betaFeedback

WorksPhysics and Astronomy

Associative memory by virtual oscillator network based on single spin-torque oscillator

Abstract

A coupled oscillator network may be able to perform an energy-efficient associative memory operation. However, its realization has been difficult because inhomogeneities unavoidably arise among the oscillators during fabrication and lead to an unreliable operation. This issue could be resolved if the oscillator network were able to be formed from a single oscillator. Here, we performed numerical simulations and theoretical analyses on an associative memory operation that uses a virtual oscillator network based on a spin-torque oscillator. The virtual network combines the concept of coupled oscillators with that of feedforward neural networks. Numerical experiments demonstrate successful associations of 6060-pixel patterns with various memorized patterns. Moreover, the origin of the associative memory is shown to be forced synchronization driven by feedforward input, where phase differences among oscillators are fixed and correspond to the colors of the pixels in the pattern.

The human brain has a sophisticated function called associative memory[1], whereby it can remember a pattern when shown a portion of that pattern. This function has been modeled in various ways with the goal of achieving a better understanding of brain activity and realizing energy-efficient bio-inspired computing. Since the development of an autocorrelation model in the 1970s[2, 3, 4], several theoretical models, such as the Hopfield model[5], have been developed that draw their inspiration from the characteristics of neural activity[6, 7, 8, 9, 10, 11, 12, 13, 14, 15]. These models have also been implemented in experimental devices. For example, the associative memory operation was recently performed in a spintronic memory consisting of a nanometer-scale ferromagnetic multilayer[16]. In addition to these efforts embodying neuronal dynamics, it has been proposed that synchronized phenomena in coupled oscillator networks can be used to perform the associative memory operation[17, 18, 19, 20, 21]. For example, a detailed analysis was conducted on an LC-circuit oscillator network performing the operation[21]. A network of spintronic oscillators, called spin-torque oscillators (STOs), has also been shown to perform an associative memory operation[22].

There are two major issues with using an oscillator network for the associative memory operation. One is unstable operation due to inhomogeneity in the oscillator’s parameters. For example, variations in frequency among the oscillators are unavoidable in experimental realizations; they prevent a synchronization between the oscillators and decrease the accuracy of the associative memory[21]. The other issue is that the required number of oscillators grows with the amount of input data. There are numerous challenges in fabricating a large number of oscillators and getting them to interact with each other. These issues might be resolved if we can construct an oscillator network virtually by using a single physical oscillator[23]. Such a network would have no inhomogeneities in its parameters as only one oscillator would have to be fabricated. However, there are questions on how such a network could be realized and how it could show synchronization phenomena.

In this work, we demonstrate an associative memory operation by a virtual oscillator network through numerical simulations and theoretical analyses. First, we provide a detailed description of the virtual oscillator network consisting of a single physical oscillator. In particular, we discuss the principles involved, i.e., those of the coupled oscillator networks and feedforward neural networks. Next, we show that a virtual oscillator network consisting of a single STO can recognize several different 60-pixel patterns by numerically simulating the motion of the STO. We reveal that the feedforward input in the virtual network forces the virtual oscillators to synchronize and that this phenomenon results in the associative memory operation.

Results

Associative memory operation of this study

The associative memory operation studied here is to associate a pattern, called the pattern to be recognized, with a pattern in a stored set of patterns, called memorized patterns. For example, suppose that the three patterns, “0”, “1”, and “2”, shown in Figure 1(a) are memorized, and the one shown in Figure 1(b) is the pattern to be recognized: we can recognize that the pattern to be recognized is similar to the memorized pattern “1”. Throughout this paper, we will suppose the memorized patterns use

Image not extracted

Figure 1. Examples of memorized patterns and a pattern to be recognized. (a) Three (Nm=3N_{\mathrm{m}} = 3) memorized patterns, “0”, “1”, and “2”. (b) The pattern to be recognized resembles memorized pattern “1”. The oscillator network tries to associate the pattern to be recognized with the pattern “1”. In an associative memory operation performed by a system consisting of NN oscillators, the color of the iith (i=1,2,…,Ni = 1, 2, \ldots, N) pixel is determined by the phase ψi\psi_i of the corresponding oscillator. The color is white (black) when the phase difference, Δψi=ψi−ψ1\Delta\psi_i = \psi_i - \psi_1, is 0 (π\pi). The color is on a gray scale when the phase difference is 0<Δψi<π0 < \Delta\psi_i < \pi.

10(rows)×6(columns)=6010(\mathrm{rows}) \times6(\mathrm{columns}) = 60-pixels patterns for memorized patterns and patterns to be recognized.

In the following subsections, we describe the concept of our virtual oscillator network after briefly reviewing a conventional oscillator network for comparison. Then, we demonstrate through numerical simulations that the virtual oscillator network can perform the associative memory operation.

Associative memory operation by conventional oscillator network

The associative memory operation by a conventional coupled oscillator network consists of two steps[21]. The first step is to give a correspondence between the phases of the oscillators and the colors of the pattern to be recognized. We prepare NN oscillators corresponding to the pixels of the pattern to be recognized, where NN is the number of oscillators (pixels). We introduce phases ψi\psi_i (i=1,…,Ni = 1, \ldots, N) and phase differences Δψi=ψi−ψ1\Delta\psi_i = \psi_i - \psi_1. The color of the iith pixel is determined by cos⁡Δψi\cos\Delta\psi_i, which is white (black) when Δψi=0\Delta\psi_i = 0 (π\pi). According to this definition, the color of the first pixel is always white (see also the Methods for the definitions of color). Initially, there are no interactions between the oscillators. Thus, their phases are arbitrary, and the colors in the pattern are random, as schematically shown on the left of Figure 2(a). When interactions between the oscillators are introduced and the interaction strengths are appropriately determined by the Hebbian rule, all the phase differences become 0 or π\pi in correspondence with the white and black pixels of the pattern to be recognized, as shown in the middle of Figure 2(a) (see also Methods for model of the conventional oscillator network). Here, the Hebbian rule means that the interaction strength between the jjth and iith oscillator is proportional to the weight,

Image not extracted

Figure 2. Schematic illustration of conventional and virtual oscillator networks. (a) In the conventional oscillator network, the oscillators are initially uncoupled (left). Therefore, the phase of each oscillator is arbitrary. When the oscillators interact with appropriate weights [wij(1)][w_{ij}^{(1)}], the phases saturate to values corresponding to the pattern to be recognized (middle). When the weight changes [wij(2)][w_{ij}^{(2)}], the phases change so that the corresponding pattern resembles one of memorized patterns (right). (b) In a virtual oscillator network, we drive an oscillation of a single oscillator and divide its output into NN parts. The iith part is regarded as an output from the iith virtual oscillator. First [top of (b)], we measure the NN outputs. The corresponding pattern in this step is arbitrary because there is no correlation among the oscillators. Second [middle of (b)], an external force is added to the oscillator. This force is a linear combination of the outputs in the first step with appropriated weights [wij(1)][w_{ij}^{(1)}]. The phase of each part eventually saturates to a value corresponding to the pixel color in the pattern to be recognized. Third [bottom of (b)], the second step is repeated while the force is a linear combination of the outputs in the second step with weights wij(2)w_{ij}^{(2)}. Eventually, the phases saturate to the values corresponding to the memorized pattern most resembling the pattern to be recognized.

wij(1)=ξiRξjR,(1)w_{ij}^{(1)} = \xi_i^{\mathrm{R}}\xi_j^{\mathrm{R}}, \tag*{(1)}

where ξiR=+(−)1\xi_i^{\mathrm{R}} = +(-)1 when the color of the pattern to be recognized at the iith pixel is white (black). Thus, wij(1)=+(−)1w_{ij}^{(1)} = +(-)1 when the colors of the iith and jjth are the same (opposite).

The second step is to replace the weights by the following ones, which can be regarded as an average of the weights among the memorized patterns,

wij(2)=1Nm∑m=1Nmξimξjm,(2)w_{ij}^{(2)} = \frac{1}{N_{\mathrm{m}}}\sum_{m=1}^{N_{\mathrm{m}}}\xi_i^m\xi_j^m, \tag*{(2)}

where NmN_{\mathrm{m}} is the number of memorized patterns. The symbol m=1,2,…,Nmm = 1, 2, \ldots, N_{\mathrm{m}} is used to distinguish the memorized patterns. For example, the memorized patterns “0”, “1”, and “2” in Figure 2(a) are labelled m=1m = 1, 22, and 33. The parameter ξim\xi_i^m is +(−)1+(-)1 when the color of the iith pixel in the mmth memorized pattern is white (black). Then, the oscillator phases change to those of the memorized pattern most resembling the pattern to be recognized, and the association is achieved, as shown in the right in Figure 2(a).

Description of associative memory operation by virtual oscillator network

The associative memory operation by a virtual oscillator network consists of three steps.

First, we measure an oscillation of a single oscillator and divide it into NN parts, as schematically shown on the first line of Figure 2(b). The iith part of the measured data is regarded as the output from the iith oscillator in a virtual network. In this step, the

phase of each part is arbitrary, and therefore, the pattern arising from it is random. The measured data should be stored in a computer in order for it to be used in the next step.

