{"id":"35b0df56-a5c7-4b50-9d48-d4b19257148a","arxiv_id":"2501.18708","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review that maps the field of Scientific Machine Learning and presents the authors' own cardiac simulation program as its flagship application.","lead":"This paper is a broad survey of Scientific Machine Learning, the field that combines physics-based simulation with neural networks to solve complex engineering and medical problems. It explains the mathematical foundations, reviews methods like physics-informed neural networks, and uses heart simulation as a detailed example.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.2.1's claim that NNs efficiently approximate high-dimensional PDE solutions contradicts the paper's own Theorem 3.2; the flagship cardiac 'success' is also inherited from self-cited work without quantitative evidence in this review.","rationale":"This is an expository review, so citing prior work without reproducing benchmarks is normal practice; that alone would not force rejection. However, the paper asserts a high-dimensional efficiency advantage for neural networks that its own Theorem 3.2 and Theorem 3.5 contradict unless an additional structural assumption is supplied. That is an internal inconsistency, not merely a disagreement with the wider literature, and it directly weakens the paper's broader claim that SciML methods are a compelling alternative for high-dimensional PDE problems. The cardiac applications are three-dimensional and may well be successful, so the inconsistency does not destroy the central claim; it does, however, mean the survey should be read conditionally, with the over-broad sentence in §4.2.1 and the unquantified 'successful' in the abstract both revised or qualified. The reader's CONDITIONAL verdict is therefore appropriate; my read does not change it, though I weight the internal contradiction more heavily than the inherited-evidence point.","tokens_in":50696,"tokens_out":5726,"duration_ms":54584,"concrete_test":"Verify the contradiction analytically: fix s=2, n=10, ε=10^{-3}; Eq. (32) requires N1 between (10^3)^{9/2} ≈ 3×10^{13} and (10^3)^5 = 10^{15}. If the §4.2.1 claim is meant for arbitrary Sobolev-regular functions, it is false. Then check the original papers cited in §5.2 and §5.3 for quantitative validation metrics; if none are reported, the abstract should hedge 'successful' to 'reported by the authors'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest advertised advantage of PINNs in this review is not supported and is internally contradicted. Section 4.2.1 says that for high-dimensional PDEs, 'PINNs offer a compelling alternative' because of 'neural networks' ability to efficiently approximate functions in high-dimensional spaces.' But Theorem 3.2/Eq. (32) gives N1 ≳ ε^{-(n-1)/s} for W^{s,2}((0,1)^n), i.e., exponential in n for fixed s; Theorem 3.5 gives a similar lower bound for deep ReLU nets. The paper itself notes in §3.2.5 that shallow FFNNs suffer the curse of dimensionality. No structural assumption (low intrinsic dimension, analyticity, or s scaling with n) is stated in §4.2.1. Separately, the abstract's 'successful application' to cardiac simulation is inherited: §5.2–§5.8 are self-cited summaries with no error bars, baselines, or reproduction in this paper, so the survey does not itself establish the flagship empirical claim. Both issues reduce the strength of the paper's conclusion; a conditional acceptance with those passages revised is appropriate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a broad survey/tutorial on Scientific Machine Learning (SciML). It first reviews digital models and their mathematical/numerical foundations (Section 2), then introduces machine learning concepts, architectures, training algorithms, and approximation theory (Section 3), and then describes SciML paradigms such as surrogate modeling, PINNs, VPINNs, Deep Ritz, operator learning, and hidden-dynamics discovery (Section 4). The final part (Section 5) presents the authors' Integrated Heart Model program and summarizes seven applications of SciML to cardiac simulation, which the abstract describes as the 'successful application of SciML to the simulation of the human cardiac function.' The central thesis is that combining physics-based models with data-driven algorithms can inject physical knowledge into ML while using data to enhance or accelerate physics-based models.","tokens_in":50778,"tokens_out":3038,"duration_ms":30681,"significance":"As a pedagogical survey, the paper has genuine merit: it provides a readable mathematical introduction to neural networks, cites approximation-theoretic results (Cybenko, Yarotsky, G\"uring-Kutyniok-Petersen) accurately, distinguishes observational/inductive/learning biases, and gives practical cost criteria for surrogate models, e.g. Eq. (74). It also explicitly acknowledges some limitations of PINNs, such as spectral bias and higher cost than FEM for forward problems. However, the paper's flagship claim—that SciML has been successfully applied to cardiac simulation—is not established in this manuscript: Sections 5.2-5.8 report the authors' own prior results without external benchmarks, error bars, or reproduction. In addition, a load-bearing statement in Section 4.2.1 about the efficiency of neural networks in high-dimensional spaces