{"id":"cadb5a74-cc50-4adb-9750-8e417b12262a","arxiv_id":"2411.18121","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tiny neural network using simple distance-based fingerprints matches the accuracy of million-parameter neural network potentials for small molecules, while evaluating faster and extrapolating more safely.","lead":"This paper introduces KerNN, a molecular energy model that uses simple mathematical fingerprints of atomic distances as inputs to a very small neural network. It reports accuracy close to much larger neural network potentials, with faster evaluation and more stable behavior far from the training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Extrapolation claim is broader than evidence: saturation of k[3,3] features only guarantees bounded inputs, not controlled outputs, and the paper itself documents a breakdown for hydrogen oxalate.","rationale":"The reader's weakest assumption identified the same core issue: the extrapolation claim rests on an empirical property of trained KerNN models rather than a guaranteed consequence of decaying kernel features, and the hydrogen oxalate trajectory breakdown is the paper's own documented counterexample. The paper is otherwise strong on interpolation: learning curves, force errors, harmonic frequencies, IR spectra, H-transfer rates, and the HeH2+ bound-state comparison versus an FCI-based RKHS PES are independent and mutually consistent evidence that the small network is competitive with PhysNet in-distribution. The computational speed advantage (Table S5) and the transfer-learning demonstration with an experimental dissociation energy are also credible. I do not see an internal inconsistency in the architecture or the loss functions; the issue is scope of the extrapolation claim relative to the evidence. The HeH2+ system does show good extrapolation along the tested atom-diatom channels, and the 5000 K H2CO test is a genuine extrapolation, so partial credit is due. But the abstract and conclusion state the solution of 'a general problem of NN-based PESs' without caveat, while the only system with a conformational/topological change beyond the training topology is acknowledged to break down. That mismatch justifies a conditional acceptance: the paper should either restrict the extrapolation claim to bond-stretching dissociation within a fixed connectivity, or provide evidence and/or data augmentation that the method survives bond-breaking and internal rotation. Code/data availability is also a condition, as the reader noted, since the paper is otherwise hard to verify independently.","tokens_in":22112,"tokens_out":1712,"duration_ms":15711,"concrete_test":"Take the trained KerNNns hydrogen oxalate model and the published MP2 data, and run a controlled set of 50 short NVE trajectories initialized at 600 K (above the reported 300 K safe regime) with the O-H bond broken or with C-C dihedral rotated beyond the training distribution. Count the fraction of trajectories that complete 5 ps without energy blow-up or atomic overlap, and compare the potential energy drift against reference MP2 energies sampled along the same trajectories. If the breakdown rate is high or energy errors exceed the 0.03 kcal/mol test-set scale, the abstract's generalized extrapolation claim is not supported and should be limited to the tested bonded/topology-preserving regimes.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central strong claim is that kernels as features improve extrapolation \"far beyond\" training coverage, solving a general problem of NN-based PESs. The mechanism offered is that k[3,3](r,r') decays smoothly and monotonically to zero for large r (Fig. S1, Eq. 9), so the NN input saturates. But saturation of inputs does not constrain the output of the subsequent softplus network: once all kernel features are near zero, the network output is essentially a constant bias term plus whatever the hidden layers produce from near-zero inputs, and nothing in the architecture forces that constant to be the correct dissociation energy. The H2CO evidence is supportive (Fig. 4), but it is one-dimensional C-H cuts plus a 5000 K sample set with a modest 130 kcal/mol range; this does not establish a general solution to extrapolation for arbitrary bond breaking or conformational change. The paper's own hydrogen oxalate section states that once the hydrogen bond breaks and rotation about C-C becomes possible, \"the trajectory breaks down in such situations\" because features are not symmetrized and the geometry leaves the training topology. That is precisely a failure in an extrapolation regime that the abstract's general claim would