Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Advancing Magnetic Materials Discovery -- A structure-based machine learning approach for magnetic ordering and magnetic moment prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A structure-only 518-feature descriptor, trained with LightGBM, classifies ferromagnetic versus ferrimagnetic ordering across 5,741 stable compounds at 82.4% accuracy and predicts magnetic moment per atom with 0.93 correlation.

desk verdict Useful descriptor extension and a genuinely broader FM/FiM dataset, but the headline comparison to Hund's matrix and OFM is not controlled, so the claimed margin of improvement is unproven. read the letter →

arxiv 2507.01913 v1 pith:2QR57H2T submitted 2025-07-02 cond-mat.mtrl-sci cs.LG

Apoorv Verma , Junaid Jami , Amrita Bhattacharya This is my paper
classification cond-mat.mtrl-scics.LG
keywords magneticorderingpredictionferrimagneticclassificationmomentperatommachinelearningdescriptorLightGBMhigh-throughputmaterialsscreeningstructure-basedProject
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 argues that a structure-derived descriptor with 518 features, paired with LightGBM, can classify ferromagnetic versus ferrimagnetic ordering across 5,741 stable binary and ternary compounds at 82.4% accuracy and predict magnetic moment per atom with a correlation coefficient of 0.93. The central motivation is that earlier machine-learning models either neglected ferrimagnetic states or were limited to narrow chemical families such as Mn-based or lanthanide-transition-metal compounds. If correct, the approach would give materials scientists a fast screening tool that needs only structural data, not expensive DFT magnetic-configuration searches, to prioritize magnetic candidates. It would also provide balanced recall between the two ordered magnetic classes, addressing a known blind spot.

What carries the argument

The carrying object is the refined 518-feature descriptor built from a $16 \times 1$ elemental vector (8 properties, each duplicated as its square). For every atomic pair within a 5 Å cutoff in a $3 \times 3 \times 3$ supercell, the outer product of the two elemental vectors is divided by the squared interatomic distance, and the element-wise mean and standard deviation of these interaction matrices across the compound supply 512 features; six more features (mean and standard deviation of bond length, electronegativity difference, and squared electronegativity difference) complete the vector. This descriptor is fed to LightGBM, a gradient-boosting tree model, which learns the mapping from structure to magnetic ordering, magnetic moment per atom, and formation energy.

What would settle it

Re-implement the Hund's matrix and orbital field matrix descriptors on the exact same 5,741-compound dataset and train/test split used here and check whether LightGBM with the proposed descriptor still beats them by roughly ten accuracy points and 0.22 in correlation; if the gap shrinks or reverses under identical protocols, the claimed superiority would not survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that enriching the elemental representation with eight ground-state properties and their squares, then forming pairwise interaction matrices scaled by inverse squared distance, captures enough chemistry and structure to outperform both the Hund's matrix and the orbital field matrix on magnetic property prediction. On a uniform dataset of 5,741 stable binary and ternary ferromagnetic and ferrimagnetic compounds from the Materials Project, the proposed descriptor gives 82.4% classification accuracy for FM versus FiM ordering, with FiM recall of 0.75 against 0.47 for the Hund's matrix baseline, and a magnetic-moment correlation coefficient of 0.9395, with MAE 0.1862 and RMSE 0.28 on the test set. The same descriptor also predicts formation energy per atom with a correlation coefficient of 0.977. The authors attribute the gain to three factors: a broader set of elemental properties, inclusion of nonlinear squared terms, and reduced matrix sparsity.

Load-bearing premise

The claimed advantage over the Hund's matrix and orbital field matrix depends on comparing against metrics quoted from earlier papers rather than re-running those descriptors on the same compounds with the same train/test split, so the size of the improvement could be an artifact of comparing different benchmarks.

