Pith. sign in

REVIEW 3 major objections 5 minor 54 references

PhononBench-MP40: a spectrum-resolved benchmark dataset for phonon stability

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper presents a benchmark dataset of 46,899 crystal phonon spectra whose stability labels are deterministic functions of the released spectra, so every stable/unstable tag can be reproduced and re-thresholded.

desk verdict A useful, honestly scoped phonon-stability dataset that needs an executed audit and a live DOI before it fully delivers on its auditable-reuse promise. read the letter →

arxiv 2607.22573 v1 pith:QK6WA3TY submitted 2026-06-04 cs.AI cond-mat.mtrl-sciphysics.comp-ph

Wen-Kao Li , Ze-Feng Gao , Zhong-Yi Lu This is my paper · ORCID
classification cs.AIcond-mat.mtrl-sciphysics.comp-ph
keywords phononstabilityspectrabenchmarkdatasetmaterialsinformaticsminimumfrequencyimaginarymodesworkflow-definedlabelingdatarelease
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper presents a spectrum-resolved benchmark dataset in which workflow-defined phonon stability is made auditable. The central claim is that 46,899 crystal records — 16,683 stable and 30,216 unstable — can be released together with the local band-frequency spectra from which those labels were derived, under one fixed rule: any sampled band frequency below −10⁻³ THz means unstable. It also separates 1,067 relaxation failures that produced no spectrum from the completed denominator, so failed calculations are not mistaken for physical instabilities. A sympathetic reader should care because large phonon datasets typically offer only binary flags, hiding the difference between a shallow soft mode, a deep imaginary branch, and a missing result; here the spectrum behind each label is the release object, making classification reproducible, threshold studies possible, and failure-aware triage a distinct task. If the data hold up, the field gains a shared target for stability classification, minimum-frequency regression, and workflow-robustness studies.

What carries the argument

The central object is the local YAML spectrum: the file storing the sampled high-symmetry-path band frequencies for each completed calculation. It carries the evidence behind every label — the dispersion, the minimum frequency ωmin, and the threshold margin — and it is the object through which all derived uses flow: reconstructing dispersions, extracting ωmin, and re-deriving labels under new thresholds. The two supporting mechanisms are the completion rule (a completed record is the intersection of a recovered workflow label and a matching YAML output, which yields 46,899 records and sends 1,067 relaxation failures to a separate table) and the deterministic label threshold (ω < −10⁻³ THz ⇒

What would settle it

Re-run the release's own audit: take the master task table and the released YAML paths, count how many structure-name tokens fail to resolve or resolve to multiple YAML files, then recompute the completed label from every YAML by applying the −10⁻³ THz rule and compare with the completed table's labels. If the pairwise counts differ from 46,899/16,683/30,216, or if any released YAML gives a different label than the table, the central audit claim is falsified.

Watch

Extended reading notes

Core claim

The core discovery is a dataset design: each completed phonon-stability record consists of a recovered workflow label joined to a matching local YAML spectrum, and the manuscript label is not an independent model output but a derived view of that spectrum. The label rule is deterministic and simple — unstable iff any sampled high-symmetry-path band frequency is below −10⁻³ THz; otherwise stable — so users can re-extract the minimum sampled frequency and reproduce the binary label from the YAML file. The completed cohort totals 46,899 records (16,683 stable, 30,216 completed-phonon unstable), with 1,067 relaxation failures listed separately. The release also supplies a formula-group split int

Load-bearing premise

The load-bearing premise is that the 46,899 completed records are correctly paired label-plus-YAML triples joined on the structure-name token, so that the released counts and threshold-derived labels are internally consistent; if token collisions occurred or the raw labels used a different threshold, the numbers would not mean what they appear to mean.

