阅读说明：本技术 储备池元件和神经形态元件
(Reservoir element and neuromorphic element
) 是由 佐佐木智生
柴田龙雄 于 2019-09-11 设计创作，主要内容包括：本发明的一个方式的储备池元件包括：包含非磁性的导电体的自旋传导层；相对于自旋传导层位于第1方向,且在从上述第1方向俯视时彼此隔着间隔地配置的多个铁磁性层；和与上述自旋传导层的上述铁磁性层电连接的多个连通配线。(A reserve tank element according to one embodiment of the present invention includes: a spin-conducting layer comprising a non-magnetic electrical conductor; a plurality of ferromagnetic layers located in a 1 st direction with respect to the spin-conductive layer and arranged at intervals from each other in a plan view from the 1 st direction; and a plurality of communication wirings electrically connected to the ferromagnetic layer of the spin-conductive layer.)
1. A reservoir element, comprising:
a spin-conducting layer comprising a non-magnetic electrical conductor;
a plurality of ferromagnetic layers located in a 1 st direction with respect to the spin-conductive layer and arranged at intervals from each other in a plan view from the 1 st direction; and
a plurality of communication wirings electrically connected to the ferromagnetic layer of the spin-conductive layer.
2. The reserve pool element of claim 1, wherein:
each of the plurality of ferromagnetic layers is located at a position overlapping each of the plurality of communication wirings when viewed from the 1 st direction in plan view.
3. A reservoir element according to claim 1 or 2, characterized in that:
a reference potential terminal electrically connected to the spin conduction layer is included.
4. A reservoir element according to any of claims 1 to 3, characterized in that:
the communication wiring includes a ferromagnetic body,
the orientation direction of magnetization of the ferromagnetic material constituting the communication wiring is opposite to the orientation direction of magnetization of the ferromagnetic layer.
5. A reservoir element according to any of claims 1 to 4, characterized in that:
there is also a 1 st tunnel barrier layer between the spin-conducting layer and the plurality of ferromagnetic layers.
6. A reservoir element according to any of claims 1 to 5, characterized in that:
there is also a 2 nd tunnel barrier layer between the spin-conductive layer and the communication wiring.
7. A reservoir element according to any of claims 1 to 6, characterized in that:
the distance between two adjacent ferromagnetic layers among the plurality of ferromagnetic layers is equal to or less than the spin transport length of a material constituting the spin-conductive layer.
8. A reservoir element according to any one of claims 1 to 7, wherein:
the distance between two adjacent ferromagnetic layers among the plurality of ferromagnetic layers is equal to or less than the spin diffusion length of the material constituting the spin-conductive layer.
9. A reservoir element according to any of claims 1 to 8, characterized in that:
the spin conductive layer comprises a metal or alloy of any element selected from Cu, Ag, Al, Mg, Zn.
10. A reservoir element according to any of claims 1 to 8, characterized in that:
the spin-conducting layer contains a simple substance or a compound of any element selected from Si, Ge, and C.
11. A reservoir element according to any one of claims 1 to 10, wherein:
the plurality of ferromagnetic layers are arranged in a hexagonal lattice shape when viewed from the 1 st direction.
12. A reservoir element according to any one of claims 1 to 10, wherein:
the plurality of ferromagnetic layers are formed with an aggregate of a plurality of ferromagnetic layers which are densely arranged when viewed from the 1 st direction,
in the aggregate, the ferromagnetic layers are arranged in a hexagonal lattice.
13. A neuromorphic element, comprising:
a reservoir element according to any one of claims 1 to 12;
an input connected to the reservoir element; and
and an output unit connected to the reserve tank element and learning a signal from the reserve tank element.
Technical Field
The present invention relates to a reservoir (reservoir) element and a neuromorphic (neuromorphic) element.
Background
Neuromorphic elements are elements that mimic the brain of a human being through a neural network. Neuromorphic elements artificially mimic the relationship of neurons (neurons) and synapses (synapses) in the human brain.
The hierarchical element is one of the neuromorphic elements. The hierarchical element has chips (neurons in the brain) arranged in a hierarchy and a transfer unit (synapse in the brain) connecting between them. The hierarchical elements increase the correct answer rate of the question by learning the transmission unit (synapse). Learning refers to finding knowledge from information that can be used in the future and weighting the input data in the neuromorphic elements. The hierarchy type element learns in each hierarchy.
