{"id":"e993e91d-0a14-41a2-9eda-bb6b6cc0ab3c","arxiv_id":"2505.12140","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Deep Potential model trained on ab initio molecular dynamics reproduces elastic, fracture, and vibrational properties of graphene in large-scale simulations.","lead":"Researchers trained a machine-learning model called a Deep Potential on small quantum simulations of graphene, then used it in large molecular dynamics runs to reproduce stiffness, fracture, and vibration properties. The work offers a practical route to near-quantum-accuracy simulations of large graphene sheets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fracture/stress–strain transferability is under-supported: training data already contain tearing, and large-sheet stress–strain validation is qualitative, so the ab initio-level precision claim is overextended.","rationale":"The reader correctly identified the small, short LDA AIMD trajectories and the trained-in tearing as the weakest premise. I agree and sharpen the concern: the missing element is a direct, held-out DFT benchmark in the deformed and fractured regime. The near-equilibrium properties are likely adequate, and the public release of data and code is a genuine strength that makes the proposed test feasible. However, the existing validation for mechanical behavior consists of a few elastic constants from the linear regime (Table 2) plus qualitative stress–strain trends (Figure 4), while the signature fracture/LAC observation is inside the training distribution. The vibrational comparisons are also indirect, with VDOS used as a Raman proxy and the ZA mode serving as a symmetry-expected check rather than a quantitative accuracy test. Therefore the central claim should remain conditional until the model is benchmarked against fresh DFT data at a nontrivial held-out cell size under deformation; the paper should also report quantitative fracture stresses and clarify which reported observations are predictions rather than training-set reproductions.","tokens_in":13465,"tokens_out":11096,"duration_ms":126074,"concrete_test":"Compute, with the published MLIP and fresh GPAW-LDA calculations under identical conditions, a quasi-static uniaxial stress–strain curve to fracture for a held-out 64-carbon cell at 300 K (straining in 0.5% increments with relaxation, or equivalent short AIMD at 10^-4 fs^-1). Compare ultimate tensile strength, fracture strain, and whether the post-fracture topology contains LAC chains. If the MLIP stress–strain curve deviates from AIMD by more than about 10% in strength or strain, or produces a qualitatively different fracture topology, the central transferability claim for large-sheet deformation is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the MLIP reproduces stress–strain behavior and fracture, including LAC chain formation, with ab initio-level precision on a 9072-atom sheet. The load-bearing support is weakest exactly in that regime. All AIMD training data come from 4- to 32-atom cells with 0.5–5 ps trajectories (Methods, Section 2; Table S1), and the NPT runs at strain rates up to 10^-2 fs^-1 already exhibited tearing. Thus the large-sheet LAC chains in Figure 5 are expected from the fitted data, not an independent emergent prediction. More importantly, the stress–strain curves in Figure 4 are validated only qualitatively against temperature/direction trends (Ref. 42); no fracture stress, fracture strain, or post-yield path is compared with DFT or experiment. Table 2 validates only the linear elastic regime (C11, C21, ν, Ym), which is near-equilibrium interpolation, not the fracture behavior highlighted in the abstract. The VDOS results are an indirect proxy for Raman spectra, and the quadratic ZA mode is largely a symmetry-dictated feature, so it is not strong evidence of accuracy. Consequently, the assertion of ab initio-level precision for large-scale mechanical deformation is not established by the reported evidence, and the model may be reproducing small-cell, high-strain-rate artifacts rather than transferable fracture physics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a DeepMD machine-learning interatomic potential for graphene trained on LDA AIMD trajectories of 4-, 8-, 16-, and 32-atom cells under NVT, NPT, and NVE conditions, including strained and torn configurations. The authors report RMSE validation of energies, forces, and virial stress; linear-scaling LAMMPS performance; stress-strain curves for a 9072-atom sheet at 1 K and 300 K that show temperature- and direction-dependent fracture; the appearance of linear acetylenic carbon (LAC) chains upon tearing; elastic constants and derived Young's modulus and Poisson's ratio; phonon