Editorial extensions

If this is right

  • Anyone can verify a label by parsing the released YAML and applying the −10⁻³ THz rule; no label table has to be trusted on its own.
  • The same released objects support minimum-frequency regression, so models can distinguish near-threshold soft modes from deep imaginary branches instead of collapsing them into one 'unstable' class.
  • Because the formula-group split keeps every formula in exactly one partition, model comparisons on the completed cohort avoid exact-formula leakage.
  • The 1,067 relaxation failures form a separate failure-aware triage target: predicting whether a workflow produces an auditable spectrum, which should not be mixed with predicting instability.
  • Threshold studies become measurable: moving the −10⁻³ THz cut and relabeling gives a quantitative map of borderline records, useful for calibration in screening pipelines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper stops short of saying that the 64.4% unstable fraction is partly a threshold artifact, but the released spectra make that testable: reporting what fraction of 'unstable' records have ωmin between −10⁻³ and, say, −0.1 THz would show how many borderline cases drive the high instability rate.
  • A natural three-class extension not in the paper is to split completed-phonon unstable into soft-mode (near-threshold) and deep-imaginary records; this would give downstream validation a more graded screening target.
  • Because all labels inherit a single machine-learned potential, a spot-check of a few hundred near-threshold YAML spectra against higher-fidelity calculations would bound the workflow's systematic error — a test the authors flag implicitly but do not run.
  • The failure table could be modeled as a missingness signal; if relaxation failures correlate with chemistry or symmetry, ignoring them would bias any classifier trained on the completed cohort, and weighting or joint modeling would be needed.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. PhononBench-MP40 is a data-descriptor paper releasing 46,899 completed records for workflow-defined phonon stability, with paired stability labels and phonopy YAML spectra, derived from an MP40-derived task cohort of 47,969 tasks. The authors report 16,683 Stable and 30,216 completed-phonon unstable records, plus 1,067 relaxation failures reported outside the completed denominator. The central design is that the stability label is derived from each released YAML spectrum by a deterministic threshold rule (unstable iff any sampled band frequency is below −1e-3 THz, S3), making the labels auditable and threshold-refinable. The paper also provides a formula-group reference split, coverage statistics, benchmark tasks, and a stated boundary that the labels are workflow-defined rather than claims of absolute physical stability.

Significance. If the dataset is correct and accessible, it would be a valuable resource for materials informatics: it is substantially larger than most open phonon-stability datasets, it pairs labels with the underlying spectral object, it explicitly separates relaxation failures from completed phonon instabilities, and it provides a deterministic label rule and split. The paper also benefits from a consistent internal accounting of task, label, YAML, and completed-record counts, and from explicit discussion of interpretation boundaries. The main scientific value depends on verifiability of the label–spectrum pairing, which is currently asserted rather than demonstrated.