Editorial extensions

If this is right

  • A structure-only screen could rank thousands of candidate compounds for ferromagnetic or ferrimagnetic ordering before any DFT magnetic-configuration calculation is run.
  • The raised FiM recall (0.75 versus 0.47) means ferrimagnetic candidates, which earlier models often merged into FM or AFM classes, would be visible to high-throughput searches.
  • Because the same descriptor also predicts formation energy per atom, a screening pipeline could filter simultaneously for magnetic class and thermodynamic stability.
  • The approach is not restricted to Mn-based or lanthanide-transition-metal chemistries, so it can be applied to unexplored composition spaces in binary and ternary systems.

Reading between the lines

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

  • A natural extension the paper does not test is transferring the classifier to antiferromagnetic compounds, where the severe class imbalance would likely require sublattice-aware features.
  • The choice of a 5 Å cutoff and a $3 \times 3 \times 3$ supercell is tuned on this dataset; testing whether the descriptor's margin persists for larger unit cells or longer-ranged exchange interactions would show how general the mechanism is.
  • Because the descriptor uses explicit elemental properties rather than learned embeddings, feature-importance analysis could reveal which property differences are most decisive for FM-versus-FiM discrimination; the paper does not report such an analysis.
  • If the reported margin over baselines survives a same-protocol re-run, the descriptor could plausibly be adapted to other magnetic targets like Curie temperature or magnetocrystalline anisotropy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a 518-feature descriptor for machine learning prediction of magnetic ordering (FM vs. FiM) and magnetic moment per atom across 5,741 stable binary and ternary ferromagnetic/ferrimagnetic compounds from the Materials Project. The descriptor combines an enriched 16-dimensional elemental representation (eight ground-state properties and their squares) with distance-scaled pairwise interaction matrices aggregated by mean and standard deviation, plus bond-length and electronegativity-derived features. Using LightGBM, the authors report 82.4% classification accuracy for FM/FiM ordering with a FiM recall of 0.75, and a correlation coefficient of 0.9395 for magnetic moment per atom, claiming superiority over the Hund's matrix and orbital field matrix descriptors. The paper also reports formation energy prediction with CC=0.977.

Significance. If the reported performance is robust, the proposed descriptor is a computationally efficient and general screening tool for ferromagnetic and ferrimagnetic materials, addressing a real gap: prior descriptors were demonstrated on narrower datasets (Mn-based compounds or bimetallic alloys) and showed poor FiM recall. The use of a large, chemically diverse dataset and a fast gradient-boosting model are practical strengths. However, the central comparative claim against Hund's matrix and OFM is not supported by the evidence as presented, because the baseline numbers are quoted from prior work on different datasets rather than re-evaluated under the same protocol. The paper also overstates its 'structure-based' scope by including explicit elemental property features.

