{"id":"ad534abe-85d8-433c-bf78-84e7f8bfab3b","arxiv_id":"2504.19352","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A machine-learning and molecular-dynamics workflow predicts inelastic neutron scattering spectra that agree with measurements for silicon, benzene, and hydrogenated scandium-doped barium titanate.","lead":"This paper builds a computational pipeline that predicts inelastic neutron scattering spectra from atomistic simulations and compares the results with measurements on three materials. It matters because such simulations could help plan and interpret neutron experiments without empirical fitting, and they could also validate machine-learned interatomic potentials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc selection of the Brillouin zone that best matches experiment in Fig. 2c makes the 1500 K Si validation circular; the predicted spectrum should be reported as a distribution over all 52 zones without using experimental data.","rationale":"The reader's verdict is CONDITIONAL, and the rationale already lists the post-hoc Brillouin zone selection as one of the main reasons. My stress-test agrees that this is the most load-bearing concern, but I would emphasize it more sharply: it is not merely a missing uncertainty estimate, but a circular validation step. Using the experimental spectrum to select the simulated spectrum that is then compared to that same experiment invalidates the comparison as a test of predictive power. The paper is otherwise honest about limitations, provides reproducible code and data, and the workflow is a plausible integration of established tools. The central claim is a demonstration of potential, not a benchmark, so this flaw does not require rejection; it does require that acceptance be conditional on reporting the full distribution over Brillouin zones or a pre-specified selection rule. Since the reader already reached CONDITIONAL and identified this issue in the rationale, my verdict is UNCHANGED. The disagreement with the reader's stated weakest_assumption is only partial: the reader focused on instrument resolution and quantum-correction approximations, which are also real but are explicitly acknowledged and bounded by the paper's own discussion, whereas the BZ selection affects the validity of the key experimental comparison itself.","tokens_in":891,"tokens_out":994,"duration_ms":48618,"concrete_test":"Recompute the 1500 K X-point intensity for all 52 Brillouin zones and compare each simulated spectrum against the experimental spectrum of Ref. 51 without any selection. Report the full distribution of mean-squared errors, overlay the full simulated spread on the experimental data, and determine whether the selected 'best' zone lies within the central bulk of the distribution or is a clear outlier. A second, cleaner check: re-run the comparison with a pre-specified selection rule that does not use the experimental spectrum, such as 'use the third Brillouin zone, q = [0.5, -1.0, -1.5],' and see whether the reported agreement survives. If the best-matching zone is an outlier or the pre-specified rule gives substantially worse agreement, the 1500 K multi-phonon and anharmonicity claim in Fig. 2c is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is in Section III A, specifically the 1500 K silicon comparison. The text states: 'we selected the Brillouin zone for which the simulated spectrum best reproduces the experimental data, based on the mean-squared error calculated over the spectrum.' This means the experimental spectrum was used to choose which of the 52 computed S(q,ω) curves to display, and the chosen curve is then presented as evidence that the workflow 'captures both the anharmonicity and multi-phonon effects present in the experimental data.' Because Fig. 4a shows that the intensity of the multi-phonon shoulder near 30 meV varies greatly between Brillouin zones, the selected curve is likely an outlier of the ensemble. The validation of temperature-dependent anharmonicity therefore rests on a selection rule that uses the very data the simulation is supposed to predict. The paper discloses this choice, but the central claim of predictive capability is weakened by it: a workflow intended to guide experiments cannot require the target experiment to select the simulation point. The other demonstrations do not use this selection, but the silicon temperature series is the only demonstration of anharmonicity as a function of temperature.