Second, we excite another oscillation and divide the measured data into NN parts again. At the initial time of each part, the phase, as well as the pattern determined from it, is arbitrary, as shown in the middle of Figure 2(b). This time, however, we apply an external force to the oscillator that is proportional to a linear combination of the measured data in the first step with weights [1]. For example, in this study, the external force comes from a torque excited by an external magnetic field, which applied during the iith part of the oscillation is given by

Hi(1)=H∑j=1Nwij(1)yj(1),(3)H_i^{(1)}=\mathcal{H}\sum_{j=1}^{N}w_{ij}^{(1)}y_j^{(1)}, \tag*{(3)}

where H\mathcal{H} denotes the amplitude and yj(1)y_j^{(1)} is the output from the jjth oscillator measured in the first step [see also Methods for the detailed definition of yj(1)y_j^{(1)} in the numerical simulations]. Therefore, Eq. (3) is an oscillating function with the frequency of the oscillator. Because of the application of the magnetic field, the phase in each part eventually saturates to a certain value, and the pattern to be recognized is output, as shown in the middle of Figure 2(b). Note that the output signal of this process should be stored in a computer.

Third, we perform a measurement similar to one in the second step but the magnetic field applied during the iith part is replaced by

Hi(2)=H′∑j=1Nwij(2)yj(2),(4)H_i^{(2)}=\mathcal{H}'\sum_{j=1}^{N}w_{ij}^{(2)}y_j^{(2)}, \tag*{(4)}

where H′\mathcal{H}' denotes the amplitude, while yj(2)y_j^{(2)} is the output from the jjth oscillator measured at the second step (see also Methods pertaining to the numerical simulations). The weights wij(2)w_{ij}^{(2)} are given by Eq. (2). The phase at the end of each part saturates to a value corresponding to the memorized pattern most resembling the pattern to be recognized, as shown in the bottom of Figure 2(b); i.e., the associative memory operation is completed.

There are several differences between the conventional and virtual oscillator networks (see also Methods for the models). For example, the oscillators in the conventional oscillator network interact instantaneously, and their phase differences saturate to values corresponding to pixel colors as a result of mutual synchronization. On the other hand, the oscillators in the virtual oscillator network do not interact each other instantaneously. As can be seen in Eqs. (3) and (4), the oscillator outputs from the previous steps are used in the magnetic field in the current step. From perspective, the virtual oscillator network is similar to a feedforward neural network because the information on the oscillator phases in one step is sent to the oscillation in the next step. At the same time, we should note that the weights in the virtual oscillator network are fixed, as in the case of the conventional oscillator network. This is in contrast with a feedforward neural network used in deep learning, in which weights are updated by backpropagation. Thus, the virtual oscillator network can be regarded as a hybrid combination of a coupled oscillator network and a feedforward neural network. In the discussion below, we will reveal that the feedforward inputs cause forced synchronization among the divided parts and result in the associative memory operation. Before that, however, we must demonstrate that this virtual oscillator network can actually perform the associative memory operation.

Equation of motion of oscillator

As an oscillator in the virtual oscillator network, we use a vortex STO, which has various advantages for practical applications and has been frequently used in spintronics experiments on bio-inspired computing[24, 25, 26, 27]. An STO consists of a ferromagnetic/nonmagnetic multilayer on the nanometer scale, as schematically shown in Figure 3(a). A vortex of magnetic moments appears when a diameter and thickness of a cylinder-shape ferromagnet are on the order of 100 and 1 nm, respectively. When an electric current and/or magnetic field are applied to the STO, magnetic moments show precessions around their equilibrium direction. According to a recent experiment on chaos excitation in an STO[28], we assume that a force added to the virtual oscillator network corresponds to a torque excited by magnetic field, as mentioned above. It has been shown both experimentally and theoretically that the dynamics in a vortex STO are well described by the Thiele equation[29, 30, 31, 32, 33, 34, 35, 36], which is the equation of motion for a center of the vortex structure, called the vortex core (see also Methods for Thiele equation):

Image not extracted

Figure 3. Description of STO and demonstration of associative memory by a virtual oscillator network. (a) Schematic illustration of vortex spin torque oscillator and (b) vortex-core dynamics driven by electric current. The STO has a cylindrical shape, and the zz axis is orthogonal to the circular plane. Magnetic moments, shown as colored arrows in top ferromagnet, form a circular structure. The black dot around which the moments turn is the vortex core. Electric current is injected into the STO; positive current flows from bottom to top in the figure. When the electric current density JJ exceeds a threshold value, the vortex core oscillates around the disk center. The output signals from the STO during the first (second) step in Figure 2(b) are stored, and their linear combination with weights wij(1)w_{ij}^{(1)} [wij(2)w_{ij}^{(2)}] defined from the pattern to be recognized (memorized patterns) is used as magnetic field during the second (third) step. For simplicity, the dynamics in the absence of the magnetic field is shown. The components of the vortex-core’s position, X/RX/R, and Y/RY/R, oscillate around the disk center, and a trajectory is approximately a circle. The distance of the vortex-core’s position from the disk center, ss, is approximately constant value, s0s_0. The phase measured from the xx axis is denoted as ψ\psi. (c) Time evolutions of the 59 phase differences, Δψi\Delta\psi_i (i=2,3,⋯ ,60i = 2, 3, \cdots, 60) and (d) snapshots of generating a pattern to be recognized on 60-pixels. (e) Time evolutions of the phase difference and (f) snapshots of the corresponding pattern for association from memorized patterns.

−Gez×X˙−∣D∣(1+ξs2)X˙−κ(1+ξs2)X+aJJpzez×X+caJR0pxex+cμ∗ez×H=0,(5)-G\boldsymbol{e}_z\times\dot{\boldsymbol{X}}-|D|(1+\xi s^2)\dot{\boldsymbol{X}}-\kappa(1+\xi s^2)\boldsymbol{X}+a_JJp_z\boldsymbol{e}_z\times\boldsymbol{X}+ca_JR_0p_x\boldsymbol{e}_x+c\mu^*\boldsymbol{e}_z\times\boldsymbol{H}=0, \tag*{(5)}

where X=(X,Y,0)\boldsymbol{X}=(X,Y,0) represents the position of the vortex core in the xyxy plane. While the physical meanings and the values of many parameters are explained in Methods, two quantities should be explained here. The first is the current density JJ, which causes a limit-cycle oscillation of the vortex core. The other is the external magnetic field H\mathbf{H}, which is used to excite a torque. It is useful to notice that Eq. (5) can be approximated as (see also Methods for the analytical solution of the Thiele equation)

s˙=as−bs3+cμ∗GRHysin⁡ψ,(6)\dot{s}=as-bs^{3}+\frac{c\mu^{*}}{GR}H_{y}\sin\psi, \tag*{(6)}
ψ˙=κG(1+ζs2)+cμ∗GRsHycos⁡ψ,(7)\dot{\psi}=\frac{\kappa}{G}\left(1+\zeta s^{2}\right)+\frac{c\mu^{*}}{GRs}H_{y}\cos\psi, \tag*{(7)}

where s=∣X∣/Rs=|\mathbf{X}|/R (0≤s≤10\leq s\leq1) is the distance of the vortex core from the center of the ferromagnet normalized by the disk radius RR, while ψ=tan⁡−1(Y/X)\psi=\tan^{-1}(Y/X) is the phase. Here, a=(∣D∣κ/G2)[(J/Jc)−1]a=(|D|\kappa/G^{2})[(J/J_{\mathrm{c}})-1] and b=(∣D∣κ/G2)(ξ+ζ)b=(|D|\kappa/G^{2})(\xi+\zeta), where Jc=∣D∣κ/(GaJpz)J_{\mathrm{c}}=|D|\kappa/(GaJ_{\mathrm{p}}z). The magnetic field H\mathbf{H} is assumed to have only a yy component HyH_y. Note that Eqs. (6) and (7) are similar to the equation of motion of the Stuart-Landau oscillator[37]. Therefore, the vortex core shows a limit-cycle oscillation around the disk center in the xyxy plane with an oscillating amplitude s0=a/bs_0=\sqrt{a/b} when JJ exceeds a threshold value JcJ_{\mathrm{c}}, while the terms related to HyH_y act as a perturbation. The connection to such a fundamental nonlinear oscillator model indicates that our results are also valid for various oscillators in nature and engineering. Figure 3(b) shows an example of nonperturbative vortex dynamics, showing an approximately circular oscillation of the vortex core around the disk center. The phase difference of the oscillation was used to define the colors in the patterns in the associative memory operation. Readers should note that the plots in Figure 3(b), as well as the results of the numerical simulations shown below, were obtained by solving Eq. (5), while the approximate equations, Eqs. (6) and (7), are used in the model analyses described below.

Demonstration of associative memory

Figure 3(c) shows the time evolution of the phase difference, Δψi\Delta\psi_i, obtained by solving Eq. (5) with Eq. (3) substituting for HyH_y. Note that this solution corresponds to the second step in Figure 2(b). The phase differences saturate to 0 or π\pi within a few hundred nanoseconds. Snapshots of patterns corresponding to this time evolution of the phases are shown in Figure 3(d). The patterns eventually settle to the one to be recognized. Figure 2(b) shows the result of solving Eq. (5) with Eq. (4) substituting for HyH_y. Here, Eq. (2) in Eq. (4) is for the three memorized patterns in Figure 1(a). Figure 3(e) and Figure 3(f) show the time evolution of the phase differences and snapshots of the corresponding patterns. Remind that the information of the phases corresponding to the colors of the pixels in the pattern to be recognized is included in the magnetic field in Eq. (4) through yj(2)y_j^{(2)}. Consequently, even though the initial pattern is random, the oscillator phases finally saturate to values corresponding to one of the memorized patterns [Figure 3(f)].

