{"id":"3e41fdde-3467-4de0-bfe9-13ed72543c5c","arxiv_id":"2412.16294","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new Bayesian force field for aluminium predicts icosahedral nanoparticle stability up to ~2,000 atoms, decahedral stability up to ~25,000 atoms, and fcc beyond, with melting/freezing hysteresis at 100 K/ns.","lead":"Researchers trained a machine-learned force field for aluminium on quantum calculations and used it to simulate melting and freezing of nanoparticles from 200 to 11,000 atoms. The potential predicts that icosahedral shapes win below 2,000 atoms, decahedra up to 25,000 atoms, and face-centred cubic shapes beyond, plus a temperature hysteresis between melting and freezing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Al-BFF2's structural crossover sizes rest on unvalidated extrapolation: no perfect-icosahedron training data and no uncertainty check at 2100–25000 atoms.","rationale":"The reader's weakest assumption—transferability of Al-BFF2 from 150-atom training clusters to 10^4–10^5 atom predictions—is the right concern, and I sharpen it in two ways. First, Table I shows that the production potential Al-BFF2 contains no perfect icosahedron at all: Ih55 is only in D1, while D2 (used for Al-BFF2) omits it. The key motif of the headline claim is therefore never explicitly learned. Second, the paper already contains its own evidence of sensitivity: the two BFFs place the crossovers at 1500 vs 2100 atoms and 10000 vs 25000 atoms (Table V), a factor of ~1.4–2.5 shift. The energy differences at these crossings are small (curves in Fig. 1 nearly touch), so even modest errors in the strain energy of interior environments could move the crossing size substantially. Since the BFF provides predictive variances, the authors have a built-in, inexpensive way to check whether the energy difference at the crossover is statistically resolved; they do not report it. The self-admitted absence of global minimization further limits the claim to 'relative stability of ideal motifs', not equilibrium shapes. These points do not invalidate the qualitative melting/freezing hysteresis or the bulk-surface validation, but they do mean the quantitative phase map is less secure than the abstract implies. The reader's CONDITIONAL verdict remains appropriate; I would add the uncertainty check as an explicit condition, hence UNCHANGED rather than a verdict move.","tokens_in":17485,"tokens_out":4570,"duration_ms":44173,"concrete_test":"Use the released Al-BFF2 checkpoint to evaluate the FLARE predictive variance on the energy of the optimized Ih, Dh, and FCC-Wulff particles at the crossover sizes (e.g., N=2057 and N≈25000 magic numbers). Compute the posterior standard deviation of ΔE(Ih−Dh) and ΔE(Dh−FCC); if the 95% confidence interval of either difference includes zero, the crossover is not resolved and the phase map should be reported with error bars or revalidated with DFT at smaller sizes. This is a one-step check using the model's own uncertainty, requiring no new simulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline phase map (Ih stable below 2100 atoms, Dh 2100–25000, FCC beyond; Table V) is computed by comparing relaxed idealized motifs with Al-BFF2. Al-BFF2's training set (Table I) contains no icosahedral cluster: Ih55 appears only in D1, and D2 consists of bulk, surfaces, Dh85, and 100/150-atom droplets. Thus the environments that decide the Ih–Dh balance—interior atoms of large fivefold particles under radial strain—are absent from training. The authors report that Al-BFF1 vs Al-BFF2 moves the Ih→Dh crossing from 1500 to 2100 and the Dh→FCC crossing from 10000 to 25000, showing the crossover is highly sensitive to the model. Yet no DFT reference is computed above 150 atoms, and the BFF's predictive uncertainties are not used to test whether the energy differences at N≈2100 and N≈25000 are larger than the model error. The conclusion also concedes the comparison is 'without any formal global minimisation', so the claimed stability ranges describe locally relaxed ideal geometries, not necessarily equilibrium shape. The melting/freezing hysteresis, by contrast, is a robust qualitative observation, and the bulk/surface validation (Table III) is a point in the paper's favor; the load-bearing weakness is specifically the size-dependence of the motif energy balance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces two FLARE Bayesian force fields for aluminium (Al-BFF1 and Al-BFF2), trained on QE-PBEsol data from bulk, low-index surfaces, and small clusters (55-150 atoms), with D2 adding melted nanodroplets to cure nonphysical sublimation. Using Al-BFF2, the authors compute excess energies of idealized icosahedral, decahedral, and FCC motifs up to 3×10^5 atoms, obtaining Ih stability up to ~2100 atoms, Dh from ~2100 to 25000 atoms, and FCC beyond. They perform iterative MD heating/cooling at 100 K/ns for nanoparticles between 