cover. Thus the extrapolation advantage is demonstrated only for cases where the reference structure remains a reasonable anchor and no new chemical topology appears; it is an empirical property of the trained models rather than a guaranteed consequence of decaying kernels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces KerNN, a neural-network potential energy surface (PES) model that uses one-dimensional reciprocal power reproducing kernels k[3,3] as fixed input features to a small fully connected network (two hidden layers, about 10^3 parameters). KerNN is applied to formaldehyde (H2CO), HeH2+, and hydrogen oxalate. The authors report that KerNN matches or exceeds the accuracy of PhysNet (which has about 10^6 parameters) and other baselines on held-out energies and forces, harmonic frequencies, DVR3D bound states for HeH2+, IR spectra from molecular dynamics, and hydrogen-transfer barriers and rates. They also claim that the kernel features improve extrapolation far beyond the training data, citing a 5000 K extrapolation set for H2CO, one-dimensional C-H dissociation scans with transfer learning to the experimental dissociation energy, and long-range HeH2+ scans. The paper concludes that considerably smaller and simpler ML models can be competitive, and that kernel features solve a general extrapolation problem of NN-based PESs.","tokens_in":22425,"tokens_out":5350,"duration_ms":47965,"significance":"If the central claim holds, KerNN is a significant contribution: it demonstrates that extremely compact NN architectures with tailored fixed features can provide state-of-the-art accuracy and orders-of-magnitude speedups, and it offers a possible route to controlling asymptotic behavior via kernel descriptors. The paper benefits from multiple independent validations: out-of-sample error statistics, harmonic frequencies, quantum bound-state calculations (DVR3D), classical IR spectra, and experimental comparison for HeH2+. The transfer-learning experiment using the experimental C-H dissociation energy is a useful demonstration. However, the extrapolation claim is more general than the evidence supports, and reproducibility is limited by the absence of released code/data and incomplete training hyperparameters. The paper is honest about the hydrogen oxalate trajectory breakdown, which partially undercuts the general extrapolation claim. Overall, the core modeling idea is sound and the empirical results are encouraging.","major_comments":[{"comment":"The claim that kernel features \"solve a general problem of NN-based PESs\" and guarantee controlled long-range behavior is not supported by the mechanism or by the full set of experiments. Equation (9) shows that k[3,3](r,r') decays to zero as r grows, but this only ensures that the NN input vector tends to a fixed (zero) point; the subsequent softplus layers can in principle produce any output at that input, so the decay of the features does not by itself control the asymptotic energy. The paper's own hydrogen oxalate section states that when the hydrogen bond breaks and rotation about the C-C bond becomes possible, \"the trajectory breaks down in such situations\" because the features are not symmetrized. That is precisely an extrapolation failure for a geometry outside the training topology. The abstract and conclusion should be qualified to claim extrapolation within the same bonding topology (or for coordinates where the reference structure remains a valid anchor), and the hydrogen oxalate limitation should be prominently stated. Suggested concrete tests: systematic bond-breaking scans for more than one coordinate in a polyatomic molecule, or a demonstration that KerNN's output in the saturated-feature region is bounded by training data constraints.","section":"Abstract; Conclusion; H2CO section (Figure 4)"},{"comment":"The comparison to PhysNet and other baselines for H2CO is taken from Reference 12, and it is not clear that the training/validation/test splits and data sizes are identical to those used for KerNN. Because the parity claim (\"on par with PhysNet\") is central, the manuscript should either retrain the baselines on the identical splits (as is done for hydrogen oxalate) or provide a clear statement that the reference errors were obtained with the same protocol. Additionally, the five repeated KerNN runs are reported only as means in Figure 2; including standard deviations or individual points would allow assessment of model variability, which matters for a model with only about 10^3 parameters.","section":"H2CO section; Table S3; Figure 2"},{"comment":"The code and data are promised \"upon publication,\" but no public repository or version is provided in the preprint, and the training hyperparameters (learning rate, batch size, number of epochs, AMSGRAD beta values) are not reported. Since the paper's contributions are empirical and rely on training a small NN, this incomplete information prevents an independent validation of the numerical results. Please provide the repository link, a data statement, and a complete hyperparameter table in the revision.","section":"Data Availability; Methods"}],"minor_comments":[{"comment":"The sentence \"the training and inference of KerNN_s is computationally more demanding than for KerNN_s\" should read \"than for KerNN_ns.