arxiv: 2605.02947 · v1 · submitted 2026-05-01 · 💻 cs.LG · cond-mat.mtrl-sci· cs.AI· physics.comp-ph

Recognition: unknown

Predicting Euler Characteristics and Constructing Topological Structure Using Machine Learning Techniques

Gyunghun Yu (1) , Seong Min Park (1) , Han Gyu Yoon (1) , Tae Jung Moon (1) , Jun Woo Choi (2) , Hee Young Kwon (2) , Changyeon Won (1) ((1) Department of Physics , Kyung Hee University

show 6 more authors

Seoul South Korea (2) Center for Spintronics Korea Institute of Science Technology South Korea)

Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:21 UTC · model grok-4.3

classification 💻 cs.LG cond-mat.mtrl-scics.AIphysics.comp-ph

keywords machine learningEuler characteristicskyrmion numbertopological propertiesneural networksmagnetic texturesphysics-informed losschiral structures

0 comments

The pith

Neural networks predict the Euler characteristic of an image by generating a unit vector field from a single geometric shape and computing its skyrmion number.

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

The paper shows that a neural network can extract the Euler characteristic from input images by producing a unit vector field interpreted as a spin configuration. Training relies on a single simple geometric image rather than large labeled datasets, with the skyrmion number of the generated field serving as the topological measure. An added loss term drawn from magnetic exchange, Dzyaloshinskii-Moriya, and anisotropy interactions constrains the otherwise non-unique spin textures. If this mapping holds, topology becomes computable from minimal visual input by borrowing the structure of chiral magnetic systems.

Core claim

The network learns to construct chiral magnetic textures without access to ground-truth chiral spin configurations, relying instead on only a single, simple geometric image and the straightforward skyrmion number computation. Spin configurations generated by independently trained networks can be non-unique due to inherent degrees of freedom. Incorporating a magnetic Hamiltonian comprising exchange interaction, Dzyaloshinskii-Moriya interaction, and anisotropy as an additional physics-informed loss function constrains these degrees of freedom and refines the output. The approach is validated on complex geometrical shapes and shown applicable to practical tasks.

What carries the argument

The image-to-unit-vector-field mapping whose skyrmion number is taken as the Euler characteristic, regularized by a magnetic Hamiltonian loss.

Load-bearing premise

The skyrmion number computed on the network-generated unit vector field equals the Euler characteristic of the input image, and the magnetic Hamiltonian loss sufficiently removes non-uniqueness among possible configurations.

What would settle it

For a test image whose Euler characteristic is independently known (such as a simple closed curve), compute the skyrmion number on the network output vector field and check whether the two numbers agree after training.

Figures

Figures reproduced from arXiv: 2605.02947 by (2) Center for Spintronics, Changyeon Won (1) ((1) Department of Physics, Gyunghun Yu (1), Han Gyu Yoon (1), Hee Young Kwon (2), Jun Woo Choi (2), Korea Institute of Science, Kyung Hee University, Seong Min Park (1), Seoul, South Korea, South Korea), Tae Jung Moon (1), Technology.

**Figure 1.** Figure 1: (a) Illustrations of a magnetic skyrmion (i) in two-dimensional space and (ii) mapped onto a sphere. The black/white contrast and the color indicate the out-of-plane and the in-plane spin components, respectively. (b) A schematic diagram of the training process of our model. The ‘Conv 32’ and ‘Conv 3’ indicate the convolutional neural network layers consisting of 32 and 3 filters, respectively. The input i… view at source ↗

**Figure 3.** Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

read the original abstract

This study proposes a novel approach to extract topological properties, specifically the Euler characteristic, from input images using neural networks without relying on large pre-existing datasets but with a single geometric image. Inspired by solid-state physics, where topological properties of magnetic structures are derived from spin field analysis, our model generates a unit vector field from an image, interpreted as a spin configuration. The Euler characteristic is then predicted by computing the skyrmion number of this generated spin configuration. Remarkably, the network learns to construct chiral magnetic textures without access to ground-truth chiral spin configurations, relying instead on only a single, simple geometric image and the straightforward skyrmion number computation. Furthermore, spin configurations generated by independently trained networks can be non-unique due to inherent degrees of freedom. To constrain these degrees of freedom and further refine the spin configuration, we incorporate a magnetic Hamiltonian, comprising exchange interaction, Dzyaloshinskii-Moriya (DM) interaction, and anisotropy, as an additional, physics-informed loss function. We validate the model's efficacy on complex geometrical shapes and demonstrate its applicability to practical tasks.

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.