major comments (4)
  1. [Section III.B, Table I] The headline claim that the proposed descriptor 'surpasses' the Hund's matrix and orbital field matrix is not established by a controlled comparison. The Hund's matrix metrics come from Ref. [45], which targeted Mn-based compounds, and the OFM metrics from Ref. [53], which targeted bimetallic alloys; neither was re-implemented on the same 5,741-compound dataset or the same train/test split as the present model. Differences in dataset composition and split can easily account for the reported margins. Please re-evaluate both baselines under an identical protocol, or, if this is not feasible, explicitly state that the comparison is not controlled and soften the abstract/conclusion claims accordingly. In addition, report the exact split procedure, number of repeated runs, and standard deviations for all metrics in Tables I and II, as the current single-point estimates do not allow the reader to assess significance.
  2. [Abstract and Section III.A] The abstract states that prediction is performed 'using only the structural information of materials,' but the descriptor described in Section II.B includes eight explicit elemental properties (atomic number, atomic radius, group number, period number, density, ionization energy, electronegativity, and number of unpaired valence electrons) and their squares. This is a direct contradiction: those features are compositional/electronic, not purely structural. Please revise the wording to accurately describe the descriptor as combining structural and elemental information, or provide an ablation showing that the elemental features do not drive the predictive performance.
  3. [Section II.B, Eq. (2); Section III.B] The descriptor includes the number of unpaired valence electrons (U) of the free atom as a feature. This quantity is a direct proxy for the free-atom magnetic moment under Hund's rules, and the target being predicted is the solid-state magnetic moment per atom. The reported CC of 0.9395 may therefore partly reflect a near-identity mapping from U to the target rather than a genuinely structural prediction. Please add an ablation study that removes U (and its square) from the feature set and reports the change in CC and MAE, and discuss the feature importance of U relative to structural features. Without this, the claim that the descriptor captures 'structure-property relationships' is not fully supported.
  4. [Section III.A] The classification accuracy of 82.4% and the balanced recalls (0.85 FM / 0.75 FiM) are reported without describing the class-imbalance handling or the train/test split. FiM compounds constitute only 22% of the dataset, so the model configuration matters for interpreting the balanced recall. Please specify how LightGBM was set up (e.g., class weights, sampling, hyperparameters), how the split was performed (random vs. stratified), and report precision and F1 scores per class in addition to recall. Cross-validated performance with standard deviations would also strengthen the reliability of the claim.
minor comments (5)
  1. [Section I] The text contains a duplicated word in 'for emerging emerging high-density, high-speed and high-efficiency spintronic applications'; please correct.
  2. [Section III.C, Fig. 8 caption area] The sentence 'Fig. 8 shows the predicted magnetic moment per atom versus the actual values for the test set using our refined descriptor' refers to formation energy per atom, not magnetic moment; please correct the wording.
  3. [Section II.B] The paper describes the elemental descriptor both as a '16×1 descriptor vector' and as a '1 × 16 elemental vector'; please unify the notation to avoid confusion.
  4. [Section II.B] The statement 'the model does not inherently recognize this periodicity' is unclear; crystal structures are periodic by definition, and the supercell construction is a standard way to include periodic images in a distance-based descriptor. Consider rephrasing to clarify what limitation the supercell addresses.
  5. [Whole paper] No data or code availability statement is provided. Given that the dataset is derived from the Materials Project and the descriptor is algorithmically defined, releasing the feature-generation code and the exact train/test split would greatly improve reproducibility; please add a statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: descriptor and target are external, and the claimed benchmark gap is a validation concern, not circularity.

full rationale

This paper is an empirical supervised-learning study. The 518-feature descriptor is constructed from elemental properties taken from PubChem (Z, R, G, P, D, I, E, U plus squared terms) and from structural distances, while the targets (FM/FiM ordering labels and magnetic moment per atom) are external DFT-derived quantities from the Materials Project. No equation or construction in the paper defines the target in terms of the descriptor or fits a parameter that is then renamed as a prediction; the model is trained and evaluated on held-out data. The only notable issue is that the baseline metrics in Table I are quoted from prior work on different datasets (Mn-based compounds for the Hund's matrix and bimetallic alloys for OFM) rather than re-evaluated under the same train/test protocol, which undermines the comparative claim but is a benchmarking validity concern, not circularity. The inclusion of the number of unpaired valence electrons as a feature is a physically motivated proxy correlated with atomic magnetic moments, but it is not the fitted target nor derived from the target, so the prediction does not reduce to the input by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. I find no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on data from the Materials Project, a descriptor built from eight neutral-atom properties, and a LightGBM model. The ledger lists the data-dependent choices (cutoff radius, supercell size, unreported hyperparameters) and the domain assumptions about label accuracy, descriptor sufficiency, and split integrity. No new physical entities are introduced.

