{"id":"e9554284-00d6-4d67-89e8-43422c747552","arxiv_id":"2608.07081","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Temperature-dependent EXAFS reveals strongly anisotropic thermal vibrations in hcp zinc across eight coordination shells, and fine-tuning the CHGNet machine-learning potential restores agreement with the measured X-ray spectra.","lead":"By measuring how zinc's X-ray absorption changes with temperature, this study maps how strongly atoms vibrate in different directions, showing much larger motion between atomic layers than within them. It then uses that data to test and improve a machine-learned atom model, finding that a fine-tuned version matches the measured spectra far better than the original.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaussian decomposition of overlapping first/second RDF peaks may bias MSRD1 and MSRD2 and thus the central anisotropy ratio; direct per-shell MSRDs from RMC configurations would test this.","rationale":"The paper's strongest claim is that RMC/EA analysis yields accurate shell-resolved MSRDs, revealing anisotropy between in-plane and out-of-plane force constants, and that fine-tuned CHGNet improves agreement with experiment. The reader identified the Gaussian decomposition of the overlapping first and second coordination shells as the weakest assumption; this is indeed the most load-bearing concern. The first (2.66 Å, A-type) and second (2.91 Å, B-type) shells overlap strongly at 300 K, and the MSRD2/MSRD1 ratio is the key quantitative evidence for anisotropy, cross-compared to diffraction-derived U33/U11. Any systematic bias in decomposing the composite RDF peak directly propagates into the force constants in Table 1 and the anisotropy ratio in Fig. 6(b). The k^3 weighting cross-check in the paper tests only the EXAFS weighting, not the Gaussian shape assumption, so it does not address this risk. The absence of reported error bars further obscures whether the anisotropy is statistically significant. A direct computation of per-shell MSRDs from the final RMC configurations would resolve the issue: if the direct values match the Gaussian-fitted values, the central claim stands; if not, the anisotropy ratio and force constants would need revision. The paper is otherwise methodologically coherent, with internal cross-checks and a plausible MLIP fine-tuning procedure, so the concern does not warrant rejection, but it justifies a conditional acceptance pending this validation. The reader's verdict of CONDITIONAL is therefore unchanged.","tokens_in":13576,"tokens_out":6920,"duration_ms":62863,"concrete_test":"From the final atomic configurations produced by the ten RMC/EA runs at 300 K (and preferably also 10 K), compute shell-resolved MSRDs without Gaussian fitting: assign each Zn–Zn pair to a coordination shell by distance windows around the ideal lattice distances or by mapping to the ideal hcp site of each neighbor, then calculate the variance of relative displacements for each shell. Compare these direct MSRDs with the Gaussian-fitted values in Fig. 5(a). If MSRD1 or MSRD2 shifts by more than ~0.002 Å^2 (5% of the 300 K value), or if the ratio MSRD2/MSRD1 changes by more than ~0.1, the decomposition bias is material and the anisotropy/force-constant conclusions need re-quantification. A supplementary synthetic check on MD trajectories with known per-shell variances would confirm the method recovers input variances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central anisotropy claim rests on MSRDs extracted by decomposing the RMC RDF into independent Gaussian peaks, one per coordination shell (Section 2.2, Section 3, Fig. 5(a)). At 300 K the first (2.66 Å, A-type) and second (2.91 Å, B-type) shells form one heavily overlapping peak in the RDF (Fig. 4). A two-Gaussian fit to such a composite peak has a known trade-off between the widths and areas of the two components, so MSRD1 and MSRD2 are not independently determined by the RDF alone. The reported force constants (Table 1) and the MSRD2/MSRD1 anisotropy ratio (Fig. 6(b)) inherit any bias from this decomposition. The k^3-weighting cross-check only changes the EXAFS weighting, not the Gaussian model, so it cannot detect this bias. No error bars are given for the MSRDs, even though the RMC runs were repeated ten times, so the statistical significance of the anisotropy is unquantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the local lattice dynamics of hcp zinc by combining temperature-dependent (10–300 K) Zn K-edge EXAFS with reverse Monte Carlo/evolutionary-algorithm (RMC/EA) simulations and molecular dynamics with a machine-learning interatomic potential (CHGNet). The authors extract mean-square relative displacements (MSRDs) for eight coordination shells, fit their temperature dependence with a correlated Einstein model to obtain effective force constants, and report a pronounced anisotropy between in-plane (A-type) and out-of-plane (B-type) interactions. They also compare the MSRD2/MSRD1 ratio with the diffraction-based U33/U11 ratio. In the