major comments (3)
  1. [§6 and S10, Table S5] The central claim of the paper is that every completed label is deterministically derived from the released YAML spectrum, so that users can re-extract ωmin and reproduce the label. However, no executed validation is reported. Section S10 lists a 'Threshold relabeling' check, a 'YAML path existence' check, a 'Split leakage check', and other integrity checks, but the paper gives no results for any of them. This is load-bearing: the 16,683/30,216 split and the auditable-reuse promise stand or fall on whether parsing the released YAMLs and applying the S3 rule reproduces the completed labels. I ask that the authors provide, in the release or supplement, a runnable audit script and a summary table reporting, over all 46,899 completed records: (a) the number of YAML paths that resolve, (b) the number of YAML files containing parseable high-symmetry-path frequencies, (c) the number of records
  2. [S2 and S3] The relationship between the raw 'workflow_label' and the derived 'completed_label' is not made explicit. Section S2 defines workflow_label as the raw PhononBench label and completed_label as the label 'after enforcing the label+YAML completion rule', while S3 states that the completed stability label is derived from high-symmetry-path frequencies using the −1e-3 THz threshold. If the raw PhononBench workflow_label was produced by a different threshold, path convention, or post-processing step, then the completed_label might not equal the threshold-derived label for some records. The paper never reports the agreement between workflow_label and threshold-derived completed_label. I request a cross-tabulation (or, at minimum, a statement that completed_label was recomputed from the YAML frequencies using S3 and that the raw labels were not used except as an audit cross-check).
  3. [Data availability statement] The data availability statement says the dataset 'is being archived' and that the DOI 'should be cited once registration is active'. This means the central artifact is not available for reviewers or readers to verify any of the claims in the paper. For a data-descriptor paper, an active and accessible archival record is not a minor detail; it is the primary evidence. I recommend that acceptance be conditioned on an active Science Data Bank DOI with the completed table, split table, failure table, metadata, checksum file, and the 46,899 YAML spectra available under the claimed paths.
minor comments (5)
  1. [Table 1 and §2] The accounting is internally consistent, but the three YAML outputs that do not have a matching label (46,902 YAMLs vs 46,899 completed records) are not discussed. A sentence explaining what those three records are would help users understand the exact matching rule.
  2. [S2, structure_name] The matching key structure_name is defined as a 'workflow key' and source_id as the first token. Given that labels, YAML paths, and table rows are joined on this token, a uniqueness audit of structure_name in the master, completed, and failure tables should be included in the S10 validation results. This is a minor presentation point only if the proposed audit passes; it is part of the major concern if it has not been run.
  3. [§9 and S8] The discussion of limitations is clear and appropriately cautious. However, the phrase 'within the adopted tolerance' for MgB2 (ωmin = 0.0000 THz) may mislead readers into thinking there is a nonzero tolerance on the stable side. The rule is a one-sided threshold (ω < −1e-3 THz), so a zero minimum is consistent with Stable. Consider stating this explicitly to avoid confusion.
  4. [S5] The SHA-256 split rule should specify the exact byte representation of the formula key (e.g., UTF-8 encoding of the parsed formula token) so that users can independently reproduce the split. This is a reproducibility detail, not a correctness issue.
  5. [§7, baseline protocol] The recommended reporting checklist is useful. A small addition would be to require reporting the exact PhononBench-MP40 version and Git commit of the parsing utilities, as already suggested in S3, so that any relabeling or threshold experiments can be traced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: labels are computed from released spectra by an explicit, externally checkable rule.

full rationale

The central derivation chain is YAML spectrum → ωmin → completed label under the fixed threshold ω < −10^-3 THz (S3), and the YAML files are material artifacts produced by a described PhononBench/MatterSim–phonopy workflow rather than objects defined by the labels. The label is therefore a data-processing view of an independently released spectrum, not an input defined in terms of the output. Reference [31] (PhononBench) is partly authored by the present authors, but it is used as workflow provenance; the paper independently specifies the relaxation, supercell, force, path and threshold settings and releases the spectra themselves, so no load-bearing mathematical claim reduces to the self-citation. The S10 validation checklist includes a threshold-relabeling check that is, by construction, a consistency check rather than an independent test, and the paper does not report executed results for that checklist; this is a completeness/audit limitation, not circularity. No specific reduction of a claimed prediction to its own inputs can be exhibited.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no physics entities and fits no parameters to data. The five listed free parameters are hand-chosen workflow constants that directly determine the labels (the -1e-3 THz threshold), the force constants (0.01 A displacement), the computational cost (2x2x2 supercell), the sampling (101 points per segment), and the split (0.80/0.90 hash cutoffs). The axioms are domain assumptions inherited from the MLIP-phonon paradigm: MatterSim force fidelity, path-sampling adequacy for detecting instabilities, structure_name key-matching integrity, and the validity of the inherited MP-40 task cohort. No invented entities (particles, forces, dimensions) are postulated.