However, if the number of chips (neurons) increases in learning in each layer, a large load is imposed on circuit design, which causes an increase in power consumption of neuromorphic devices. Pool computers have been investigated as a way to alleviate this burden.
Reservoir computers are one type of neuromorphic element. The reserve tank computer includes a reserve tank element and an output. The reservoir element comprises a plurality of chips interacting with each other. The plurality of chips output signals by interacting with each other through the input signals. The transfer units connecting the plurality of chips are fixed in weight without learning. The output unit learns the signal from the reserve tank element and outputs the signal to the outside. The pool computer increases the correct answer rate of the question by compressing the data in the pool element and weighting the data in the output section. The learning in the pool computer is performed only by the output section. A reserve pool computer is expected as a method capable of simplifying the circuit design of neuromorphic elements and improving power consumption.
Non-patent document 1 describes a neuromorphic device using a Spin Torque Oscillation (STO) device as a chip (neuron).
Disclosure of Invention
Drawings
Fig. 1 is a conceptual diagram of a neuromorphic element of embodiment 1.
Fig. 2 is a perspective view of the reservoir member of embodiment 1.
Fig. 3 is a side view of the reservoir member of embodiment 1.
Fig. 4 is a plan view of the reservoir member of embodiment 1.
Fig. 5 is a plan view of another example of the reservoir member according to embodiment 1.
Fig. 6 is a plan view of another example of the reservoir member according to embodiment 1.
Fig. 7 is a plan view of another example of the reservoir member according to embodiment 1.
Fig. 8A is a sectional view showing a method of manufacturing a reservoir element according to embodiment 1.
Fig. 8B is a sectional view showing a method of manufacturing a reservoir element according to embodiment 1.
Fig. 8C is a sectional view showing a method of manufacturing a reservoir element according to embodiment 1.
Fig. 8D is a sectional view showing a method of manufacturing a reservoir element according to embodiment 1.
Fig. 9 is a diagram showing another example of the usage of the reserve tank element according to embodiment 1.
Fig. 10 is a sectional view of another example of the reservoir member according to embodiment 1.
Fig. 11 is a perspective view of the reservoir member of embodiment 2.
Fig. 12 is a side view of the reservoir member of embodiment 3.
Fig. 13 is a side view of the reservoir member of embodiment 4.
Fig. 14 is a schematic diagram for explaining an example of the operation of the neuromorphic device.
Fig. 15 is a schematic diagram for explaining another example of the operation of the neuromorphic element.
Fig. 16 is a schematic diagram for explaining another example of the neuromorphic device.
Fig. 17 is a schematic diagram for explaining another example of the neuromorphic device.
Fig. 18 is a plan view of the reservoir member of the 5 th embodiment.
Description of the reference numerals
1. 1', 1A, 1B, 1C, 1D, 1E, 1F, 1G ferromagnetic layer
1A input terminal
1B output terminal
2. 2' spin-conducting layer
3. 3m connected wiring
3C common electrode layer
3G reference potential terminal
4 st tunnel barrier layer
5 nd 2 nd tunnel barrier layer
10. Reservoir elements 10A, 10B, 10C, 10D, 10E, 11, 12, 13, 14
20 input part
21. 22, 23, 24 input terminal
30 output part
31. 32 output terminal
40 2 nd output part
41. 42, 43 terminal
100 neuromorphic element
Aggregate A
Cp chip
HM hard mask (hard mask)
I interlayer insulating film
Sb substrate
Sp synapse
Detailed Description
Hereinafter, the present embodiment will be described in detail with reference to the accompanying drawings as appropriate. The drawings used in the following description may be enlarged to show the characteristic portions for easier understanding of the characteristics, and the dimensional ratios of the respective components may be different from those in reality. The materials, dimensions, and the like exemplified in the following description are examples, and the present invention is not limited thereto, and can be modified as appropriate within a range in which the effects of the present invention are achieved.
"embodiment 1
Fig. 1 is a conceptual diagram of a neuromorphic element of embodiment 1. The neuromorphic element 100 has: an input 20, a reservoir element 10 and an output 30. The input 20 and the output 30 are connected to the reservoir element 10.