second part, they fine-tune the CHGNet universal machine-learning potential on DFT relaxation trajectories and compare simulated EXAFS and RDFs with experiment at 300 K, finding that fine-tuning substantially improves agreement relative to the vanilla potential.","tokens_in":13929,"tokens_out":3996,"duration_ms":35742,"significance":"If the central methodological concern is resolved, the paper offers a valuable shell-resolved view of anisotropic thermal motion in a strongly anisotropic hcp metal, together with a concrete demonstration that fine-tuning a universal machine-learning potential on DFT data improves its fidelity to EXAFS measurements. The work includes several commendable internal checks: the RMC/EA analysis explicitly treats multiple-scattering paths up to fourth order and is repeated with k3 weighting at 10 and 300 K, and the MD-EXAFS comparison is a genuine independent test because the potential was trained on DFT energies/forces/stresses rather than on the experimental EXAFS. The paper also reports quantitative training metrics (MAEs of 1 meV/atom, 22 meV/Å, 0.07 GPa) for the fine-tuned potential, which strengthens the credibility of the MLIP part. However, the absence of error bars on all extracted MSRDs and the reliance on a Gaussian decomposition of overlapping coordination shells leave the statistical significance of the central anisotropy claim unverified.","major_comments":[{"comment":"The MSRDs for the first and second coordination shells are obtained by decomposing the total RDF into independent Gaussian peaks, but at 300 K the first (2.66 Å) and second (2.91 Å) shells form a single heavily overlapping peak (Fig. 4). A two-Gaussian fit to such a composite peak has a known trade-off between the widths and areas of the two components, so MSRD1 and MSRD2 are not independently determined by the RDF alone. The k3-weighting cross-check described in Section 2.2 changes only the EXAFS weighting, not the Gaussian decomposition model, so it cannot detect this bias. Please provide a direct test: compute per-shell MSRDs from the RMC configurations by assigning atoms to coordination shells from the initial ideal hcp structure (or an equivalent neighbor-based assignment) and compare those values with the Gaussian-decomposition results. If the direct per-shell MSRDs confirm the anisotropy, report them; otherwise, the central anisotropy claim requires qualification.","section":"Section 2.2 and Section 3, Figs. 4 and 5(a)"},{"comment":"No error bars are shown for the MSRDs in Fig. 5(a), for the MSRD-versus-distance curves in Fig. 6(a), or for the MSRD2/MSRD1 ratio in Fig. 6(b), although the RMC/EA simulations were repeated ten times at each temperature. Without a measure of run-to-run or fit uncertainty, the statistical significance of the reported anisotropy (approximately 1.2–1.6 times) and of the Einstein-model force constants in Table 1 cannot be assessed. Please add error bars derived from the ten independent RMC runs (or another defensible uncertainty estimate) to all reported MSRDs and ratios, and state explicitly how the uncertainties of the force constants in Table 1 were obtained.","section":"Section 2.2 and Figs. 5–6"},{"comment":"The claim that the MSRD2/MSRD1 ratio 'closely matches' the diffraction-based U33/U11 ratio is based on a qualitative visual comparison of two curves without uncertainty bands. Since MSRD includes correlation between atoms while MSD is a single-atom displacement, the two ratios are not expected to be equal in general; the text itself later states that the ratios indicate displacements only 'approximately 1.2–1.6 times larger,' which is a weaker statement than 'closely matches.' Please provide a quantitative comparison (e.g., a confidence interval for the difference, or a reduced chi-square value) and clarify the expected relationship between MSRD2/MSRD1 and U33/U11, or soften the wording accordingly.","section":"Section 3, Fig. 6(b)"},{"comment":"The improvement of the fine-tuned CHGNet over the vanilla potential is demonstrated only through visual comparison of the EXAFS spectra and RDFs in Fig. 7. Since the abstract and conclusions state that fine-tuning 'substantially improves' agreement with experiment, please include a quantitative metric for both the EXAFS (e.g., R-factor or RMSD in k-space) and the RDF (e.g., integrated absolute difference over the plotted range) so that the degree of improvement is numerically grounded.","section":"Section 2.3 and Fig. 7"}],"minor_comments":[{"comment":"The phrase 'reverse Monte Carlo method enable' should be 'reverse Monte Carlo method enables'.","section":"Abstract"},{"comment":"The sentence 'demonstrating the robustness of the results with respect to the choice of EXAFS weighting and and showing that they are not biased by the reduced contribution of the high-k region' contains a duplicated 'and'.","section":"Section 2.2"},{"comment":"The sentence beginning 'The anisotropy of thermal vibrations is clearly evidenced by the difference between the out-off-plane MSD' contains a grammatical error and the typo 'out-off-plane'; the intended phrase is 'out-of-plane'.","section":"Section 3"},{"comment":"The text refers to the Supplementary Material for details of the Gaussian decomposition of the RDFs, but no supplementary material is included with the arXiv submission, so the decomposition procedure (initial parameters, fitting ranges, constraints) cannot be checked. Please provide the supplement or include the relevant details in an appendix.","section":"Section 2.2 and Section 3"},{"comment":"The abbreviation 'uMLIPs' appears with inconsistent capitalization (e.g., 'uMLIPS' in the conclusions); please use one consistent form.