free parameters (5)
  • stability threshold = -1e-3 THz
    Hand-chosen threshold defining the completed label (S3): unstable iff any sampled band frequency < -1e-3 THz. Label counts are threshold-dependent; the paper explicitly invites threshold studies.
  • finite-displacement step = 0.01 Å
    Chosen displacement distance in the finite-displacement phonon workflow (S3); affects force-constant and frequency accuracy.
  • supercell matrix = 2x2x2
    Chosen supercell construction (S3); sets practical cost (up to 320-atom force cells) and limits which phonon wavelengths are sampled.
  • band-path sampling density = 101 points per segment
    Chosen sampling along seekpath/spglib high-symmetry paths (S3); determines which q-points are checked for imaginary modes and therefore which instabilities the label can see.
  • reference split hash cutoffs = train < 0.80, validation < 0.90
    Hand-set SHA-256 hash cutoffs on formula tokens (S5); define the fixed train/validation/test partition and the reported split counts.
assumptions (4)
  • domain assumption MatterSim/uMLIP forces approximate interatomic forces accurately enough for the phonon-stability purpose across the MP40 cohort
    All 46,899 spectra inherit the MLIP's force fidelity. The paper cites prior validation (refs 30, 31, 48) and explicitly disclaims model-independent physical stability (S7, S8), making this a stated domain assumption rather than a hidden one.
  • domain assumption Instabilities relevant to the label are captured by the sampled high-symmetry path at 101 points per segment
    The label rule in S3 uses only sampled band frequencies; an imaginary mode at a q-point off the released path would be missed. The paper names 'path sampling procedure' among its boundaries (S8).
  • domain assumption structure_name key matching between MP40 task records, PhononBench workflow labels, and phonopy YAML outputs is lossless
    The completed-cohort definition (label intersect YAML, S2) presupposes correct key joining. Counts are internally consistent (46,899 + 1,067 = 47,966) but no collision or disagreement audit is reported in the text.
  • domain assumption The inherited MP-40 task cohort is a valid, correctly scoped atom-count window
    The 47,969-task audit scope is taken from MP-20 / MP-21-40 / MP-40 benchmark settings (refs 1, 26, 27); N_atoms <= 40 recorded cells with 2x2x2 supercells define the release domain (S1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PhononBench-MP40: a spectrum-resolved benchmark dataset for phonon stability." pith.science (2026). https://pith.science/paper/QK6WA3TY

@misc{pith2026260722573,
  author       = {Pith},
  title        = {Pith review of: PhononBench-MP40: a spectrum-resolved benchmark dataset for phonon stability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QK6WA3TY}},
  note         = {Machine review of arXiv:2607.22573}
}
read the original abstract

Imaginary phonon modes remain a practical bottleneck in computational materials screening because otherwise plausible structures can be locally dynamically unstable under a chosen workflow. Here we present PhononBench-MP40, a spectrum-resolved benchmark dataset of Materials Project-derived crystals for workflow-defined phonon stability. The dataset starts from 47,969 MP40 workflow tasks and provides 46,899 completed records with paired stability labels and local phonopy YAML spectra, including 16,683 Stable records and 30,216 completed-phonon unstable records. A further 1,067 relaxation failures are reported separately rather than merged into the completed phonon denominator. The release centers on the local YAML spectrum: the stability label, the lowest sampled frequency and any threshold-dependent relabeling are derived from that spectrum. The dataset is openly available through Science Data Bank at https://doi.org/10.57760/sciencedb.38735. A companion GitHub repository provides the calculation code and lightweight access utilities. PhononBench-MP40 provides an auditable reference for workflow-defined stability classification, minimum-frequency analysis, threshold studies and failure-aware triage, while keeping the reference workflow, data schema and interpretation boundaries explicit.

Figures

Figures reproduced from arXiv: 2607.22573 by the authors.

Figure 1
Figure 1. Workflow and release object of PhononBench-MP40. The dataset starts from 47,969 MP40 tasks and follows the PhononBench route through structural relaxation, finite-displacement supercell construction, MatterSim/uMLIP force evaluation and phonopy processing. The local YAML output is the primary release object because it enables dispersion reconstruction, minimum-frequency extraction and stability-label derivation. The… view at source ↗
Figure 2
Figure 2. Representative phonon spectra behind the stability labels. Three YAML-derived dispersions illustrate the label rule and the information retained beyond the binary decision. a, MgB2 is labeled Stable, with ωmin = 0.0000 THz within the adopted tolerance. b, RbPbCl3 is a soft-mode completed-phonon unstable example with ωmin = −1.4984 THz. c, BaTiO3 is a deep-imaginary completed-phonon unstable example with ωmin = −7.91… view at source ↗
Figure 3
Figure 3. Chemical and structural coverage of the completed cohort. All panels are computed over the 46,899 completed label+YAML records. a, Distribution of recorded unit-cell atom counts within the MP40 scope, with prominent peaks at 8, 20 and 40 atoms. b, Element occurrence, counted once per completed record when an element appears in the parsed formula. c, Ranked space-group occurrence, showing broad but uneven crystallogr… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references

