{"id":"aae05de0-0512-4699-b6fc-b01e8b5aa162","arxiv_id":"2602.16250","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Facet-resolved MD simulations of hcp Co nanoparticles show stepped {01¯11} facets premelt ~200 K before flat {0001} facets, with facet melting temperatures following a 2D Gibbs–Thomson scaling.","lead":"Molecular dynamics simulations of hcp cobalt nanoparticles show that surface premelting is facet-dependent: stepped {01¯11} facets disorder and melt about 200 K before flat {0001} facets, at temperatures as low as ~20% of the bulk melting point. The authors extend the Gibbs–Thomson relation to individual facets and propose a size-dependent critical liquid shell that triggers complete melting.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Facet 'melting' temperatures are classifier-based: atoms leaving a facet class due to surface diffusion are counted as melted, so the 205 K offset may be a mobility artifact, not thermodynamic premelting.","rationale":"The reader's weakest assumption is the physical fidelity of the self-cited q-SNAP potential, which is a legitimate concern. However, the paper reports that direct DFT-PBE surface energies closely match the q-SNAP values (2.13/2.40 vs 2.13/2.38 J/m2), so the qualitative facet ordering is not solely inherited from the ML potential. A more load-bearing and more internal weakness is that the facet 'melting' temperatures are not direct thermodynamic measurements but classifier outputs: an atom is scored as no longer belonging to a facet whenever its environment deviates from a low-temperature reference, which conflates surface mobility and vertex diffusion with melting. The paper's own description of the method admits that non-unimodal classes violate the key assumption of the Mahalanobis-distance approach. Because the central claim is precisely about which facet 'melts' first and by how much, this conflation threatens the quantitative headline even if the potential is perfect. The recommended concrete test would settle this by an independent order parameter. The verdict remains CONDITIONAL: the paper deserves publication only if the facet-specific analysis is shown to reflect genuine liquid-like disordering, not just mobility. This is a change in emphasis rather than a change in the reader's overall conditional assessment, so I mark UNCHANGED.","tokens_in":15446,"tokens_out":6178,"duration_ms":67379,"concrete_test":"Recompute the {01¯11} and {0001} facet characteristic temperatures using an independent, facet-resolved liquid-order parameter instead of the classifier: e.g., time-averaged Steinhardt q6 or a-CNA on atoms whose initial positions define each facet, with thresholds calibrated on bulk liquid at the same temperature. If the resulting Tc difference is not ~205 K and size-independent, the headline claim is a classifier artifact. A cheaper check: at the reported T80 for {01¯11} (e.g., 787 K for the 5333-atom NP), compute the fraction of atoms in the initial {01¯11} region that are liquid-like by q6; if this fraction is small (<20%), the facet is not premelted but merely mobile.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—{01¯11} facets melt ~205 K below {0001} regardless of size—rests on the facet-specific characteristic temperatures Tc derived from the fraction of atoms that remain assigned to each facet class under a Mahalanobis-distance classifier trained at 50–400 K (Fig. 7, Eq. 3). An atom is removed from its facet class as soon as its local environment deviates from the low-temperature reference; this happens not only upon melting but also upon surface diffusion, vertex migration, and thermal broadening. The paper itself states that a migrating vertex atom and nearby surface atoms are classified as outliers (Sec. II.A, Fig. 1B). Because stepped {01¯11} facets are more open and weakly coordinated, atoms there become mobile at lower temperatures; the classifier will therefore report an earlier 'melting' even if the facet remains crystalline. No independent facet-resolved order parameter (e.g., bond-order, Lindemann, or a-CNA) is used to validate that atoms leaving the {01¯11} class are actually liquid-like. The global melting point is validated against the heat-capacity maximum (Fig. 3), but the facet characteristic temperatures are not. Thus the 205 K offset and its size independence could be an artifact of classifying mobility as melting, independent of the fidelity of the q-SNAP potential.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports molecular dynamics simulations of hcp Co nanoparticles (587–5333 atoms per Table S1) heated to melting with a q-SNAP machine-learned potential, and uses a hierarchical GMM/Mahalanobis-distance classifier to label atoms as bulk, facet, edge, vertex, or outlier. The authors find that stepped {01-11} facets lose their crystalline class at lower temperatures than flat {0001} facets, with a reported offset of about 205 K that is claimed to be size-independent. They also fit global melting temperatures to a 3D Gibbs–Thomson relation with a bulk intercept of 1771 K, fit facet characteristic temperatures to a 2D Gibbs–Thomson relation, and propose a model for the critical outlier fraction τc based on a critical liquid-layer thickness lc that grows linearly with nanoparticle size. The central thesis is that premelting is facet-selective and anisotropic, contradicting the isotropic uniform-liquid-shell