free parameters (3)
  • Cutoff radius r_c = 5 Å
    The paper tests cutoffs from 3 Å to 10 Å and selects 5 Å because 'best results were obtained' at that value (Section II.B). This is a data-dependent choice, not a parameter derived from first principles.
  • Supercell size = 3x3x3
    The 3x3x3 supercell is a modeling choice for periodic interactions; no justification is given for why this size is sufficient (Section II.B).
  • LightGBM hyperparameters = not reported
    No hyperparameters, number of estimators, learning rate, or regularization are reported. These are fitted to data and affect all reported metrics.
assumptions (4)
  • domain assumption Materials Project labels for magnetic ordering (FM/FiM) and magnetic moments are accurate ground truth for training and evaluation.
    The paper uses 5,741 stable binary and ternary compounds from the Materials Project and treats their FM/FiM labels and magnetic moment values as targets (Section II.A). If these DFT-derived labels are noisy or inconsistent, the reported accuracy and CC are not reliable.
  • domain assumption The descriptor built from neutral-atom properties (Z, radius, group, period, density, ionization energy, electronegativity, unpaired valence electrons) is sufficient to capture magnetic interactions in solids.
    The method assumes these eight atomic properties, together with pairwise distances, encode the physics relevant for magnetic ordering and moment magnitude (Section II.B). No first-principles argument is given for sufficiency.
  • domain assumption LightGBM gradient boosting generalizes from the training split to unseen compositions without leakage.
    The paper reports a single train/test result without describing the split, chemical holdout, or repeated validation (Section III). The validity of the accuracy estimate depends on this unstated assumption.
  • ad hoc to paper Using the number of unpaired valence electrons of a free atom as a descriptor feature is a valid proxy for solid-state magnetic moments.
    This property directly tracks the maximum spin moment of the free atom, so it partially encodes the target magnetic moment. The paper introduces it as one of eight elemental features (Section II.B) but does not address this proxy relationship.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Magnetic Materials Discovery -- A structure-based machine learning approach for magnetic ordering and magnetic moment prediction." pith.science (2026). https://pith.science/paper/2QR57H2T

@misc{pith2026250701913,
  author       = {Pith},
  title        = {Pith review of: Advancing Magnetic Materials Discovery -- A structure-based machine learning approach for magnetic ordering and magnetic moment prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QR57H2T}},
  note         = {Machine review of arXiv:2507.01913}
}
read the original abstract

Accurately predicting magnetic behavior across diverse materials systems remains a longstanding challenge due to the complex interplay of structural and electronic factors and is pivotal for the accelerated discovery and design of next-generation magnetic materials. In this work, a refined descriptor is proposed that significantly improves the prediction of two critical magnetic properties -- magnetic ordering (Ferromagnetic vs. Ferrimagnetic) and magnetic moment per atom -- using only the structural information of materials. Unlike previous models limited to Mn-based or lanthanide-transition metal compounds, the present approach generalizes across a diverse dataset of 5741 stable, binary and ternary, ferromagnetic and ferrimagnetic compounds sourced from the Materials Project. Leveraging an enriched elemental vector representation and advanced feature engineering, including nonlinear terms and reduced matrix sparsity, the LightGBM-based model achieves an accuracy of 82.4% for magnetic ordering classification and balanced recall across FM and FiM classes, addressing a key limitation in prior studies. The model predicts magnetic moment per atom with a correlation coefficient of 0.93, surpassing the Hund's matrix and orbital field matrix descriptors. Additionally, it accurately estimates formation energy per atom, enabling assessment of both magnetic behavior and material stability. This generalized and computationally efficient framework offers a robust tool for high-throughput screening of magnetic materials with tailored properties.

Figures

Figures reproduced from arXiv: 2507.01913 by the authors.

Figure 1
Figure 1. FIG. 1. Workflow for predicting material properties using ML. Structural data is extracted from a materials database, followed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. In the Hund’s matrix method, elements [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. FIG. 4. (a) Distribution of magnetic moments per atom for [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: FIG. 5. Illustration of the generation of a 16 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Confusion matrix for FM vs. FiM magnetic ordering [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The plot shows the predicted magnetic moment per [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. The plot shows the predicted formation energy per [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 51 canonical work pages

  1. [45]

    Khatri and A

    Y. Khatri and A. Kashyap, Journal of Magnetism and Magnetic Materials 597, 172026 (2024)

    work page 2024

  2. [53]

    T. L. Pham, H. Kino, K. Terakura, T. Miyake, I. Taki- gawa, K. Tsuda, and H. C. Dam, Science and Technology of Advanced Materials 18 (2017)

    work page 2017

  3. [1]

    Ghorbani, I

    Y. Ghorbani, I. Ilankoon, N. Dushyantha, and G. T. Nwaila, Resources, Conservation and Recycling 212, 107966 (2025)

    work page 2025

  4. [2]