257 and 10179 atoms, reporting a size-dependent melting/freezing hysteresis (with Tm up to ~200 K above Tf at the largest size), a Gibbs-Thomson constant C=1.58, and a CNA/clustering analysis that shows no surface melting and possible surface freezing. The paper concludes that annealing liquid droplets stabilizes icosahedral solidification outcomes up to 5000-6000 atoms.","tokens_in":17771,"tokens_out":11115,"duration_ms":91099,"significance":"If the phase map is correct, the paper would be an important demonstration that a small DFT dataset can produce a transferable MLP for size-dependent structural transitions in aluminium nanoparticles, with practical implications for the community. The strengths are real: the BFF gives bulk melting at 906±1 K versus 933 K experiment, the force/energy MAEs on held-out bulk/surface test sets are small (~0.026 eV/Å, ~0.005 eV), and the potentials, training database, and trajectories are openly available. The methodological point that melted nanodroplets must be included in the training set to avoid unphysical high-temperature behavior is useful. However, the central quantitative predictions—the crossover sizes and the extended icosahedral range after annealing—rest on unvalidated extrapolation and on energy comparisons of idealized motifs without global optimization, so the significance is conditional on additional validation.","major_comments":[{"comment":"The central structural phase map is derived from energy differences between locally relaxed idealized motifs, and the Conclusion explicitly concedes that this is done 'without any formal global minimisation.' The abstract's statements that 'Al-BFF predicts an icosahedral stability range of up to 2000 atoms... a region of stability for decahedra, up to 25000 atoms' therefore overstate what is actually computed: these are relative stabilities of chosen ideal geometries under Al-BFF2, not equilibrium stability ranges. Please either perform global optimization (e.g., basin hopping or replica exchange) for representative sizes, or systematically rephrase the abstract and conclusion to describe the result as the relative stability of idealized motifs.","section":"III (Fig. 1, Table V) and Conclusion"},{"comment":"The crossover locations are an extrapolation that is never validated. The D2 dataset used for all MD and for the Al-BFF2 numbers in Table V contains no icosahedral cluster (Ih55 appears only in D1), and the largest DFT reference is 150 atoms, yet the energy comparisons extend to 3×10^5 atoms and the MD to 10,179 atoms. No DFT calculation is reported on relaxed large motifs near the crossover sizes. The BFF predictive uncertainty (Eq. 22) is available but is not used to test whether the energy differences between motifs at N≈2100 and N≈25000 exceed model error. The sensitivity of the crossings to the training set is visible in the shift between Al-BFF1 (1500/10000) and Al-BFF2 (2100/25000) in Table V. Please add single-point DFT checks on relaxed motifs at least at the two crossover sizes and report uncertainty estimates on the motif energy differences.","section":"Table I, Table V, Eq. (22)"},{"comment":"The claim that 'the annealing of a liquid droplet further stabilizes icosahedral structures, extending their stability range to 5000 atoms' conflates kinetic freezing outcomes with thermodynamic stability. The MD runs at 100 K/ns are non-equilibrium, and the same section notes that the annealed structures 'are not the global minimum.' The defective Ih and Dh solids obtained after freezing are kinetic products; they do not establish an equilibrium stability range. Please either compute free energies of the competing solid phases or reframe this statement as a description of solidification outcomes at the imposed cooling rate.","section":"III.A and Conclusion"},{"comment":"The Gibbs-Thomson fit is used to argue that the good agreement with experiment 'corroborates the accuracy' of Al-BFF2. However, the fit uses (Tm+Tf)/2 from 100 K/ns simulations, and no uncertainty is reported for the fitted constant C=1.58. Without confidence intervals and a statement of the number and size range of fitting points, the comparison with the experimental C=1.76 is not quantitative. Moreover, the arithmetic mean of non-equilibrium hysteresis endpoints is not obviously the equilibrium melting temperature. Please report confidence intervals for C and, if possible, compare with a quasi-static or coexistence-based melting temperature for at least one nanoparticle size.","section":"Eq. (8), III.A"},{"comment":"The Wulff construction and the FCC-motif energies used in the crossover analysis rely on the surface-energy ratio γ100/γ111=1.17. Table III shows that Al-BFF2's γ111 (0.94 J/m2) is 19% below the quoted experimental value (1.16 J/m2); because the BFF reproduces the PBEsol value (0.96 J/m2), the discrepancy is inherited from the reference functional. Since the Ih→Dh→FCC balance