picture.","tokens_in":15638,"tokens_out":7519,"duration_ms":82314,"significance":"If the central claim is correct, the paper would be a valuable step beyond isotropic premelting models, showing facet-resolved surface melting in a metallic nanoparticle and a facet-resolved extension of the Gibbs–Thomson relation. The global melting analysis is well cross-checked: the structural melting points correlate with heat-capacity maxima (Fig. 3), and the 3D Gibbs–Thomson fit is clean, giving a bulk intercept of 1771 K close to the experimental Co value. The descriptor-based classification approach is also potentially useful for analyses where standard order parameters struggle. However, the main quantitative claim—the size-independent 205 K offset—rests on the facet-classifier response and is not independently validated against a facet-resolved order parameter. In addition, the τc model is partly circular, and several internal numerical inconsistencies weaken the presentation. The manuscript is therefore promising but needs substantial revision before the central claim can be accepted.","major_comments":[{"comment":"The facet 'melting' temperatures are defined from the fraction of atoms that remain assigned to a facet class under a Mahalanobis-distance classifier. The paper itself states (§II.A, Fig. 1B) that a migrating vertex atom and nearby surface atoms are classified as outliers, and that surface diffusion begins near 400 K on stepped facets. Thus, atoms leave the {01-11} class not only upon melting but also upon surface diffusion and thermal broadening. Since stepped facets are more open and lower-coordinated, they will lose class membership at lower temperatures even if they remain crystalline. No independent facet-resolved order parameter (bond-order, Lindemann, or a-CNA) is provided to show that the atoms leaving the {01-11} class are liquid-like. The 205 K offset and its size independence may therefore be a mobility artifact. Please add a facet-resolved structural order parameter or otherw","section":"§II.C, Fig. 7, Methods c"},{"comment":"The headline number is inconsistent. The abstract states that stepped {01-11} facets melt 'nearly 150 K below' flat {0001} facets, while the Introduction says 'nearly 200 Kelvin lower', §II.C reports a value of 205 K, and the Conclusion repeats 'nearly 200 Kelvin'. Since this offset is the central quantitative claim, the discrepancy must be corrected and the single value used consistently.","section":"Abstract vs. §II.C and Conclusion"},{"comment":"The claim that the difference between characteristic temperatures 'remains constant irrespective of the nanoparticle size, with a value of 205 K' is not supported by the tabulated data. From Table S1, Tc({0001}) − Tc({01-11}) equals 236, 210, 169, 219, 188, 201, and 210 K for the seven sizes. The spread is 169–236 K, i.e., about ±30 K around the mean and a 67 K range. This should be reported with uncertainties and the 'constant' claim reframed, e.g., as a weak or no size trend, rather than an invariant offset.","section":"§II.C, Table S1"},{"comment":"The τc 'model' is largely circular. lc is defined via Eq. (10) from the same τc and ns/N values it is intended to explain, then fitted linearly to N (Fig. 11), and finally inserted into Eq. (18) with a hand-tuned δ = 1.13 to 'reproduce' τc(N) in Fig. 5. Because lc is derived from τc, the agreement is not an independent test. The non-monotonic minimum near N ≈ 3000 follows algebraically from a linearly growing lc multiplied by N^{-1/3}, and is not a prediction. The authors already note that the linear lc(N) relation is not verified for N > 6000; this should be stated more prominently and the model presented as an empirical parametrization rather than a validated theory.","section":"§II.D, Eqs. (10)–(18), Figs. 5 and 11"},{"comment":"The 2D Gibbs–Thomson fit for the {0001} facet excludes the two smallest nanoparticles without explanation (shaded points in Fig. 9). The extrapolated infinite-facet temperature Tc,∞ = 1371 K and the resulting ratio Tc,∞/TM,∞ = 0.77 depend on this exclusion. Please justify the exclusion (e.g., finite-size crossover) or show that the fit is robust to including all points, or report the sensitivity of the intercept to this choice.","section":"§II.D, Fig. 9"}],"minor_comments":[{"comment":"The nanoparticle size range is inconsistent: the full-text abstract says '500 to 6000 atoms', the Introduction says '587 to 6847 atoms', and Table S1 lists sizes 587 to 5333. Please use one consistent range and ensure the figures and tables agree.","section":"Abstract and Introduction"},{"comment":"Typo: 'strong depnedency' should be 'strong dependence'.","section":"§II.B"},{"comment":"Typo: 'Distorsion' should be 'Distortion'.","section":"Fig. 13 caption"},{"comment":"The text states that for the 587-atom NP 'Tc ≈ T20' for {0001}, but Table S1 gives Tc = 1059 K and T20 = 1204 K. The intended statement appears to be that T20 ≈ TM (1204 vs 1220 K). Please correct.","section":"§II.C, Fig. 8"},{"comment":"The classifier is trained at temperatures up to 400 K, yet the text says that at 400 K vertex atoms begin to diffuse. If 400 K frames are included in training, diffusing atoms may contaminate the reference distributions. Please clarify whether such frames were excluded or justify their inclusion.","section":"Methods c"},{"comment":"The unimodal-Gaussian assumption for the