    Bhushan, Microsyst

    B. Bhushan, Microsyst. Technol. 29, 1529–1546 (2023)

    work page 2023

  5. [3]

    Grochowski and R

    E. Grochowski and R. D. Halem, IBM Systems Journal 42, 338 (2003)

    work page 2003

  6. [4]

    R. H. Dee, Proceedings of the IEEE 96, 1775 (2008)

    work page 2008

  7. [5]

    Lee, K.-C

    H.-W. Lee, K.-C. Kim, and J. Lee, IEEE Transactions on Magnetics 42, 1917 (2006)

    work page 2006

  8. [6]

    Mellari, International Journal of Air-Conditioning and Refrigeration 31 (2023)

    S. Mellari, International Journal of Air-Conditioning and Refrigeration 31 (2023)

    work page 2023

Show all 58 references
  1. [7]

    Leuenberger and D

    M. Leuenberger and D. Loss, Nature 410, 789 (2001)

    2001

  2. [8]

    R.-J. M. van Geuns, P. A. Wielopolski, H. G. de Bruin, B. J. Rensing, P. M. van Ooijen, M. Hulshoff, M. Oud- kerk, and P. J. de Feyter, Progress in Cardiovascular Dis- eases 42, 149 (1999), advances in Coronary Imaging

    1999

  3. [9]

    V. I. Belotelov, L. Bi, A. Kalashnikova, M. Levy, and N. Maccaferri, Opt. Mater. Express 12, 2087 (2022)

    2022

  4. [10]

    Ma˜ nas-Valero, T

    S. Ma˜ nas-Valero, T. van der Sar, R. A. Duine, and B. van Wees, Newton 1, 100018 (2025)

    2025

  5. [11]

    Giannozzi, O

    P. Giannozzi, O. Andreussi, T. Brumme, O. Bunau, M. Buongiorno Nardelli, M. Calandra, R. Car, C. Cavaz- zoni, D. Ceresoli, M. Cococcioni, N. Colonna, I. Carn- imeo, A. Dal Corso, S. de Gironcoli, P. Delugas, R. A. DiStasio, A. Ferretti, A. Floris, G. Fratesi, G. Fugallo, R. Ge...

    2017

  6. [12]

    M. J. Donahue and D. G. Porter, Oommf user’s guide: Version 1.0 (1999)

    1999

  7. [13]

    G. P. M¨ uller, M. Hoffmann, C. Dißelkamp, D. Sch¨ urhoff, S. Mavros, M. Sallermann, N. S. Kiselev, H. J´ onsson, and S. Bl¨ ugel, Phys. Rev. B99, 224414 (2019)

    2019

  8. [14]

    R. F. L. Evans, W. J. Fan, P. Chureemart, T. A. Ostler, M. O. A. Ellis, and R. W. Chantrell, Journal of Physics: Condensed Matter 26, 103202 (2014)

    2014

  9. [15]

    Blaha, K

    P. Blaha, K. Schwarz, F. Tran, R. Laskowski, G. K. H. Madsen, and L. D. Marks, The Journal of Chemical Physics 152, 074101 (2020), https://pubs.aip.org/aip/jcp/article- pdf/doi/10.1063/1.5143061/16727313/074101 1 online.pdf

    2020 doi

  10. [16]

    Lovesey, Theory of neutron scattering from condensed matter (Clarendon Press., 2025)

    S. Lovesey, Theory of neutron scattering from condensed matter (Clarendon Press., 2025)

    2025

  11. [17]

    L. J. P. Ament, M. van Veenendaal, T. P. Devereaux, J. P. Hill, and J. van den Brink, Rev. Mod. Phys. 83, 705 (2011)

    2011

  12. [18]

    Bushra Munir, A

    H. Bushra Munir, A. Afaq, A. Bakar, I. Asif, and N. ul Aarifeen, Materials Today Communications 26, 101763 (2021)

    2021

  13. [19]

    D. King, S. Middleburgh, P. Burr, T. Whiting, P. Fossati, and M. Wenman, Physical Review B 98 (2018)

    2018

  14. [20]