The associative memory operation becomes more difficult when there are similar memorized patterns. To clarify this point, let us examine what happens when the number of the memorized patterns is increased, as shown in Figure 4(a) from the three in Figure 1(a). The added patterns do not affect the second step in Figure 2(b). For the association corresponding to the third step in Figure 2(b), the magnetic field, defined by Eq. (4), is changed by these new memorized patterns. As a result, the final pattern output resembles none of the memorized ones [Figure 4(b)].

Image not extracted

Figure 4. Problem of associative memory operation when the similarity between the memorized patterns is high and the number of patterns is large. (a) Ten (Nm=10N_m = 10) memorized patterns, “0”, “1”, ⋯\cdots, “9”. (b) Time evolution of the phase difference during the association and snapshots of the corresponding pattern. In this case, the memorized patterns include both “1” and “7”. Because of their similarity, the pattern does not finally saturate to “1”. (c) When “7” is removed from the memorized patterns (Nm=9N_m = 9), the association is successful, even though there are nine remaining memorized patterns. (d) The association is successful when the memorized patterns include only “1” and “7”.

This failure of the associative memory operation is due to two reasons. The first is that the pattern “7” is similar to the pattern “1”, which should be the one associated. When “7” is excluded from the memorized patterns, the association succeeds, as shown in Figure 4(c). The second reason is that the number of memorized patterns is large. As shown in Figure 4(d), the association succeeds when the memorized patterns include only “1” and “7”, the association is succeeded. Therefore, we conclude that an association may fail when the memorized patterns include similar patterns and the number of memorized patterns is large.

To quantify the similarity between patterns AA and BB, we introduce the degree of overlap:

O(ξA,ξB)≡1N∣∑i=1NξiAξiB∣,(8)\mathcal{O}(\boldsymbol{\xi}^{A},\boldsymbol{\xi}^{B})\equiv\frac{1}{N}\left|\sum_{i=1}^{N}\xi_i^{A}\xi_i^{B}\right|, \tag*{(8)}

where ξA=(ξ1A,…,ξNA)\boldsymbol{\xi}^{A}=(\xi_1^{A},\ldots,\xi_N^{A}) is defined from the color of the iith pixel of pattern AA [ξiA=+(−)1\xi_i^{A}=+(-)1 when the iith pixel is white [black]]. The overlap becomes 1 when the two patterns are completely identical or their black and white colors are all exchanged (see also Methods for the definitions of color and overlap). For example, in the example shown in Figure 1 and Figure 3, the degree of overlap between the pattern to be recognized and the memorized pattern “0” is O(ξR,ξ1)=18/60=0.30\mathcal{O}(\boldsymbol{\xi}^{R},\boldsymbol{\xi}^{1})=18/60=0.30. It is O(ξR,ξ2)=44/60≃0.73\mathcal{O}(\boldsymbol{\xi}^{R},\boldsymbol{\xi}^{2})=44/60\simeq0.73 for pattern “1”, and O(ξR,ξ3)=6/60=0.10\mathcal{O}(\boldsymbol{\xi}^{R},\boldsymbol{\xi}^{3})=6/60=0.10 for pattern “2” (the memorized patterns are labelled as m=1,2,3,⋯m=1,2,3,\cdots while the examples of memorized patterns in this work are “0”, “1”, “2”, etc.; thus, the label mm and the corresponding number are off by one). Since the degree of overlap of the pattern to be recognized and “1” is large in the examples in Figure 1 and Figure 3, pattern “1” should be associated in this case. On the other hand, in the example shown in Figure 4,

the overlap between the pattern to be recognized [Fig. 1(b)] and “7” is also relatively large, i.e., O(ξR,ξ8)=32/60≃0.53\mathcal{O}(\boldsymbol{\xi}^{R},\boldsymbol{\xi}^{8})=32/60\simeq0.53. In addition, the overlap between the memorized patterns “1” and “7”, O(ξ2,ξ8)=28/60≃0.47\mathcal{O}(\boldsymbol{\xi}^{2},\boldsymbol{\xi}^{8})=28/60\simeq0.47, is also relatively large compared with those between the other patterns; for example, the overlap between “1” and “8” is O(ξ2,ξ9)=2/60≃0.03\mathcal{O}(\boldsymbol{\xi}^{2},\boldsymbol{\xi}^{9})=2/60\simeq0.03 (see also Supplementary Information, where the overlaps of the ten memorized patterns are summarized). Accordingly, when the memorized patterns include “1” and “7”, the virtual oscillator network cannot associate a correct pattern, and the final pattern produced corresponds to none of the memorized ones. Similarly, when the number of memorized patterns is large, there might be patterns having large overlaps and the association fails.

In summary, we have shown that the virtual oscillator network based on the algorithm in Figure 2(b) can perform the associative memory operation. Its accuracy, however, is low when the memorized patterns include some patterns having large overlaps and there is a large number of memorized patterns. Note that the maximum number of patterns that can be memorized by neural network is approximately N/(2log⁡N)N/(2\log N) [8]. It would be of interest if such a formula can be derived for virtual oscillator networks in future.

We examined the associative memory operation for various cases, i.e., for different patterns to be recognized, and studied the rate of the accurate association; see Supplementary Information.

Discussion

Here we discuss the principles of the associative memory operation analytically by using Eqs. (6) and (7). As mentioned above, the operation consists of three steps, and in each step, the oscillator output is divided into NN parts. In what follows, we denote the phase of the vortex core during the iith part of the kkth step as ψi(k)\psi_i^{(k)}. We also assume that the oscillation amplitude s0s_0 is approximately constant because the current density is fixed. Therefore, the oscillation frequency, f=Ω/(2π)=[κ/(2πG)](1+ζs02)f=\Omega/(2\pi)=[\kappa/(2\pi G)](1+\zeta s_0^2), is also approximately constant (see also Methods for the analytical solution of the Thiele equation).

The phase in the second step obeys,

ψ˙i(2)=Ω+cμ∗GRs0ξiR∑ℓ=1NξℓRyℓ(1)cos⁡ψi(2).(9)\dot{\psi}_i^{(2)}=\Omega+\frac{c\mu^*}{GRs_0}\xi_i^R\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)}\cos\psi_i^{(2)}. \tag*{(9)}

Thus, the phase difference between the iith and jjth parts obeys,

ψ˙i(2)−ψ˙j(2)=cμ∗GRs0H(∑ℓ=1NξℓRyℓ(1))(ξiRcos⁡ψi(2)−ξjRcos⁡ψj(2)).(10)\dot{\psi}_i^{(2)}-\dot{\psi}_j^{(2)}=\frac{c\mu^*}{GRs_0}\mathcal{H}\left(\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)}\right)\left(\xi_i^R\cos\psi_i^{(2)}-\xi_j^R\cos\psi_j^{(2)}\right). \tag*{(10)}

The steady state condition on the phase difference leads to

ξiRcos⁡ψi(2)−ξjRcos⁡ψj(2)=0.(11)\xi_i^R\cos\psi_i^{(2)}-\xi_j^R\cos\psi_j^{(2)}=0. \tag*{(11)}

Note that ξiR=+(−)1\xi_i^R=+(-)1 when the color at the iith pixel of the pattern to be recognized is white (black). Therefore, ψi(1)\psi_i^{(1)} and ψj(1)\psi_j^{(1)} will be in-phase ψi(1)=ψj(1)\psi_i^{(1)}=\psi_j^{(1)} [anti-phase ψi(1)=ψj(1)±π\psi_i^{(1)}=\psi_j^{(1)}\pm\pi] when the colors of the iith and jjth pixels are the same (opposite). As a result, the phase differences in the second step saturate to 00 or π\pi corresponding to the white or black in the pattern to be recognized. Note that this synchronization is caused by a feedforward input from the first step, which corresponds to the second term on the right-hand side in Eq. (9). Here, the term ∑ℓ=1NξℓRyℓ(1)\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)} in Eq. (9) is the sum of the NN oscillator outputs yℓ(1)y_\ell^{(1)} in the first step, multiplied by the factor ξℓR\xi_\ell^R determining the pixel color of the pattern to be recognized, and is common for all ii of Eq. (9). Equation (9) also includes a factor ξiR\xi_i^R, which determines the sign of the input. Regarding these facts, the feedforward input has only two values, depending on the value of ξiR\xi_i^R. The phase synchronization among the NN parts in the second step is the result of forced synchronization with respect to this feedforward input, and the phase difference has only two values, 00 or π\pi, depending on the value of ξiR\xi_i^R. This mechanism is in contrast with that of the previous work [21], where a mutual synchronization is the origin of the associative memory operation. Also, the method is different from the previous works [38, 39]. In Ref. [38], a forced synchronization of frequency with respect to an external signal was studied, while the input signal in the present work is generated by the oscillator output itself and the phase synchronization plays the central role in the associative memory operation. In Ref. [39], a delayed-feedback was used to generate input signal, while the input signal in the present work is generated by multiplying appropriated weight to perform the associative memory operation.