dispersion; and strain-dependent vibrational density of states. The stated claims are that the potential reproduces mechanical and vibrational properties with ab initio-level precision and can serve as a transferable surrogate for DFT in large-scale graphene molecular dynamics.","tokens_in":13719,"tokens_out":4381,"duration_ms":47054,"significance":"If fully supported, the potential would be a useful open resource for graphene MD: the elastic constants are in reasonable agreement with several DFT and experimental references, the phonon dispersion has the expected qualitative structure, and the public release of training data and model parameters on NOMAD and GitHub is a clear strength that supports reproducibility. The linear-scaling benchmark is also a useful practical result. However, the load-bearing evidence for the fracture and transferability claims is incomplete: the training set already contains torn high-strain configurations and the loss function includes virial stress, so the 'emergent' LAC-chain and fracture behavior is at least partly fitted rather than independently predicted; the stress-strain validation is qualitative; and the vibrational validation is largely visual. The contribution is real, but the claimed scope exceeds what the evidence currently establishes.","major_comments":[{"comment":"The training dataset already includes torn graphene structures: the authors state 'During these latter runs, tearing of the graphene structures was observed' for NPT runs with strain rates up to 10^-2 fs^-1, and the 32-atom entries in Table S1 include 5x10^-3 and 10^-2 fs^-1 runs. Since the loss function also fits the virial stress, the fracture behavior and LAC chains in Figures 4 and 5 are at least in part direct reproduction of training configurations, not emergent predictions. The abstract's wording 'emergence of linear acetylenic carbon chains upon tearing' and the conclusion's characterization of LAC formation as 'an emergent phenomenon' are therefore not supported unless a test is provided, e.g., retraining without tearing frames or a quantitative comparison of the large-sheet fracture path with AIMD data.","section":"Methods, Section 2; Supporting Information Table S1"},{"comment":"The stress-strain curves are discussed only qualitatively in terms of temperature and direction trends relative to Ref. 42. No quantitative comparison of fracture stress, fracture strain, or post-yield stress-strain path with DFT or experiment is provided, despite the abstract claiming 'ab initio-level precision' for mechanical deformation. Please add quantitative benchmarks for the 9072-atom sheet, including ideal strength and fracture strain at the simulated strain rates, with appropriate error estimates and a discussion of strain-rate effects.","section":"Section 3, Figure 4"},{"comment":"The elastic validation covers only the near-equilibrium linear regime, which is interpolation from the small-strain AIMD training data. The reported C11 and C21 values differ from the cited MD/MM results by large margins (e.g., C21 differs by roughly 20-46% from the AMBER, Tersoff, and MM values), while only Young's modulus and Poisson's ratio agree well with experiment and DFT. This table does not validate the fracture behavior highlighted in the abstract, and the phrase 'ab initio-level precision' should be restricted to the properties actually benchmarked unless additional anharmonic/fracture validation is provided.","section":"Section 3, Table 2"},{"comment":"The 'excellent agreement' with experimental phonon data is based on visual comparison, and the quadratic ZA mode near the Gamma point is a symmetry-dictated feature of a 2D membrane rather than a stringent test of the potential. Similarly, the VDOS peak shifts are compared qualitatively to Raman literature; the reported in-plane red shift of about 100 cm^-1 at 3% strain appears substantially larger than typical experimental G-band shifts for uniaxial strain, so a quantitative comparison with experimental or DFT-computed strain coefficients is needed. Please report quantitative phonon-frequency errors at high-symmetry points and a numerical comparison of strain-dependent VDOS peak positions.","section":"Section 3, Figures 6 and 7"}],"minor_comments":[{"comment":"There is a typo in the first paragraph: 'sytems' should be 'systems'.","section":"Introduction"},{"comment":"The phrase 'a total of 108 steps' should likely read '10^8 steps'; please use standard mathematical formatting for the learning schedule.","section":"Methods, MLIP training paragraph"},{"comment":"The switching function definition is confusing because the first branch also contains 1/r_ij and the variable r is not defined relative to r_ij; please clarify the notation and ensure the switching function is continuous at r_s and r_c.","section":"Methods, Eq. (2)"},{"comment":"Figure 5 colors atoms by von Mises stress but no color bar or scale is provided, making the snapshots difficult to interpret quantitatively.","section":"Figure 5"},{"comment":"The stress-strain figure caption does not state the units or the axis labels; please include them explicitly and note which curve corresponds to which orientation.