arxiv: 2605.01360 · v1 · submitted 2026-05-02 · ❄️ cond-mat.dis-nn

Recognition: unknown

Reservoir computing by thin film embedded with magnetic impurities

Shuto Kamakura , Tomi Ohtsuki , Jun-ichiro Ohe

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:50 UTC · model grok-4.3

classification ❄️ cond-mat.dis-nn

keywords reservoir computingmagnetic thin filmdipole-dipole interactionmagnetization dynamicshandwritten digit recognitionlong-range interactionsneuromorphic computing

0 comments

The pith

A thin film with magnetic impurities performs reservoir computing by encoding spatial patterns into time signals via long-range interactions.

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

The paper simulates magnetization dynamics in a thin film embedded with magnetic impurities, including dipole-dipole interactions, and tests it on handwritten digit recognition. Even when the output signal is spatially averaged across the sample, the system reaches high classification accuracy. The key demonstration is that long-range interactions convert complex spatial inputs into distinguishable temporal variations that a simple readout can use. A sympathetic reader would care because this removes the need for spatially resolved sensors, potentially making magnetic reservoir hardware simpler to build and operate.

Core claim

The authors establish that the long-range dipole-dipole interaction in the thin film with magnetic impurities effectively encodes the complex spatial input pattern into the time domain. This encoding allows high-accuracy handwritten digit classification even when only a spatially averaged output is accessible, showing that the magnetization dynamics can function as a reservoir without requiring full spatial readout.

What carries the argument

Magnetization dynamics driven by long-range dipole-dipole interactions among randomly placed magnetic impurities in a thin film, which map spatial input patterns onto temporal evolution of the averaged magnetization.

If this is right

The long-range interaction encodes spatial input information into accessible time-domain signals.
High classification accuracy remains possible without spatially resolved output measurements.
The approach supports simpler hardware realizations of reservoir computing in magnetic materials.

Where Pith is reading between the lines

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

This encoding mechanism could lower the experimental barrier for testing magnetic reservoir computers by removing the need for imaging-based readouts.
Similar long-range coupling effects might be explored in other condensed-matter systems to achieve temporal encoding of spatial data.
Scaling tests with varied impurity densities or larger input datasets would check whether the encoding capacity grows with system size.

Load-bearing premise

The numerical model of magnetization dynamics, including the treatment of dipole-dipole interactions and impurity placement, faithfully represents the behavior of a physically realizable thin-film system.

What would settle it

Fabricate the described thin film with embedded magnetic impurities, drive it with spatial input patterns such as handwritten digits, measure only the time-dependent spatially averaged magnetization, and test whether standard training on those signals yields classification accuracy comparable to the reported simulations.

Figures

Figures reproduced from arXiv: 2605.01360 by Jun-ichiro Ohe, Shuto Kamakura, Tomi Ohtsuki.

**Figure 2.** Figure 2: FIG. 2. Classification accuracy for two magnetic materials [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 1.** Figure 1: The maximum magnetic field is set to 1 T, and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Classification accuracy for several impurities conc [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Classification accuracy for the Ni impurities (50%). [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

read the original abstract

The reservoir computing based on the thin film embedded with magnetic impurities in the presence of the long-range (the dipole-dipole) interaction is numerically investigated. We simulated the magnetization dynamics by taking into account the dipole-dipole interaction and performed the handwritten-digit recognition task. Although the training data is prepared by taking spatial average in the sample, the high classification accuracy is achieved. Our result demonstrates that the long range interaction effectively encodes the complex spatial input pattern into the time domain, even when only a spatially averaged output is accessible. The proposed system paves the way for easily realizable magnetic reservoir computing.

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 runs a standard micromagnetic simulation of an impurity-doped thin film with dipole-dipole coupling and shows it can classify handwritten digits from the time series of the spatially averaged magnetization, but the absence of any control run without the dipole term leaves the claimed encoding mechanism unisolated.

read the letter

The main result is a numerical integration of the LLG equation on a thin film containing magnetic impurities. Spatial input patterns are applied, the dipole-dipole field is included, and the time-dependent spatially averaged magnetization is used as the reservoir state for a handwritten-digit classification task. The authors state that high accuracy is reached even though only the average is observed, and they attribute this to the long-range interaction mapping spatial structure into temporal dynamics.