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a computational workflow for predicting inelastic neutron scattering (INS) spectra from first principles. The workflow combines density functional theory, neuroevolution-potential (NEP) machine-learned interatomic potentials, molecular dynamics (including path-integral MD), autocorrelation analysis via dynasor, instrument-specific resolution functions and kinematic constraints, and a first-order quantum correction factor. The authors demonstrate the workflow on three systems: crystalline silicon at several temperatures compared with ARCS data, crystalline benzene at 127 K compared with TOSCA data, and hydrogenated Sc-doped BaTiO3 at various Sc concentrations compared with IN1 Lagrange, TOSCA, and MAPS data. The paper reports good agreement in peak positions and many relative intensities, while also identifying limitations for higher-order multi-phonon features and systematic shifts attributed to the underlying DFT functionals.","tokens_in":16366,"tokens_out":4002,"duration_ms":42933,"significance":"If the validation concerns are addressed, this is a valuable methods contribution. The workflow is carefully assembled from open-source components, the models and training data are made available on Zenodo, and the central pipeline from MD trajectories to instrument-specific INS predictions is plausible and reproducible. The paper does not reduce any derived equation to a fitted parameter; the free parameters are limited to normalization/scaling choices and the selection of a specific Brillouin zone in one comparison. The demonstrations span three chemically different systems and four instruments, which gives the workflow breadth. The explicit recognition of the limitations of the first-order quantum correction and the comparison with harmonic AbINS calculations is a strength. The main weakness is that the strongest validation claim, particularly for temperature-dependent anharmonicity in silicon, relies on a partly circular selection procedure that must be repaired before the predictive claim can be accepted.","major_comments":[{"comment":"The 1500 K silicon validation is partly circular and therefore cannot support the claim that the workflow captures anharmonicity and multi-phonon effects as a function of temperature. The text states that the Brillouin zone shown in Fig. 2c was selected because it minimizes the mean-squared error against the experimental spectrum, and Fig. 4a shows that the multi-phonon intensity near 30 meV varies substantially among the 52 zones. With this selection rule, the displayed curve is not an independent prediction; it is the best-fitting zone chosen using the target data. Please report the full distribution over all 52 zones, for example as an ensemble band or as an average with spread, and draw the validation conclusion from the distribution without using the experimental spectrum to select the displayed simulation.","section":"III A, Fig. 2c and Fig. 4"},{"comment":"The paper acknowledges that the quantum correction in Eq. (6) is valid only for first-order Stokes scattering, and Section III C shows the consequence: the 250 meV overtone is underestimated and the 550 meV combination mode is absent from the MD workflow. Given these results, the Discussion's statement that 'almost all experimental features . . . are faithfully reproduced' and that relative intensities are reproduced 'after applying correction factors for quantum statistics' overstates what is demonstrated. Please restrict the validation claim to first-order features, state explicitly that higher-order multi-phonon intensities are not yet predictive in the MD-based workflow, and temper the wording in Section III D accordingly.","section":"II C, Eq. (6); III C; III D"},{"comment":"The benzene comparison uses per-spectrum intensity rescaling: the caption states that the simulated spectra are individually scaled to match the experiment above 50 meV. Because the relative intensity between high- and low-energy regions is part of the quantitative validation, please clarify whether this is a single fixed scale factor for each spectrum or a free normalization, and report how the scaling affects the comparison. The current description makes it difficult to judge how much of the apparent agreement comes from this adjustable normalization.","section":"III B, Fig. 5"}],"minor_comments":[{"comment":"The sentence 'The first step is the the construction' contains a duplicated article; please correct it.","section":"II, first paragraph"},{"comment":"The axis label 'Freuency' should be 'Frequency'.","section":"Fig. 5(b)"},{"comment":"The shared color scale clips the intensity in panel (a); consider using per-panel normalization or an unclipped supplementary version so that the fundamental-mode intensity is visible.","section":"Fig. 7"},{"comment":"The phrase 'the dispersion in (a) and intensity in (b) is aggregated' has a subject-verb disagreement; it