We also note that, when yℓ(1)y_\ell^{(1)} is a simple trigonometric function, its linear combination, ∑ℓ=1NξℓRyℓ(1)\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)}, is also a trigonometric function with the same frequency and a different phase. According to the above discussion, the phase of the term ∑ℓ=1NξℓRyℓ(1)\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)} does not play any role to excite forced synchronization among the NN parts. Thus, the term ∑ℓ=1NξℓRyℓ(1)\sum_{\ell=1}^{N}\xi_\ell^R y_\ell^{(1)} could be replaced by, for example, y1(1)y_1^{(1)}. In this case, it is unnecessary to measure other (N−1)(N-1) yℓ(1)y_\ell^{(1)} (ℓ=2,3,⋯ ,N\ell= 2,3,\cdots,N) in the first step in Figure 2(b), although we solved the equation of motion for NN virtual oscillators to clarify similarities and differences between the second and third step. When (N−1)(N-1) parts in the first step are omitted for simplicity, the power consumption to drive the oscillator in the virtual oscillator network is proportional to 2N+12N+1, where 2N2N comes from the second and third steps in Figure 2(b). On the other hand, the power consumption in the conventional oscillator network is proportional to 2N2N because NN oscillators are driven two times, as implied in Figure 2(a). For a large NN, the power consumption of two oscillator networks are comparable. The time required for the operation increases linearly as NN increases, which is not suitable for practical applications, although the same might be true for a conventional oscillator network because the relaxation time of the phase will depend on the number of the oscillators. the time of a conventional (coupled) oscillator network might also increase as NN increases. However, the virtual oscillator network has an advantage from a viewpoint of reliability, as discussed below.

Next, we focus on the third step, where the phase during the iith part obeys

ψi(3)=Ω+cμ∗GRs0H′1Nm∑m=1Nm∑ℓ=1Nξℓmξℓmyℓ(2)cos⁡ψi(3).(12)\psi_i^{(3)}=\Omega+\frac{c\mu^*}{GRs_0}\mathcal{H}'\frac{1}{N_m}\sum_{m=1}^{N_m}\sum_{\ell=1}^{N}\xi_\ell^m\xi_\ell^m y_\ell^{(2)}\cos\psi_i^{(3)}. \tag*{(12)}

Since the oscillators in the second step are in the synchronized state, the output yℓ(2)y_\ell^{(2)} can be expressed as yℓ(2)=ξℓRξ1Ry1(2)y_\ell^{(2)}=\xi_\ell^{\mathcal{R}}\xi_1^{\mathcal{R}}y_1^{(2)}, where y1(2)y_1^{(2)} is the output of the first part in the second step. We substitute this relation into Eq. (12) and assume that

∑ℓ=1NξℓmξℓA≃δm,A∑ℓ=1NξℓAξℓA,(13)\sum_{\ell=1}^{N}\xi_\ell^m\xi_\ell^{\mathcal{A}}\simeq\delta_{m,\mathcal{A}}\sum_{\ell=1}^{N}\xi_\ell^{\mathcal{A}}\xi_\ell^{\mathcal{A}}, \tag*{(13)}

where the symbol A\mathcal{A} corresponds to a pattern in the memorized patterns that resembles the pattern to be recognized. The assumption (13) means that only a pattern having a large degree of overlap with the pattern to be recognized contributes to the feedforward input. The other memorized patterns, which are greatly different from the pattern to be recognized, do not contribute to the feedforward input because of their small overlap. When the assumption is satisfied, Eq. (12) becomes

ψi(3)=Ω+cμ∗GRs0H′1Nmy1(2)ξ1A(∑ℓ=1NξℓAξℓA)ξiAcos⁡ψi(3).(14)\psi_i^{(3)}=\Omega+\frac{c\mu^*}{GRs_0}\mathcal{H}'\frac{1}{N_m}y_1^{(2)}\xi_1^{\mathcal{A}}\left(\sum_{\ell=1}^{N}\xi_\ell^{\mathcal{A}}\xi_\ell^{\mathcal{A}}\right)\xi_i^{\mathcal{A}}\cos\psi_i^{(3)}. \tag*{(14)}

Equation (14) is similar to Eq. (9), and therefore, the steady-state condition of the phase difference between the iith and jjth parts in the third step is given by

ξiAcos⁡ψi(3)−ξjAcos⁡ψj(3)=0.(15)\xi_i^{\mathcal{A}}\cos\psi_i^{(3)}-\xi_j^{\mathcal{A}}\cos\psi_j^{(3)}=0. \tag*{(15)}

Equation (15) means that in-phase or anti-phase synchronization between the NN parts occurs, and the phase differences in the third step saturate to 0 or π\pi corresponding to the white or black colors in a memorized pattern most resembling the one to be recognized.

The operation principle is based on Eq. (13). Equation (13) is satisfied if there is only one pattern that has a large degree of overlap with the pattern to be recognized. On the other hand, if there are other patterns having large overlaps with the pattern to be recognized, Eq. (13) is not satisfied. In this case, Eq. (15) is not necessarily satisfied, and the colors in the steady state in the third step might be different from the pattern most resembling the one to be recognized or they might be gray (neither black nor white); see also Supplementary Information.

Our analysis also assumed that the oscillation frequencies of the NN parts are the same. This assumption is a natural one because each part is obtained from a single oscillator. Technically speaking, the oscillation frequency in each part is varied by changing the magnitude of the electric current. If the oscillation frequencies of the iith and jjth parts, denoted as Ωi/(2π)\Omega_i/(2\pi) and Ωj/(2π)\Omega_j/(2\pi), are different, the right-hand side of Eq. (10) has an additional term Ωi−Ωj\Omega_i-\Omega_j. In such a case, the phase difference is not well defined because ψi\psi_i and ψj\psi_j oscillate with different frequencies. Even if we introduce an instantaneous phase by, for example, making a Hilbert transformation, as was done in experiments[40], the phase difference still does not necessarily saturate to 0 or π\pi. In such a case, the associative memory operation fails. Therefore, there is no reason to change the oscillation frequency in each part. This fact also indicates an advantage to using the virtual oscillator network. In the conventional oscillator network, variations in the oscillation frequency naturally appear because inhomogeneities in the parameters of the oscillators are unavoidable, and such variations lead to the failure of the associative memory operation[21]. The virtual oscillator network does not have such variation and thus would be a more reliable associative memory. A weak point of the present proposal is, on the other hand, that the method requires a computer to store the output signal in each step, which is not preferable for practical applications. We would like to keep this issue as a future work.

In conclusion, we described the concept of the associative memory operation by a virtual oscillator network and performed numerical simulations. The operation consists of three steps, where the output of one step is sent to the next step with weights defined by the Hebbian rule. In this sense, the virtual oscillator network can be regarded as a hybrid combination of a coupled oscillator network and a feedforward neural network. The network successfully associated black-and-white patterns with a few memorized patterns. However, it failed to make an association when the number of memorized patterns was large (ten compared to three) and some of the memorized patterns resembled each other. We also developed a theoretical analysis and clarified that the origin of the associative memory operation is forced synchronization driven by feedforward input. Either in-phase or anti-phase synchronization was excited among the oscillators and provides appropriate correspondence between the oscillator phases and the colors in the patterns. The virtual oscillator network is more reliable than a conventional oscillator network, which is affected by unavoidable inhomogeneities among the oscillators.

Methods

Definitions of color and overlap

By convention, the first pixel (the pixel in the top left-hand corner of a pattern) is always white. The pattern should be regarded as the same even when all of the black and white pixels are swapped for each other, Mathematically, this means that ∑i=1NξiAξiB=N\sum_{i=1}^{N} \xi_i^A \xi_i^B=N when the patterns A and B are completely the same, and ∑i=1NξiAξiB=−N\sum_{i=1}^{N} \xi_i^A \xi_i^B=-N when patterns A and B represent the same pattern but their black and white colors are completely swapped. According to this definition of the same figure, the maximum number of the difference between two patterns is N/2N/2; in this case, the degree of overlap is zero (see also the discussion on noise in Supplementary Information).

Models of conventional and virtual oscillator networks

The conventional oscillator network for the associative memory operation[21] is based on the Kuramoto model[37]. The Kuramoto model describes the oscillator dynamics with a generalized phase, θ\theta. Moreover, the oscillators interact instantaneously, and the phase of the iith oscillator obeys

θ˙i=ω+Q∑j=1Nwijsin⁡(θi−θj),(16)\dot{\theta}_i=\omega+\mathcal{Q}\sum_{j=1}^{N}w_{ij}\sin(\theta_i-\theta_j), \tag*{(16)}

where ω/(2π)\omega/(2\pi) is the oscillation frequency while Q\mathcal{Q} is the interaction strength. For simplicity, we will assume that all oscillators share the same values of ω\omega and Q\mathcal{Q}. The weight wijw_{ij} is given by Eq. (1) or (2) depending on the step of the procedure. In the LC-circuit model[21], Qwij\mathcal{Q}w_{ij} is proportional to the transconductance. The phase difference between the iith and jjth oscillators obeys

θ˙i−θ˙j=Q[∑ℓ=1Nwiℓsin⁡(θi−θℓ)−∑ℓ=1Nwjℓsin⁡(θj−θℓ)].(17)\dot{\theta}_i-\dot{\theta}_j=\mathcal{Q}\left[\sum_{\ell=1}^{N}w_{i\ell}\sin(\theta_i-\theta_\ell)-\sum_{\ell=1}^{N}w_{j\ell}\sin(\theta_j-\theta_\ell)\right]. \tag*{(17)}

In a limiting case of only two oscillators (N=2N=2), the phase difference obeys

θ˙1−θ˙2=2Qw12sin⁡(θ1−θ2),(18)\dot{\theta}_1-\dot{\theta}_2=2\mathcal{Q}w_{12}\sin(\theta_1-\theta_2), \tag*{(18)}

and the in-phase (anti-phase) synchronization of θ1\theta_1 and θ2\theta_2 is a stable fixed point when 2Qw122\mathcal{Q}w_{12} is negative (positive). The phase differences of θi−θj=0,π\theta_i-\theta_j=0,\pi are always fixed points even when there are a large number of oscillators (N≥3N\geq3). Accordingly, the phase differences in the conventional oscillator network saturate to the in-phase or anti-phase state, which thereby enables the associative memory operation.