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central finding hinges on the Gaussian decomposition of overlapping first and second coordination shells. The absence of error bars and the lack of a direct per-shell MSRD calculation are important but fixable within the manuscript's scope. Please ask the authors to provide the supplementary material during revision, since the Gaussian decomposition details are only referenced there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives the community the first eight-shell MSRD dataset for hcp zinc and a clean test of CHGNet fine-tuning against EXAFS. The fine-tuning validation is genuinely independent: the potential was trained on DFT energies/forces/stresses, not on the EXAFS, so the improved MD-EXAFS agreement at 300 K is real evidence. The RMC/EA workflow is the group's own established method, so the novelty is material-specific rather than methodological, but the new numbers are useful and the paper is honest about what it is.\n\nWhat it does well: the EXAFS data are strong (k up to 18 Å^-1), the RMC fits are shown with imaginary parts at four temperatures, the MSRD temperature curves look physically sensible, the Einstein force constants show the expected A/B anisotropy, and the MSRD2/MSRD1 ratio tracks the diffraction U33/U11 ratio. The k^3-weighting cross-check at 10 and 300 K is a legitimate robustness test. The paper also states plainly that the C-type shell is too poorly constrained to report.\n\nThe soft spots, in order. First, the Gaussian decomposition of the RDF into per-shell peaks is the load-bearing step for the central MSRDs, and it is not directly validated. At 300 K the first two shells merge into one broad peak, and a two-Gaussian fit has a known width-area trade-off, so MSRD1 and MSRD2 (and their ratio) could be biased. The k^3 cross-check does not address this because it changes only the EXAFS weighting, not the Gaussian model. Since the RMC runs produce full atomic configurations, per-shell MSRDs could be computed directly from the coordinates; the authors should do this in revision. Second, no error bars are reported for the MSRDs even though the RMC runs were repeated ten times; the anisotropy ratio has no quantified uncertainty. Third, data and code are not deposited, which weakens the 'reusable workflow' message. These are addressable concerns, not fatal ones. The central argument survives.\n\nThis paper is for two audiences: the EXAFS local-structure community and people building or validating MLIPs. It deserves a serious referee. Send it out, and ask for direct per-shell MSRDs from the RMC configurations and error bars. I would also push for data/code deposit.","headline":"A useful new shell-resolved MSRD dataset and an independent CHGNet fine-tuning check, but the central anisotropy ratio hinges on an unvalidated Gaussian decomposition and missing error bars.","tokens_in":14486,"tokens_out":2748,"would_cite":true,"duration_ms":23777,"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":"By fitting temperature-dependent EXAFS with reverse Monte Carlo simulations, this paper extracts shell-resolved mean-square relative displacements for eight coordination shells of hcp zinc and shows that fine-tuning the CHGNet…","keywords":["hcp zinc","EXAFS","reverse Monte Carlo","mean-square relative displacement","lattice dynamics","machine-learning interatomic potential","CHGNet","anisotropy"],"falsifier":"A decisive check would be to rerun the 300 K reverse Monte Carlo fit with a non-Gaussian or two-shell-correlated decomposition of the first two overlapping peaks and see whether the $\\mathrm{MSRD}_2/\\mathrm{MSRD}_1$ anisotropy moves outside the quoted errors; a complementary check is to compare the shell-resolved MSRDs with values from inelastic neutron scattering or first-principles phonon calculations.","tokens_in":2045,"feed_emoji":"🔬","tokens_out":6041,"duration_ms":121524,"temperature":0.7,"pith_summary":"This paper tries to show that temperature-dependent EXAFS, analysed with reverse Monte Carlo simulations, can resolve how thermal vibrations differ between the in-plane and out-of-plane directions in hexagonal close-packed zinc, and that the resulting shell-by-shell measures provide a stringent test for machine-learning interatomic potentials. The authors extract mean-square relative displacements for eight coordination shells, fit them with a correlated Einstein