    Sarkar and P

    S. Sarkar and P. Kratzer, Phys. Rev. B 103, 224421 (2021)

    2021

  15. [21]

    J. Ma, V. I. Hegde, K. Munira, Y. Xie, S. Keshavarz, D. T. Mildebrath, C. Wolverton, A. W. Ghosh, and W. H. Butler, Phys. Rev. B 95, 024411 (2017)

    2017

  16. [22]

    Opahle, H

    I. Opahle, H. K. Singh, J. Zemen, O. Gutfleisch, and H. Zhang, Phys. Rev. Res. 2, 023134 (2020)

    2020

  17. [23]

    Payne, G

    A. Payne, G. Avenda˜ no-Franco, E. Bousquet, and A. H. Romero, Journal of Chemical Theory and Computation 14, 4455 (2018), pMID: 29966084, https://doi.org/10.1021/acs.jctc.8b00404

    2018 doi

  18. [24]

    Horton, J

    M. Horton, J. Montoya, M. Liu, and K. Persson, npj Computational Materials 5, 2 (2019)

    2019

  19. [25]

    Zheng and P

    F. Zheng and P. Zhang, Computer Physics Communica- tions 259, 107659 (2021)

    2021

  20. [26]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Materials 1, 011002 (2013)

    2013

  21. [27]

    Curtarolo, W

    S. Curtarolo, W. Setyawan, G. L. Hart, M. Jahnatek, R. V. Chepulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy, M. J. Mehl, H. T. Stokes, D. O. Demchenko, and D. Morgan, Computational Materials Science58, 218 (2012)

    2012

  22. [28]

    Draxl and M

    C. Draxl and M. Scheffler, MRS Bulletin 43, 676 (2018)

    2018

  23. [29]

    J. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolver- ton, JOM 65, 1501 (2013)

    2013

  24. [30]

    Nieves, S

    P. Nieves, S. Arapan, J. Maudes-Raedo, R. Marticorena- S´ anchez, N. Del Br ´ ıo, A. Kovacs, C. Echevarria-Bonet, D. Salazar, J. Weischenberg, H. Zhang, O. Vekilova, R. Serrano-L´ opez, J. Barandiaran, K. Skokov, O. Gut- fleisch, O. Eriksson, H. Herper, T. Schrefl, and S. Cues...

    2019

  25. [31]

    Mal and P

    S. Mal and P. Sen, Journal of Magnetism and Magnetic Materials 589, 171590 (2024)

    2024

  26. [32]

    Dutta and P

    A. Dutta and P. Sen, J. Mater. Chem. C 10, 3404 (2022)

    2022

  27. [33]

    Balachandran, Journal of Materials Research 35, 1 (2020)

    P. Balachandran, Journal of Materials Research 35, 1 (2020)

    2020

  28. [34]

    T. Liao, W. Xia, M. Sakurai, R. Wang, C. Zhang, H. Sun, K.-M. Ho, C.-Z. Wang, and J. R. Chelikowsky, Phys. Rev. Mater. 6, 024402 (2022)

    2022

  29. [35]

    Broyles, W

    C. Broyles, W. Charles, and S. Ran, Phys. Rev. Mater. 8, 114405 (2024)

    2024

  30. [36]

    Merker, H

    H. Merker, H. Heiberger, L. Nguyen, T. Liu, Z. Chen, N. Andrejevic, N. Drucker, R. Okabe, S. Kim, Y. Wang, T. Smidt, and M. Li, iScience 25, 105192 (2022)

    2022

  31. [37]

    N. C. Frey, M. K. Horton, J. M. Munro, S. M. Griffin, K. A. Persson, and V. B. Shenoy, Science Advances 6, eabd1076 (2020), https://www.science.org/doi/pdf/10.1126/sciadv.abd1076

    2020 doi

  32. [38]

    Nakamura, S

    Y. Nakamura, S. B. S. Chauhan, and P. B. Lim, Photon- ics 11, 10.3390/photonics11100931 (2024)

    2024 doi

  33. [39]