should read 'are aggregated'.","section":"Fig. 2 caption"},{"comment":"The data availability statement gives Zenodo links for the benzene and Sc-doped BaTiO3 models but not for the silicon NEP model; since the silicon model is taken from Ref. [17], please state explicitly where that model can be obtained.","section":"VI, Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong methods demonstration and the central workflow is credible, but the silicon validation at 1500 K uses a post-hoc Brillouin-zone selection that undermines the strongest predictive claim. The other two demonstrations do not suffer from this circularity and can be reported more cautiously. The quantum-correction limitation is disclosed but the Discussion overstates the agreement; I would ask the authors to align the claims with the evidence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look: this is an integration paper, not a method breakthrough, but it is a good integration. The authors chain DFT to NEP machine-learned potentials, gpumd MD (including PIMD), dynasor correlation functions, instrument resolution and kinematic masks, and a quantum correction, then test the chain on Si, benzene, and Sc-doped BaTiO3 against four spectrometers. They also ship new NEP potentials, training data, and a fresh TOSCA benzene spectrum on Zenodo. That is real work and a useful reference point for anyone planning beamtime or interpreting INS spectra. The benzene section makes a nice didactic point: raw MD S(q,omega) is not enough; resolution, kinematics, and quantum statistics each matter.\n\nThe central claim, however, is stronger than the validation. The 1500 K Si comparison is partly circular: the text says they selected the Brillouin zone that gave the best mean-squared error against the experimental spectrum, and Fig. 4 shows the multi-phonon shoulder near 30 meV varies strongly across the 52 zones. So the simulation is not predicting the temperature-dependent anharmonic signal; it is selecting one of 52 postdictions to match the experiment. The paper discloses this choice, which is to the authors' credit, but disclosure does not make it a prediction. The other two systems do not have this problem, and the peak positions in benzene and BaTiO3 look credible.\n\nThe secondary soft spots are in proportion. The spectra are individually scaled, so the validation of relative intensities is weaker than it first appears. The instrument resolution function is a 1D Gaussian, although the authors note the true function is 4D and non-Gaussian. And Eq. (6) is a first-order Stokes correction; the paper itself shows the 250 meV overtone in BaTiO3 is underestimated and the 550 meV combination mode is not captured. These are honest limitations, and they do not kill the workflow, but they should temper the 'remarkable quantitative agreement' language in Section III D and the abstract.\n\nWho is this for? Someone building or using MLIP-based scattering prediction pipelines, and neutron facility users who want a practical starting point. It is not reshaping a field, but it is a solid, citable reference. It deserves peer review: the integration is reproducible, the data are out, and the flaws are fixable in revision. I would send it to a serious referee, with a request to report the full distribution of spectra over Brillouin zones or pre-specify a selection rule that does not use the experimental spectrum, and to soften the agreement claims.","headline":"A useful, reproducible integration of MLIP-based MD with instrument-specific INS prediction, held back by one partly circular validation choice and overstated conclusions.","tokens_in":767,"tokens_out":834,"would_cite":true,"duration_ms":26404,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Simulated neutron spectra from first-principles dynamics match measurements on four spectrometers.","keywords":["inelastic neutron scattering","machine-learned interatomic potentials","molecular dynamics","dynamic structure factor","quantum correction","instrument resolution function","kinematic constraint","anharmonicity"],"falsifier":"Measure the absolute intensity of a clean, isolated overtone band in a crystal and compare it with the workflow's prediction: the first-order quantum correction fixes its intensity, so a factor-of-two disagreement would falsify that correction layer. The paper already shows this test partially failing in hydrogenated scandium-doped barium titanate, where the 250 meV overtone is underestimated and the 550 meV combination mode is absent.","tokens_in":15950,"feed_emoji":"⚛️","tokens_out":7304,"duration_ms":72097,"temperature":0.7,"pith_summary":"This paper establishes that a purely computational workflow can predict neutron