conflicts with the paper's own approximation-theoretic lower bounds in Section 3.2.5. These issues materially affect the strength of the conclusions and require revision.","major_comments":[{"comment":"The paragraph claiming that 'in cases involving PDEs in high-dimensional spaces... PINNs offer a compelling alternative. This is due to the neural networks' ability to efficiently approximate functions in high-dimensional spaces' is not supported by the paper's own approximation theory. Theorem 3.2 (Eq. (32)) gives N1 ≳ ε^{-(n-1)/s} for shallow networks on W^{s,2}((0,1)^n), and Theorem 3.5 gives a similar exponential-in-dimension lower bound for deep ReLU networks. The paper itself states in Section 3.2.5 that 'shallow FFNNs suffer from the curse of dimensionality' and that deep FFNNs 'do not break the curse of dimensionality.' Since Section 4.2.1 states no structural assumptions (such as low intrinsic dimension, analyticity, or special solution classes), the high-dimensional-efficiency claim is internally inconsistent. The authors should either remove the claim or qualify it with a concrete problem class for which positive results exist.","section":"4.2.1"},{"comment":"The abstract's claim of 'the successful application of SciML to the simulation of the human cardiac function' is inherited from seven summaries of the authors' own prior publications in Sections 5.2-5.8. The survey provides no error bars, no comparisons against independent baselines or full-order models, and no reproduction or external validation for these results. Because this is the paper's central demonstration, the authors should either include quantitative evidence (e.g., accuracy metrics against FOM solutions or clinical data) or explicitly reframe Section 5 as a programmatic description of the authors' ongoing research rather than an established, independently verified success.","section":"5.2-5.8"}],"minor_comments":[{"comment":"The word 'shortbreaking' should be 'shortcomings'.","section":"2.2"},{"comment":"In the density-estimation bullet, 'unveal' should be 'unveil'.","section":"3.2.1"},{"comment":"The activation function is spelled 'Heavyside' but should be 'Heaviside'.","section":"3.2.4"},{"comment":"In the RMSProp paragraph, 'RSMProp' is a typo for 'RMSProp'.","section":"3.2.6"},{"comment":"The phrase 'euclidian distance' should be 'Euclidean distance'.","section":"3.2.3"},{"comment":"The text uses 'LMM' for 'Large Language Models'; the standard abbreviation is 'LLM'.","section":"3.3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a useful survey, but its novelty and impact rest heavily on Section 5, which is almost entirely a summary of the authors' own prior work. This is acceptable in a field survey by the developers, provided the framing is explicitly programmatic and the abstract does not overstate established success. The internal contradiction between Section 4.2.1 and the approximation bounds in Section 3 should be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a review paper, not a new-results paper. Sections 2 through 4 are accurate restatements of FEM and ML theory with correct citations; Section 5 is a summary of your own cardiac program. The paper will serve as an orientation piece for newcomers to SciML.\n\nWhat it does well: the mathematics I checked is quoted correctly—Cybenko, Yarotsky, the GKP bounds, and the VC-dimension estimate are all faithful. The PINN exposition names real limitations, including spectral bias and the higher cost compared with FEM for standard forward problems. The surrogate-model cost comparison in Section 4.1 is a genuinely useful rule of thumb. The authors are also honest about the scope: they state that they had to select topics and that their classification is not rigid.\n\nThe soft spots are two, and I think the reader's note flags the right ones. First, Section 4.2.1 says PINNs are a compelling alternative for high-dimensional PDEs because neural networks efficiently approximate functions in high dimensions. That sits in direct tension with the paper's own Theorem 3.2 and Theorem 3.5, both of which give exponential dependence on dimension. Unless you add a structural assumption—low intrinsic dimension, analyticity, or an underlying low-dimensional manifold—that claim is not justified. As written, it is the standard PINN marketing phrase, and it is wrong in this paper. Second, the abstract promises a 'successful application' to cardiac simulation, but Section 5.2 through 5.8 are all self-cited summaries of the authors' prior work, without error bars, baselines, or independent reproduction. That does not make the underlying results false, but the survey itself does not demonstrate the success it advertises. A careful reader should treat Section 5 as a pointer to the primary literature, not as evidence.