In the presence of frequency variations, the right-hand side of Eq. (17) has an additional term ωi−ωj\omega_i-\omega_j. In this case, the phase difference is not stabilized, and this instability leads to an inaccurate associative memory operation[21].

The Thiele equation is slightly different from the Kuramoto model in the following ways. First, the Thiele equation uses the phase ψ\psi, which describes the vortex core’s position in the xyxy plane, instead of a generalized phase. This is because the quantity measured in experiments is the vortex core’s position, and the phase synchronization studied in the experiments[40] corresponds to that of ψ\psi, not a generalized phase θ\theta. Note that we can introduce a generalized phase analytically as θ=ψ+[ζκ/(Gb)]ln⁡(s/s0)\theta=\psi+\left[\zeta\kappa/(Gb)\right]\ln(s/s_0) with a phase sensitivity function Z=(−sin⁡θ+[ζκ/(Gb)]cos⁡θ,cos⁡θ+[ζκ/(Gb)]sin⁡θ,0)/s0\mathbf{Z}=(-\sin\theta+[\zeta\kappa/(Gb)]\cos\theta,\cos\theta+[\zeta\kappa/(Gb)]\sin\theta,0)/s_0. The analysis is mostly unchanged with the generalized phase, so we decided to use ψ\psi for simplicity. Second, the equation of motion for the phase difference, Eq. (10), includes a term cos⁡ψi−cos⁡ψj\cos\psi_i-\cos\psi_j whereas the Kuramoto model often uses an interacting term proportional to sin⁡(θi−θj)\sin(\theta_i-\theta_j). More generally, the interaction term in the Kuramoto model can be assumed to be a function of the phase difference, θi−θj\theta_i-\theta_j after applying an averaging technique with respect to a fast variable (see Ref. 37 for details). The difference between the two models might however be insignificant; notice that, by using formulas, cos⁡x−cos⁡y=−2sin⁡[(x+y)/2]sin⁡[(x−y)/2]\cos x-\cos y=-2\sin[(x+y)/2]\sin[(x-y)/2] and cos⁡x+cos⁡y=2cos⁡[(x+y)/2]cos⁡[(x−y)/2]\cos x+\cos y=2\cos[(x+y)/2]\cos[(x-y)/2] and applying the averaging technique, the interaction term in our model can be approximated as a function of θi−θj\theta_i-\theta_j. Third, as mentioned above, the input term in the virtual oscillator network consists of the oscillator output from the previous step, while the interaction in the Kuramoto model is instantaneous. Because of these differences, the associative memory operation by the virtual oscillator network is significantly different from those of conventional coupled oscillator networks on which previous experiments and the theoretical analyses have been conducted.

Parameters in the Thiele equation

Spin torque oscillators (STOs) mainly consist of a ferromagnetic metal/insulating layer/ferromagnetic metal trilayer. The first ferromagnetic layer of the trilayer is called the free layer and is where the magnetic vortex forms. The second ferromagnetic layer having a uniform magnetization is called the reference layer. When electric current is injected into STOs, spin-transfer torque[41, 42, 43] is excited on the magnetic moments in the free layer and drives their dynamics[35, 36]. The output signal from the STOs depends on the relative angle between the magnetizations in the free and reference layers.

The definitions and physical meanings of the parameters in Eq. (5) are as follows. The parameters G=2πpML/γG=2\pi pML/\gamma and D=−(2παML/γ)[1−(1/2)ln⁡(R0/R)]D=-(2\pi\alpha ML/\gamma)[1-(1/2)\ln(R_0/R)] consist of the polarity p(=±1)p(=\pm1) of the vortex core, the saturation magnetization MM, the thickness LL of the ferromagnet, the gyromagnetic ratio γ\gamma, the Gilbert damping constant α\alpha, and the vortex radius R0R_0. The chirality c(±1)c(\pm1) of the vortex core also appears in Eq. (5). The parameters κ\kappa and ζ\zeta relate to a magnetic potential energy defined as W=(κ/2)[1+(ζ/2)s2]∣X∣2W=(\kappa/2)[1+(\zeta/2)s^2]|X|^2. The dimensionless parameter ξ\xi is introduced to describe the nonlinear damping in a highly excited state[35]. The parameter κ\kappa relates to the material parameters as κ=(10/9)4πM2L2/R35\kappa=(10/9)4\pi M^2L^2/R^{35}. The parameter aJ=πℏP/(2e)a_J=\pi\hbar P/(2e) includes the reduced Planck constant ℏ\hbar, spin polarization PP of the electric current, and the elementary charge e(>0)e(>0). The vector p=(px,0,pz)\mathbf{p}=(p_x,0,p_z) is the unit vector pointing in the magnetization direction in the reference layer. Here, we assume that p\mathbf{p} lies in the xzxz plane, by convention. As a result, the output signal from the vortex STO is proportional to the yy component of the vortex core’s position. The parameter μ∗=πMLR\mu^*=\pi ML R.

The material parameters used in the numerical study were taken from typical experiments and simulations[35, 36, 44]: M=1300M=1300 emu/cm3^3, γ=1.764×107\gamma=1.764\times10^7 rad/(Oe s), α=0.01\alpha=0.01, L=5L=5 nm, R=187.5R=187.5 nm, R0=10R_0=10 nm, P=0.7P=0.7, ξ=2.0\xi=2.0, and ζ=0.1\zeta=0.1. The polarity and chirality were assumed to be p=+1p=+1 and c=+1c=+1, for simplicity. The magnetization direction in the reference layer was p=(sin⁡60∘,0,cos⁡60∘)\mathbf{p}=(\sin60^\circ,0,\cos60^\circ). An electric current II of 1 mA corresponded to a current density JJ of 0.9 MA/cm2^2. The electric current in the numerical simulations was set to 4.0 mA.

We do not include field-like torque in the Thiele equation, which is expressed as −cbJJRpxey-cb_J J R p_x\mathbf{e}_y in Eq. (5); see, for example, Ref.[45]. This is because its magnitude was not visible in an experiment using CoFeB/MgO based STO[23]. One might consider to inject the input through the field-like torque, instead of the torque due to the external magnetic field as we have done. However, the modulation of the field-like torque requires that of electric current, which leads to the modulation of the frequency of the STO. Since the advantage of our proposal is that the frequency is unique during the operation, we do not prefer to use the field-like torque for injecting the input.

Analytical solution of the Thiele equation

The Gilbert damping constant α\alpha is often small, in such cases, ∣D∣/G≃α≪1|D|/G\simeq\alpha\ll1. Also, the radius R0R_0 of the vortex core is much shorter than the disk radius, RR. Therefore, by neglecting terms related to R0R_0 and higher-order terms of α\alpha, we can approximate Eq. (5) as Eqs. (6) and (7) in terms of s=∣X∣/Rs=|X|/R and ψ=tan⁡−1(Y/X)\psi=\tan^{-1}(Y/X). The approximated Thiele equation without magnetic field is

s˙=as−bs3,(19)\dot{s}=as-bs^3, \tag*{(19)}
ψ˙=κG(1+ξs2).(20)\dot{\psi}=\frac{\kappa}{G}(1+\xi s^2). \tag*{(20)}

These equations are identical to the Stuart-Landau equation[37], which was introduced by Landau to describe the evolution of turbulence phenomenologically and was derived from hydrodynamics by Stuart. This equation provides one of the simplest example of Hopf bifurcation. A stable solution of ss is s0=a/bs_0=\sqrt{a/b} (00 for a>(<)0a>(<)0, or equivalently, J/Jc>(<)1J/J_c>(<)1. When J/Jc>1J/J_c>1, i.e., the current density JJ exceeds a threshold value JcJ_c, the vortex core oscillates around the disk center with an oscillation amplitude s0s_0 and the frequency f=[κ/(2πG)](1+ξs02)f=[\kappa/(2\pi G)](1+\xi s_0^2). Note that the oscillation frequency is proportional to the current density JJ through the term s02=a/bs_0^2=a/b (a∝Ja\propto J), which has been confirmed by both experiments and simulations[35, 36]. Even in the presence of the magnetic field, the oscillation frequency remains ff, if the input strength is weak.

The solution of ss obtained from the exact Thiele equation, Eq. (5), shows a small oscillation around s0s_0[46]. This means that the trajectory of a limit-cycle oscillation is approximately circular but also has a small amplitude modulation. This small modulation is caused by the term caJR0pxexca_JR_0p_x\mathbf{e}_x in Eq. (5), which breaks the axial symmetry of the dynamics around the zz-axis. The deviation of ss from s0s_0 is, however, negligible, and the oscillation trajectory is approximately circular, as shown in Figure 3(b). Therefore, it is reasonable to omit the term from Eqs. (6) and (7). Note that this term arises from the in-plane component pxp_x of the magnetization in the reference layer. pxp_x plays a role in experiments for the following reason. Recall that the output signal measured in experiments depends on the relative angle of the magnetizations in the free and reference layers. Since the vortex core is located in the xyxy plane, a finite pxp_x is necessary to detect its position. On the other hand, the zz component pzp_z is also necessary because the spin-transfer torque originating from it excites the limit-cycle oscillation of the vortex core. In fact, the threshold current density Jc=∣D∣κ/(GaJpz)J_c = |D|\kappa/(Ga_Jp_z) is inversely proportional to pzp_z; therefore, if pzp_z is zero, JcJ_c becomes infinite and the oscillation cannot be excited. In experiments[28, 40], the magnetization initially pointed in an in-plane direction, where pz=0p_z = 0. A finite pzp_z was induced by applying an external magnetic field in the zz direction.