\"","section":"Conclusion"},{"comment":"In the data generation description, the temperature list \"100, 300, 500, 100, 1500, 2000 K\" appears to contain a typo; \"100\" should likely be \"1000\".","section":"Hydrogen Oxalate (Methods)"},{"comment":"The timing entries are difficult to parse because of mixed units and the bold-face markers for FORTRAN implementations; please reformat with explicit columns for each implementation and clear units.","section":"Table S5"},{"comment":"The loss function in Eq. (7) is missing the double vertical bars used for the norm in Eq. (6); please use consistent norm notation throughout.","section":"Equation (7)"},{"comment":"The label \"KerNN_s^TL\" is introduced in the caption but not explained in the main text at first use; please define the transfer-learning procedure explicitly when the figure is referenced.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a genuinely interesting and potentially useful method, and the multi-property validation is a strength. The main reason for major revision is the overstatement of the extrapolation claim in the abstract and conclusion, which is contradicted by the paper's own hydrogen oxalate observation; this needs qualification and ideally additional targeted tests. I also recommend asking the authors to make the H2CO baseline comparison protocol explicit and to provide code/data and full training details, as these are essential for a methods contribution in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this one: it is a clean, well-executed demonstration that a tiny feed-forward network with fixed 1D reproducing-kernel features can match a much larger, modern NN potential on accuracy while being orders of magnitude cheaper. On H2CO, HeH2+, and hydrogen oxalate, KerNN (roughly 10^3 parameters) lands on par with or better than PhysNet (10^6 parameters) on test-set errors, harmonic frequencies, DVR3D bound states, and IR spectra. That is a genuinely useful result for anyone building cheap ML potentials for dynamics.\n\nWhat is actually new is the specific hybrid: using k[3,3] kernel values relative to a reference structure as fixed features inside a small softplus network. That is a natural combination, but I do not see it in the cited prior work, and the paper shows it works. The benchmarking is solid: multiple independent evaluations, comparison against PhysNet, RKHS, and FCHL, and the HeH2+ bound-state comparison against an FCI/RKHS surface is a strong validation. They also disclose the hydrogen oxalate breakdown honestly, which tells me they are not hiding the rough edges.\n\nThe soft spot is the extrapolation claim. The abstract says kernels improve extrapolation “far beyond” training coverage and “solve a general problem” of NN PESs. The mechanism they offer is that the kernel features decay smoothly to zero, so the NN input saturates. That bounds the input, but it does not control what the network outputs once saturated; the output is just whatever the bias and hidden layers produce. The H2CO evidence is supportive, but it is mostly 1D cuts plus a 5000 K test set, and the paper itself reports that hydrogen oxalate trajectories “break down” once the hydrogen bond breaks and rotation about the C–C bond becomes possible. That is exactly an extrapolation regime the abstract would cover. So the extrapolation advantage is real for near-reference geometries but should be framed as empirical and system-dependent, not a guaranteed consequence of the descriptors.\n\nTwo smaller issues: code and data are not yet released (the GitHub link is for “upon publication”), and training hyperparameters like learning rate and epochs are incomplete, though Table S1 gives the architecture and loss weights. Neither is fatal.\n\nThis paper deserves a serious referee. It is a credible, parsimonious method with clear practical value for molecular simulation. I would recommend acceptance after the authors scope the extrapolation claims and ideally release the code. I would cite it if I work on lightweight ML potentials.","headline":"A small hybrid kernel/NN potential (KerNN, ~10^3 params) matches PhysNet on accuracy for the tested systems and is much faster; the extrapolation claim is real but should be scoped more carefully than the abstract does.","tokens_in":22926,"tokens_out":1565,"would_cite":true,"duration_ms":17085,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tiny neural networks match million-parameter molecular energy models","keywords":["potential energy surface","neural network","reproducing kernel","extrapolation","parsimony","molecular dynamics","infrared spectroscopy","reactive scattering"],"falsifier":"Train KerNN on a molecule that dissociates into ions, e.g., NaCl, where the energy approaches the Coulomb law -1/R rather than a constant; if the predicted potential levels off instead of following the 1/R curve for R beyond the training data, the claimed extrapolation advantage fails.","tokens_in":21935,"feed_emoji":"⚛️","tokens_out":8940,"duration_ms":73373,"temperature":0.7,"pith_summary":"The paper claims that accurate molecular potential energy surfaces do not require the huge neural networks currently in use. It introduces KerNN, which feeds one-dimensional reproducing kernel functions of interatomic distances into a small feed-forward network with about a thousand parameters. On formaldehyde, HeH2+, and hydrogen oxalate, this tiny network matches or beats PhysNet—a network with roughly a million parameters—on energies, forces, dipole moments, and computed spectra, and it extrapolates more reliably beyond the training data. If correct, this would make machine-learned molecular dynamics far cheaper and more robust, especially in dissociation and bond-breaking regions. The paper's central insight is that simpler, smaller models can match or exceed larger ones when the descriptors encode the right physics.","feed_headline":"Tiny neural networks match million-parameter molecular energy models","feed_subtitle":"Roughly 1,000 parameters match million-parameter accuracy, plus reliable extrapolation.","key_machinery":"The carrying object is the one-dimensional reciprocal-power reproducing kernel k[3,3](r, r'), a smooth similarity function of two interatomic distances that decays as $r^{{-4}}$ for large arguments. It converts raw distances into features that enter a two-hidden-layer feed-forward network with softplus activations; the kernel's monotonic decay toward zero is what gives the network a sensible long-range asymptote. A symmetrized variant built from fundamental invariants imposes permutational invariance for like atoms.","core_discovery":"The central discovery is that a neural-network potential energy surface can be built from about $10^{3}$ parameters without losing accuracy. KerNN uses one-dimensional reciprocal-power reproducing kernels k[3,3](r, r') as input features, each measuring the similarity between an interatomic distance and a reference distance, and a two-hidden-layer feed-forward network with softplus activations maps those features to the total energy. On the same ab initio reference data, this small network reaches test-set errors comparable to PhysNet's $10^{6}$-parameter model for energy, forces, and dipole moments across three very different molecules. The kernel features also improve extrapolation: because k[3,3] decays monotonically toward zero for large distances, the network is smoothly anchored as bonds are stretched far beyond the training range. This is demonstrated on 5000 K-sampled geometries, one-dimensional dissociation cuts, and a transfer-learned correction to the experimental dissociation energy of formaldehyde, and on spectroscopy and reactivity for HeH2+ and hydrogen oxalate.","pith_inferences":["A stronger version of the paper's claim would be that descriptor physics, not network capacity, sets the accuracy ceiling; this could be tested by ablating the kernel feature set against raw distances or exponentials while holding the architecture fixed.","The reliable long-range behavior holds for dissociations whose true asymptote is flat; for ionic or other slowly decaying long-range tails, the current k[3,3] features would saturate and the potential would plateau, so the extrapolation guarantee is specific to the tested systems.","The hydrogen oxalate breakdown after hydrogen-bond breaking suggests that for floppy, large-amplitude systems, unsymmetrized descriptors need data augmentation or approximate permutational invariance before the approach can be used without restrictions."],"forward_implications":["Training and evaluation costs drop by roughly two orders of magnitude for the systems tested, making long-time molecular dynamics simulations with machine-learned potentials practical.","Because the kernel features decay monotonically outside the training range, the potential remains physical in dissociation and bond-breaking regions, where typical neural-network potentials drift.","The same tiny network can be adapted to predict dipole moments and infrared spectra, so accelerated simulations feed directly into spectroscopic observables.","Unlike kernel ridge regression, KerNN's evaluation cost does not grow with the number of training points, so it can scale to larger reference data sets."],"supporting_citations":[{"why":"Defines the reciprocal-power reproducing kernels whose one-dimensional form becomes KerNN's input features.","marker":"[64]"},{"why":"Establishes the long-range behavior of RKHS potentials, supporting the claim that the decaying kernel controls the far-field asymptote.","marker":"[67]"},{"why":"PhysNet, the large neural-network baseline that KerNN matches with far fewer parameters.","marker":"[16]"},{"why":"Supplies the H2CO reference data, learning curves, extrapolation set, and prior comparisons to PhysNet, RKHS+F, and FCHL.","marker":"[12]"},{"why":"Permissionally invariant RKHS approach used as a comparison and as a source of the RKHS feature concept.","marker":"[17]"},{"why":"Provides fundamental invariants used to construct the permutationally symmetrized descriptor Ds.","marker":"[59]"},{"why":"DVR3D program used to compute ro-vibrational bound states on the KerNN HeH2+ PES.","marker":"[29]"},{"why":"FCI/RKHS HeH2+ PES whose bound-state energies serve as the reference for validation.","marker":"[30]"}],"fun_headline_variants":["Minimal neural net matches million-parameter energy models","1000 parameters rival million-parameter molecular models","Kernel features extend tiny neural network's reach","Smaller is better: neural PES with 1000 parameters","Extrapolation beyond training: tiny neural network does it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that a feed-forward network whose kernel features have saturated at zero will still produce physical energies outside the training range; this is an empirical property of the trained models, not a mathematical guarantee, so the extrapolation claim could fail for a different molecule or training set.","fun_headline_variants_meta":{"raw":{"variants":["Minimal neural net matches million-parameter energy models","1000 parameters rival million-parameter molecular models","Kernel features extend tiny neural network's reach","Smaller is better: neural PES with 1000 parameters","Extrapolation beyond training: tiny neural network does it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001057,"raw_usage":{"total_tokens":4417,"prompt_tokens":907,"completion_tokens":3510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":3432}},"tokens_in":523,"tokens_out":3510,"duration_ms":22587,"temperature":1.0,"reasoning_tokens":3432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:30:17.008189+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train KerNN on a molecule that dissociates into ions, e.g., NaCl, where the energy approaches the Coulomb law -1/R rather than a constant; if the predicted potential levels off instead of following the 1/R curve for R beyond the training data, the claimed extrapolation advantage fails.","supporting_citations":[{"cited_title":"A general method for constructing multidimensional molecular potential energy surfaces from ab initio calculations","cited_arxiv_id":null,"evidence_quote":"Defines the reciprocal-power reproducing kernels whose one-dimensional form becomes KerNN's input features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the long-range behavior of RKHS potentials, supporting the claim that the decaying kernel controls the far-field asymptote."},{"cited_title":"S.; von Lilienfeld, O","cited_arxiv_id":null,"evidence_quote":"Supplies the H2CO reference data, learning curves, extrapolation set, and prior comparisons to PhysNet, RKHS+F, and FCHL."},{"cited_title":"Permutationally invariant, reproducing kernel-based potential energy surfaces for polyatomic molecules: From formaldehyde to acetone","cited_arxiv_id":null,"evidence_quote":"Permissionally invariant RKHS approach used as a comparison and as a source of the RKHS feature concept."},{"cited_title":"A.; Barletta, P.; Harris, G","cited_arxiv_id":null,"evidence_quote":"DVR3D program used to compute ro-vibrational bound states on the KerNN HeH2+ PES."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FCI/RKHS HeH2+ PES whose bound-state energies serve as the reference for validation."}],"review_version":1}