The neuromorphic device 100 compresses the signal input from the input unit 20 in the reservoir device 10, weights (learns) the signal compressed in the output unit 30, and outputs the signal to the outside.
The input section 20 transmits a signal input from the outside to the reservoir member 10. The input unit 20 includes, for example, a plurality of sensors. The plurality of sensors sense information outside the neuromorphic element 100, and input the information as a signal to the reservoir element 10. The signal may be input to the reservoir unit 10 continuously with time, or may be input to the reservoir unit 10 divided into predetermined time zones.
Reservoir element 10 has a plurality of chips Cp. The plurality of chips Cp interact with each other. The signal input to the reservoir unit 10 contains a large amount of information. The signal contains a large amount of information that is compressed into the desired information by the interaction of the chips Cp. The compressed signal is transmitted to the output section 30. The reserve pool element 10 does not learn. That is, the plurality of chips Cp interact with each other only, and signals transmitted between the plurality of chips Cp are not weighted.
Output 30 receives a signal from chip Cp of reservoir element 10. The output unit 30 performs learning. The output unit 30 weights each signal from each chip Cp by learning. The output unit 30 includes, for example, a nonvolatile memory. The nonvolatile memory is, for example, a magnetoresistive effect element. The output unit 30 outputs a signal to the outside of the neuromorphic device 100.
The neuromorphic element 100 increases the correct answer rate of the question by compressing the data in the reservoir element 10 and weighting the data in the output unit 30.
In addition, the neuromorphic device 100 is excellent in power consumption. In the neuromorphic device 100, learning is performed only by the output unit 30. Learning is to adjust the weight of the signal transferred from each chip Cp. The weight of the signal is determined according to the importance of the signal. If the weight of the signal is adjusted at any time, the circuit between the chips Cp is activated (active). The more circuits that are activated, the more power the neuromorphic element 100 consumes. The neuromorphic device 100 is excellent in power consumption as long as it is learned only by the output unit 30 at the final stage.
Fig. 2 is a perspective view of the reservoir member 10 of embodiment 1. Fig. 3 is a side view of the reservoir member 10 of embodiment 1. Fig. 4 is a plan view of the reservoir member 10 of embodiment 1.
The reserve tank element 10 includes: a plurality of ferromagnetic layers 1, a spin-conductive layer 2, and a plurality of communication wirings 3. The plurality of ferromagnetic layers 1 correspond to the chip Cp in fig. 1.
Here, the direction is defined. An arbitrary direction in the expanded plane of the spin conduction layer 2 is defined as an x-direction, a direction intersecting (e.g., substantially orthogonal to) the x-direction in the expanded plane of the spin conduction layer 2 is defined as a y-direction, and a direction intersecting (e.g., substantially orthogonal to) the expanded plane of the spin conduction layer 2 is defined as a z-direction. The z direction is an example of the 1 st direction.
The spin conduction layer 2 extends continuously in the xy plane. The spin conduction layer 2 is made of a nonmagnetic conductive material. The spin conduction layer 2 propagates spin current infiltrated from the ferromagnetic layer 1.
The spin conduction layer 2 is made of a material having a long spin diffusion length and a long spin transport length. The spin diffusion length is a distance until information of injected spins is halved by spin diffusion of the injected spins into the spin conduction layer 2. The spin transport length is the distance until the spin current of a spin-polarized current flowing in the nonmagnetic body is halved. When the applied voltage to the spin conduction layer 2 is small, the spin diffusion length and the spin transport length substantially coincide. On the other hand, when the applied voltage to the spin conduction layer 2 is large, the spin transport length is longer than the spin diffusion length due to the drift effect.
The spin-conductive layer 2 is, for example, a metal or a semiconductor. The metal used for the spin conductive layer 2 is, for example, a metal or an alloy containing any one element selected from Cu, Ag, Al, Mg, Zn. The semiconductor used for the spin-conductive layer 2 is, for example, a monomer or an alloy of an element selected from any of Si, Ge, and C. Examples thereof include Si, Ge, Si-Ge compounds, and graphene.
