IMR KINKEN Research Highlights　2021

概要

IMR KINKEN Research Highlights 2021
著者
journal or
publication title
year
URL

Institute for Materials Research Tohoku
University
IMR KINKEN Research Highlights
2021-04
http://hdl.handle.net/10097/00135223

Electronic Materials

IMR KINKEN Research Highlights 2021

Research

Electronic Materials

Hopfield Neural Network with Intrinsic Learning in Magnetic Films

M

acroscopic spin ensembles possess brain-like features, which offer opportunities for neuromorphic
computing by spintronics devices. We propose a four-node Hopfield network based on magnetic textures. Its
synapses do not require external input to update weights as this occurs intrinsically via physical feedback
mechanisms built into the texture.
Over the last decade, tremendous progress has
propelled neuromorphic computing to the forefront
of information technology. However, brain-inspired
algorithms are mainly emulated by conventional von
Neumann architectures with physically separated
computing and storage units, thereby limiting the
power of artificial intelligence algorithms.
Spintronics is a leading candidate for hardwareimplemented neuromorphic computing, in which
analogue devices emulate neurons and synapse
functions. However, practical implementations
require the improved efficiency of the training and
inference processes. To the best of our knowledge,
the main problem of existing approaches is the need
for external algorithms running on conventional
computers to update the weights of the synapses
during training.
We propose a neuromorphic computing
platform with superior performance based on the
plasticity of magnetic textures [1]. The physical
requirements are a magnetic film with chiral domain
patterns generated by the Dzyaloshinskii–Moriya
interaction, anisotropic magnetoresistance, and
current-induced spin-transfer torque [1]. We
simulated the device performance by finite element
simulations, which have been successfully applied
to problems in spin cavitronics [2].
We show that an electrically conducting
magnetic film can operate as a collection of artificial
synapses with weights encoded by the conductance
between external electrodes in a matrix configuration.
A voltage applied to the magnetic thin film (Fig. 1(a))
drives the magnetic textures by the spin-transfer
torque. With changes in the magnetic configuration,
the conductance (or resistance) is modified by the
anisotropic magnetoresistance. Figure 1(b) shows
the enhanced horizontal conductance over time
under a current, as the vertical conductance
decreases. This “plasticity” encodes “weights” into
magnetic textures.
We implemented this concept as a four-node
Hopfield network in Fig.1(c). The value of four
neurons were encoded as voltage applied to four
electrodes. The weight matrix was encoded into the

Fig. 1 (a) Configuration of magnetic texture at an initial
time. Black (white) color represents up (down)
magnetization perpendicular to the film and the arrows
(inset)
represent
in-plane
magnetization.
(b)
Conductance change in either the horizontal or vertical
direction in the presence of an electric current. (c)
Schematic four-node Hopfield network.

effective conductance matrix inside the magnetic
film. The four-node Hopfield network can operate as
an associative memory to memorize and recall a
four-pixel binary picture. In the training stage, only
the pattern to be memorized was applied in terms of
binary voltages, and the network naturally updated
the weights without requiring external computations.
In the inferring stage, the network is fixed, and the
stored patterns can be recalled by standard
methods.
The learning rule is based on physical laws
rather than human interference. The concept can be
generalized to other materials with current-induced
plasticity. Our work paves the way for hardwarebased neuromorphic computing devices with onchip learning.
References
[1] W. Yu, J. Xiao, and G.E.W. Bauer, arXiv: 2101.03016 (2021).
[2] W. Yu, T. Yu, and G.E.W. Bauer, Phys. Rev. B 102, 064416
(2020).

Keywords: spin dynamics, magnetoresistance, spintronics
Weichao Yu (Theory of Solid State Physics Research Laboratory)
E-mail: wcyu@imr.tohoku.ac.jp
URL: http://www.bauer-lab.imr.tohoku.ac.jp/

26

IMR KINKEN Research Highlights 2021

IMR KINKEN Research Highlights 2021

High Magnetic Field Element Specific Magnetization to Resolve the
Magnetization Process of a Hard-Intermetallic Ferrimagnet

Nonmonotonic magnetic-field-dependence of magnetic moments in TmFe Al were determined using high magnetic field
5

7

X-ray circular magnetic dichroism. The complex behavior was caused by the competition between two different magnetic
anisotropies and exchange coupling.

Rare-earth transition metal intermetallic compounds
have been studied intensively for their strong anisotropies
and large magnetic couplings. These are an important
class of materials for hard magnets used in a wide range
of applications, such as energy conversion and transport.
High magnetic fields have been used to investigate the
anisotropies and exchange couplings of these magnets,
and important material parameters have been evaluated.
TmFe5Al7 has a ThMn12-type structure with tetragonal
symmetry. The alignment of the magnetic moments of Fe
and Tm is determined by the balance between the easyplane type anisotropy of Fe, the easy-axis type anisotropy
of Tm, and the exchange coupling between Fe and Tm. It
should be noted that the easy-axis type ferrimagnet
ground state at low temperature is rare in the RFe5Al7
family, when R (rare earth element) is heavy rare-earth.
A magnetization curve in high magnetic fields was
measured to examine the magnetic parameters of the
system at the Dresden High Magnetic Field Laboratory.
The bulk magnetization showed an upturn curvature up to
12 T, where a small jump in magnetization was observed.
The curvature then became negative above 12 T. This
indicates that there is a flop of magnetic moments at
approximately 12 T. Despite the characteristic features
found in the magnetization curve, the magnetic parameters
or the alignment of magnetic moments could not be
inferred.
The key to understanding the magnetization process
is element-specific magnetization measurement with X-ray
circular magnetic dichroism (XMCD) in a pulsed high
magnetic field, which was developed by the collaboration
between the magnetism division of the Institute for
Materials Research, Tohoku University, and BL25 at
SPring-8 [1]. In XMCD, the magnetic moment of a specific
magnetic ion is deduced by tuning the X-ray energy to the
related absorption edge. When soft X-rays are employed,
the final states of the edge absorptions are the 3d orbitals
of the transition element and the 4f orbital of the rare earth
element, respectively. This is the advantage of using soft
X-rays in XMCD despite the difficulty of using ultra-high
vacuum. In BL25 at SPring8, the XMCD can be measured
up to 40 T using a compact bipolar pulsed field generator.
Figure 1 presents the results of the element-specific

Fig. 1 Element-specific magnetization of Tm and Fe and the
alignment of magnetic moment determined by XMCD [2].

magnetization curve and the change in the magnetic
moment alignment in the magnetization process. It shows
that Fe moments always align along the magnetic field with
field-dependent tilting, while the Tm moment changes
from collinear to spin-flopped-like arrangements around
the transition found by the small bulk magnetization jump.
This result can be understood as the balance between the
magnetic anisotropies and exchange coupling. Such
complex nonmonotonic magnetic-field-dependence of the
Fe moment can be evaluated only by the XMCD technique
in high magnetic fields. This method contributes to the
study of various intermetallic magnetic compounds.
References
[1] H. Yasumura, Y. Narumi, T. Nakamura, Y. Kotani, A. Yasui, E. Kishaba,
A. Mitsuda, H. Wada, K. Kindo, and H. Nojiri, J. Phys. Soc. Jpn. 86,
054706 (2017).
[2] S. Yamamoto, D. I. Gorbunov, H. Akai, H., Yasumura, Y. Kotani, T.
Nakamura, T. Kato, N. V. Mushnikov, A. V. Adreev, and H. Nojiri, Phys.
Rev. B 101, 174430 (2020).

Keywords: high magnetic field, hard magnet, XMCD
Hiroyuki Nojiri (Magnetism Research Laboratory)
E-mail: nojiri@imr.tohoku.ac.jp
URL: http://www.hfpm.imr.tohoku.ac.jp

IMR KINKEN Research Highlights 2021

27

Electronic Materials

Conductive Oxide with Surface Polarity for Device Applications

A layered delafossite, PdCoO , is one of the most conductive oxides with a bulk conductivity comparable with
2

those of Au metals. Charged layers of Pd+ and [CoO2]− alternate in the PdCoO2 crystal structure, generating
surface polarities that strongly influence the surface/interface properties. Using pulsed laser deposition, we
successfully grew high-quality thin films of PdCoO2 and fabricated functional devices: a nanodevice showing
quantum transport and a diode that can operate at elevated temperatures. The superior properties of these
devices are attributed to the anisotropic electrical conductivity of PdCoO2 that originates from its unique layered
crystal structure.
8

Conductivity σ (107 S/m)

The layered oxide PdCoO2 has attracted
significant attention as a highly conductive metal
comparable to the noble metals (Fig. 1). Using
PdCoO2 thin films [1], we are currently exploring
novel electronic functionalities that originate from
two key features of PdCoO2: the high electron
mobility (~51000 cm2/Vs for bulk single crystals)
and surface polarity induced by the ionic layered
structure (inset of Fig. 1). By fabricating submicronscale Hall-bar devices using electron beam
lithography, we carried out the first observation of
quantum transport in PdCoO2 thin films [2]. We
found that the surface polarity of PdCoO2
significantly affects the surface electronic states,
giving the inherently nonmagnetic PdCoO2 unique
magnetic states in the vicinity of the Pd-terminated
surfaces [3]. ...

この論文で使われている画像