According to Eqs. (6) and (7), one might consider that the magnetic field changes the value of ss from s0s_0 and modifies the oscillation frequency. Such a frequency shift is, however, negligibly small, which can be discussed accordingly. First, remind that the frequency of the magnetic field applied during the second step is the frequency of the vortex core without the magnetic field because it consists of the output during the first step. The fact that the phases in the second step are saturated to 0 or π\pi, as shown in Figure 3(c), indicates that the forced phase synchronization occurs, and the frequency of the vortex core in the second step is the same with that in the first step. Second, let us roughly estimate the frequency shift by the application of the magnetic field. The change of ss by the magnetic field will be maximized when the phase of the magnetic field HyH_y in Eq. (6) is the same with ψ\psi. In this case, the magnitude of the last term in Eq. (6), averaged over a precession period τ=1/f\tau= 1/f, is about [cμ∗/(2GR)]Hyτ∼(γ/2)Hyτ[c\mu^*/(2GR)]H_y\tau\sim(\gamma/2)H_y\tau. The period τ\tau is about 5 ns while HyH_y is on the order of 1 Oe; see next section. Accordingly, the shift Δs\Delta s of ss by the application of the magnetic field is less than 0.1 at maximum. As mentioned, the oscillation frequency is proportional to 1+ζs21+\zeta s^2. Using ζ=0.1\zeta= 0.1 and s0≃0.6s_0 \simeq0.6, estimated from Figure 3(b), the frequencies with and without Δs\Delta s, which are proportional to 1+ζs021+\zeta s_0^2 and 1+ζ(s0+Δs)21+\zeta(s_0+\Delta s)^2, respectively, differ only 1 % at maximum. Therefore, we consider that the frequency modulation by the application of the magnetic field is negligible.

One might be interested in the applicability of the Thiele equation. While the original Thiele equation assumes a translation symmetry in an infinite space, a finite-size effect of nanostructured may restrict the applicability of the equation. Therefore, the Thiele equation had been applied to analyses on small-amplitude dynamics[47]. There have been, at the same time, several efforts to make the equation applicable to large-amplitude dynamics. For example, adding nonlinear frequency and damping terms is one approach[35, 36], which is also used in the present work, where the additional terms are characterized by the dimensionless parameters ξ\xi and ζ\zeta. Adding further higher-order nonlinear terms is also investigated recently[48, 49, 50]. It was also shown that the Thiele equation is applicable to analyze small-amplitude dynamics, and effort has been made to extrapolating it to a large-amplitude dynamics, such as vortex-core expulsion, although there are some limitations[51]. In the present study, we use the model developed in Refs.[35, 36] due to the following reasons. First, the applicability of the model to wide ranges of parameters has been verified by comparison with experiments[35, 36, 44]. Second, adding higher-order nonlinear terms does not change main conclusion in this work. These terms might change, for example, current dependence of the oscillation frequency. In the present work, however, the frequency is kept constant, and thus, adding such terms do not play a central role in the associative memory operation. Third, the Thiele equation with the present approximation clarifies the connection between spintronics and other research fields such as nonlinear science and computer science. This is because the equation can be reduced to the Stuart-Landau equation, as mentioned above. The Stuart-Landau equation has a long history, as in the case of the Thiele equation, and has been frequently used in nonlinear science[37, 52]. The present work indicates that the Stuart-Landau oscillator can be emulated in nanostructures and therefore, prompts communications between spintronics and other research fields. Therefore, although we understand that there have been great efforts[48, 49, 50, 53] for the validity and applicability of the Thiele equation, we use the model developed in Refs.[35, 36]. Note that the Oersted field generated in the current, discussed in these previous works, does not play a role in the associative memory operation because the current magnitude is kept constant during the operation. Also, since the external magnetic field induces forced synchronization, a frequency shift due to an external magnetic field studied in the previous work[48] does not exist in the present algorithm.

Details of the numerical simulations

The associative memory operation in the virtual oscillator network consists of three steps. The initial state of the vortex core in each step is prepared by adding a thermal activation to the Thiele equation and solving it in the absence of magnetic field, as is done in Ref.[45]. The torque due to the thermal activation gives an additional term, −ηxex−ηyey-\eta_x\mathbf{e}_x-\eta_y\mathbf{e}_y, to the left-hand side of Eq. (5), which obeys the fluctuation-dissipation theorem,

⟨ηi(t)ηj(t′)⟩=2kBT∣D∣δijδ(t−t′),(21)\langle\eta_i(t)\eta_j(t')\rangle= 2k_{\mathrm{B}}T|D|\delta_{ij}\delta(t-t'), \tag*{(21)}

where the temperature TT is 300 K. The solution of the Thiele equation in each step is divided into N=60N = 60 parts, where the time width of each part is denoted as τ~\tilde{\tau}. In the experiment[23], a certain time period was inserted between these parts to remove their correlation. In contrast, our numerical simulations used parallel computations, wherein the initial state of each part was randomly prepared using the method described above. The value of t~\tilde{t} was changed depending on the number of memorized patterns, as well as the number of noisy pixels in the pattern to be recognized. For example, t~\tilde{t} is 750 ns in Figure 3(c). For all cases, t~\tilde{t} was divided into n~=t~/tp\tilde{n}=\tilde{t}/t_{\mathrm{p}} parts, where tp=0.125t_{\mathrm{p}}=0.125 ns.

Now let us explain the meanings of yℓ(1)y_{\ell}^{(1)} and yℓ(2)y_{\ell}^{(2)} in Eqs. (3) and (4). Since they are defined in a similar manner, we will describe only yℓ(1)y_{\ell}^{(1)}. When defining the magnetic field in Eq. (3), it is convenient to reset the time origin for each part; i.e., each of the NN parts runs from t=0t=0 to t=t~t=\tilde{t}. Remember that the output from the STO is proportional to the yy component of the vortex core’s position, YY. We denote the solution of the normalized yy component, y=Y/Ry=Y/R (0≤y≤10\leq y\leq1), during the ℓ\ellth part in the first step as yℓy_{\ell}. Then, yℓ(1)y_{\ell}^{(1)} is made from yℓy_{\ell} as follows,

yℓ(1)=∑n=0n~−1yℓ(ntp){Θ(t−ntp)−Θ[t−(n+1)tp]},(22)y_{\ell}^{(1)}=\sum_{n=0}^{\tilde{n}-1}y_{\ell}(nt_{\mathrm{p}})\left\{\Theta(t-nt_{\mathrm{p}})-\Theta\left[t-(n+1)t_{\mathrm{p}}\right]\right\}, \tag*{(22)}

where Θ(t)\Theta(t) is a step function. Note that Θ(t−ntp)−Θ[t−(n+1)tp]\Theta(t-nt_{\mathrm{p}})-\Theta\left[t-(n+1)t_{\mathrm{p}}\right] is 1 for ntp≤t<(n+1)tpnt_{\mathrm{p}}\leq t<(n+1)t_{\mathrm{p}} and is zero for the other times; thus, it has a pulse shape. Equation (22) means that the input strength is constant for ntp≤t<(n+1)tpnt_{\mathrm{p}}\leq t<(n+1)t_{\mathrm{p}} and is proportional to yℓ(t)y_{\ell}(t) at t=ntpt=nt_{\mathrm{p}}. n~\tilde{n} is the number of input pulses. There are two reasons to shape the output into a pulse. The first one relates to numerical simulations. In this work, the Thiele equation was solved within a time increment of Δt=0.005\Delta t=0.005 ns, which is shorter than the pulse width tpt_{\mathrm{p}}. It was, however, impractical to store the output at each Δt\Delta t step because the amount would have been huge. Second, there is a technical limitation in real experiments on a measurable time step. The value we used, tp=0.125t_{\mathrm{p}}=0.125 ns, is close to shortest possible time step in an experiment[23]. Because of these reasons, we define yℓ(1)y_{\ell}^{(1)} used in the magnetic field, Eq. (3), as a pulse input. At the same time, we emphasize that tpt_{\mathrm{p}} is much shorter than an oscillation period of the vortex core, 1/f=4.481/f=4.48 ns (f=223f=223 MHz). In addition, the pulse-shaped yℓ(1)y_{\ell}^{(1)}s are continuously injected. Therefore, the magnetic field can be approximately regarded as a continuously oscillating signal with respect to the STO.

The strength of the input H\mathcal{H} in the second step is 1.0 Oe, while that in the third step is H′=Nm×0.2\mathcal{H}'=N_{\mathrm{m}}\times0.2 Oe. Here, we increase H′\mathcal{H}' as the number NmN_{\mathrm{m}} of memorized patterns increases. This is because the time necessary to reach a steady state becomes long as NmN_{\mathrm{m}} increases; therefore, to perform the numerical simulations efficiently, the input strength should be made to increase with NmN_{\mathrm{m}}.

Acknowledgements

The authors are grateful to Sumito Tsunegi and Yuki Hibino for discussion. Some of the results were obtained from a project “Innovative AI Chips and Next-Generation Computing Technology Development/(2) Development of Next-generation Computing Technologies/Exploration of Neuromorphic Dynamics towards Future Symbiotic Society” commissioned by NEDO. T. T. was supported by a JPS KAKENHI Grant, Number 20H05655.