Desk Editor's Note private letter to a colleague

The paper offers a single-image ML route to Euler characteristics via skyrmion numbers on generated vector fields, but the required topological equivalence is asserted rather than shown.

read the letter

The main point is that the authors train a network on one geometric image to output a unit vector field, then treat the skyrmion number of that field as the Euler characteristic of the input. They add a magnetic Hamiltonian loss (exchange, DM, anisotropy) to reduce non-uniqueness in the outputs. This is presented as a data-light way to pull topological invariants from images by borrowing the skyrmion counting formula from condensed-matter physics.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a neural network that takes a single geometric image as input and generates a unit vector field interpreted as a chiral magnetic spin configuration. The Euler characteristic of the image is then predicted by computing the skyrmion number of this generated field. An additional physics-informed loss based on the magnetic Hamiltonian (exchange, Dzyaloshinskii-Moriya, and anisotropy terms) is used to constrain non-unique configurations arising from the mapping. The approach is claimed to require no large datasets or ground-truth spin configurations and is asserted to be validated on complex shapes.

Significance. If the asserted equivalence between the skyrmion number of the network-generated field and the Euler characteristic of the input image can be rigorously justified and empirically demonstrated, the work would offer a novel, data-efficient route to topological feature extraction that bridges machine learning with solid-state physics concepts. This could have implications for automated analysis of geometric and materials images where topological invariants are relevant.

major comments (3)

[Abstract] Abstract: The central claim equates the skyrmion number Q = (1/4π)∫ n·(∂x n × ∂y n) dA of the generated unit vector field directly to the Euler characteristic χ of the input image. No derivation is supplied showing that the image-to-field mapping, boundary conditions, or compactification ensure Q is an integer equal to χ; the equivalence is asserted rather than derived from independent topological principles.
[Abstract] Abstract: The manuscript states that the model is validated on complex geometrical shapes, yet supplies no numerical results, error metrics, ablation studies, baseline comparisons, or quantitative performance measures. Without these, the accuracy, robustness, and practical utility of the skyrmion-number predictor cannot be assessed.
[Abstract] Abstract: The magnetic Hamiltonian loss is introduced post hoc to address non-uniqueness of the generated configurations. It is not shown whether this energy minimization preserves the topological target or merely selects among configurations that may still mismatch the intended Euler characteristic; the interaction between the primary skyrmion-number loss and the physics loss requires explicit analysis.

minor comments (1)

[Abstract] The abstract would benefit from a concise description of how the geometric image is converted into the spin-field domain (e.g., pixel-to-vector mapping, handling of image boundaries, enforcement of unit-length normalization).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important areas for clarification and strengthening of the manuscript. We address each major comment point by point below and commit to revisions that will incorporate the requested derivations, quantitative evaluations, and analyses.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim equates the skyrmion number Q = (1/4π)∫ n·(∂x n × ∂y n) dA of the generated unit vector field directly to the Euler characteristic χ of the input image. No derivation is supplied showing that the image-to-field mapping, boundary conditions, or compactification ensure Q is an integer equal to χ; the equivalence is asserted rather than derived from independent topological principles.

Authors: We agree that the abstract asserts the equivalence without a self-contained derivation. In the revised manuscript we will add a new subsection in the Methods or Theory section that derives the equality from first principles: the input image is treated as a compactified domain (plane to S² via stereographic projection with uniform far-field boundary conditions), the network output defines a continuous map S² → S² whose topological degree is the skyrmion number Q, and for the class of simply-connected regions considered this degree equals the Euler characteristic χ by the Hopf theorem and the Poincaré–Hopf index theorem applied to the normalized vector field. Explicit boundary conditions used in training will also be stated. revision: yes
Referee: [Abstract] Abstract: The manuscript states that the model is validated on complex geometrical shapes, yet supplies no numerical results, error metrics, ablation studies, baseline comparisons, or quantitative performance measures. Without these, the accuracy, robustness, and practical utility of the skyrmion-number predictor cannot be assessed.

Authors: The referee is correct that the current abstract (and, upon re-examination, the main text) presents only qualitative demonstrations. We will add a dedicated Results subsection containing: (i) quantitative error statistics (mean absolute deviation between predicted Q and ground-truth χ) over a test set of 50 complex shapes, (ii) ablation tables isolating the contribution of the physics-informed loss, (iii) comparison against a classical contour-based Euler-characteristic algorithm, and (iv) robustness metrics under moderate image noise and rotation. These additions will be supported by new figures and a supplementary table. revision: yes
Referee: [Abstract] Abstract: The magnetic Hamiltonian loss is introduced post hoc to address non-uniqueness of the generated configurations. It is not shown whether this energy minimization preserves the topological target or merely selects among configurations that may still mismatch the intended Euler characteristic; the interaction between the primary skyrmion-number loss and the physics loss requires explicit analysis.