model, and obtain effective force constants that are roughly two to three times larger for in-plane (A-type) than for out-of-plane (B-type) neighbours. They also show that the ratio of the second-shell to first-shell MSRDs tracks the out-of-plane/in-plane displacement ratio measured by single-crystal diffraction, confirming the anisotropy with a local probe. Finally, molecular dynamics with the CHGNet universal machine-learning potential overestimates thermal disorder, while a version fine-tuned on a small set of DFT relaxation trajectories brings the simulated EXAFS spectrum and radial distribution function close to experiment. The payoff is a workflow in which EXAFS, which sees around specific atoms, can validate and refine machine-learning potentials for anisotropic metals.","feed_headline":"Thermal motion in zinc's eight shells resolved by EXAFS","feed_subtitle":"In-plane bonds are two to three times stiffer; machine-learning potentials match the measured anisotropy.","key_machinery":"The central machinery is a reverse Monte Carlo loop over a periodically repeated supercell: an evolutionary algorithm moves zinc atoms while a real-space multiple-scattering code computes the configuration-averaged EXAFS spectrum, and the fit is scored in both wavenumber and distance space against measured spectra at each temperature. The final atomic configurations give zinc–zinc radial distribution functions, which are decomposed into Gaussian peaks whose variances are the shell-resolved mean-square relative displacements; fitting their temperature dependence to the correlated Einstein model yields effective force constants. For the potential validation, the same multiple-scattering calculation is applied to snapshots from constant-temperature molecular dynamics driven by the CHGNet universal machine-learning interatomic potential, so the experimental EXAFS spectrum itself becomes the arbiter of whether the potential produces the right thermal disorder.","core_discovery":"At 10–300 K, the Zn K-edge EXAFS of hcp zinc was modelled by reverse Monte Carlo in a 288-atom supercell, with configuration-averaged spectra computed by real-space multiple scattering over all absorbing atoms. From the final configurations, the Zn–Zn radial distribution function was decomposed into Gaussian peaks, and the variance of each peak gives the mean-square relative displacement (MSRD) for that coordination shell. Fitting each $\\mathrm{MSRD}(T)$ to the correlated Einstein model gives effective force constants: $34\\pm1$ N/m for the six in-plane nearest neighbours at 2.66 Å, versus $12\\pm1$ N/m for the six second-shell neighbours at 2.91 Å, with the ratio $\\mathrm{MSRD}_2/\\mathrm{MSRD}_1$ rising with temperature in line with the diffraction-derived $U_{33}/U_{11}$. The paper further shows that the original CHGNet universal machine-learning potential, used in 300 K NVT molecular dynamics, softens the lattice and broadens the RDF, damping the computed EXAFS; fine-tuning CHGNet on roughly 400 DFT frames of strained and displaced zinc supercells substantially corrects this, bringing both EXAFS and RDF into agreement with the RMC reference and experiment.","pith_inferences":["The paper does not test cadmium, but the same RMC/EXAFS pipeline could be applied to Cd, whose even larger $c/a \\approx 1.89$ should produce a stronger MSRD anisotropy; whether current machine-learning potentials capture that would be a direct test of the method's resolving power.","A natural next step is to use the RMC-derived radial distribution function at 300 K as a training target for fine-tuning other universal machine-learning potentials, since it encodes thermal disorder directly rather than through energies and forces.","The reported force-constant pattern implies a minimal two-parameter anisotropic Einstein model for zinc's local dynamics; one consequence the authors do not pursue is that this model should reproduce diffraction Debye–Waller factors at all temperatures, not just their ratio.","The unresolved $c$-axis C-type shell could likely be recovered with a longer simulation or symmetry-restrained RMC sampling, completing the full anisotropy map of the first coordination shell."],"forward_implications":["The eight-shell MSRD dataset gives a quantitative, shell-resolved picture of thermal disorder in a strongly anisotropic metal, not just the usual first-shell picture.","Effective force constants from the correlated Einstein model show which crystallographic directions are stiff: in-plane A-type neighbours ($\\kappa \\approx 34$ N/m at 2.66 Å) resist thermal motion much more than out-of-plane B-type neighbours ($\\kappa \\approx 12$ N/m at 2.91 Å).","The $\\mathrm{MSRD}_2/\\mathrm{MSRD}_1$ anisotropy ratio reproduces the temperature trend of the diffraction-measured $U_{33}/U_{11}$, strengthening the case that local EXAFS and global diffraction are measuring the same anisotropic displacements.","The original CHGNet potential's softening is visible as over-broad RDF peaks and a damped EXAFS spectrum; fine-tuning on a few hundred