The ferromagnetic layer 1 is formed on one surface of the spin-conductive layer 2. The ferromagnetic layers 1 protrude in the z direction, and are present in plural at a distance from each other in the xy plane. The ferromagnetic layers 1 are present in plural relative to one spin-conductive layer 2. The adjacent ferromagnetic layers 1 are insulated by, for example, an interlayer insulating film.
The plurality of ferromagnetic layers 1 are arranged in a hexagonal grid pattern when viewed from the z direction, for example (see fig. 4). A signal input to the ferromagnetic layer 1 propagates as a spin current in the spin-conductive layer 2. When the ferromagnetic layers 1 are arranged in a hexagonal lattice, a signal input from a ferromagnetic layer 1 easily interferes with a signal input from another ferromagnetic layer 1.
The arrangement of the plurality of ferromagnetic layers 1 is not limited to the case of fig. 4. Fig. 5 to 7 are plan views of another example of the reserve tank element according to embodiment 1.
The plurality of ferromagnetic layers 1 of the reserve cell element 10A shown in fig. 5 are arranged in a square grid. The distances between adjacent ferromagnetic layers 1 are equal to homogenize the input signal.
The plurality of ferromagnetic layers 1 of the reservoir element 10B shown in fig. 6 are densely packed in a hexagonal grid. Since the density of the ferromagnetic layers 1 increases, signals input to different ferromagnetic layers 1 easily interfere with each other. Further, even in this case, the ferromagnetic layers 1 are insulated from each other.
The reservoir element 10C shown in fig. 7 forms an aggregate a in which a plurality of ferromagnetic layers 1 are densely packed. In the aggregate a, the ferromagnetic layers 1 are arranged in a hexagonal lattice shape. Adjacent ferromagnetic layers 1 are insulated from each other. In the reservoir element 10C, the mutual interference of signals input to the ferromagnetic layers 1 constituting one aggregate a and the mutual interference of signals input to the ferromagnetic layers 1 constituting a different aggregate a are different from each other in terms of the mutual interference conditions. In the reservoir element 10C, by adjusting the condition of mutual interference, the reservoir element 10C emphasizes a specific signal and transmits it to the output section 30.
Each ferromagnetic layer 1 has a cylindrical shape, for example (see fig. 1). The shape of the ferromagnetic layer 1 is not limited to the cylindrical shape. The ferromagnetic layer 1 may have, for example, an elliptic cylindrical shape, a quadrangular prism, a circular cone, an elliptic cone, a truncated cone, a quadrangular frustum, or the like.
The ferromagnetic layer 1 contains a ferromagnetic body. The ferromagnetic layer 1 contains, for example, a metal selected from Cr, Mn, Co, Fe, and Ni, an alloy containing one or more of these metals, an alloy containing these metals and B, C, and at least one or more elements of N. The ferromagnetic layer 1 is, for example, Co-Fe-B, Ni-Fe, Co-Ho alloy (CoHo)2) Sm-Fe alloy (SmFe)12). When the ferromagnetic layer 1 contains a Co-Ho alloy (CoHo)2) Sm-Fe alloy (SmFe)12) In this case, the ferromagnetic layer 1 is likely to be an in-plane magnetization film.
The distance between two adjacent ferromagnetic layers 1 is preferably equal to or less than the spin transport length, and preferably equal to or less than the spin diffusion length of the material constituting the spin-conductive layer 2.
The communication wiring 3 is electrically connected to the surface of the spin-conductive layer 2 opposite to the ferromagnetic layer 1. The communication wiring 3 may be connected directly to the spin conduction layer 2 or may be connected via another layer. The communication wiring 3 shown in fig. 1 and 3 protrudes from the spin-conductive layer 2 in the-z direction, and a plurality of the wiring are present apart from each other in the xy plane.
The communication wiring 3 includes a conductor. The interconnection line 3 is, for example, Cu, Al, or Au. The adjacent communication wiring 3 is insulated.
The communication wirings 3 shown in fig. 1 and 3 are disposed at positions corresponding to the plurality of ferromagnetic layers 1, respectively. That is, each ferromagnetic layer 1 and each communication line 3 overlap each other when viewed from the z direction in plan view.
