens":1463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1393}},"tokens_in":496,"tokens_out":1463,"duration_ms":11224,"temperature":1.0,"reasoning_tokens":1393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:41:40.287735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a clean synthetic Burgers solution with known coefficients, add 5% uniform noise to a sparse training subset and run DL-PDE; if the automatic-differentiation derivatives from the noisy-trained network differ from a high-order finite-difference reference on a clean grid by more than the finite-difference error, the recovered equation will pick up spurious terms, falsifying the noise-robustness claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the STRidge algorithm and the PDE-discovery formulation that the paper adopts for its final sparse regression step."},{"cited_title":"Schaeffer, Learning partial differential equations via data discovery and sparse optimization, Proc","cited_arxiv_id":null,"evidence_quote":"Provides an alternative sparse-optimization route for learning PDE terms, cited as one of the methods the DL-PDE step can use."},{"cited_title":"Güneş Baydin, B","cited_arxiv_id":null,"evidence_quote":"Underpins the claim that automatic differentiation gives stable, noise-robust derivatives compared with numerical differentiation."},{"cited_title":"Berg and K","cited_arxiv_id":null,"evidence_quote":"Previous neural-network automatic-differentiation PDE discovery that DL-PDE extends by generating meta-data and testing on limited, noisy data."}],"review_version":1}