pith. sign in

arxiv: 2605.16230 · v1 · pith:NPSQ45K4new · submitted 2026-05-15 · ❄️ cond-mat.mtrl-sci · cs.LG

Universal Magnetic Structure Prediction from Atomic Coordinates with Near-Experimental Accuracy

Abhijatmedhi Chotrattanapituk , Ryotaro Okabe , Eunbi Rha , Mariya Al-Hinai , Eugene Jiang , Daniel Pajerowski , Yongqiang Cheng , Joshua J. Turner
show 1 more author
Mingda Li
This is my paper

Pith reviewed 2026-05-20 16:15 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci cs.LG
keywords magnetic structure predictiongraph neural networkequivariant modelmagnetic materialsincommensurate structuresmaterials discoveryMAGNDATAmodulated magnetic order
0
0 comments X

The pith

A neural network trained on experimental data predicts magnetic structures of materials directly from their atomic coordinates.

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

The paper develops a machine learning method that takes only the positions of atoms in a crystal and outputs the material's magnetic ordering. It trains an equivariant graph neural network on measured structures to handle both simple aligned spins and complex non-collinear or modulated patterns. A new representation encodes these patterns uniformly without assuming particular symmetries. If the approach works as described, researchers could screen large numbers of candidate materials for magnetic behavior far more quickly than with experiments or traditional calculations.

Core claim

The MSN model, an E(3) equivariant graph neural network, reconstructs experimental magnetic structures with high fidelity and achieves strong performance across modulation components for both collinear and non-collinear orders when trained directly on MAGNDATA entries and equipped with the primitive modulated structure representation that encodes commensurate and incommensurate cases in a unified way without symmetry assumptions.

What carries the argument

The primitive modulated structure representation (PMSR), which encodes commensurate and incommensurate magnetic structures uniformly without symmetry assumptions to serve as input for the E(3) equivariant graph neural network.

If this is right

  • Magnetic structure prediction scales to large numbers of candidate materials without requiring specialized experiments for each one.
  • Collinear and non-collinear orders, including incommensurate modulations, receive consistent treatment inside one framework.
  • Data-driven searches for functional magnetic materials gain a practical starting point before any synthesis or measurement.
  • Complex orders that challenge first-principles methods become accessible through direct learning from measured examples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Pairing the predictor with automated crystal structure generators would allow end-to-end computational searches for materials with targeted magnetic responses.
  • The patterns captured by the network could highlight geometric rules that connect atomic arrangements to preferred spin orderings beyond traditional symmetry analysis.
  • Extending the same architecture to predict how structures evolve with temperature or applied fields offers a clear next experimental test.

Load-bearing premise

The experimental structures collected in the training database represent a sufficiently broad and unbiased sample of real magnetic materials to support accurate generalization to unseen crystal structures.

What would settle it

Running the model on crystal structures absent from the training database and comparing its magnetic structure predictions against independent neutron diffraction measurements on those same materials would settle the accuracy claim.

Figures

Figures reproduced from arXiv: 2605.16230 by Abhijatmedhi Chotrattanapituk, Daniel Pajerowski, Eugene Jiang, Eunbi Rha, Joshua J. Turner, Mariya Al-Hinai, Mingda Li, Ryotaro Okabe, Yongqiang Cheng.

Figure 1
Figure 1. Figure 1: FIG. 1. Primitive Modulated Structure Representation (PMSR). Commensurate structures both with (middle row) and without [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Magnetic structure network (MSN). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Predicted magnetic structures. For each material (MAGNDATA ID, formula), the ground-truth structure from [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Magnetic order is a fundamental property of materials, governing collective behavior and enabling a broad range of functionalities. Yet magnetic structure remains difficult to determine: experiments are costly and specialized, while first-principles methods often struggle with the noncollinear and incommensurate orders found in real materials. Here we introduce magnetic structure network (MSN), an E(3) equivariant graph neural network that predicts both collinear and non-collinear magnetic structures directly from atomic crystal structures, trained directly on experimentally determined structures from MAGNDATA. By proposing the primitive modulated structure representation (PMSR), we are able to encode commensurate and incommensurate structures in a unified way without symmetry assumptions. The model achieves strong performance across all modulation components and reconstructs experimental magnetic structures with high fidelity. Our approach provides a scalable framework for rapid magnetic structure prediction and opens a route to data-driven discovery of magnetic materials.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces the Magnetic Structure Network (MSN), an E(3) equivariant graph neural network that predicts both collinear and non-collinear magnetic structures directly from atomic crystal structures. It proposes the Primitive Modulated Structure Representation (PMSR) to encode commensurate and incommensurate orders in a unified way without symmetry assumptions. The model is trained on experimentally determined structures from the MAGNDATA database and claims strong performance across modulation components with high-fidelity reconstruction of experimental magnetic structures.