is a competition between surface energy and strain, the quantitative crossover sizes should be discussed with this systematic uncertainty in mind.","section":"Table III, III"}],"minor_comments":[{"comment":"The definition of Tm and Tf as 'the temperatures showing the nearest larger and smaller DKL to the discontinuity' is not reproducible as written; please specify the exact algorithm (e.g., the largest jump in DKL, with a window or threshold) used to locate the transition.","section":"II.B, Eq. (7)"},{"comment":"The sentence 'we note an increment in the IH percentage sharper than that of SH, indicating that the surface organises first than the inner part' is internally inconsistent: if IH (inner ordered) increases more sharply, the core organizes more abruptly, not the surface. Please clarify the evidence for surface-first ordering during freezing.","section":"III.A, Fig. 6"},{"comment":"The test set used for the MAE values in Table IV comprises only bulk and surface configurations (33 entries in Table I), so these low MAEs do not directly validate the potential on nanoparticle environments; this limitation should be stated when the MAEs are presented.","section":"II.A, Table IV"},{"comment":"Because both E(N) and ε_b are negative, the formula for the excess energy is ambiguous without an explicit statement of the sign convention; please state clearly that ΔE is defined as the positive energy cost per surface atom.","section":"II.B, Eq. (2)"},{"comment":"The phrase 'the nano-slush state is at least up to 3.4 nm' introduces a term and a size bound that are not defined earlier; please specify which descriptor supports this bound and what 'nano-slush' means.","section":"Conclusion"},{"comment":"There is a typo: 'We re grateful' should read 'We are grateful.'","section":"Acknowledgments"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a promising contribution for a materials/physics journal, and the open release of potentials and trajectories is commendable. The main risk is that the headline phase map is stated as a prediction of equilibrium stability, whereas the underlying evidence is restricted to relaxed ideal geometries plus non-equilibrium MD. I would advise the editor to request a revision that either adds targeted DFT validation and global optimization, or substantially softens the claims, before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is a solid application of FLARE Bayesian force fields to aluminium nanoparticles, and it makes one genuinely useful methodological point: without adding melted nanodroplets to the training set, the potential produces unphysical trajectories (atoms subliming off the cluster below melting). Al-BFF 2 fixes that, and the bulk validation is respectable—melting temperature 906 K vs 933 K experiment, surface energies and lattice constants in line with PBEsol.\n\nThe new results are the size-dependent stability map (Ih up to ~2100 atoms, Dh between 2100 and ~25000, FCC beyond) and the melting-freezing hysteresis as a function of size. The hysteresis picture is likely robust: it shows up in every descriptor, grows with size, and the Gibbs-Thomson constant they fit (1.58) is close to the experimental fit (1.76). That part is good evidence the potential behaves sensibly at the sizes simulated.\n\nThe load-bearing weakness is exactly where the stress-test says it is. The crossover sizes come from comparing relaxed ideal motifs—the paper concedes this, no global optimisation—and the training set has no icosahedral cluster at all in D2, so the environments that decide the Ih–Dh balance at 2000–25000 atoms are extrapolations. Al-BFF 1 vs Al-BFF 2 shifts the crossovers from 1500 to 2100 and from 10000 to 25000, which tells you the model is sensitive in exactly that region, and no DFT reference or uncertainty check is provided there. FLARE gives predictive variances for free; the authors should have used them to say whether the energy differences at the crossings are larger than the model error. This is fixable but it means the headline phase map is a prediction from an extrapolated potential, not a validated result.\n\nOther soft spots are minor: the larger MD runs (5083, 5096, 5341, 10179) are single trajectories, so no error bars on the phase-transition temperatures at the sizes that matter most; and the surface-melting conclusion rests on a clustering analysis that is qualitative. The surface-freezing claim is mild and phrased as such, which is fine.\n\nOverall this deserves a serious referee. The potential is public, the bulk checks are good, and the methodological message about training data is useful to the community. The referee should push on the crossovers and ask for uncertainty quantification or a few DFT checks at intermediate sizes. I'd cite the potential if I worked on Al NPs, but I'd treat the crossover sizes as tentative.