Mahalanobis distance is not rigorously verified, as acknowledged in §II.D. This limitation should also be stated in the Conclusion, since it directly affects the interpretation of Tc as a melting temperature.","section":"Methods c, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper has a compelling qualitative observation and a clean global-melting analysis, but the central quantitative claim—the size-independent 205 K facet offset—is not yet established due to the classifier/mobility concern and the inconsistencies in the reported numbers. I do not think rejection is warranted; the authors can address the issues with additional facet-resolved order-parameter analysis, uncertainty quantification, and a re-framed τc model. The self-citation to the q-SNAP potential is appropriate, but the potential's near-melting surface behavior remains a residual risk."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"We should talk about this one. The paper does something genuinely useful: it shows that surface premelting in hcp Co nanoparticles is facet-resolved, not a uniform liquid shell, and that facet characteristic temperatures follow a 2D Gibbs–Thomson scaling with facet size. The observation is directly visible in snapshots, and the global melting points track heat-capacity maxima, which is a good internal check. The unsupervised GMM/Mahalanobis classifier is a nice contribution for analyzing nanoparticle surfaces under thermal noise.\n\nThe soft spots are real but not fatal. First, the 205 K offset between {01¯11} and {0001} facets comes from the classifier: an atom is removed from its facet class as soon as its local environment deviates from the low-T reference. That includes surface diffusion and vertex migration, which the paper explicitly says are classified as outliers. Stepped facets are more mobile at lower temperatures, so part of the offset may be mobility, not thermodynamic premelting. The paper doesn't validate facet Tc with an independent order parameter (per-facet Lindemann or bond-order), so this should be flagged. Second, the entire hierarchy rests on the self-cited q-SNAP potential; no second potential or DFT melting cross-check. The DFT surface energies quoted are close, which reduces the concern, but a single potential remains a single witness. Third, the τc model in Eqs. 10–18 is more of a fit than a prediction: lc is defined from τc, linearly fitted, then reinserted with an adjustable δ to reproduce the τc minimum. The authors are honest about the linear lc assumption, but the \"model\" doesn't add independent evidence.\n\nMinor issues: one trajectory per size, no error bars, a 10^11 K/s ramp; and the abstract disagrees with the text on the offset (150 vs 205 K) and size range (587–13047 vs 500–6000 atoms). The {0001} 2D Gibbs–Thomson fit drops the two smallest particles without strong justification.\n\nNet: the central observation is probably right, but the quantitative claims are overstated and need tightening. I'd send this to peer review with a request for careful revision: cross-check the facet Tc with a local order parameter, add a second potential or DFT data for the surface-energy ordering, and reword \"melting\" for facets unless they can back it up. It's a solid contribution for the nanoparticle MD community, and the methods section will be reused.\n\nRecommend: send to peer review, but expect major revision.","headline":"Facet-resolved premelting in Co nanoparticles is a genuine and useful observation, but the 205 K offset and the τc model are weaker than the abstract implies.","tokens_in":16362,"tokens_out":3094,"would_cite":true,"duration_ms":33454,"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":"In faceted cobalt nanoparticles, stepped {01-11} facets begin to melt roughly 200 K before flat {0001} facets, a size-independent offset that overturns the usual picture of premelting as an isotropic liquid shell.","keywords":["nanoparticle melting","surface premelting","facet-dependent melting","Gibbs-Thomson relation","machine-learned interatomic potential","molecular dynamics","cobalt","Gaussian Mixture Model"],"falsifier":"Run the same heating simulations with an independently fitted interatomic potential (or direct ab initio molecular dynamics below the melting point) and check whether the {01-11} facets still melt ~200 K before {0001}; alternatively, resolve facet disorder in faceted Co nanoparticles by in-situ transmission electron microscopy and look for the predicted ordering.","tokens_in":15115,"feed_emoji":"🔥","tokens_out":10227,"duration_ms":99113,"temperature":0.7,"pith_summary":"The paper argues that surface premelting in hexagonal close-packed cobalt nanoparticles is not a uniform liquid shell growing around a solid core, but a facet-by-facet process that depends on the crystallographic orientation of each surface. Using molecular dynamics simulations of particles with 587 to 5333 atoms, the authors show that stepped {01-11} facets disorder and melt at characteristic temperatures about 205 K lower than the flat {0001} facets, and that this gap does not change with particle size. Surface atoms on the stepped facets start diffusing as low as ~400 K, about 20% of the bulk melting point. Both the global melting point and the facet melting temperatures scale with size through the Gibbs-Thomson relation, and complete melting requires a critical liquid layer thickness that grows