  1. [1]

    Crystal diffusion variational autoencoder for periodic material generation, 2022

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crystal diffusion variational autoencoder for periodic material generation, 2022

    2022

  2. [2]

    Crystal structure prediction by joint equivariant diffusion, 2023

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal structure prediction by joint equivariant diffusion, 2023

    2023

  3. [3]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin D

    Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin D. Cubuk. Scaling deep learning for materials discovery.Nature, 624:80–85, 2023

    2023

  4. [4]

    A generative model for inorganic materials design.Nature, 639:624–632, 2025

    Claudio Zeni, Robert Pinsler, Daniel Zugner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbe, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Bichlien Nguyen, Hannes Schulz, Sarah Lewis, Chin-Wei Huang, Ziheng Lu, Yichi Zhou, Han Yang, Hongxia Hao, Jielan Li, Chu Yang, et al. A generative model for inorgani...

    2025

  5. [5]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The materials project: A materials genome approach to accelerating materials innovation.APL Materials, 1(1):011002, 2013

    2013

  6. [6]

    Stefano Curtarolo, Wahyu Setyawan, Gus L. W. Hart, Michal Jahnatek, Roman V. Chepulskii, Richard H. Taylor, Shidong Wang, Junkai Xue, Kesong Yang, Ohad Levy, Michael J. Mehl, Harold T. Stokes, Denis O. Demchenko, and Dane Morgan. Aflow: An automatic framework for high-throughput materials discovery.Computational Materials Science, 58:218–226, 2012

    2012

  7. [7]

    Saal, Scott Kirklin, Muratahan Aykol, Bryce Meredig, and C

    James E. Saal, Scott Kirklin, Muratahan Aykol, Bryce Meredig, and C. Wolverton. Materials de- sign and discovery with high-throughput density functional theory: The open quantum materials database (oqmd).JOM, 65(11):1501–1509, 2013

    2013

  8. [8]

    Saal, Bryce Meredig, Alex Thompson, Jeff W

    Scott Kirklin, James E. Saal, Bryce Meredig, Alex Thompson, Jeff W. Doak, Muratahan Aykol, Stephan Ruhl, and C. Wolverton. The open quantum materials database (oqmd): assessing the accuracy of dft formation energies.npj Computational Materials, 1:15010, 2015

    2015

Show all 54 references
  1. [9]

    Garrity, Andrew C

    Kamal Choudhary, Kevin F. Garrity, Andrew C. E. Reid, Brian DeCost, Adam J. Biacchi, An- gelaR.HightWalker, ZacharyTrautt, JasonHattrick-Simpers, A.GiladKusne, AndreaCentrone, Albert Davydov, Jie Jiang, Ruth Pachter, Gowoon Cheon, Evan Reed, Ankit Agrawal, Xiaofeng Qian, Vishu...

    2020

  2. [10]

    Nomad: The fair concept for big data-driven materials science.MRS Bulletin, 43(9):676–682, 2018

    Claudia Draxl and Matthias Scheffler. Nomad: The fair concept for big data-driven materials science.MRS Bulletin, 43(9):676–682, 2018

    2018

  3. [11]

    Yakutovich, Valeria Granata, Fer- nando Gargiulo, Marco Borelli, Martin Uhrin, Sebastiaan P

    Leopold Talirz, Snehal Kumbhar, Elsa Passaro, Aliaksandr V. Yakutovich, Valeria Granata, Fer- nando Gargiulo, Marco Borelli, Martin Uhrin, Sebastiaan P. Huber, Spyros Zoupanos, Carl S. Adorf, Casper W. Andersen, Ole Schutt, Carlo A. Pignedoli, Daniele Passerone, Joost VandeVon...

    2020

  4. [12]

    Chevrier, Kristin A

    Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open-source python library for materials anal- ys...