Next, an example of a method for manufacturing the reservoir device 10 in the neuromorphic device 100 will be described. Fig. 8A to 8D are sectional views showing a method of manufacturing the reservoir element 10 according to embodiment 1.
First, a hole (ホール) is formed in the substrate Sb, and the hole is filled with an electric conductor (fig. 8A). The substrate Sb is, for example, a semiconductor substrate. The substrate Sb is preferably Si or AlTiC, for example. Si and AlTiC easily provide a surface having excellent flatness. For example, Reactive Ion Etching (RIE) is used to form the holes. The conductor charged in the hole serves as a communication wiring 3.
Next, the surfaces of the substrate Sb and the communication wiring 3 are planarized by Chemical Mechanical Polishing (CMP). The spin conductive layer 2 and the ferromagnetic layer 1' are stacked in this order on the planarized substrate Sb and the interconnection 3 (fig. 8B). The spin-conductive layer 2 and the ferromagnetic layer 1' are laminated using, for example, a Chemical Vapor Deposition (CVD) method.
Next, a hard mask HM is formed at a predetermined position on the surface of the ferromagnetic layer 1' (fig. 8C). The portion not covered by the hard mask HM of the ferromagnetic layer 1' is removed by RIE or ion milling. By removing unnecessary portions, the ferromagnetic layer 1' becomes a plurality of ferromagnetic layers 1. Finally, the ferromagnetic layers 1 are protected by an interlayer insulating film I (fig. 8D). The neuromorphic device 100 of embodiment 1 was obtained through the above-described procedure.
Next, the function of the neuromorphic device 100 will be described. Fig. 14 is a schematic diagram for explaining an example of the operation of the neuromorphic device 100. The input unit 20 has a plurality of input terminals 21, 22, 23, and 24 …. The plurality of input terminals 21, 22, 23, and 24 … constituting the input unit 20 are connected to one of the ferromagnetic layers 1 of the reservoir element 10. An input signal is input to the input unit 20 from the outside. The input signal is divided, for example, for each time domain, and is input to the input unit 20 as a plurality of signals S1, S2, S3, S4, S5, and S6 …. Although an example in which the input signal is not processed and input is shown, a signal subjected to Fast fourier transform Analysis (FFT Analysis) may be input. The FFT analysis is able to filter signals of small amplitude caused by noise. When the input terminals 21, 22, 23, and 24 receive an external signal, a current flows from the corresponding ferromagnetic layers 1A, 1B, 1C, and 1D to the communication wiring 3, and signals S1, S2, S3, S4, S5, and S6 are input to the reservoir element 10. The input signal is divided in time series in the order of the signal S1, the signal S2, the signal S3, and the signal S4, for example. A write current flows through the ferromagnetic layer 1A to which the signal S1 is input, the ferromagnetic layer 1B to which the signal S2 is input, the ferromagnetic layer 1C to which the signal S3 is input, and the ferromagnetic layer 1D to which the signal S4 is input in this order. When the communication lines 3 are arranged at positions corresponding to the plurality of ferromagnetic layers 1, respectively, a current flows in the z direction in many cases.
The current is spin-polarized by the ferromagnetic layers 1A, 1B, 1C, and 1D, and reaches the spin-conductive layer 2. The electric charges flow through the connection wiring 3 and hardly flow in the spin-conductive layer 2. A spin current flows in the spin conduction layer 2. That is, spins are injected from the ferromagnetic layer 1 into the spin-conductive layer 2 near the ferromagnetic layer 1, and spins are accumulated in the spin-conductive layer 2. The accumulated spins propagate as spin currents in the spin conduction layer 2. The spins injected from the ferromagnetic layer 1 diffuse around the ferromagnetic layer 1 in the spin-conductive layer 2.
The range of spin diffusion varies depending on the application time and the application amount of the current. This is because the amount of spin and the spin transport length accumulated in the vicinity of the ferromagnetic layer 1 change. If the amount of current applied is large, the spin transfer length becomes long, and spin current is transmitted from the vicinity of the ferromagnetic layers 1A, 1B, 1C, and 1D to a wide range.