scattering experiments closely enough to guide and interpret them. The chain starts with density functional theory data, trains machine-learned interatomic potentials, runs large molecular dynamics simulations, and converts the trajectories into instrument-specific inelastic neutron scattering spectra. Tested on crystalline silicon, crystalline benzene, and hydrogenated scandium-doped barium titanate, the simulated spectra reproduce nearly every vibrational peak seen on four different neutron spectrometers. The result matters because it gives materials scientists a way to plan, validate, and interpret scattering experiments without waiting for beamline access.","feed_headline":"Neutron spectra from first principles match four spectrometers","feed_subtitle":"DFT-trained potentials plus molecular dynamics reproduce vibrational peaks, anharmonicity, and crystal-phase differences.","key_machinery":"The load-bearing object is the dynamic structure factor $S(\\mathbf{q},\\omega)$, the time Fourier transform of the intermediate scattering function $F(\\mathbf{q},t)$, which is the autocorrelation of the Fourier-transformed particle density. $S(\\mathbf{q},\\omega)$ is proportional to the intensity measured in a scattering experiment, and molecular dynamics supplies it directly, including anharmonic and multi-phonon content that harmonic phonon models must add by hand. What turns a generic spectrum into a prediction is the instrument-specific layer: a kinematic mask from the time-of-flight geometry, an energy-dependent Gaussian resolution function, and the first-order Stokes-Raman quantum correction $\\beta\\hbar\\omega/(1-e^{-\\beta\\hbar\\omega})$ applied to correct for classical phonon statistics. Each component is needed; the benzene comparison shows the raw spectrum improves markedly only after both the kinematic constraint and the quantum correction are applied.","core_discovery":"The paper's central claim is that inelastic neutron scattering spectra can be computed from first principles with quantitative fidelity, provided every step of the simulation is made to speak the instrument's language. Molecular dynamics trajectories from machine-learned potentials are transformed into the dynamic structure factor $S(\\mathbf{q},\\omega)$; the result is weighted by species-dependent neutron scattering lengths, masked by the spectrometer's kinematic constraint, convolved with an energy-dependent Gaussian resolution function, and multiplied by a first-order quantum correction factor that restores Bose statistics to the classically occupied phonon modes. Across silicon, benzene, and hydrogenated scandium-doped barium titanate, this procedure reproduces essentially all experimental vibrational peaks, including anharmonic mode softening with temperature, multi-phonon intensity between acoustic and optical branches, and the spectral differences between competing crystal phases. The paper states the agreement as quantitative and notes that almost all experimental features are faithfully reproduced.","pith_inferences":["A natural next stress test is a disordered or liquid system, where the same dynamic structure factor pipeline should capture diffusive quasi-elastic broadening; the paper asserts applicability but does not demonstrate it.","If the workflow is adopted at user facilities, simulated instrument responses could be inverted against desired scientific outcomes to optimize which spectrometer, energy range, and counting time a proposal should request.","The high-order combination-mode failure suggests a targeted extension: apply higher-order quantum corrections only in energy windows where an independent harmonic analysis identifies multi-phonon contributions, leaving first-order regions unchanged.","General-purpose machine-learned potentials could replace bespoke training for new materials, turning the workflow into a nearly turnkey predictive tool at the cost of heavier per-simulation compute."],"forward_implications":["Simulated spectra can be used to identify and assign measured vibrational modes, since the workflow yields the spectrum and the underlying phonon dispersion from the same potential.","The method differentiates crystal phases by comparing simulated spectra of each phase against experiment, demonstrated by distinguishing hexagonal from cubic scandium-doped barium titanate as doping varies.","Anharmonicity, thermal expansion, and multi-phonon effects are included inherently, so the workflow predicts mode softening and finite-temperature spectral changes without perturbative corrections.","The same dynamic-structure-factor pipeline extends to diffraction and to X-ray or electron scattering by changing the weighting factors, as the