with particle size. If correct, the results imply that models of nanoparticle melting must resolve facet orientation rather than treat the surface as one entity.","feed_headline":"Stepped facets melt 200 K before flat facets in cobalt nanoparticles","feed_subtitle":"Simulations find the gap is size-independent, recasting premelting as a facet-resolved process.","key_machinery":"The key tool is an unsupervised machine-learning classifier of local atomic environments. Each atom's neighborhood is represented by a bispectrum descriptor vector; hierarchical Gaussian Mixture Models built from low-temperature snapshots define reference classes for bulk, {0001} facets, {01-11} facets, edges, and vertices; and a Mahalanobis-distance 'distortion score' labels atoms as in-class or outlier along the heating trajectory. The outlier fraction's derivative maximum defines the global melting temperature, while the temperature where each facet class loses its atoms defines facet-specific melting temperatures. This classification -- rather than a geometric order parameter -- is what","core_discovery":"On the paper's own terms, the central discovery is that melting of a faceted hcp Co nanoparticle begins on the least stable facets: stepped {01-11} surfaces disorder first, with vertex atoms diffusing around 400 K, and their characteristic melting temperature sits consistently ~205 K below that of the flat {0001} facets, regardless of particle size. This ordering matches the surface-energy hierarchy (2.13 versus 2.38 J/m^2): the facet that costs less energy to expose melts later. Extrapolating facet melting temperatures to infinite size gives 1065 K and 1371 K for the two facet types, while global melting temperatures collapse onto a Gibbs-Thomson line with a bulk intercept of 1771 K. The pa","pith_inferences":["If the size-independent facet gap is generic, other hexagonal close-packed metals with similar surface-energy ordering (e.g., Ti, Zr, Mg, Zn) should show the same hierarchy; running the same protocol on those metals would test the mechanism.","Because the outlier-based classification is parameter-light and unsupervised, it could be reused to track facet-resolved evolution in other thermally activated phenomena, such as sublimation, oxidation, or catalytic restructuring, without retuning thresholds.","If the 2D Gibbs-Thomson scaling is confirmed by experiment, a single measurement at one particle size would fix facet melting temperatures for all sizes, giving a practical design rule for the thermal stability of faceted nanocrystals."],"forward_implications":["Premelting in faceted nanoparticles must be modeled facet-by-facet, not as a uniform liquid shell; the isotropic-shell picture misses a ~200 K spread in local melting temperatures.","The facet-melting offset is independent of particle size, so the anisotropy persists down to the smallest particles studied (587 atoms).","The 2D Gibbs-Thomson extension gives facet-specific melting temperatures as a function of facet area, with infinite-surface limits of 1065 K and 1371 K.","The critical liquid-layer thickness grows linearly with particle size, which explains the non-monotonic size dependence of the critical fraction of disordered atoms (minimum near 3000 atoms).","For the smallest nanoparticles, the flat {0001} facets gate global melting: the whole particle cannot melt until those facets disorder."],"fun_headline_variants":["Stepped facets premelt 205 K before flat facets in Co nanoparticles","Facet-dependent premelting: stepped {01-11} melt first in Co nanoparticles","Melting starts on stepped facets, 205 K before flat, in Co nanoparticles","Co nanoparticle premelting is facet-specific: stepped surfaces go first","Size-independent 205 K gap: stepped facets lead melting in Co nanoparticles"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire conclusion depends on the machine-learned interatomic potential's relative surface energies of the two facet families; the paper provides no independent validation against a second potential or experiment, and if the potential mis-orders or exaggerates the 0.25 J/m² difference, the ~205 K facet gap is a simulation artifact.","fun_headline_variants_meta":{"raw":{"variants":["Stepped facets premelt 205 K before flat facets in Co nanoparticles","Facet-dependent premelting: stepped {01-11} melt first in Co nanoparticles","Melting starts on stepped facets, 205 K before flat, in Co nanoparticles","Co nanoparticle premelting is facet-specific: stepped surfaces go first","Size-independent 205 K gap: stepped facets lead melting in Co nanoparticles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000962,"raw_usage":{"total_tokens":3916,"prompt_tokens":713,"completion_tokens":3203,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":3118}},"tokens_in":457,"tokens_out":3203,"duration_ms":23734,"temperature":1.0,"reasoning_tokens":3118,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:58:28.935742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same heating simulations with an independently fitted interatomic potential (or direct ab initio molecular dynamics below the melting point) and check whether the {01-11} facets still melt ~200 K before {0001}; alternatively, resolve facet disorder in faceted Co nanoparticles by in-situ transmission electron microscopy and look for the predicted ordering.","supporting_citations":[],"review_version":1}