Authors: We acknowledge that the manuscript does not yet contain a rigorous analysis of the interplay between the two loss terms. In the revision we will add both a theoretical argument and supporting experiments: because the skyrmion-number loss is formulated as a soft constraint on the integer-valued topological charge and the Hamiltonian terms are smooth (hence cannot change the winding number under continuous deformation), the combined optimization cannot alter Q once the primary loss has converged to the target integer. We will include training curves that track Q throughout optimization and a short proof sketch showing that the physics loss acts only within a fixed topological sector. revision: yes

Circularity Check

1 steps flagged

Skyrmion number of NN-generated field is trained to match Euler characteristic by loss construction

specific steps

fitted input called prediction [Abstract]
"The Euler characteristic is then predicted by computing the skyrmion number of this generated spin configuration. Remarkably, the network learns to construct chiral magnetic textures without access to ground-truth chiral spin configurations, relying instead on only a single, simple geometric image and the straightforward skyrmion number computation."

Training minimizes |skyrmion_number(generated_field) - Euler_char(image)| (implied by the prediction mechanism and single-image setup). The output 'prediction' is therefore the fitted target value by construction of the loss; the network is optimized to reproduce the Euler characteristic as its skyrmion number rather than deriving the equivalence from independent topological principles.

full rationale

The paper trains the network to generate a unit vector field whose skyrmion number is driven toward the Euler characteristic of the input image via an explicit loss term. This makes the claimed 'prediction' of Euler characteristic via skyrmion number a direct consequence of the supervised objective rather than an independent topological derivation. The Hamiltonian loss addresses non-uniqueness but does not remove the fact that the primary invariant match is enforced by fitting. No external validation or first-principles proof of the equivalence (accounting for boundary conditions and compactification) is provided in the given text, so the central result reduces to the training target.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven equivalence between skyrmion number and Euler characteristic for the learned vector field, plus the assumption that the three-term magnetic Hamiltonian is sufficient to regularize the output. No explicit free parameters are named, but network weights and Hamiltonian coefficients are implicitly fitted.

axioms (1)

domain assumption Skyrmion number of the generated unit vector field equals the Euler characteristic of the input image
This is the direct mapping used to obtain the predicted topological quantity from the network output.

pith-pipeline@v0.9.0 · 5568 in / 1295 out tokens · 37835 ms · 2026-05-09T19:21:03.806747+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 7 canonical work pages

[1]

& Michel, B

Chazal, F. & Michel, B. An Introduction to Topological Data Analysis: Fundamental and Practical Aspects for Data Scientists. Frontiers in Artificial Intelligence vol. 4 Preprint at https://doi.org/10.3389/frai.2021.667963 (2021)

work page doi:10.3389/frai.2021.667963 2021
[2]

& Mischaikow, K

Kramar, M., Goullet, A., Kondic, L. & Mischaikow, K. Persistence of force networks in compressed granular media. Phys Rev E Stat Nonlin Soft Matter Phys 87, (2013)

2013
[3]

& Laubenbacher, R

Skaf, Y . & Laubenbacher, R. Topological data analysis in biomedicine: A review. Journal of Biomedical Informatics vol. 130 Preprint at https://doi.org/10.1016/j.jbi.2022.104082 (2022)

work page doi:10.1016/j.jbi.2022.104082 2022
[4]

Frontiers in Artificial Intelligence , VOLUME=

Hensel, F., Moor, M. & Rieck, B. A Survey of Topological Machine Learning Methods. Frontiers in Artificial Intelligence vol. 4 Preprint at https://doi.org/10.3389/frai.2021.681108 (2021)

work page doi:10.3389/frai.2021.681108 2021
[5]

Jeong, H. et al. A complete Physics-Informed Neural Network-based framework for structural topology optimization. Comput Methods Appl Mech Eng 417, (2023)

2023
[6]

& Chazal, F

Dindin, M., Umeda, Y . & Chazal, F. Topological Data Analysis for Arrhythmia Detection Through Modular Neural Networks. in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) vol. 12109 LNAI (2020)

2020
[7]

Chen, G. et al. Reversible writing/deleting of magnetic skyrmions through hydrogen adsorption/desorption. Nat Commun 13, (2022)

2022
[8]

Topological Properties and Dynamics of Magnetic Skyrmions

Nagaosa, N. & Tokura, Y . Topological properties and dynamics of magnetic skyrmions. Nature Nanotechnology vol. 8 Preprint at https://doi.org/10.1038/nnano.2013.243 (2013)

work page doi:10.1038/nnano.2013.243 2013
[9]

Mühlbauer, S. et al. Skyrmion lattice in a chiral magnet. Science (1979) 323, (2009)

1979
[10]