    ¨Ozg¨ ur, Y

    ¨U. ¨Ozg¨ ur, Y. Alivov, and H. Morko¸ c, Journal of materials science: Materials in electronics 20, 789 (2009)

    2009

  34. [40]

    Zhang, X

    Y. Zhang, X. Feng, Z. Zheng, Z. Zhang, K. Lin, X. Sun, G. Wang, J. Wang, J. Wei, P. Vallobra, Y. He, Z. Wang, L. Chen, K. Zhang, Y. Xu, and W. Zhao, Applied Physics Reviews 10, 011301 (2023)

    2023

  35. [41]

    S. K. Kim, G. S. D. Beach, K.-J. Lee, T. Ono, T. Rasing, and H. Yang, Nature Materials 21, 24 (2022)

    2022

  36. [42]

    T. Xu, Y. Cheng, Y. Dong, H. Bai, H.-A. Zhou, X. Shu, P. Gargiani, M. Valvidares, P. Yu, and W. Jiang, Phys. Rev. Appl. 19, 034088 (2023)

    2023

  37. [43]

    Finley and L

    J. Finley and L. Liu, Applied Physics Letters116, 110501 (2020)

    2020

  38. [44]

    Nayak, M

    A. Nayak, M. Nicklas, S. Chadov, P. Khuntia, C. Shekhar, A. Kalache, M. Baenitz, Y. Skourski, V. Guduru, A. Puri, U. Zeitler, J. M. D. Coey, and C. Felser, Nature materials (2015)

    2015

  39. [46]

    A. Seko, A. Togo, and I. Tanaka, Descriptors for machine learning of materials data (2018) pp. 3–23

    2018

  40. [47]

    M. Rupp, A. Tkatchenko, K.-R. M¨ uller, and O. A. von Lilienfeld, Phys. Rev. Lett. 108, 058301 (2012)

    2012

  41. [48]

    Faber, A

    F. Faber, A. Lindmaa, O. A. von Lilien- feld, and R. Armiento, International Jour- nal of Quantum Chemistry 115, 1094 (2015), https://onlinelibrary.wiley.com/doi/pdf/10.1002/qua.24917

    2015 doi

  42. [49]

    Huo and M

    H. Huo and M. Rupp, Machine Learning: Science and Technology 3, 045017 (2022)

    2022

  43. [50]

    Behler, The Journal of Chemical Physics 134, 074106 (2011), https://pubs.aip.org/aip/jcp/article- pdf/doi/10.1063/1.3553717/15435271/074106 1 online.pdf

    J. Behler, The Journal of Chemical Physics 134, 074106 (2011), https://pubs.aip.org/aip/jcp/article- pdf/doi/10.1063/1.3553717/15435271/074106 1 online.pdf

    2011 doi

  44. [51]

    Eckhoff and J

    M. Eckhoff and J. Behler, npj Computational Materials 7, 170 (2021)

    2021

  45. [52]

    A. P. Bart´ ok, R. Kondor, and G. Cs´ anyi, Phys. Rev. B 87, 184115 (2013)

    2013

  46. [54]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman, Basis ex- pansions and regularization, in The Elements of Statis- tical Learning: Data Mining, Inference, and Prediction (Springer New York, New York, NY, 2009) pp. 139–189

    2009

  47. [55]

    National Center for Biotechnology Information, Periodic table of elements, https://pubchem.ncbi.nlm.nih.gov/ periodic-table/ (2025), accessed: 2025-05-08. 11

    2025

  48. [56]

    P. F. Lang, Monatshefte f¨ ur Chemie - Chemical Monthly 155, 507 (2024)

    2024

  49. [57]

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, in Advances in Neural Information Processing Systems, Vol. 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vish- wanathan, and R. Garnett (Curran Associates, Inc., 2017)

    2017

  50. [5741]

    Compounds with an energy above hull equal to zero have been consid- ered stable

    have been used for the present work. Compounds with an energy above hull equal to zero have been consid- ered stable. Notably, no specific elemental filter or sym- metry criteria were applied in the compound selection process from the database, resulting in a highly diverse an...

    1979

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.