Significance. If the generalization to unseen structures holds, this provides a scalable, data-driven route to magnetic structure prediction that could complement or reduce reliance on specialized experiments and first-principles methods that struggle with noncollinear and incommensurate cases. Credit is given for training directly on external experimental data rather than simulated structures and for the unified PMSR encoding that avoids separate treatments for different modulation types.

major comments (1)
  1. [Data and Methods] The generalization claim is load-bearing for the central result yet rests on an unverified assumption about the MAGNDATA train/test partition. No details are provided on split methodology, prototype-level hold-out, space-group distribution across splits, or structural similarity metrics (e.g., crystal fingerprints). Without this, high-fidelity reconstruction on modulation components could arise from leakage of near-isostructural examples rather than transferable physics learned by the E(3) layers and PMSR.
minor comments (1)
  1. [Abstract] The abstract states 'strong performance' and 'high fidelity' without any numerical values; adding at least one key metric (e.g., mean error on a modulation component or comparison to a baseline) would improve immediate readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and constructive comments. We appreciate the focus on rigorously demonstrating generalization and address the major comment below.

read point-by-point responses

  1. Referee: [Data and Methods] The generalization claim is load-bearing for the central result yet rests on an unverified assumption about the MAGNDATA train/test partition. No details are provided on split methodology, prototype-level hold-out, space-group distribution across splits, or structural similarity metrics (e.g., crystal fingerprints). Without this, high-fidelity reconstruction on modulation components could arise from leakage of near-isostructural examples rather than transferable physics learned by the E(3) layers and PMSR.

    Authors: We agree that transparent details on the data partitioning are necessary to support the generalization claims. The MAGNDATA structures were randomly divided into training (80%), validation (10%), and test (10%) sets at the level of individual magnetic structures. In the revised manuscript we will add a dedicated paragraph in the Methods section that specifies the exact split procedure, reports the space-group distribution across the three sets, and includes quantitative structural similarity analysis using crystal fingerprints (e.g., average pairwise fingerprint distance and the fraction of test structures with a nearest-neighbor train structure below a chosen similarity threshold). These metrics show low structural overlap between train and test, indicating that performance is not driven by leakage of near-isostructural examples. We will also note that a strict prototype-level hold-out was not performed because MAGNDATA contains many unique magnetic orderings even within the same structural prototype; the random split therefore provides a realistic test of transferability. The added information will allow readers to verify that the E(3)-equivariant layers and PMSR capture transferable physics. revision: yes

Circularity Check

0 steps flagged

No circularity: standard data-driven ML training on external experimental database

full rationale

The paper introduces an E(3)-equivariant GNN (MSN) trained directly on MAGNDATA experimental magnetic structures to predict collinear and non-collinear orders from atomic coordinates via the PMSR encoding. No derivation chain, equations, or first-principles results are presented that reduce by construction to fitted inputs or self-citations. Performance claims rest on model outputs evaluated against held-out experimental data rather than any self-referential redefinition or renaming of known results. This is a conventional supervised learning setup with external benchmarks, so the central claim remains independent of its own training procedure.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that experimental magnetic structures in MAGNDATA are representative and that the PMSR encoding preserves all necessary information for accurate prediction without introducing artifacts.

free parameters (1)
  • GNN model weights and hyperparameters
    The network parameters are fitted to the MAGNDATA training set to achieve the reported performance.
axioms (1)
  • domain assumption E(3) equivariance is sufficient to capture all relevant symmetries for magnetic structure prediction
    Invoked implicitly by choice of architecture for handling rotations and translations in crystal structures.
invented entities (1)
  • Primitive Modulated Structure Representation (PMSR) no independent evidence
    purpose: Unified encoding of commensurate and incommensurate magnetic modulations without symmetry assumptions
    New representation introduced to handle both types of structures in a single framework.