    2013

  5. [13]

    Groves, Bjork Hammer, Cory Hargus, Eric D

    AskHjorthLarsen, JensJorgenMortensen, JakobBlomqvist, IvanoE.Castelli, RuneChristensen, Marcin Dulak, Jesper Friis, Michael N. Groves, Bjork Hammer, Cory Hargus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kols- bjerg, ...

    2017

  6. [14]

    Anubhav Jain, Shyue Ping Ong, Wei Chen, Bharat Medasani, Xiaohui Qu, Michael Kocher, Miriam Brafman, Guido Petretto, Gian-Marco Rignanese, Geoffroy Hautier, Dan Gunter, and Kristin A. Persson. Fireworks: a dynamic workflow system designed for high-throughput appli- cations.Con...

    2015

  7. [15]

    Kiran Mathew, Joseph H. Montoya, Alireza Faghaninia, Shyam Dwaraknath, Muratahan Aykol, Hanmei Tang, Iek-Heng Chu, Tess Smidt, Brandon Bocklund, Matthew Horton, John Dagdelen, Brandon Wood, Zi-Kui Liu, Jeffrey B. Neaton, Shyue Ping Ong, Kristin A. Persson, and Anubhav Jain. At...

    2017

  8. [16]

    Aiida: automated interactive infrastructure and database for computational science.Computational Materials Science, 111:218–230, 2016

    Giovanni Pizzi, Andrea Cepellotti, Riccardo Sabatini, Nicola Marzari, and Boris Kozinsky. Aiida: automated interactive infrastructure and database for computational science.Computational Materials Science, 111:218–230, 2016

    2016

  9. [17]

    Guido Petretto, Shyam Dwaraknath, Henrique P. C. Miranda, Donald Winston, Matteo Gi- antomassi, Michiel J. van Setten, Xavier Gonze, Kristin A. Persson, Geoffroy Hautier, and Gian- Marco Rignanese. High-throughput density-functional perturbation theory phonons for inorganic ma...

    2018

  10. [18]

    Plata, Pinku Nath, Demet Usanmaz, Jesus Carrete, Cormac Toher, Maarten de Jong, Mark Asta, Marco Fornari, Marco Buongiorno Nardelli, and Stefano Curtarolo

    Jose J. Plata, Pinku Nath, Demet Usanmaz, Jesus Carrete, Cormac Toher, Maarten de Jong, Mark Asta, Marco Fornari, Marco Buongiorno Nardelli, and Stefano Curtarolo. An efficient and accurate framework for calculating lattice thermal conductivity of solids: Aflow–aapl automatic ...

    2017

  11. [19]

    Finding un- precedentedly low-thermal-conductivity half-heusler semiconductors via high-throughput materi- als modeling.Physical Review X, 4:011019, 2014

    Jesus Carrete, Wu Li, Natalio Mingo, Shidong Wang, and Stefano Curtarolo. Finding un- precedentedly low-thermal-conductivity half-heusler semiconductors via high-throughput materi- als modeling.Physical Review X, 4:011019, 2014

    2014

  12. [20]

    Castelli, Andrea Cepellotti, Giovanni Pizzi, and Nicola Marzari

    Nicolas Mounet, Marco Gibertini, Pedro Schwaller, Andrius Merkys, Ivano E. Castelli, Andrea Cepellotti, Giovanni Pizzi, and Nicola Marzari. Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds.Nature Nanotechnology, 13:246–...

    2018

  13. [21]

    Katcho, and Natalio Mingo

    Wu Li, Jesus Carrete, Nebil A. Katcho, and Natalio Mingo. Shengbte: A solver of the boltzmann transport equation for phonons.Computer Physics Communications, 185(6):1747–1758, 2014

    2014

  14. [22]

    Anharmonic force constants extracted from first-principles molecular dynamics: applications to heat transfer simulations.Journal of the Physical Society of Japan, 83:074702, 2014

    Terumasa Tadano, Yoshihiro Gohda, and Shinji Tsuneyuki. Anharmonic force constants extracted from first-principles molecular dynamics: applications to heat transfer simulations.Journal of the Physical Society of Japan, 83:074702, 2014

    2014

  15. [23]