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serviceable, openly released DeepMD potential for graphene, and the elastic/phonon validations are honest and mostly fine. But the abstract oversells 'ab initio-level precision' for fracture. The stress-strain and LAC-chain behavior are trained-in, not emergent predictions, so the headline claim should be softened and supported with quantitative benchmarks.\n\nWhat is actually new: a fresh AIMD dataset (4-32 C atoms, NVT/NPT/NVE, strained and torn frames) and a trained DeepMD potential with a local embedding descriptor. The linear scaling test is real, the elastic constants in Table 2 align reasonably with experiment/DFT, the phonon dispersion looks plausible, and the VDOS redshift under uniaxial strain matches the known experimental trend. The authors also made the training data and model publicly available, which is a concrete plus for reproducibility.\n\nWhere the soft spots are: the main one is transferability. The Methods state that tearing was already observed in the NPT training runs, so the large-sheet LAC chains in Figure 5 are learned interpolation from small, high-strain-rate cells, not an emergent prediction. The stress-strain curves are only checked qualitatively against temperature/direction trends from Ref. 42; no fracture stress, fracture strain, or post-yield path is compared to DFT or experiment. Table 2 covers only the linear elastic regime, which is near-equilibrium interpolation. The short trajectories (0.5-5 ps) and strain rates up to 1e-2 fs^-1 make it hard to claim the model transfers to quasi-static deformation of a 9072-atom sheet. The paper should benchmark against the prior graphene MLIPs it cites (Rowe et al., Singh & Li) and give error bars or repeat runs. Two minor points: the quadratic ZA mode is symmetry-dictated, so it is weak evidence; and the VDOS-Raman link is indirect. Also, the 'first comprehensive MLIP' claim is contradicted by refs 15 and 16, which already treated mechanical and thermal properties.\n\nFor a reader who wants a freely available reactive graphene MLIP for production MD, this is a useful asset. It deserves serious peer review and likely publication after revision: tone down the abstract, add quantitative fracture benchmarks, compare against prior MLIPs, and report uncertainties.","headline":"A useful, openly released DeepMD graphene potential, but the headline 'ab initio-level precision' for fracture outstrips the evidence since the tearing and stress–strain behavior are trained-in rather than independently predicted.","tokens_in":14285,"tokens_out":3112,"would_cite":false,"duration_ms":32108,"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":"A machine-learned potential trained on tiny ab initio cells reproduces fracture, elasticity, and phonons in large graphene sheets.","keywords":["machine learning interatomic potential","graphene","molecular dynamics","stress-strain","elastic constants","phonon dispersion","vibrational density of states","fracture"],"falsifier":"Retrain the exact same architecture on the same AIMD dataset with every NPT frame that shows torn edges or carbon-chain fragments removed, then deform a pristine 9,072-atom sheet at 300 K. If the linear acetylenic chains no longer appear, the paper's claim that they emerge during tearing is actually a memory of the training data; if they still appear, the potential is generating new reactive configurations beyond its training set.","tokens_in":13251,"feed_emoji":"🧪","tokens_out":10169,"duration_ms":100800,"temperature":0.7,"pith_summary":"This paper claims that a machine-learning interatomic potential (MLIP) trained exclusively on ab initio molecular dynamics of tiny graphene cells (4 to 32 atoms) can replace density functional theory in large-scale molecular dynamics of graphene. The potential reproduces the stress-strain response up to fracture, the elastic constants, the phonon dispersion including the quadratic ZA mode, and the vibrational density of states, and it does so at computational cost that grows linearly with the number of atoms. If true, this gives a drop-in surrogate for DFT in simulations of roughly 9,000-atom graphene sheets, enabling studies of fracture, strain