When a current flows from the ferromagnetic layers 1A, 1B, 1C, and 1D to the interconnection 3, a spin current diffuses in the spin-conductive layer 2 from each position near the ferromagnetic layers 1A, 1B, 1C, and 1D to which the current is applied. The spin currents transferred from different positions interfere with each other, respectively. For example, the spin current transferred from the ferromagnetic layer 1A and the spin current transferred from the ferromagnetic layer 1B interfere with each other. The spin current transferred from the ferromagnetic layer 1A and the spin current transferred from the ferromagnetic layer 1B easily interfere with each other than the spin current transferred from the ferromagnetic layer 1A and the spin current transferred from the ferromagnetic layer 1C. This is because the distance between the ferromagnetic layer 1A and the ferromagnetic layer 1B is shorter than the distance between the ferromagnetic layer 1A and the ferromagnetic layer 1C. That is, as the distance between the ferromagnetic layers 1 is closer, the input signals S1, S2, S3, S4, S5, and S6 … are more likely to interfere with each other. The spin life is several hundred psec in the case of a metal such as Ta or Pt, and several nsec in the case of a semiconductor such as Si. Information of spins injected into the spin-conductive layer 2 cannot be read to the order of several hundreds of psec to several nsec. Therefore, the closer the time series of signals S1, S2, S3, S4, S5, and S6 … are, the more likely to interfere with each other. Therefore, it is preferable that the signals S1, S2, S3, S4, S5, and S6 … in the time series are input to the ferromagnetic layer 1 at a shorter distance. For example, the distance between the ferromagnetic layer 1A to which the signal S1 is input and the ferromagnetic layer 1B to which the signal S2 is input is preferably shorter than the distance between the ferromagnetic layer 1A to which the signal S1 is input and the ferromagnetic layer 1C to which the signal S3 is input.
Finally, the signal is output from the reservoir element 10 to the output section 30. The output unit 30 has, for example, a plurality of output terminals 31, 32, …. The plurality of output terminals 31 and 32 … are connected to any ferromagnetic layer 1. Fig. 14 shows an example in which the ferromagnetic layers 1E and 1F different from the ferromagnetic layers 1A, 1B, 1C, and 1D to which the input terminals 21, 22, 23, and 24 … for input are connected. The input and output can be switched, and the output terminals 31, 32 … for output may be connected to the ferromagnetic layers 1A, 1B, 1C, 1D to which the input terminals 21, 22, 23, 24 … for input are connected. The signal is read as a potential difference between the communication wiring 3 and the ferromagnetic layer 1. Although a current does not flow in the spin conduction layer 2, a spin current flows. When spin current is generated, the potential (potential) of the spin conductive layer 2 with respect to the ferromagnetic layer 1 changes, generating a potential difference. The potential difference is read as a potential difference between the reference potential and each ferromagnetic layer 1 with an arbitrary connecting line 3 as a reference potential.
The potential of the spin-conductive layer 2 in the vicinity of each ferromagnetic layer 1 is influenced by the spin current diffused from a different position. A signal read out from one ferromagnetic layer 1 as a potential difference contains information written to the other ferromagnetic layer 1, and the information is compressed.
The finally compressed signal is transmitted to the output section 30 via the plurality of output terminals 31, 32, …. The output unit 30 weights the signals read from the ferromagnetic layers 1 by learning. Fig. 15 is a schematic diagram for explaining another example of the operation of the neuromorphic device 100. Fig. 15 shows a method of dividing an input signal, and the connection positions of the input terminals 21, 22, 23, 24, 31, and 32 … are different from those of the example shown in fig. 14. In the example shown in fig. 15, the input signal is not overlapped with the signals S1, S2, S3, and S4 … and is time-sequentially divided. In the neuromorphic device shown in fig. 15, the ferromagnetic layers 1A, 1B, 1C, and 1D to which the signals S1, S2, S3, and S4 … are input are spaced apart from each other as the time series increases.
Fig. 16 is a schematic diagram for explaining another example of the neuromorphic device. The neuromorphic device shown in fig. 16 differs from the example shown in fig. 15 in that it has the 2 nd output unit 40. The terminals 41, 42, 43 of the 2 nd output unit 40 are connected to the output terminals 31, 32 of the output unit 30 via synapses Sp. When information is transmitted from the output terminals 31 and 32 of the output unit 30 to the terminals 41, 42, and 43 of the 2 nd output unit 40, data is weighted by the synapse Sp. The neuromorphic element shown in fig. 16 learns between the output unit 30 and the 2 nd output unit 40. That is, the neuromorphic element shown in fig. 16 is a structure of a deep neural network. By having the neuromorphic element as a deep neural network structure, more complex information can be identified.
Fig. 17 is a schematic diagram for explaining another example of the neuromorphic device. The neuromorphic element shown in fig. 17 differs from the example shown in fig. 15 in that reservoir elements 10 are arranged in parallel and an output unit 30 connected to each reservoir element 10 is shared. With such a configuration, it is possible to simultaneously recognize signals having different signal speeds and outputs from a plurality of sensors, and to realize a multi-mode (multimodal) reservoir element.
As described above, the reservoir element 10 of embodiment 1 is such that the spin currents transferred from the respective ferromagnetic layers 1 interfere with each other in the spin conduction layer 2. The signals input from the input unit 20 interfere with each other in the spin-conductive layer 2, and a specific state is generated in the spin-conductive layer 2. That is, a signal input from the input unit 20 is compressed into one state in the spin-conductive layer 2. Thus, the neuromorphic element 100 of embodiment 1 utilizes reservoir element 10 to properly compress the signal. By compressing the signal, only the output unit 30 needs to learn, and the power consumption of the neuromorphic device 100 is reduced.
The reservoir element 10 according to embodiment 1 can be variously modified.
Fig. 9 is a perspective view of another example of the reserve tank element according to embodiment 1. The reservoir element 10D shown in fig. 9 has a reference potential terminal 3G, and the ferromagnetic layer 1 is divided into an input terminal 1A and an output terminal 1B.
The reference potential terminal 3G is electrically connected to the spin conductive layer 2. The reference potential terminal 3G is preferably located sufficiently away from each of the output terminals 1B. The reference potential terminal 3G is made of the same material as the communication wiring 3.
The ferromagnetic layer 1 is divided into an input terminal 1A for inputting a signal and an output terminal 1B for outputting a signal. When a current flows from each input terminal 1A to the connection wiring 3, spin currents flow in the spin conduction layer 2 and interfere with each other. The output terminal 1B outputs, as a potential difference, a difference in potential between the spin in the spin-conductive layer 2 near the output terminal 1B and the magnetization of the output terminal 1B at a certain instant. The potentials V1, V2, and V3 of the respective output terminals 1B are measured with reference to the reference potential terminal 3G. The potentials V1, V2, and V3 become output signals. By fixing the reference potential by the reference potential terminal 3G, the potentials V1, V2, and V3 can be relatively evaluated.
The shortest distance between the input terminal 1A and the output terminal 1B is preferably equal to or less than the spin transport length, and preferably equal to or less than the spin diffusion length of the material constituting the spin-conductive layer 2. By sufficiently transmitting the spin current to the output terminal 1B, the SN (Signal/Noise) ratio of the output Signal is increased.
In the case of dividing the input terminal 1A and the output terminal 1B, the communication line 3 may not be provided at a position facing the output terminal 1B. As shown in fig. 10, the communication lines 3 facing the input terminal 1A or the output terminal 1B may be connected to each other through the common electrode layer 3C.
"embodiment 2
Fig. 11 is a sectional view of the reservoir member of embodiment 2. The reservoir element 11 of embodiment 2 differs from the reservoir element 10 of embodiment 1 in that the communication wiring 3m includes a magnetic material. The other structures are the same as those of the reservoir member 10 of embodiment 1, and description thereof is omitted. In fig. 11, the same components as those in fig. 1 are denoted by the same reference numerals.
The communication wiring 3m includes a magnetic body. The through wiring 3m may include a magnetic body at a position close to the spin conductive layer 2. The connection wiring 3m may have a structure in which a ferromagnetic layer and a conductive layer are stacked in this order from a position close to the spin-conductive layer 2, for example. The same material as that of the ferromagnetic layer 1 can be used for the magnetic body.
The orientation direction of the magnetization of the communication wiring 3m is opposite to the orientation direction of the magnetization of the ferromagnetic layer 1. When a current flows between the ferromagnetic layer 1 and the connecting wiring 3m, which have different magnetization orientation directions, spins in the same direction as the spin-conductive layer 2 can be efficiently injected.