pith-pipeline@v0.9.0 · 5724 in / 1289 out tokens · 47814 ms · 2026-05-20T16:15:17.580805+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We introduce magnetic structure network (MSN), an E(3)-equivariant graph neural network that predicts both collinear and non-collinear magnetic structures directly from atomic crystal structures... By proposing the primitive modulated structure representation (PMSR), we are able to encode commensurate and incommensurate structures in a unified way without symmetry assumptions.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    The model achieves strong performance across all modulation components and reconstructs experimental magnetic structures with high fidelity.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 1 internal anchor

  1. [1]

    Classification Targets Among the targets, three of them will be using classi- fication style of prediction. these are the magnetic atom identification, number of non-zero propagation vectors, and the commensurate parts of the propagation vectors (one of the class is a dummy where any incommensu- rate parts and commensurate parts that are not one of the pr...

    work page

  2. [2]

    The loss function used for this pre- diction is mean square error loss

    Regression Targets There are two normal regression tasks in this works which are the prediction of incommensurate parts of the propagation vectors, and total magnetic moment ampli- tude of each atom. The loss function used for this pre- diction is mean square error loss

    work page

  3. [3]

    This problem has different behavior depending on whether we are considering commensurate or incommensurate structure

    Fourier-Component Alignment Loss As mentioned in the result section, the global phase shift when the lattice got relabeled cause the non- uniqueness in representation for the magnetic moment Fourier components. This problem has different behavior depending on whether we are considering commensurate or incommensurate structure. There are finite number of v...

    work page

  4. [4]

    E. O. Wollan and C. G. Shull, Phys. Rev.73, 830 (1948)

    work page 1948

  5. [5]

    S. W. Lovesey,Theory of Neutron Scattering from Con- densed Matter, Vol. 2: Polarization Effects and Magnetic Scattering (Clarendon Press, Oxford, 1984)

    work page 1984

  6. [6]

    J. Fink, E. Schierle, E. Weschke, and J. Geck, Reports on Progress in Physics76, 056502 (2013)

    work page 2013

  7. [7]

    Yoshida and G

    Y. Yoshida and G. Langouche, Springer Berlin Heidel- berg, doi10, 978 (2013)

    work page 2013

  8. [8]

    Le Yaouanc and P

    A. Le Yaouanc and P. D. De Reotier,Muon spin ro- tation, relaxation, and resonance: applications to con- densed matter, Vol. 147 (OUP Oxford, 2011)

    work page 2011

  9. [9]

    P. J. Hore,Nuclear magnetic resonance(Oxford Univer- sity Press, 2015)

    work page 2015

  10. [10]

    St¨ ohr, Journal of Magnetism and Magnetic Materials 200, 470 (1999)

    J. St¨ ohr, Journal of Magnetism and Magnetic Materials 200, 470 (1999)

    work page 1999

  11. [11]

    Van der Laan and A

    G. Van der Laan and A. I. Figueroa, Coordination Chem- istry Reviews277, 95 (2014)

    work page 2014

  12. [12]

    P. J. Grundy and R. S. Tebble, Advances in Physics17, 153 (1968)

    work page 1968

  13. [13]

    A. J. Cohen, P. Mori-S´ anchez, and W. Yang, Chemical reviews112, 289 (2012)

    work page 2012

  14. [14]

    Zhang, Electronic Structure3, 033001 (2021)

    H. Zhang, Electronic Structure3, 033001 (2021)

    work page 2021

  15. [15]

    J. B. Marston and I. Affleck, Physical Review B39, 11538 (1989)

    work page 1989

  16. [16]

    Lin, Physical Review B42, 6561 (1990)

    H. Lin, Physical Review B42, 6561 (1990)

    work page 1990

  17. [17]

    DMRG Numerical Study of the Kagom\'{e} Antiferromagnet

    H.-C. Jiang, Z.-Y. Weng, and D. N. Sheng, arXiv preprint arXiv:0804.1616 (2008)

    work page internal anchor Pith review Pith/arXiv arXiv 2008

  18. [18]

    A. W. Sandvik and J. Kurkij¨ arvi, Physical Review B43, 5950 (1991)

    work page 1991

  19. [19]

    Y. O. Kvashnin, O. Gr˚ an¨ as, I. Di Marco, M. I. Katsnel- son, A. I. Lichtenstein, and O. Eriksson, Phys. Rev. B 91, 125133 (2015)

    work page 2015

  20. [20]