paper argues.","The remaining quantitative gap, underestimated overtone and missing combination-mode intensities, can be reduced by higher-order quantum corrections, although the paper recommends the lowest-order correction for general guidance."],"supporting_citations":[{"why":"Supplies the molecular dynamics engine and the silicon machine-learned potential used for the large-scale trajectories.","marker":"[17]"},{"why":"Provides the tool that extracts the dynamic structure factor from molecular dynamics trajectories.","marker":"[18]"},{"why":"Extends the autocorrelation analysis to the multi-species, instrument-ready dynamic structure factors used here.","marker":"[19]"},{"why":"Applies instrument-specific resolution functions and kinematic constraints to the calculated spectra.","marker":"[29]"},{"why":"Defines the energy-dependent resolution functions for the four spectrometers.","marker":"[30]"},{"why":"Justifies the first-order quantum correction factor applied to all simulated spectra.","marker":"[33]"},{"why":"Supplies the Stokes-Raman scattering basis for the quantum correction factor.","marker":"[34]"},{"why":"Provides the experimental silicon spectrum measured on a wide-range spectrometer that the silicon simulation is compared against.","marker":"[51]"},{"why":"Provides the experimental spectra for hydrogenated scandium-doped barium titanate on three spectrometers used for comparison.","marker":"[56]"},{"why":"Describes the spectrometer on which the new crystalline benzene measurement was taken.","marker":"[42]"}],"fun_headline_variants":["ML workflow predicts neutron spectra from first principles","Neutron spectra from first principles: ML workflow matches experiments","DFT + ML potentials reproduce neutron scattering peaks","Simulated INS spectra agree with four spectrometers","First-principles neutron spectra match experiments across four instruments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that simplified corrections—a one-dimensional Gaussian resolution function and a first-order quantum factor—accurately convert classical molecular dynamics spectra into measured neutron intensities; the paper itself notes the true resolution is four-dimensional and non-Gaussian and that higher-order scattering processes require corrections this workflow does not apply.","fun_headline_variants_meta":{"raw":{"variants":["ML workflow predicts neutron spectra from first principles","Neutron spectra from first principles: ML workflow matches experiments","DFT + ML potentials reproduce neutron scattering peaks","Simulated INS spectra agree with four spectrometers","First-principles neutron spectra match experiments across four instruments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2308,"prompt_tokens":937,"completion_tokens":1371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1297}},"tokens_in":553,"tokens_out":1371,"duration_ms":10704,"temperature":1.0,"reasoning_tokens":1297,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:54:48.496884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the absolute intensity of a clean, isolated overtone band in a crystal and compare it with the workflow's prediction: the first-order quantum correction fixes its intensity, so a factor-of-two disagreement would falsify that correction layer. The paper already shows this test partially failing in hydrogenated scandium-doped barium titanate, where the 250 meV overtone is underestimated and the 550 meV combination mode is absent.","supporting_citations":[{"cited_title":"Cardona and G","cited_arxiv_id":null,"evidence_quote":"Supplies the Stokes-Raman scattering basis for the quantum correction factor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the molecular dynamics engine and the silicon machine-learned potential used for the large-scale trajectories."},{"cited_title":"Dynasor 2: From Simulation to Experiment Through Correlation Functions","cited_arxiv_id":"2503.21957","evidence_quote":"Extends the autocorrelation analysis to the multi-species, instrument-ready dynamic structure factors used here."},{"cited_title":"Turanyi, A","cited_arxiv_id":null,"evidence_quote":"Defines the energy-dependent resolution functions for the four spectrometers."},{"cited_title":"Rosander, E","cited_arxiv_id":null,"evidence_quote":"Justifies the first-order quantum correction factor applied to all simulated spectra."},{"cited_title":"Monacelli, R","cited_arxiv_id":null,"evidence_quote":"Provides the experimental spectra for hydrogenated scandium-doped barium titanate on three spectrometers used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the spectrometer on which the new crystalline benzene measurement was taken."}],"review_version":1}