\n\nI agree with the reader's conditional verdict. The paper deserves a serious referee, because it is a competent survey by leading authors with a clear pedagogical value. It is not a research advance, and it should not be cited as one. Who gets value from it: graduate students, applied mathematicians moving into SciML, and anyone needing a map of the field with canonical references. My recommendation is to engage with it as a survey, and if it comes back for revision, push the authors to qualify the high-dimensional claim and to label Section 5 as a summary of their own prior work with pointers. Both are easy fixes, and they would make the paper significantly more defensible.","headline":"A competent, well-referenced survey that will help newcomers, but it repeats the common high-dimensional PINN overstatement that its own complexity theorems contradict, and it presents the authors' prior cardiac results as established without benchmarks.","tokens_in":51479,"tokens_out":2506,"would_cite":true,"duration_ms":25435,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Scientific Machine Learning—hybrid models that inject physics into neural networks and use data to enrich PDE solvers—is, this survey argues, a mature and effective strategy for complex problems governed by partial differential equations…","keywords":["Scientific Machine Learning","physics-informed neural networks","operator learning","surrogate models","partial differential equations","cardiac simulation","digital twins","neural networks"],"falsifier":"Take one of the seven cardiac tasks, for instance the multifidelity PINN estimation of ionic parameters: run it on a public dataset with known ground-truth parameters and compare the recovered values and their uncertainty to a standard finite-element calibration. If the PINN does not meet or beat the baseline within the reported accuracy, the flagship demonstration of SciML fails. Alternatively, on a high-dimensional (say $d=10$) linear PDE with known solution, measure PINN test error versus number of collocation points; if the error shows no better than exponential-in-dimension scaling, the paper's high-dimensional efficiency premise is false.","tokens_in":50330,"feed_emoji":"❤️","tokens_out":5756,"duration_ms":55520,"temperature":0.7,"pith_summary":"This survey sets out to establish that Scientific Machine Learning (SciML)—the deliberate combination of physics-based mathematical models with data-driven machine learning—is a workable and valuable strategy for problems governed by partial differential equations. It argues that neither approach alone is enough: physics-based models are robust but computationally heavy and need unknown constitutive laws and parameters, while pure data-driven models ignore causality and generalize poorly when data are sparse. The paper builds a unified mathematical description of both sides, then surveys the main hybrid families: surrogate models trained on high-fidelity simulations, physics-informed neural networks that put PDE residuals into the loss, and operator learning that approximates entire input-to-solution maps. The capstone claim is that these methods have been successfully applied to the Integrated Heart Model, a multiphysics simulation of human cardiac function, in seven concrete tasks from parameter estimation to latent-dynamics acceleration. If the survey is right, SciML is the route to patient-specific digital twins and to affordable many-query simulations.","feed_headline":"Hybrid physics-ML modelling works for PDEs and the heart","feed_subtitle":"A survey shows how neural networks and PDE solvers cooperate to make cardiac simulation tractable.","key_machinery":"The load-bearing object is the physics-informed loss function: a neural-network hypothesis space whose training objective combines the standard data-fitting term with residuals of the PDE, boundary and initial conditions, and optionally a regularization term. This is what lets the same architecture serve as forward solver, inverse parameter estimator, or surrogate: one changes only which variables are trainable and which terms enter the loss. Secondary machinery includes operator networks, which learn maps between function spaces rather than pointwise values, and reduced-order or latent-dynamics models that compress high-fidelity cardiac simulations into a few latent coordinates. The paper also leans on a formal analogy between finite-element trial spaces and neural-network hypothesis spaces, which frames the whole survey.","core_discovery":"On its own terms, the paper's central claim is that physics and machine learning are not competing paradigms but complementary components of a single modelling strategy. The mechanism is to take the loss function of a neural network and add terms that encode the residual of the governing differential equations, boundary conditions, or data from a high-fidelity solver, so the network is pushed toward physically admissible solutions. The survey classifies the field into three families—surrogate modelling of full-order models, physics-informed learning (PINNs, VPINNs, Deep Ritz), and operator learning (DeepONet, neural operators, neural ODEs)—and then presents seven cardiac applications (ionic parameter estimation by multifidelity PINNs, physics-aware inverse electrocardiography, multiscale learning of microscopic dynamics, time-dependent operator learning for multiphysics coupling, NN surrogates for sensitivity analysis and Bayesian estimation, and Latent Dynamics Networks to accelerate electrophysiology) as evidence that the hybrid approach delivers. The paper does not present new numerical results; it curates and frames existing ones to make the case that SciML is mature enough for flagship biomedical simulation.","pith_inferences":["An implication the paper leaves implicit is that the taxonomy itself is the contribution: by ordering the