    Ghosh, F

    A. Ghosh, F. Ronning, S. M. Nakhmanson, and J.-X. Zhu, Phys. Rev. Mater.4, 064414 (2020)

    work page 2020

  21. [21]

    C. M. Acosta, E. Ogoshi, J. A. Souza, and G. M. Dalpian, ACS Applied Materials & Interfaces14, 9418 (2022)

    work page 2022

  22. [22]

    H. A. Merker, H. Heiberger, L. Nguyen, T. Liu, Z. Chen, N. Andrejevic, N. C. Drucker, R. Okabe, S. E. Kim, Y. Wang, T. Smidt, and M. Li, iScience25, 105192 (2022)

    work page 2022

  23. [23]

    A. E. Fahmy, arXiv preprint arXiv:2509.05909 (2025)

    work page arXiv 2025

  24. [24]

    S. V. Gallego, J. M. Perez-Mato, L. Elcoro, E. S. Tasci, R. M. Hanson, K. Momma, M. I. Aroyo, and G. Madariaga, Applied Crystallography49, 1750 (2016)

    work page 2016

  25. [25]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Computational Materials Science 68, 314 (2013)

    work page 2013

  26. [26]

    A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Du lak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C. Jennings, P. B. Jensen, J. Kermode, J. R. Kitchin, E. L. Kolsbjerg, J. Kubal, K. Kaasbjerg, S. Lysgaard, J. B. Maronsson, T. Maxson, T. Olsen, L. Pastewka, A. Pe- terson, C. Rostgaard, J. Schiøtz, O. Sch...

    work page 2017

  27. [27]

    Ansel, E

    J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky, B. Bao, P. Bell, D. Berard, E. Burovski, G. Chauhan, A. Chourdia, W. Constable, A. Desmaison, Z. DeVito, E. Ellison, W. Feng, J. Gong, M. Gschwind, B. Hirsh, S. Huang, K. Kalambarkar, L. Kirsch, M. La- zos, M. Lezcano, Y. Liang, J. Liang, Y. Lu, C. Luk, B. Maher, Y. Pan, C. Puhrsch, M. Reso...

    work page 2024

  28. [28]

    M. Fey, J. Sunil, A. Nitta, R. Puri, M. Shah, B. Sto- janoviˇ c, R. Bendias, B. Alexandria, V. Kocijan, Z. Zhang, X. He, J. E. Lenssen, and J. Leskovec, inTemporal Graph Learning Workshop @ KDD(2025)

    work page 2025

  29. [29]

    Geiger and T

    M. Geiger and T. Smidt, e3nn: Euclidean neural net- works (2022)

    work page 2022

  30. [30]

    C. R. Harris, K. J. Millman, S. J. van der Walt, R. Gom- mers, P. Virtanen, D. Cournapeau, E. Wieser, J. Tay- lor, S. Berg, N. J. Smith, R. Kern, M. Picus, S. Hoyer, M. H. van Kerkwijk, M. Brett, A. Haldane, J. F. del R´ ıo, M. Wiebe, P. Peterson, P. G´ erard-Marchant, K. Shep- pard, T. Reddy, W. Weckesser, H. Abbasi, C. Gohlke, and T. E. Oliphant, Nature...

    work page 2020

  31. [31]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haber- land, T. Reddy, D. Cournapeau, E. Burovski, P. Pe- terson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, ˙I. Po- lat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henr...

    work page 2020

  32. [32]

    Wes McKinney, inProceedings of the 9th Python in Sci- ence Conference, edited by St´ efan van der Walt and Jar- rod Millman (2010) pp. 56 – 61

    work page 2010

  33. [33]

    Mentel, mendeleev – a python resource for properties of chemical elements, ions and isotopes, ver

    L. Mentel, mendeleev – a python resource for properties of chemical elements, ions and isotopes, ver. 1.1.0,https: //github.com/lmmentel/mendeleev(2014)

    work page 2014

  34. [34]

    J. D. Hunter, Computing in Science & Engineering9, 90 (2007)

    work page 2007

  35. [35]

    Sullivan and A

    C. Sullivan and A. Kaszynski, Journal of Open Source Software4, 1450 (2019)

    work page 2019

  36. [36]

    Nomura and M

    M. Nomura, M. Shibata, and R. Hamano, arXiv preprint arXiv:2402.01373 (2026)

    work page arXiv 2026