Author contributions statement

T.T. designed the project. Y.I. performed the numerical simulations based on a program code developed by T.T. Y.I. also developed a model analysis with help of T.T. Y.I. and T.T. wrote the manuscript and prepared the figures.

Competing interests

The authors declare no competing interests.

Data availability

The datasets used and/or analyses during the current study available from the corresponding author on reasonable request.

Additional information

Supplementary information is available for this paper. Correspondence and requests for materials should be addressed to T.T.

Degree of overlap between the numerical patterns

Table S1 lists the overlap degrees between the ten memorized patterns, “0”, “1”, ⋯\cdots, “9”, used in the main text. Remind that we discuss the similarly between the patterns “1” and “7” in the main text. As can be seen in the table, the pattern having the largest overlap with respect to “1” is certainly “7”.

There are other pairs of numbers having large degrees of overlap. For example, the overlap between “3” and “9” is 0.54, which is larger than that between “1” and “7”. Moreover “8” has large overlaps with several other patterns; the overlap with “3”, “5”, “6”, and “9” are 0.73, 0.50, 0.73, and 0.60, respectively. Recall that we concluded in the main text that the associative memory operation fails when the number of the memorized patterns is large and some of the memorized patterns have large overlaps. We also showed that the association becomes accurate when the number of the memorized patterns is reduced or similar patterns are removed from the memorized ones. In the next section, we show that these conclusions are valid even for the memorized pattern “8”.

Associative memory operation for the pattern of “8”

Here, we made a pattern to be recognized by adding noise to the pattern “8” as shown in Figure S1, and then examined the associative memory operation with it. The amplitude of the magnetic field during the association was H′=Nm×0.17H' = N_{\mathrm{m}} \times0.17 Oe. First, we used ten memorized patterns, “0”, “1”, ⋯\cdots, “9”. The finally obtained pattern is shown in Figure S1; it is slightly different from “8” and does not match any of the memorized patterns exactly. This result is similar to the one shown in the

Image not extracted

Figure S1. (a) The pattern to be recognized was obtained by adding noise to the pattern “8”. (b) When the memorized patterns included ten patterns, the association of “8” failed. (c) When the memorized patterns included only “3” (or “6”) and “8”, the association succeeded, even though these patterns have a large overlap.

overlap0123456789
01.000.370.070.270.130.170.270.300.200.27
10.371.000.030.100.030.130.100.470.030.10
20.070.031.000.130.200.300.270.170.070.13
30.270.100.131.000.130.430.470.300.730.53
40.130.030.200.131.000.230.200.100.330.07
50.170.130.300.430.231.000.700.070.500.43
60.270.100.270.470.200.701.000.100.730.60
70.300.470.170.300.100.070.101.000.030.10
80.200.030.070.730.330.500.730.031.000.60
90.270.100.130.530.070.430.600.100.601.00

Table S1. Table of overlaps between ten memorized patterns, “0”, “1”, ⋯\cdots, “9”.

main text, where the pattern to be recognized was similar to “1” but had noise in it, and the finally obtained pattern did not correspond to any of the memorized patterns when the number of the memorized patterns was ten. Next, we examined the association by removing some of the memorized patterns. For example, when we used only “3” and “8” as the memorized patterns, the finally obtained pattern eventually becomes “8”, as shown in Figure S1; thus, the association succeeded. Moreover, the association succeeded even when we used “6” instead of “3”. These results support the conclusions in the main text.

Dependence of associative memory operation on noise

Here, we investigate the dependence of the accuracy of the associative memory operation on the number of noisy pixels in the pattern. Before showing the results, however, we should define what we mean by noise (or noisy pixel) and accuracy.

Noise is defined as swapping black and white colors in a memorized pattern to make a new pattern to be recognized. For example, the pattern to be recognized in Figure S1 is obtained by swapping the colors of 8 pixels from the pattern in Figure S1; therefore, there are 8 noisy pixels in this case. If the virtual oscillator network can associate the pattern in Figure S1 with that Figure S1, we say it is an accurate association. If the finally obtained pattern still includes some noise, as in the case of Figure S1, it is classified as an inaccurate association.

Here, one might imagine other definitions of accuracy. As mentioned above, we prepared the patterns to be recognized by adding a noise to one of the memorized patterns. Let us call the pattern to be recognized RR and the original memorized pattern AA. If the amount of noisy pixel is large, RR might become more similar to one of the other memorized patterns, call it pattern BB. Mathematically, this means that the overlap between RR and BB is larger than that between RR and AA, even though the pattern RR is generated from the pattern AA. In this case, the finally associated pattern might be BB. Similarly, even when the overlap between RR and AA is the larger than those between RR and any of the other memorized patterns, the finally obtained pattern might become one of the other memorized patterns. We regarded such associations to be inaccurate, even though the finally obtained pattern is amongst the memorized ones. This is because noise reduction (or pattern recovery) remains a challenging aspect of the associative memory operation, so we aimed to associate RR with the pattern it is derived from, i.e., AA.

With these definitions of noise (noisy pixel) and accuracy in hand, let us evaluate the accuracy as a function of the number of noisy pixels. Here, let us consider associating “1” from three memorized patterns, “0”, “1”, and “2”, as in the case studied in the main text. As mentioned in the Methods, the maximum number of noisy pixels is N/2N/2, where NN is the number of the pixels, 60 in the present study.

Figure S2 shows an example of the association in which the pattern to be recognized was generated by adding 30 noisy pixels to pattern “1”. In this case, the finally associated pattern is not “1”. Furthermore, some of the pixels are neither black nor white. This means that the phase differences of some parts did not saturate to either π\pi or 0 [see Figure S2]. Such a gray pattern is obtained when the number of noisy pixels is relatively large; in contrast, when the number of noisy pixels is relatively small, the finally obtained pattern is black-and-white even if the association fails, as shown in Figure S1.

Image not extracted

Figure S2. (a) Example of an inaccurate association. The top left and right patterns are the pattern to be recognized and the final pattern obtained after the association. In this case, the final pattern does not match any of the memorized ones (“0”, “1”, and “2”). In particular, some of its pixels are colored neither black nor white. This means that some phase differences did not saturate to π\pi or 0, as shown in the bottom figure. (b) Other examples of inaccurate associations. In these cases, the finally associated patterns are “0” and “2”, even though the patterns to be recognized were generated by adding noise to “1”. (c) Dependence of accuracy of the associative memory operation on the number of noisy pixels.

Figure S2 shows other examples for which the patterns to be recognized were obtained by adding 30 noisy pixels to “1”. The finally associated patterns are other memorized patterns, “0” and “2”. Here, the overlaps between the patterns to be recognized and “0” and “2” are relatively small; the overlaps between the patterns to be recognized and the finally obtained patterns were 0.1. Even for such a small overlap, the association became inaccurate because the overlaps between the patterns to be recognized and “1” are zero. Moreover, the finally obtained pattern has the largest overlap with the pattern to be recognized. For example, in the top example in Figure S2, the overlaps between the pattern to be recognized and the memorized patterns are 0.1 for “0”, 0 for “1”, and 0.03 for “2”. Similarly, in the bottom example in Figure S2, the overlaps are 0.03 for “0”, 0 for “1”, and 0.1 for “2”. Therefore, the finally associated pattern is the one with the largest overlap. According to the definition of accuracy above, however, these examples should be regarded as inaccurate associations.

Figure S2 summarizes the accuracy as a function of the number of noisy pixels. Here, we quantified the accuracy as follows. We prepared ten (Ne=10N_{\mathrm{e}} = 10) patterns to be recognized for a certain number of noisy pixels, where noisy pixel was added randomly to the ten patterns. For example, the patterns shown in Figure S2 and Figure S2 are examples of such patterns with 30 noisy pixels. The accuracy in Figure S2 is the number of the correctly associated patterns divided by Ne=10N_{\mathrm{e}} = 10. Therefore, the accuracy is 1 when the association succeeds for Ne=10N_{\mathrm{e}} = 10 prepared patterns to be recognized, while the accuracy is zero when the association fails for all prepared patterns. We used “0”, “1”, and “2” as the memorized patterns and generated the patterns to be recognized by adding noisy pixels to pattern “1”. The accuracy was 1 for relatively small number of noisy pixels. It deviated from 1 when the number of noisy pixels was close to half of the maximum value, i.e., the association was inaccurate when the number of noisy pixels was approximately larger than N/4N/4. The accuracy was zero when the number of noisy pixels was close to the maximum number, N/2N/2, because the overlap between the pattern to be recognized and the original pattern becomes approximately zero.