field into surrogate, physics-informed, and operator-learning families, the survey shapes which combinations of methods are tried next.","The paper's high-dimensional efficiency claim is worth testing directly: its own approximation bound $N_1 \\sim (1/\\varepsilon)^{n/s}$ grows exponentially in the input dimension $n$, so a controlled experiment on a $d=10$ PDE would show whether the surveyed successes rely on problem-specific structure rather than general neural-network power.","A testable extension would be to run the seven cardiac tasks on a standardized public benchmark with error bars and baseline comparisons, turning the survey's inherited evidence into independently reproducible results."],"forward_implications":["Cardiac simulations can move toward real-time and patient-specific use, because neural surrogates and latent-dynamics models reduce the cost of repeatedly solving the same multiphysics model.","Inverse problems that are ill-posed in classical terms—like estimating ionic parameters from body-surface potentials—become tractable when the PDE residual is embedded in the loss function.","Operator learning means a model trained on one set of inputs (geometries, material laws, boundary data) can predict solutions for unseen inputs without retraining.","Sensitivity analysis and Bayesian parameter estimation become affordable because a cheap neural-network surrogate replaces thousands of full-order solves.","The same hybrid recipes should transfer to other multiscale, multiphysics problems beyond cardiology, such as fluid-structure interaction or materials design."],"supporting_citations":[{"why":"Supplies the Integrated Heart Model, the simulation framework whose seven tasks serve as the paper's proof of concept.","marker":"[75]"},{"why":"Introduced Physics-Informed Neural Networks, the core physics-informed method around which Section 4.2 is built.","marker":"[194]"},{"why":"Framed physics-informed learning and the taxonomy of observational, inductive, and learning biases used throughout the paper.","marker":"[127]"},{"why":"Surveys PINN formulations and extensions, supplying the step-by-step construction used in Section 4.2.1.","marker":"[222]"},{"why":"First proposed embedding equation residuals into a neural-network loss, the mechanism on which PINNs rely.","marker":"[141, 140]"},{"why":"Introduced Neural ODEs, cited as an architecture for learning hidden or intrinsic dynamics.","marker":"[51]"},{"why":"Provides a foundation model for operator learning, used to illustrate transformer-based SciML.","marker":"[104]"},{"why":"Supports the claim that low-dimensional input-output network representations generalize better than high-dimensional ones.","marker":"[209]"}],"fun_headline_variants":["Physics and ML, together: SciML for PDEs and cardiac simulation","Neural networks meet PDE solvers in a new hybrid modelling survey","From governing equations to heart models: the SciML promise","How physics-aware ML tames cardiac complexity in PDE simulations","SciML: a crossroad of physics and data for cardiac computing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's flagship conclusion—that SciML has been successfully applied to simulate the human cardiac function—is inherited from the authors' earlier publications and is presented without independent reproduction, error bars, or comparison to baselines in this paper.","fun_headline_variants_meta":{"raw":{"variants":["Physics and ML, together: SciML for PDEs and cardiac simulation","Neural networks meet PDE solvers in a new hybrid modelling survey","From governing equations to heart models: the SciML promise","How physics-aware ML tames cardiac complexity in PDE simulations","SciML: a crossroad of physics and data for cardiac computing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1799,"prompt_tokens":1048,"completion_tokens":751,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":663}},"tokens_in":664,"tokens_out":751,"duration_ms":7764,"temperature":1.0,"reasoning_tokens":663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:46:44.554746+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the seven cardiac tasks, for instance the multifidelity PINN estimation of ionic parameters: run it on a public dataset with known ground-truth parameters and compare the recovered values and their uncertainty to a standard finite-element calibration. If the PINN does not meet or beat the baseline within the reported accuracy, the flagship demonstration of SciML fails. Alternatively, on a high-dimensional (say $d=10$) linear PDE with known solution, measure PINN test error versus number of collocation points; if the error shows no better than exponential-in-dimension scaling, the paper's high-dimensional efficiency premise is false.","supporting_citations":[{"cited_title":"Theoretical foundations of physics-informed neural networks and deep neural operators: A brief review","cited_arxiv_id":null,"evidence_quote":"Surveys PINN formulations and extensions, supplying the step-by-step construction used in Section 4.2.1."},{"cited_title":"Learning the intrinsic dynamics of spatio-temporal processes through Latent Dynamics Networks","cited_arxiv_id":null,"evidence_quote":"Supports the claim that low-dimensional input-output network representations generalize better than high-dimensional ones."}],"review_version":1}