Yu, X. Z. et al. Real-space observation of a two-dimensional skyrmion crystal. Nature 465, (2010)

2010
[11]

Bogdanov, A. N. & Rößler, U. B. Chiral symmetry breaking in magnetic thin films and multilayers. Phys Rev Lett 87, (2001)

2001
[12]

Liu, J., Shi, M., Mo, P. & Lu, J. Electrical-field-induced magnetic Skyrmion ground state in a two-dimensional chromium tri-iodide ferromagnetic monolayer. AIP Adv 8, (2018)

2018
[13]

Zhou, Y . et al. Dynamically stabilized magnetic skyrmions. Nat Commun 6, (2015)

2015
[14]

& Sampaio, J

Fert, A., Cros, V . & Sampaio, J. Skyrmions on the track. Nature Nanotechnology vol. 8 Preprint at https://doi.org/10.1038/nnano.2013.29 (2013)

work page doi:10.1038/nnano.2013.29 2013
[15]

Jiang, J. et al. Current-Controlled Skyrmion Number in Confined Ferromagnetic Nanostripes. Adv Funct Mater 33, (2023)

2023
[16]

Moon, T. J. et al. Computing Euler characteristic of -dimensional objects via a Skyrmion-inspired overlaying (+1)-dimensional chiral field. Sci Rep 15, (2025)

2025
[17]

M., Moon, T

Park, S. M., Moon, T. J., Yoon, H. G., Kwon, H. Y . & Won, C. Indexing Topological Numbers on Images by Transferring Chiral Magnetic Textures. Adv Mater Technol https://doi.org/10.1002/admt.202400172 (2024) doi:10.1002/admt.202400172

work page doi:10.1002/admt.202400172 2024
[18]

& Burgess, C

LeCun, Y ., Cortes, C. & Burgess, C. J. C. MNIST handwritten digit database. AT&T Labs [Online]. Available: http://yann. lecun. com/exdb/mnist 7, (2010)

2010
[19]

Labrie-Boulay, I. et al. Machine-learning-based detection of spin structures. Phys Rev Appl 21, (2024)

2024
[20]

Albarracín, F. A. G. Unsupervised machine learning for the detection of exotic phases in skyrmion phase diagrams. http://arxiv.org/abs/2404.10943 (2024)

work page arXiv 2024
[21]

Kwon, H. Y . et al. Magnetic Hamiltonian parameter estimation using deep learning techniques. Sci Adv 6, (2020)

2020
[22]

Finazzi, M. et al. Laser-induced magnetic nanostructures with tunable topological properties. Phys Rev Lett 110, (2013)

2013
[23]

Yang, S. et al. Reversible conversion between skyrmions and skyrmioniums. Nat Commun 14, (2023)

2023
[24]

A., Silva, T

Hoefer, M. A., Silva, T. J. & Keller, M. W. Theory for a dissipative droplet soliton excited by a spin torque nanocontact. Phys Rev B Condens Matter Mater Phys 82, (2010)

2010
[25]

Mohseni, S. M. et al. Spin torque-generated magnetic droplet solitons. Science (1979) 339, 1295–1298 (2013)

1979
[26]

& Kondo, K

Ishida, Y . & Kondo, K. Theoretical comparison between skyrmion and skyrmionium motions for spintronics applications. Jpn J Appl Phys 59, (2020)

2020
[27]

Suzuki, S. & be, K. A. Topological structural analysis of digitized binary images by border following. Comput Vis Graph Image Process 30, (1985)

1985
[28]

M., Horng, L

Kao, Y . M., Horng, L. & Cheng, C. H. Analytical studies of the magnetic domain wall structure in the presence of non-uniform exchange bias. AIP Adv 11, (2021)

2021
[29]

Anisotropic superexchange interaction and weak ferromagnetism

Moriya, T. Anisotropic superexchange interaction and weak ferromagnetism. Physical Review 120, (1960)

1960
[30]

A thermodynamic theory of ‘weak’ ferromagnetism of antiferromagnetics

Dzyaloshinsky, I. A thermodynamic theory of ‘weak’ ferromagnetism of antiferromagnetics. Journal of Physics and Chemistry of Solids 4, (1958)

1958
[31]

Chen, G. et al. Tailoring the chirality of magnetic domain walls by interface engineering. Nat Commun 4, (2013)

2013
[32]

Woo, S. et al. Current-driven dynamics and inhibition of the skyrmion Hall effect of ferrimagnetic skyrmions in GdFeCo films. Nat Commun 9, (2018)

2018
[33]

Ibrahim, I. a. M., Zikry, a. a. F. & Sharaf, M. a. Preparation of spherical silica nanoparticles: Stober silica. Journal of American Science 6, (2010)

2010