\n\nRecommendation: send it to peer review with the request for a strengthening of the structural-stability section.","headline":"Useful Al-BFF potential with a solid training-data message, but the headline structural crossover map rests on an extrapolation that needs uncertainty checks before it can be trusted.","tokens_in":18312,"tokens_out":2321,"would_cite":true,"duration_ms":19673,"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":"Aluminium nanoparticles switch shape at 2,100 and 25,000 atoms","keywords":["aluminium nanoparticles","Bayesian force field","machine-learned potential","melting-freezing hysteresis","icosahedral stability","decahedral stability","molecular dynamics","structural crossover"],"falsifier":"Compute DFT excess energies for relaxed icosahedral, decahedral, and fcc nanoparticles at sizes straddling the predicted crossovers, for example 2,000-3,000 and 20,000-30,000 atoms, and check whether the energy crossings match 2,100 and 25,000 atoms; alternatively, count the structural motifs by transmission electron microscopy for aluminium nanoparticles between 2 and 5 nm and see whether icosahedra dominate at the small end as claimed.","tokens_in":17276,"feed_emoji":"🌡️","tokens_out":5605,"duration_ms":45665,"temperature":0.7,"pith_summary":"This paper claims that a newly trained Bayesian machine-learned potential for aluminium, augmented with melted nanodroplet configurations in its training data, can drive classical molecular dynamics of nanoparticles from 200 to 11,000 atoms and produce a thermodynamically consistent melting-freezing cycle. On its predictions, icosahedral aluminium nanoparticles are most stable up to about 2,000 atoms (~2 nm), decahedral shapes take over between about 2,100 and 25,000 atoms, and face-centred-cubic motifs win beyond 25,000 atoms. The same simulations show a melting-freezing hysteresis that widens with size, with melting temperatures up to about 200 K above freezing at a 100 K/ns rate, and no surface melting. A sympathetic reader would care because these size-dependent structural crossovers and transition temperatures are the quantities that matter for applications such as catalysis, energy storage, and additive manufacturing, and standard empirical potentials misjudge them.","feed_headline":"Aluminium nanoparticles switch shape at 2,100 and 25,000 atoms","feed_subtitle":"A Bayesian force field trained on melted droplets predicts which crystal motifs win and a growing melt-freeze gap.","key_machinery":"The load-bearing object is the Al-BFF 2 potential: a Gaussian-process Bayesian force field whose local energies are functions of atomic cluster expansion three-body descriptors, trained on energies, forces, and stresses and equipped with an active-learning loop that calls DFT when predictive uncertainty is high. The decisive technical addition is the inclusion of melted nanodroplet configurations in the training set, which removes non-physical sublimation seen with a first potential trained only on bulk, surfaces, and cold clusters. The analysis then runs on iterative heating and cooling schedules, with structural classification by common-neighbour signatures and unsupervised k-means clustering of local environments to separate core from surface ordering.","core_discovery":"The central claim, stated on the paper's own terms, is that the Bayesian force field Al-BFF 2, trained on density-functional-theory energies, forces, and stresses from bulk aluminium, low-index surfaces, and small clusters together with deliberately included liquid nanodroplets, reproduces bulk properties such as a melting temperature of 906 ± 1 K (experimental value 933 K) and then predicts the size-dependent structural map of aluminium nanoparticles. Relaxed excess energies place the icosahedral-to-decahedral crossover at 2,100 atoms and the decahedral-to-fcc crossover at 25,000 atoms, and iterative molecular dynamics at 100 K/ns adds the dynamical claim that hysteresis loops enlarge with size, that annealing a liquid droplet extends icosahedral stability to about 5,000 atoms, and that there is no surface melting although the surface shows local order in the liquid. The paper argues this is a complete thermodynamic cycle simulated with one potential that stays close to ab initio accuracy at classical speed.","pith_inferences":["The practice of seeding training data with melted nanodroplets is likely transferable to other metals; it addresses a generic failure mode where machine-learned potentials trained on cold solids produce unphysical vaporisation near melting.","Because the crossover sizes are set by the competition between surface energy and lattice strain, the same type of size map should hold for other aluminium-group metals, though the precise atom counts will shift with surface-energy ratios.","A slower cooling rate should shrink the observed hysteresis and may shift the solidification outcome toward the energy-preferred motif, offering a direct test of the kinetic origin the paper