Referee Report

2 major / 2 minor

Summary. The manuscript numerically investigates reservoir computing in a thin magnetic film with embedded impurities, simulating magnetization dynamics via the Landau-Lifshitz-Gilbert equation including dipole-dipole interactions. It applies the system to handwritten digit recognition and reports high classification accuracy using only the spatially averaged magnetization output, attributing this to long-range interactions that map complex spatial input patterns into the time domain. The work concludes that the approach offers an easily realizable platform for magnetic reservoir computing.

Significance. If the central result holds after isolating the mechanism, the work would demonstrate a concrete physical implementation of reservoir computing that exploits intrinsic long-range magnetic interactions to handle spatial data with minimal readout complexity. This could advance hardware realizations of reservoir computing in condensed-matter systems by reducing the need for spatially resolved measurements.

major comments (2)

[results on handwritten digit recognition] The central claim that long-range dipole-dipole interactions encode spatial input patterns into temporally varying averaged magnetization (abstract and results section) is not supported by an isolating control. No matched simulation is reported with the dipole-dipole term disabled while retaining local exchange, anisotropy, and the identical spatial input protocol; without this ablation it remains possible that accuracy arises from local precession, drive temporal structure, or impurity disorder alone. This control is load-bearing for the novelty and the phrase 'effectively encodes'.
[methods and results sections] Quantitative details required to evaluate the 'high classification accuracy' claim are insufficient. The manuscript does not report the sizes of the training and test sets, the precise mapping from digit images to spatial input fields, the number of independent runs, or error bars on accuracy; these omissions prevent assessment of robustness and reproducibility of the reported performance.

minor comments (2)

[abstract] The abstract states that high accuracy is achieved but supplies no numerical values, baselines, or parameter ranges; moving at least one quantitative result (e.g., accuracy percentage) into the abstract would improve immediate readability.
[methods] Notation for the impurity density, dipole interaction scaling, and the precise form of the spatially averaged output signal should be defined explicitly in the methods section to allow direct reproduction of the numerics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the insightful comments that have helped improve the manuscript. We address each major comment below and have made revisions accordingly.

read point-by-point responses

Referee: [results on handwritten digit recognition] The central claim that long-range dipole-dipole interactions encode spatial input patterns into temporally varying averaged magnetization (abstract and results section) is not supported by an isolating control. No matched simulation is reported with the dipole-dipole term disabled while retaining local exchange, anisotropy, and the identical spatial input protocol; without this ablation it remains possible that accuracy arises from local precession, drive temporal structure, or impurity disorder alone. This control is load-bearing for the novelty and the phrase 'effectively encodes'.

Authors: We agree that an isolating control is necessary to substantiate the role of long-range interactions. We have carried out additional simulations with the dipole-dipole term turned off, while keeping the local exchange, anisotropy, impurity configuration, and spatial input protocol the same. These control runs yield significantly lower classification accuracy, confirming that the long-range interactions are crucial for effectively encoding the spatial patterns into the time-dependent averaged magnetization. The revised manuscript will include these results and a dedicated discussion of the control study. revision: yes
Referee: [methods and results sections] Quantitative details required to evaluate the 'high classification accuracy' claim are insufficient. The manuscript does not report the sizes of the training and test sets, the precise mapping from digit images to spatial input fields, the number of independent runs, or error bars on accuracy; these omissions prevent assessment of robustness and reproducibility of the reported performance.

Authors: We apologize for not including these essential details in the original submission. The revised manuscript will report that we used a subset of the MNIST dataset consisting of 5000 training samples and 1000 test samples. The spatial input fields are generated by mapping the grayscale pixel values (0-255) to the local magnetic field amplitude applied across the film area corresponding to each pixel. We performed 10 independent runs with different random realizations of the impurity positions and initial magnetization states, and the accuracy is reported as the mean with standard deviation as error bars. These additions will allow readers to fully assess the robustness of our results. revision: yes