    Abrikosov, and Sergei I

    Olle Hellman, Igor A. Abrikosov, and Sergei I. Simak. Temperature dependent effective potential method for accurate free energy calculations of solids.Physical Review B, 84:180301, 2011

    2011

  16. [24]

    Katsnelson, and Sven P

    Petros Souvatzis, Olle Eriksson, Mikhail I. Katsnelson, and Sven P. Rudin. The self-consistent ab initio lattice dynamical method.Computational Materials Science, 44(3):888–894, 2009. 11

    2009

  17. [25]

    Masato Ohnishi, Tianqi Deng, Pol Torres, Zhihao Xu, Terumasa Tadano, Haoming Zhang, Wei Nong, Masatoshi Hanai, Zeyu Wang, Michimasa Morita, Zhiting Tian, Ming Hu, Xiulin Ruan, Ryo Yoshida, Toyotaro Suzumura, Lucas Lindsay, Alan J. H. McGaughey, Tengfei Luo, Kedar Hippalgaonkar...

    2026

  18. [26]

    An in- vertible, invariant crystal representation for inverse design of solid-state materials using generative deep learning.Nature Communications, 14:7027, 2023

    Hang Xiao, Rong Li, Xiaoyang Shi, Yan Chen, Liangliang Zhu, Xi Chen, and Lei Wang. An in- vertible, invariant crystal representation for inverse design of solid-state materials using generative deep learning.Nature Communications, 14:7027, 2023

    2023

  19. [27]

    Exploration of crystal chemical space using text-guided generative artificial intelligence.Nature Communications, 16:4379, 2025

    Hyunsoo Park, Anthony Onwuli, and Aron Walsh. Exploration of crystal chemical space using text-guided generative artificial intelligence.Nature Communications, 16:4379, 2025

    2025

  20. [28]

    First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    Atsushi Togo and Isao Tanaka. First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    2015

  21. [29]

    First-principles phonon calculations with phonopy and phono3py.Journal of the Physical Society of Japan, 92:012001, 2023

    Atsushi Togo. First-principles phonon calculations with phonopy and phono3py.Journal of the Physical Society of Japan, 92:012001, 2023

    2023

  22. [30]

    Mattersim: A deep learning atomistic model across elements, temperatures and pressures, 2024

    Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, Matthew Horton, Robert Pinsler, Andrew Fowler, Daniel Zugner, Tian Xie, Jake Smith, Lixin Sun, Qian Wang, Lingyu Kong, Chang Liu, Hongxia Hao, and Ziheng Lu. Ma...

    2024

  23. [31]

    Phononbench: A large-scale phonon- based benchmark for dynamical stability in crystal generation, 2025

    Xiao-Qi Han, Peng-Jie Guo, Ze-Feng Gao, and Zhong-Yi Lu. Phononbench: A large-scale phonon- based benchmark for dynamical stability in crystal generation, 2025. GitHub repository and manuscript resource

    2025

  24. [32]

    Band structure diagram paths based on crystallography.Computational Materials Science, 128:140–184, 2017

    Yoyo Hinuma, Giovanni Pizzi, Yu Kumagai, Fumiyasu Oba, and Isao Tanaka. Band structure diagram paths based on crystallography.Computational Materials Science, 128:140–184, 2017

    2017

  25. [33]

    Spglib: a software library for crystal symmetry search.Science and Technology of Advanced Materials: Methods, 4(1):2384822, 2024

    Atsushi Togo, Kohei Shinohara, and Isao Tanaka. Spglib: a software library for crystal symmetry search.Science and Technology of Advanced Materials: Methods, 4(1):2384822, 2024

    2024

  26. [34]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical Review Letters, 120:145301, 2018

    2018

  27. [35]

    Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    2019

  28. [36]

    Atomistic line graph neural network for improved materials property predictions.npj Computational Materials, 7:185, 2021

    Kamal Choudhary and Brian DeCost. Atomistic line graph neural network for improved materials property predictions.npj Computational Materials, 7:185, 2021

    2021

  29. [37]

    Benchmarking mate- rials property prediction methods: the matbench test set and automatminer reference algorithm