References

  1. [1]Kohonen, T. Self-Organization and Associative Memory, vol. 8 (Springer Science & Business Media, 2012).
  2. [2]Nakano, K. Associatron-A Model of Associative Memory. IEEE Trans. Syst. Man Cybern. 2, 380–388 (1972).DOI
  3. [3]Kohonen, T. Correlation Matrix Memories. IEEE Trans. Comput. 100, 353–359 (1972).
  4. [4]Anderson, J. A. A simple neural network generating an interactive memory. Math. Biosci. 14, 197–220 (1972).DOI
  5. [5]Hopfield, J. J. Neural network and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. United States Am. 79, 2554–2558 (1982).
  6. [6]Amari, S.-I. Neural theory of association and concept-formation. Biol. Cybern. 26, 175–185 (1977).pubmed.ncbi.nlm.nih.gov/901864
  7. [7]van Hemmen, J. L. & Kühn, R. Nonlinear Neural Networks. Phys. Rev. Lett. 57, 913 (1986).
  8. [8]McEliece, R. J., Posner, E., Rodemich, E. & Venkatesh, S. The Capacity of the Hopfield Associative Memory. IEEE Trans. Inf. Theory 33, 461–482 (1987).
  9. [9]Waugh, F. R., Marcus, C. M. & Westervelt, R. M. Fixed-Point Attractors in Analog Neural Computation. Phys. Rev. Lett. 64, 1986 (1990).pubmed.ncbi.nlm.nih.gov/10041545
  10. [10]Morita, M. Associative memory with nonmonotone dynamics. Neural Netw. 6, 115–126 (1993).DOI
  11. [11]Yoshizawa, S., Morita, M. & Amari, S. Capacity of associative memory using a nonmonotonic neuron model. Neural Netw. 6, 167–176 (1993).
  12. [12]Bollé, D. & Huyghebaert, J. Mixture states and storage of biased patterns in potts-glass neural networks. Phys. Rev. E 48, 2250 (1993).pubmed.ncbi.nlm.nih.gov/9960845
  13. [13]Krebs, P. R. & Theumann, W. K. Categorization in the symmetrically dilute hopfield network. Phys. Rev. E 60, 4580 (1999).pubmed.ncbi.nlm.nih.gov/11970317
  14. [14]McGraw, P. N. & Menzinger, M. Bistable gradient networks. i. attractors and pattern retrieval at low loading in the thermodynamic limit. Phys. Rev. E 67, 016118 (2003).pubmed.ncbi.nlm.nih.gov/12636575
  15. [15]Zhao, H. Designing asymmetric neural networks with associative memory. Phys. Rev. E 70, 066137 (2004).pubmed.ncbi.nlm.nih.gov/15697464
  16. [16]Borders, W. A. et al. Analogue spin–orbit torque device for artificial-neural-network-based associative memory operation. Appl. Phys. Express 10, 013007 (2017).DOI
  17. [17]Hoppensteadt, F. C. & Izhikevich, E. M. Associative memory of weakly connected oscillators. In Proceedings of International Conference on Neural Networks (ICNN’97), vol. 2, 1135–1138 (IEEE, 1997).
  18. [18]Hoppensteadt, F. C. & Izhikevich, E. M. Oscillatory Neurocomputers with Dynamic Connectivity. Phys. Rev. Lett. 82, 2983 (1999).DOI
  19. [19]Fernando, C., Michele, B. & Marco, G. Weakly Connected Oscillatory Network Models for Associative and Dynamic Memories. Int. J. Bifurc. Chaos 17, 4365–4379 (2007).DOI
  20. [20]Mirchev, M., Basnarkov, L., Corinto, F. & Kocarev, L. Cooperative Phenomena in Networks of Oscillators with Non-Identical Interactions and Dynamics. IEEE Transactions on Circuits Syst. I: Regul. Pap. 61, 811–819 (2013).DOI
  21. [21]Maffezzoni, P., Bahr, B., Zhang, Z. & Daniel, L. Oscillator Array Models for Associative Memory and Pattern Recognition. IEEE Transactions on Circuits Syst. I: Regul. Pap. 62, 1591–1598 (2015).DOI
  22. [22]Prasad, N., Mukim, P., Madhavan, A. & Stiles, M. D. Associative memories using complex-valued hopfield networks based on spin-torque oscillator arrays. Neuromorph. Comput. Eng. 2, 034003 (2022).arxiv.org/abs/2112.03358
  23. [23]Tsunegi, S., Taniguchi, T., Yakushiji, K., Yuasa, S. & Kubota, H. Associative memory operation using spin torque oscillator (The 83rd Japan Society of Applied Physics Autumn Meeting, 2022).
  24. [24]Locatelli, N., Cros, V. & Grollier, J. Spin-torque building blocks. Nat. Mater. 13, 11 (2014).DOI
  25. [25]Grollier, J., Querlioz, D. & Stiles, M. D. Spintronic Nanodevices for Bioinspired Computing. Proc. IEEE 104, 2024 (2016).pubmed.ncbi.nlm.nih.gov/27881881
  26. [26]Torrejon, J. et al. Neuromorphic computing with nanoscale spintronic oscillators. Nature 547, 428 (2017).
  27. [27]Grollier, J. et al. Neuromorphic spintronics. Nat. Electron. 3, 360 (2020).DOI
  28. [28]Kamimaki, A. et al. Chaos in spin-torque oscillator with feedback circuit. Phys. Rev. Res. 3, 043216 (2021).DOI
  29. [29]Thiele, A. A. Steady-State Motion of Magnetic Domains. Phys. Rev. Lett. 30, 230 (1973).DOI
  30. [30]Guslienko, K. Y., Han, X. F., Keavney, D. J., Divan, R. & Bader, S. D. Magnetic Vortex Core Dynamics in Cylindrical Ferromagnetic Dots. Phys. Rev. Lett. 96, 067205 (2006).pubmed.ncbi.nlm.nih.gov/16606041
  31. [31]Guslienko, K. Y. Low-frequency vortex dynamic susceptibility and relaxation in mesoscopic ferromagnetic dots. Appl. Phys. Lett. 89, 022510 (2006).DOI
  32. [32]Ivanov, B. A. & Zaspel, C. E. Excitation of Spin Dynamics by Spin-Polarized Current in Vortex State Magnetic Disks. Phys. Rev. Lett. 99, 247208 (2007).pubmed.ncbi.nlm.nih.gov/18233483
  33. [33]Khvalkovskiy, A. V., Grollier, J., Dussaux, A., Zvezdin, K. A. & Cros, V. Vortex oscillations induced by spin-polarized current in a magnetic nanopillar: Analytical versus micromagnetic calculations. Phys. Rev. B 80, 140401(R) (2009).DOI
  34. [34]Guslienko, K. Y., Aranda, G. R. & Gonzalez, J. Spin torque and critical currents for magnetic vortex nano-oscillator in nanopillars. J. Phys. Conf. Ser. 292, 012006 (2011).arxiv.org/abs/0912.5521
  35. [35]Dussaux, A. et al. Field dependence of spin-transfer-induced vortex dynamics in the nonlinear regime. Phys. Rev. B 86, 014402 (2012).DOI
  36. [36]Grimaldi, E. et al. Response to noise of a vortex based spin transfer nano-oscillator. Phys. Rev. B 89, 104404 (2014).
  37. [37]Kuramoto, Y. Chemical Oscillations, Waves, and Turbulence (Dover, New York, 2003).
  38. [38]Romera, M. et al. Vowel recognition with four coupled spin-torque nano-oscillators. Nature 563, 230 (2018).DOI
  39. [39]Skowronski, W., Cieciński, J., Ziętek, S., Yakushiji, K. & Yuasa, S. Microwave magnetic field modulation of spin torque oscillator based on perpendicular magnetic tunnel junctions. Sci. Rep. 9, 19091 (2019).arxiv.org/abs/1909.09503
  40. [40]Tsunegi, S. et al. Physical reservoir computing based on spin torque oscillator with forced synchronization. Appl. Phys. Lett. 114, 164101 (2019).DOI
  41. [41]Slonczewski, J. C. Current-driven excitation of magnetic multilayers. J. Magn. Magn. Mater 159, L1 (1996).DOI
  42. [42]Berger, L. Emission of spin waves by a magnetic multilayer traversed by a current. Phys. Rev. B 54, 9353 (1996).pubmed.ncbi.nlm.nih.gov/9984672
  43. [43]Slonczewski, J. C. Currents, torques, and polarization factors in magnetic tunnel junctions. Phys. Rev. B 71, 024411 (2005).DOI
  44. [44]Tsunegi, S. et al. Control of the stochastic response of magnetization dynamics in spin-torque oscillator through radio-frequency magnetic fields. Sci. Rep. 11, 16285 (2021).DOI
  45. [45]Imai, Y., Tsunegi, S., Nakajim, K. & Taniguchi, T. Noise-induced synchronization of spin-torque oscillators. Phys. Rev. B 105, 224407 (2022).DOI
  46. [46]Yamaguchi, T. et al. Periodic structure of memory function in spintronics reservoir with feedback current. Phys. Rev. Res. 2, 023389 (2020).DOI
  47. [47]Goto, M. et al. Electric spectroscopy of vortex states and dynamics in magnetic disks. Phys. Rev. B 84, 064406 (2011).DOI
  48. [48]Araujo, F. A., Chopin, C. & de Werfigosse, S. Ampere-Oersted field splitting of the nonlinear spin-torque vortex oscillator dynamics. Sci. Rep. 12, 10605 (2022).
  49. [49]De Werfigosse, S., Chopin, C. & Araujo, F. A. arXiv:2206.13438.
  50. [50]Araujo, F. A., Chopin, C. & de Werfigosse, S. arXiv:2206.13596.
  51. [51]Jenkins, A. S. et al. Spin-torque resonant expulsion of the vortex core for an efficient radiofrequency detection scheme. Nat. Nanotechnol. 11, 360 (2016).DOI
  52. [52]Pikovsky, A., Rosenblum, M. & Kurths, J. Synchronization: A universal concept in nonlinear sciences (Cambridge University Press, Cambridge, 2003), first edn.
  53. [53]Guslienko, K., Sukhostavets, O. V. & Berkov, D. V. Nonlinear magnetic vortex dynamics in a circular nanodot excited by spin-polarized current. Nanoscale Res. Lett. 9, 386 (2014).DOI

Paper details

vlnc:W_jzju4hhc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents