{"id":"8809c41d-f5d5-4745-b96a-33e8e0e926eb","arxiv_id":"2607.21779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A graph-fragmentation neural network that directly predicts CCSD-level nuclear force vectors reproduces the structural and vibrational features of machine-learned AIMD trajectories for H13O6+.","lead":"This paper trains neural networks to predict nuclear force vectors directly from fragment geometries, using principal axes of inertia as a covariant frame, and produces machine-learned AIMD for a solvated Zundel cation at fragment-based CCSD accuracy. It matters because it points toward coupled-cluster-quality molecular dynamics for fluxional water clusters without paying full correlated-electronic-structure cost at every step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed CCSD accuracy is only validated against a fragment-based reference; the paper's own graph error (0.056 mEh/Bohr) exceeds the claimed 10 µEh/Bohr accuracy.","rationale":"The most load-bearing condition for the paper's central claim is that the fragmentation approximation, especially the dynamic rank-1 protocol used for the reference trajectory, provides forces sufficiently close to full-system CCSD. The paper never directly compares against full-system CCSD forces, and its own Table II places the graph error at 0.056 mEh/Bohr—above the claimed 10 µEh/Bohr. Since the ML errors at the dynamic cutoff equal the graph error, the ML cannot be said to achieve CCSD accuracy beyond the reference. This is not a challenge to internal consistency of the ML pipeline, which is well constructed and convincingly demonstrates reproduction of the fragment-based reference, but it directly undermines the headline accuracy statement. The reader's conditional verdict is appropriate; this concern specifies the exact validation needed. The test proposed—full-system CCSD gradients on sampled geometries—would settle the question definitively. I agree with the reader's identification of the fragment-based reference as the weakest assumption.","tokens_in":24030,"tokens_out":9200,"duration_ms":89314,"concrete_test":"Compute full-system CCSD/6-31+G(d,p) gradients for ~50 geometries sampled from the 9326-frame reference trajectory (or from the ML trajectory). Compare (i) the graph-assembled forces from Eq. (10) with the dynamic rank-1 protocol, (ii) the ML-predicted forces, and (iii) the 'All ES' reference forces that served as training labels, against these exact CCSD gradients using the metric in Eq. (24). If the per-component MAE of (i) and (ii) relative to exact CCSD is ≲0.01 mEh/Bohr, the 10 micro claim is supported; if it is ~0.056 mEh/Bohr, the accuracy claim must be restated as 'relative to a fragment-based CCSD reference'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Introduction: '10 micro-Hartree/Bohr accuracy with respect to CCSD') rests on the graph-fragmented reference being a quantitatively accurate proxy for full-system CCSD. The reference trajectory ('All ES') and all ML training labels are generated from fragment CCSD/B3LYP calculations via Eq. (10), with R=1 and a dynamic edge cutoff; no full-system CCSD force for H13O6+ is reported anywhere in the paper. Table II shows the force error of this dynamic rank-1 graph relative to the (also fragment-based) reference is 0.056 mEh/Bohr = 56 µEh/Bohr, an order of magnitude larger than the claimed 10 µEh/Bohr. Full-system ML force errors at this cutoff are nearly identical (0.056–0.057 mEh/Bohr for 10%, 20%, and 40% training), indicating the ML model reproduces the approximate reference without correcting it. Thus the total deviation of the final forces from true CCSD is at least the fragmentation error, and the 'coupled cluster accuracy' label is not established by the data presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a machine-learning framework for direct prediction of post-Hartree-Fock nuclear forces aimed at enabling coupled-cluster-level AIMD. The method combines graph-theoretic fragmentation with fragment-level CCSD/B3LYP corrections, a principal-axis-of-inertia covariant descriptor, and vector-valued neural networks that predict force components directly instead of differentiating a learned energy surface. Training configurations are selected by mini-batch k-means tessellation. The approach is validated on the solvated Zundel cation H13O6+, using a 9326-frame reference trajectory generated from graph-fragmented CCSD/B3LYP at rank R=1. Reported fragment-level force errors are below 0.02 mEh/Bohr for most fragment types, full-system errors for the production dynamic rank-1 protocol are about 0.056 mEh/Bohr, and ML-generated trajectories are compared with the reference through radial distribution functions and velocity autocorrelation spectra.","tokens_in":24360,"tokens_out":8796,"duration_ms":89683,"significance":"If the central accuracy claim were fully supported, the combination of graph-theoretic fragmentation and direct vector-valued force learning would be a practically useful step toward correlated-level AIMD for medium-sized water clusters, with a notable reduction in trainable parameters. The paper has clear strengths: explicit vector-valued force training with shared basis functions, the k-means-based space tessellation for data selection, and systematic error decomposition by fragment type, rank, and geometry class. However, the headline 'CCSD accuracy' is not yet established relative to full-system CCSD, and several validation details—train/test separation, trajectory agreement metrics, and the handling of principal-axis sign degeneracy—are missing. These issues affect the paper's main claims but are addressable with additional analysis and reframing.","major_comments":[{"comment":"The central claim of '10 micro-Hartree/Bohr accuracy with respect to CCSD' is not supported by the reported comparisons. Table II gives the dynamic rank-1 graph error for ∂∆E/∂x as 0.056 mEh/Bohr (56 µEh/Bohr), an order of magnitude larger than 10 µEh/Bohr. The full-system ML errors at the same cutoff are 0.056–0.057 mEh/Bohr for all training fractions, meaning the ML model reproduces the graph-fragmented reference but does not correct its error. If the 'exact value' axis in Fig. 7 is a full-system CCSD/B3LYP gradient, the total ML force deviation from CCSD is at least the graph error; if it is not, the manuscript never validates against a full-system CCSD force. Please state the exact reference used, report the total error against that reference, and revise the abstract/Introduction claims accordingly.","section":"Introduction; §IV.C, Table II"},{"comment":"No train/test split is described for Eq. (24). The text says models are trained on k-means-selected 10% of fragment geometries and then errors are computed over NConfig fragments, but it is not stated whether the evaluation set is disjoint from the training set. If the same configurations are used for training and evaluation, the errors in Table I and Figures 5–6 are in-sample and optimistic. Also, Table I is internally inconsistent: H2O has NConfig=8910 and training size 3891 (~44%), whereas all other fragment types use ~10%. Please clarify the data-selection procedure, report held-out errors, and correct the table.","section":"§IV.B, Eq. (24); Table I"},{"comment":"The abstract claims that using only 10–20% of reference configurations yields successful fully predicted trajectories, but Section V.A reports that the 10%-training trajectory shrinks the second RDF peak near 2.7 Å and broadens the distribution near 4 Å, and that only the 40%-training trajectory aligns well with the reference. This directly qualifies the headline efficiency claim. Please provide quantitative agreement measures (e.g., integrated absolute differences or peak positions/heights) for the 10%, 20%, and 40% models and state which training fraction supports each claim.","section":"§V.A, Figures 10–11; Abstract"},{"comment":"The principal-axis descriptor is central to the covariant force representation, but the manuscript does not specify how the arbitrary signs of inertia eigenvectors and possible axis degeneracy (near-symmetric fragments) are handled. Without a canonical sign convention or a treatment of axis switching, the fragment-fixed frame can flip discontinuously between nearby configurations, producing discontinuous force predictions and trajectories. Please document the sign/order resolution and assess its effect on force continuity, especially for the fluxional water fragments studied here.","section":"§III.A, steps 1–6"},{"comment":"Because forces are trained directly as vector outputs, the predicted force field is not guaranteed to be conservative (curl-free). The paper does not report energy conservation along the ML-driven trajectories, which is a standard check for AIMD. For NVE simulations, non-conservative forces can lead to systematic drift and biased distributions over longer times, directly relevant to the 'long-timescale' goal. Please report total-energy drift for the predicted trajectories or explain why non-conservativity is not a concern at the simulated time scales.","section":"§III.B, Eq. (22); §V"}],"minor_comments":[{"comment":"Typos and unclear wording: 'machine trajectories' (Section V opener), 'distraibution' (Fig. 10 caption), 'calcualtions' (p.10), 'aa manifested' (p.8), and 'th-milli-Hartree/Bohr' (Section IV.B) are confusing.","section":"Various"},{"comment":"The caption is truncated: 'Errors in fragment forces as per' appears to be incomplete. Please provide a complete caption including the reference for the error.","section":"Table II"},{"comment":"References [37] and [44] appear to be the same paper (Zhu and Iyengar, Phys. Rev. X 2026, 16, 011012). Please consolidate or distinguish them.","section":"References"},{"comment":"The statement that errors below 0.02 mEh/Bohr are 'well within' 10 micro-Hartree/Bohr is arithmetically inconsistent (0.02 mEh/Bohr = 20 µEh/Bohr). Please correct the convergence-criterion comparison or the reported accuracy claim.","section":"§IV.B"},{"comment":"Details of the ML-driven trajectories (integration scheme, time step, thermostat if any, length) are not given, which hinders reproducibility of the AIMD validation.","section":"§V"}],"recommendation":"major_revision","confidential_remarks":"I do not recommend rejection: the methodological core—direct vector-valued force training combined with graph-based fragmentation—is promising, and the reported trajectory comparisons, while incomplete, are not implausible. The main fixes are to reframe the 'CCSD accuracy' claim, add a small set of full-system CCSD gradient checks, and provide a proper train/test split and quantitative trajectory-agreement metrics. With those additions the paper could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method part is worth taking seriously. Bypassing automatic differentiation of a learned energy surface and predicting force vectors directly, using fragment-fixed principal-axis frames for covariance, is a genuine extension of the authors' graph-fragmentation program. The parameter-efficiency result is concrete: vector-valued training shares the hidden basis across force components and buys an order of magnitude fewer weights at comparable error. The k-means tessellation of fragment geometries as a training-subset selector is also sensible, and the fragment-level ML errors (0.000–0.020 mEh/Bohr) are small. The fully predicted 40%-training trajectory reproduces the reference RDFs and the VDOS reasonably well, which is not nothing for a fluxional protonated water cluster.\n\nThe soft spot is the central claim. The '10 micro-Hartree/Bohr accuracy with respect to CCSD' in the introduction is not actually measured against full-system CCSD. The reference trajectory and all ML labels come from the graph-fragmented CCSD/B3LYP model of Eq. (10). Table II reports the dynamic rank-1 graph error itself as 0.056 mEh/Bohr = 56 µEh/Bohr, roughly five times the claimed accuracy. The full-system ML errors (0.056–0.057 mEh/Bohr) are essentially identical to the graph error, which tells you the ML is reproducing the approximate reference, not correcting it. So the total deviation from true CCSD is at least the fragmentation error, and the 'coupled cluster accuracy' phrasing overstates what is demonstrated. That is the main thing a referee should pressure.\n\nTwo smaller issues. First, Eq. (24) reports fragment force error, but the paper never states whether those numbers are computed on held-out fragments or on the training fragments themselves; the k-means selection makes the split ambiguous. Second, the abstract's '10% to 20% of reference configurations' framing is the wrong headline: the 10% trajectory visibly degrades (the second O–O peak broadens), and the 40% model is the one that matches. So the 'only 10-20%' claim doesn't align with the headline result.\n\nCitation pattern is heavy self-citation, but that's a natural consequence of building on the same fragmentation framework; the earlier energy papers are cited and the delta is clear. The 'CCSD accuracy' framing is load-bearing and overstated. But the underlying method is real and the paper deserves a serious referee, with a request for full-system CCSD validation (even a few gradients at representative frames), an explicit train/test split, and toned-down claims.","headline":"Genuine method advance in direct vector force learning, but the 'CCSD accuracy' claim is not established: validation uses the authors' own fragment reference, whose 56 µEh/Bohr error is larger than the claimed 10 µEh/Bohr.","tokens_in":24798,"tokens_out":3933,"would_cite":false,"duration_ms":36049,"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":"The paper claims that directly predicting nuclear force vectors on graph-theoretic fragments, instead of differentiating a learned energy surface, reproduces CCSD forces for the fluxional solvated Zundel cation H13O6+ well enough that fully","keywords":["machine-learned force fields","graph-theoretic fragmentation","coupled cluster (CCSD)","direct force vector learning","principal axes descriptors","ab initio molecular dynamics","solvated Zundel cation","k-means tessellation"],"falsifier":"Compute full-system CCSD (or CCSD(T)) nuclear gradients for a set of representative solvated-Zundel configurations, including geometries with stretched and shared protons, and compare them with the graph-fragmentation reference used here. If the per-component deviation exceeds roughly 10 micro-Hartree/Bohr, or if trajectories driven by the full-system gradients diverge from those driven by the fragmented reference, the central accuracy claim fails.","tokens_in":23947,"feed_emoji":"⚛️","tokens_out":5500,"duration_ms":56640,"temperature":0.7,"pith_summary":"This paper tries to establish that machine-learned molecular dynamics can run at coupled-cluster accuracy for systems where cheap density functionals are unreliable, using a fragment-based neural network that predicts force vectors directly rather than deriving them from a learned energy. The authors report force errors near 10 micro-Hartree/Bohr against their CCSD-derived reference and show that fully ML-predicted trajectories of the solvated Zundel cation reproduce key structural and dynamical signatures. The main contribution is the combination of covariant principal-axis force descriptors, vector-valued training that shares hidden features across force components, graph-theoretic inclusion-exclusion assembly of fragment corrections, and mini-batch k-means tessellation for training-set selection, which together stretch expensive reference data much farther than scalar force fitting.","feed_headline":"Fragment nets match CCSD forces for fluxional water cluster","feed_subtitle":"Direct force-vector learning plus graph fragmentation reproduce dynamics at a fraction of reference cost","key_machinery":"The central identity is the graph-theoretic fragmentation of post-Hartree-Fock forces, Eq. (13)/(22): full-system force corrections are assembled from fragment-level corrections weighted by inclusion-exclusion coefficients M^R derived from simplex counts, then mapped back through the Jacobian ∂x_fragment/∂x. The other key object is the covariant descriptor: each fragment is rotated so its force components align with its principal axes of inertia, producing three distance-matrix inputs and three vector outputs that preserve translational, rotational, and permutational symmetry. Vector-valued force networks share hidden-layer features across force components, cutting trainable parameters by ov","core_discovery":"On the authors' terms, the central discovery is that nuclear forces at CCSD quality can be learned as direct vector outputs of fragment neural networks using an interatomic-distance descriptor rotated into each fragment's principal moment-of-inertia frame and projected back after prediction. This covariant construction, combined with the graph-theoretic fragmentation sum (Eq. 13) and a mini-batch k-means tessellation that selects only 10-20% of reference frames, yields full-system force errors below 0.06 mEh/Bohr for H13O6+. The resulting fully ML-predicted AIMD trajectory reproduces O-O and O-H radial distributions and the velocity autocorrelation power spectrum of the reference, indicating","pith_inferences":["If the graph-fragmentation reference (Eq. 10) is a faithful surrogate for full CCSD on fluxional systems, the same direct-force protocol could be applied to other strongly polarized proton-transfer clusters by retraining only the relevant fragment networks.","The observed slow drift of the 10%-trained trajectory toward sparsely sampled regions suggests an active-learning loop—updating k-means centroids from the ML trajectory itself—as a natural next step, which the authors mention only as future work.","Because force components in each principal-axis block share hidden features, the method could in principle support transfer learning between fragments of different protonation states; the paper leaves this implicit.","A decisive external check would compare the graph-fragmented CCSD reference against full-system CCSD gradients at several off-equilibrium configurations; the paper does not perform this comparison, so the 'CCSD accuracy' label should be read as accuracy relative to its own fragmentation reference."],"forward_implications":["Coupled-cluster-quality AIMD becomes feasible for medium water clusters and similar fluxional systems that currently require DFT-level compromise.","Training cost for each fragment type is independent of full system size, so the approach scales to larger clusters by reusing fragment models.","Direct force learning sidesteps the link-atom Jacobian and multi-fragment coupling issues that arise when energy models are differentiated after covalent bond cleavage.","The k-means tessellation means 80-90% of expensive CCSD fragment reference calculations can be skipped without degrading trajectory statistics.","Because the reference trajectories themselves use graph-fragmented CCSD rather than full-system CCSD, the method is calibrated to the fragmentation approximation, so its practical accuracy inherits that approximation's fidelity."],"fun_headline_variants":["CCSD-grade forces from graph-fragmented nets","Direct force learning cracks fluxional water dynamics","Graph nets hit CCSD force accuracy at 20% of data","Covariant force nets reproduce fluxional spectra","Fragment nets beat back Hessian limits for AIMD"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the paper's graph-based recipe for cutting the molecule into overlapping fragments and stitching fragment CCSD forces back together exactly reproduces what a full CCSD calculation on the whole cluster would give; the paper never runs such a full calculation, so this remains untested.","fun_headline_variants_meta":{"raw":{"variants":["CCSD-grade forces from graph-fragmented nets","Direct force learning cracks fluxional water dynamics","Graph nets hit CCSD force accuracy at 20% of data","Covariant force nets reproduce fluxional spectra","Fragment nets beat back Hessian limits for AIMD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1147,"prompt_tokens":802,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":546,"tokens_out":345,"duration_ms":4079,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:42:44.790953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute full-system CCSD (or CCSD(T)) nuclear gradients for a set of representative solvated-Zundel configurations, including geometries with stretched and shared protons, and compare them with the graph-fragmentation reference used here. If the per-component deviation exceeds roughly 10 micro-Hartree/Bohr, or if trajectories driven by the full-system gradients diverge from those driven by the fragmented reference, the central accuracy claim fails.","supporting_citations":[],"review_version":1}