    Alexander Dunn, Qi Wang, Alex Ganose, Daniel Dopp, and Anubhav Jain. Benchmarking mate- rials property prediction methods: the matbench test set and automatminer reference algorithm. npj Computational Materials, 6:138, 2020

    2020

  30. [38]

    Janosh Riebesell, Rhys E. A. Goodall, Philipp Benner, Yuan Chiang, Bowen Deng, Gerbrand Ceder, Mark Asta, Alpha A. Lee, Anubhav Jain, and Kristin A. Persson. A framework to evaluate machine learning crystal stability predictions.Nature Machine Intelligence, 7:836–847, 2025. 12

    2025

  31. [39]

    Lawrence Zitnick, and Zachary Ulissi

    Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. The open...

    2020

  32. [40]

    A general-purpose machine learning framework for predicting properties of inorganic materials.npj Computational Materials, 2:16028, 2016

    Logan Ward, Ankit Agrawal, Alok Choudhary, and Christopher Wolverton. A general-purpose machine learning framework for predicting properties of inorganic materials.npj Computational Materials, 2:16028, 2016

    2016

  33. [41]

    Logan Ward, Alexander Dunn, Alireza Faghaninia, Nils E. R. Zimmermann, Saurabh Bajaj, Qi Wang, Joseph Montoya, Jiming Chen, Kyle Bystrom, Maxwell Dylla, Kyle Chard, Mark Asta, Kristin Persson, G. Jeffrey Snyder, and Ian Foster. Matminer: An open source toolkit for materials da...

    2018

  34. [42]

    Schutt, Huziel E

    Kristof T. Schutt, Huziel E. Sauceda, Pieter-Jan Kindermans, Alexandre Tkatchenko, and Klaus- Robert Muller. Schnet: A deep learning architecture for molecules and materials.The Journal of Chemical Physics, 148:241722, 2018

    2018

  35. [43]

    A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science, 2:718–728, 2022

    Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science, 2:718–728, 2022

    2022

  36. [44]

    Bartel, and Gerbrand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. Chgnet as a pretrained universal neural network potential for charge- informed atomistic modelling.Nature Machine Intelligence, 5:1031–1041, 2023

    2023

  37. [45]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature Communications, 13:2453, 2022

    2022

  38. [46]

    Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields, 2022

    2022

  39. [47]

    Elena, David P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, David P. Kovacs, Janosh Riebesell, Xavier R. Advincula, Mark Asta, William J. Baldwin, Fabian Berger, et al. A foundation model for atomistic materials chemistry.The Journal of Chemical Physics, 163:172501, 2025

    2025

  40. [48]

    Antoine Loew, Dewen Sun, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Universal machine learning interatomic potentials are ready for phonons.npj Computational Materials, 11:178, 2025. 13 Supplementary Information PhononBench-MP40: a spectrum-resolved benchmark data...

    2025

  41. [50]

    Download the archived release from Science Data Bank and verify the release tables against the accompanying checksum file, when provided

  42. [51]

    Read the completed table and select one completed record

  43. [52]

    Useyaml_relative_pathto locate the corresponding phonopy YAML file

  44. [53]

    Parse the YAML frequencies, compute the minimum sampled frequencyωmin and apply the thresholdω <−10 −3 THz to reproduce the completed label

  45. [54]

    For label-threshold studies, users should keep the original completed label unchanged and store any threshold-dependent relabeling in a new derived column

    Join the completed table with the split table before model training so that train, validation and test records follow the reference formula-group split. For label-threshold studies, users should keep the original completed label unchanged and store any threshold-dependent rela...

  46. [160]

    Chemical complexity is measured by the number of distinct elements in the parsed formula

    Their occurrence percentages are 16.1%, 5.9%, 4.6%, 4.4%, 3.9%, 3.8%, 3.7%, 3.6%, 3.3%, 3.3%, 3.0%, 2.9%, 2.3%, 1.9%, 1.6% and 1.5%. Chemical complexity is measured by the number of distinct elements in the parsed formula. The completed cohort contains 0.7% unary, 17.8% binary...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.