engineering, and phonon-limited transport that would otherwise be too expensive. The paper also claims the trained potential is reactive: it captures temperature-dependent tearing and the appearance of linear acetylenic carbon chains after fracture.","feed_headline":"Graphene force field matches DFT stress, fracture, and phonons","feed_subtitle":"One trained potential handles fracture, elastic constants, and vibrational spectra in 9,000-atom simulations.","key_machinery":"The load-bearing object is a deep neural-network interatomic potential whose total energy is a sum of per-atom energies, each computed from a local environment descriptor. For atom $i$, neighbors inside a cutoff radius $r_c = 8$ Å are weighted by a switching function $s_{ij}$ that smoothly goes to zero at the cutoff, forming a coordinate matrix $R_i$; an embedding network turns $R_i$ into descriptor $D_i = G_i^T R_i R_i^T G_i$, and a second network maps it to $E_i$. Forces and the virial tensor are obtained by differentiating the total energy with respect to atomic positions and cell deformations, so the model can be trained against energy, force, and stress labels simultaneously. This construction is what lets a potential trained on small strained cells carry over to slow fracture and thermal vibrations in large sheets.","core_discovery":"The paper's central discovery is that a deep neural-network potential with local environment descriptors can serve as a reactive surrogate for density functional theory across both mechanical and vibrational response of graphene. Trained on AIMD frames from 4-, 8-, 16-, and 32-atom cells in NVT, NPT, and NVE ensembles with strain rates from $10^{-6}$ to $10^{-2}$ fs$^{-1}$, the model reproduces, in a 9,072-atom sheet, the stress-strain curves at 1 K and 300 K; the elastic constants ($C_{11} = 1.093$ TPa, $C_{21} = 0.199$ TPa, shear modulus $0.446$ TPa, Poisson's ratio $0.183$, Young's modulus $1.056$ TPa); the phonon dispersion with a quadratic ZA branch; and the vibrational density of states with strain-induced red shifts in-plane and a blue shift out-of-plane. It also reports the appearance of linear acetylenic carbon chains in the torn regions after fracture, at all studied temperatures and more prominently under $x$-direction strain.","pith_inferences":["Inference: because the training data are LDA-based, the model's quantitative elastic constants plausibly inherit LDA's well-known tendency to overbind; retraining on a different exchange-correlation functional would show how much of the DFT agreement is functional-specific.","Inference: the linear acetylenic chains seen in the large sheet may be replaying torn configurations already present in the small-cell NPT training runs; deleting those frames and retraining is a clean test of whether the chains are emergent.","Inference: the same descriptor-plus-loss recipe is portable to other two-dimensional materials, but portability is not guaranteed by architecture alone; the training set must include the same breadth of strain rates, ensembles, and near-fracture geometries.","Inference: the VDOS is computed from only 1 ps of velocity autocorrelation, so low-frequency features near the ZA mode are likely under-resolved; longer correlation windows would sharpen them and test the model in precisely the regime its training data sample least."],"forward_implications":["A single potential can now replace separate classical force fields for strained and vibrating graphene, so simulations that need both mechanical and thermal response no longer have to switch models.","Fracture studies on graphene sheets of thousands of atoms become feasible at DFT-level fidelity, including crack-tip chemistry and the formation of carbon-chain products.","The strain-dependent shifts in the vibrational spectrum predicted by the model line up with Raman experiments, so the potential can guide strain-engineering experiments before they are run.","Because the potential reproduces both 1 K and 300 K failure, temperature-dependent strength and ductility can be studied in large sheets without expensive ab initio molecular dynamics.","The described training workflow, with data and parameters made public, gives other groups a template for building similar reactive potentials for new two-dimensional materials."],"supporting_citations":[{"why":"Provides the software implementation and training workflow for the deep potential model used in this work.","marker":"[17]"},{"why":"Defines the deep-potential energy model as a sum of atomic energies with forces and virial obtained by differentiation.","marker":"[18]"},{"why":"Supplies the local frame two-body embedding descriptor that makes the potential smooth and symmetry-correct.","marker":"[29]"},{"why":"The molecular dynamics engine used to run the stress-strain, phonon, and VDOS production simulations.","marker":"[19]"},{"why":"Earlier MD force-field results that the MLIP aims to improve on, used as a comparison for elastic constants.","marker":"[9]"},{"why":"DFT calculation of ideal strength and phonon stability under tension, used as a benchmark for the MLIP's mechanical and vibrational predictions.","marker":"[51]"},{"why":"Reference experimental data for Poisson's ratio and Young's modulus used to assess the MLIP's elastic properties.","marker":"[50]"},{"why":"Experimental phonon data from high-resolution electron energy-loss spectroscopy used to benchmark the MLIP phonon dispersion.","marker":"[54]"},{"why":"Baseline study of temperature- and strain-rate-dependent fracture strength of graphene, the reference for the 1 K versus 300 K behavior.","marker":"[42]"}],"fun_headline_variants":["One ML potential nails graphene's stress, fracture, phonons","Graphene MLIP: stress-strain, phonons, and fracture in one","Reactive MLIP captures graphene's mechanics and vibrations","Graphene potential: from elastic constants to tearing","DFT-grade graphene potential scales to 9,000 atoms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's transferability stands or falls on the premise that a few short, sub-5-picosecond LDA trajectories of 4- to 32-atom cells, many of which are strained until they tear, cover the same potential energy surface that a 9,072-atom sheet explores during slow fracture and vibration.","fun_headline_variants_meta":{"raw":{"variants":["One ML potential nails graphene's stress, fracture, phonons","Graphene MLIP: stress-strain, phonons, and fracture in one","Reactive MLIP captures graphene's mechanics and vibrations","Graphene potential: from elastic constants to tearing","DFT-grade graphene potential scales to 9,000 atoms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00096,"raw_usage":{"total_tokens":4100,"prompt_tokens":967,"completion_tokens":3133,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":3046}},"tokens_in":583,"tokens_out":3133,"duration_ms":18656,"temperature":1.0,"reasoning_tokens":3046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:40:37.376094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the exact same architecture on the same AIMD dataset with every NPT frame that shows torn edges or carbon-chain fragments removed, then deform a pristine 9,072-atom sheet at 300 K. If the linear acetylenic chains no longer appear, the paper's claim that they emerge during tearing is actually a memory of the training data; if they still appear, the potential is generating new reactive configurations beyond its training set.","supporting_citations":[{"cited_title":"L.; Cai, C.; Lin, Y.; Wang, B.; Xu, J.; Zhu, J.-X.; Luo, C.; Zhang, Y.; Goodall, R","cited_arxiv_id":null,"evidence_quote":"Provides the software implementation and training workflow for the deep potential model used in this work."},{"cited_title":"Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics","cited_arxiv_id":null,"evidence_quote":"Defines the deep-potential energy model as a sum of atomic energies with forces and virial obtained by differentiation."},{"cited_title":"End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the local frame two-body embedding descriptor that makes the potential smooth and symmetry-correct."},{"cited_title":"Characterizing mechanical properties of graphite using molecular dynamics simulation","cited_arxiv_id":null,"evidence_quote":"Earlier MD force-field results that the MLIP aims to improve on, used as a comparison for elastic constants."},{"cited_title":"Ab initio calculation of ideal strength and phonon instability of graphene under tension","cited_arxiv_id":null,"evidence_quote":"DFT calculation of ideal strength and phonon stability under tension, used as a benchmark for the MLIP's mechanical and vibrational predictions."},{"cited_title":"Numerical Investigation of the Fracture Mechanism of Defective Graphene Sheets","cited_arxiv_id":null,"evidence_quote":"Reference experimental data for Poisson's ratio and Young's modulus used to assess the MLIP's elastic properties."},{"cited_title":"Analysis of phonons in graphene sheets by means of HREELS measurement and ab initio calculation","cited_arxiv_id":null,"evidence_quote":"Experimental phonon data from high-resolution electron energy-loss spectroscopy used to benchmark the MLIP phonon dispersion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline study of temperature- and strain-rate-dependent fracture strength of graphene, the reference for the 1 K versus 300 K behavior."}],"review_version":1}