assigns to the loop.","If the structural crossovers are reproduced at larger sizes, the potential could replace empirical potentials in studies of nanoscale welding, sintering, and 3D printing of aluminium, where melting and resolidification are the controlling steps."],"forward_implications":["Aluminium nanoparticles below about 2,000 atoms should be treated as icosahedral rather than fcc-like when modelling catalysis or other size-sensitive properties.","The predicted size window for decahedral stability, roughly 2,100 to 25,000 atoms, gives a concrete target for electron-microscopy checks and for synthesis strategies aiming at fivefold shapes.","At 100 K/ns heating, melting and freezing temperatures differ by up to 200 K, so kinetic effects must be folded into any comparison between simulated and measured transition temperatures.","Annealing a liquid droplet before cooling shifts icosahedral stability upward to about 5,000 atoms, implying thermal history changes the resulting nanoparticle shape.","The absence of surface melting, with ordered local environments persisting at the liquid surface, is a directly testable structural prediction."],"supporting_citations":[{"why":"supplies the structural-motif taxonomy and the surface-strain competition argument that frames the size crossovers.","marker":"[1]"},{"why":"provides the data-driven MLP melting methodology and the hierarchical clustering approach reused here.","marker":"[6]"},{"why":"defines the active-learning Bayesian force field framework that Al-BFF is built on.","marker":"[12–14]"},{"why":"provides the DFT reference calculations (energies, forces, stresses) used for training and testing.","marker":"[21]"},{"why":"defines the atomic cluster expansion descriptors used to featurise local atomic environments.","marker":"[22]"},{"why":"is the empirical embedded-atom potential whose melting temperature benchmark the new potential improves on.","marker":"[26]"},{"why":"supplies the Gibbs-Thomson scaling relation used to fit the size-dependent transition temperatures.","marker":"[33]"},{"why":"provides the experimental nanocalorimetry melting points used as the comparison for the Gibbs-Thomson fit.","marker":"[34]"}],"fun_headline_variants":["Al nanoparticles change shape at 2,100 and 25,000 atoms","Machine-learnt force field maps Al nanoparticle shape flips at 2,100 and 25,000 atoms","Bayesian force field predicts Al nanoparticle shape switches at 2,100 and 25,000 atoms","Shape crossover in Al nanoparticles: 2,100 and 25,000 atoms","Al nanoparticle motifs flip at 2,100 and 25,000 atoms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The force field is trained on DFT data for clusters of only 55 to 150 atoms, yet it is used to predict behaviour of particles up to 300,000 atoms without any direct DFT check at those larger sizes.","fun_headline_variants_meta":{"raw":{"variants":["Al nanoparticles change shape at 2,100 and 25,000 atoms","Machine-learnt force field maps Al nanoparticle shape flips at 2,100 and 25,000 atoms","Bayesian force field predicts Al nanoparticle shape switches at 2,100 and 25,000 atoms","Shape crossover in Al nanoparticles: 2,100 and 25,000 atoms","Al nanoparticle motifs flip at 2,100 and 25,000 atoms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001936,"raw_usage":{"total_tokens":7600,"prompt_tokens":994,"completion_tokens":6606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":6491}},"tokens_in":610,"tokens_out":6606,"duration_ms":37439,"temperature":1.0,"reasoning_tokens":6491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:45:45.165464+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute DFT excess energies for relaxed icosahedral, decahedral, and fcc nanoparticles at sizes straddling the predicted crossovers, for example 2,000-3,000 and 20,000-30,000 atoms, and check whether the energy crossings match 2,100 and 25,000 atoms; alternatively, count the structural motifs by transmission electron microscopy for aluminium nanoparticles between 2 and 5 nm and see whether icosahedra dominate at the small end as claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the data-driven MLP melting methodology and the hierarchical clustering approach reused here."},{"cited_title":"Delgado-Callico , author K","cited_arxiv_id":null,"evidence_quote":"provides the DFT reference calculations (energies, forces, stresses) used for training and testing."},{"cited_title":"Rossi , author L","cited_arxiv_id":null,"evidence_quote":"defines the atomic cluster expansion descriptors used to featurise local atomic environments."},{"cited_title":"Gaudoin , author W","cited_arxiv_id":null,"evidence_quote":"is the empirical embedded-atom potential whose melting temperature benchmark the new potential improves on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the experimental nanocalorimetry melting points used as the comparison for the Gibbs-Thomson fit."}],"review_version":1}