Circularity Check

0 steps flagged

No circularity: results follow from direct numerical integration of standard micromagnetic equations

full rationale

The paper reports numerical solution of the Landau-Lifshitz-Gilbert equation augmented by dipole-dipole sums over impurity sites, followed by spatial averaging to produce reservoir time series that are then fed to a linear readout for digit classification. No equation, parameter, or performance metric is defined in terms of the target accuracy or the spatial-to-temporal encoding claim; the reported accuracies are computed outputs of the simulation rather than re-expressions of fitted inputs. No self-citation is invoked to establish uniqueness or to justify an ansatz that would otherwise be arbitrary. The derivation chain is therefore self-contained and externally falsifiable by independent micromagnetic codes.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The claim depends on standard assumptions of micromagnetic modeling and on the numerical integration being representative of real thin-film behavior; no new entities are postulated.

free parameters (2)

impurity density and spatial distribution
Chosen in the simulation to produce the reported dynamics; exact values not stated in abstract.
dipole interaction strength scaling
Parameter controlling long-range coupling strength, implicitly set to enable the encoding effect.

axioms (1)

domain assumption Magnetization evolves according to the Landau-Lifshitz-Gilbert equation augmented by dipole-dipole fields
Standard equation set invoked for the numerical simulation of spin dynamics.

pith-pipeline@v0.9.0 · 5395 in / 1282 out tokens · 56737 ms · 2026-05-10T15:50:21.107614+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references

[1]

Jaeger and H

H. Jaeger and H. Haas, Science 304, 78 (2004)

2004
[2]

Maass, T

W. Maass, T. Natschl\" a ger, and H. Markram, Neural Comput 14, 2531 (2002)

2002
[3]

Tanaka, T

G. Tanaka, T. Yamane, J. B. H\' e roux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Netw 115, 100 (2019)

2019
[4]

Torrejon et al., Nature 547, 428 (2017)

J. Torrejon et al., Nature 547, 428 (2017)

2017
[5]

Yokouchi, S

T. Yokouchi, S. Sugimoto, K. Yoshioka, Y. Onose, S. Kasai, and S. Ono, Sci. Adv. 8, eabq5652 (2022)

2022
[6]

Nakane, G

R. Nakane, G. Tanaka, and A. Hirose, IEEE Access 6, 4462 (2018)

2018
[7]

Nomura, T

H. Nomura, T. Furuta, K. Tsujimoto, Y. Kuwabiraki, F. Peper, E. Tamura, S. Miwa, and Y. Suzuki, Jpn. J. Appl. Phys. 58, 070901 (2019)

2019
[8]

A. J. Edwards et al., Commun. Phys. 6, 215 (2023)

2023
[9]

D. A. Allwood et al., Appl. Phys. Lett. 122, 040501 (2023)

2023
[10]

I. T. Vidamour et al., Commun. Phys. 6, 230 (2023)

2023
[11]

Lee and M

M.-K. Lee and M. Mochizuki, Phys. Rev. Appl. 18, 014074 (2022)

2022
[12]

Lee and M

M.-K. Lee and M. Mochizuki, Sci. Rep. 13, 19423 (2023)

2023
[13]

S. Das, R. Mansell, L. Flaj s man, M.-A. Syskaki, J. Langer, and S. van Dijken, Phys. Rev. Appl. 23, 054043 (2025)

2025
[14]

Tsunegi, T

S. Tsunegi, T. Taniguchi, K. Nakajima, S. Miwa, K. Yakushiji, A. Fukushima, S. Yuasa, and H. Kubota, Appl. Phys. Lett. 114, 164101 (2019)

2019
[15]

Jiang, L

W. Jiang, L. Chen, K. Zhou, L. Li, Q. Fu, Y. Du, and R. H. Liu, Appl. Phys. Lett. 115, 192403 (2019)

2019
[16]

K. Hon, Y. Kuwabiraki, M. Goto, R. Nakatani, Y. Suzuki, and H. Nomura, Appl. Phys. Express 14, 033001 (2021)

2021
[17]

M. J. Donahue, and D. G. Porter, OOMMF User's Guide National Institute of Standards and Technology (1999)

1999
[18]

Deng, IEEE Signal Process

L. Deng, IEEE Signal Process. Mag. 29, 141 (2012)

2012