DFT frames repairs most of this, so EXAFS can act as an experimental check on machine-learning potentials.","Using the RMC reference RDF, the paper establishes a workflow where EXAFS data, not just energies and forces, decide whether a machine-learning potential is physically trustworthy."],"supporting_citations":[{"why":"Supplies the temperature-dependent lattice parameters $a$ and $c$ and the $U_{33}/U_{11}$ anisotropic displacement ratio used to build the simulation box and to validate the MSRD anisotropy.","marker":"[3]"},{"why":"Provides the reverse Monte Carlo approach for modeling thermal disorder in crystals from EXAFS, the core extraction method.","marker":"[23]"},{"why":"Describes the evolutionary-algorithm enhancement used here to fit multiple atomic configurations against experimental EXAFS.","marker":"[24]"},{"why":"Underlies the real-space multiple-scattering theory used to compute configuration-averaged EXAFS spectra within the RMC and molecular-dynamics loops.","marker":"[25]"},{"why":"Supplies the correlated Einstein model whose fit to MSRD($T$) yields the reported effective force constants.","marker":"[32]"},{"why":"Introduces the CHGNet universal machine-learning interatomic potential used for the molecular dynamics simulations.","marker":"[34]"},{"why":"Prior demonstration that fine-tuning CHGNet on additional structures improves EXAFS agreement for layered compounds, motivating the same fine-tuning strategy here.","marker":"[53]"},{"why":"Documents the systematic softening of universal machine-learning potentials that explains why the vanilla CHGNet overestimates thermal disorder.","marker":"[54]"}],"fun_headline_variants":["Zinc's eight shells of thermal motion mapped by EXAFS and ML","In-plane bonds 2-3x stiffer: EXAFS quantifies zinc's anisotropy","Fine-tuned CHGNet matches zinc's EXAFS-derived lattice dynamics","Shell-by-shell view of zinc's atomic jitter from EXAFS and ML","EXAFS plus refined CHGNet pin down zinc's anisotropic forces"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The load-bearing premise is that decomposing the total zinc–zinc radial distribution function into independent Gaussian peaks, one per coordination shell, does not bias the extracted mean-square relative displacements, even when the first shell (2.66 Å) and the second shell (2.91 Å) overlap at high temperature.","fun_headline_variants_meta":{"raw":{"variants":["Zinc's eight shells of thermal motion mapped by EXAFS and ML","In-plane bonds 2-3x stiffer: EXAFS quantifies zinc's anisotropy","Fine-tuned CHGNet matches zinc's EXAFS-derived lattice dynamics","Shell-by-shell view of zinc's atomic jitter from EXAFS and ML","EXAFS plus refined CHGNet pin down zinc's anisotropic forces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3778,"prompt_tokens":991,"completion_tokens":2787,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2682}},"tokens_in":607,"tokens_out":2787,"duration_ms":19174,"temperature":1.0,"reasoning_tokens":2682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:39.974112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be to rerun the 300 K reverse Monte Carlo fit with a non-Gaussian or two-shell-correlated decomposition of the first two overlapping peaks and see whether the $\\mathrm{MSRD}_2/\\mathrm{MSRD}_1$ anisotropy moves outside the quoted errors; a complementary check is to compare the shell-resolved MSRDs with values from inelastic neutron scattering or first-principles phonon calculations.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the temperature-dependent lattice parameters $a$ and $c$ and the $U_{33}/U_{11}$ anisotropic displacement ratio used to build the simulation box and to validate the MSRD anisotropy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the reverse Monte Carlo approach for modeling thermal disorder in crystals from EXAFS, the core extraction method."},{"cited_title":"Timoshenko, A","cited_arxiv_id":null,"evidence_quote":"Describes the evolutionary-algorithm enhancement used here to fit multiple atomic configurations against experimental EXAFS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Underlies the real-space multiple-scattering theory used to compute configuration-averaged EXAFS spectra within the RMC and molecular-dynamics loops."},{"cited_title":"Sevillano, H","cited_arxiv_id":null,"evidence_quote":"Supplies the correlated Einstein model whose fit to MSRD($T$) yields the reported effective force constants."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the CHGNet universal machine-learning interatomic potential used for the molecular dynamics simulations."},{"cited_title":"Žguns, I","cited_arxiv_id":null,"evidence_quote":"Prior demonstration that fine-tuning CHGNet on additional structures improves EXAFS agreement for layered compounds, motivating the same fine-tuning strategy here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the systematic softening of universal machine-learning potentials that explains why the vanilla CHGNet overestimates thermal disorder."}],"review_version":1}