The case where the magnetization of the ferromagnetic layer 1 is oriented in the + x direction and the magnetization of the connecting wiring 3m is oriented in the-x direction will be described as an example. For example, a current flows through the ferromagnetic layer 1, the spin-conductive layer 2, and the connecting wiring 3m in this order. When a current flows from the ferromagnetic layer 1 to the spin-conductive layer 2, spins in the-x direction are injected from the ferromagnetic layer 1 to the spin-conductive layer 2. On the other hand, when a current flows from the spin-conductive layer 2 to the communication wiring 3m, since the magnetization of the communication wiring 3m is oriented in the-x direction, a spin in the-x direction flows from the communication wiring 3m to the spin-conductive layer 2. Therefore, when the communication wiring 3m contains a ferromagnetic body, spins in the same direction as the spin-conductive layer 2 can be efficiently injected.
The reservoir member 11 of embodiment 2 can be applied to the neuromorphic member 100. Further, the reserve tank element 11 according to embodiment 2 achieves the same effects as those of the reserve tank element 10 according to embodiment 1. In addition, the reservoir element 11 of embodiment 2 efficiently supplies spins to the spin-conductive layer 2. Therefore, interference of spin flow in the spin conduction layer 2 is promoted, and the reservoir element 11 can exhibit a more complicated phenomenon.
"embodiment 3
Fig. 12 is a sectional view of the reservoir member of embodiment 3. The reservoir element 12 of embodiment 3 differs from the reservoir element 10 of embodiment 1 in that it has the 1 st tunnel barrier layer 4. The other structures are the same as those of the reservoir member 10 of embodiment 1, and description thereof is omitted. In fig. 10, the same components as those in fig. 1 are denoted by the same reference numerals.
The 1 st tunnel barrier layer 4 is located between the ferromagnetic layer 1 and the spin-conductive layer 2. The 1 st tunnel barrier layer 4 extends continuously in the xy plane, for example. The 1 st tunnel barrier layer 4 may be dispersed in the xy plane only at a position between the ferromagnetic layer 1 and the spin conduction layer 2.
The 1 st tunnel barrier layer 4 is made of a non-magnetic insulator. The 1 st tunnel barrier layer 4 is, for example, Al2O3、SiO2、MgO、MgAl2O4And the like. The 1 st tunnel barrier layer 4 may Be formed of a material in which a part of Al, Si, and Mg among the above materials is replaced with Zn, Be, or the like. MgO, MgAl2O4A coherent tunneling phenomenon can be realized between the ferromagnetic layer 1 and the spin-conductive layer 2, and spins can be efficiently injected from the ferromagnetic layer 1 into the spin-conductive layer 2.
The thickness of the 1 st tunnel barrier layer 4 is preferably below 3 nm. If the resistance of the thickness of the 1 st tunnel barrier layer 4 is high, the spin current from the spin conduction layer 2 can be suppressed from flowing back. However, when the thickness of the 1 st tunnel barrier layer 4 is 3nm or more, the spin scattering effect as the spin filtering of the 1 st tunnel barrier layer 4 is not increased but only the resistance is increased, and the noise is increased.
The 1 st tunnel barrier layer 4 has a larger spin resistance than the spin conduction layer 2. Spin resistance is a quantity that quantitatively represents the ease of flow of spin current (the difficulty of spin relaxation).
The spin resistance Rs is defined by the following equation (see non-patent document 1).
[ formula 1 ]
Where λ is the spin diffusion length of the material, ρ is the resistivity of the material, and a is the cross-sectional area of the material.
In the nonmagnetic material, when the cross-sectional areas a are equal, the magnitude of the spin resistance is determined by the value of ρ λ, which is the spin resistance ratio in equation (1).
The spin flows in a portion where the spin resistance is low. Since the 1 st tunnel barrier layer 4 includes an insulator, the resistivity is large and the spin resistance is large. The 1 st tunnel barrier layer 4 suppresses the spin return ferromagnetic layer 1 from reaching the spin conduction layer 2.
The reservoir member 12 of embodiment 3 can be applied to the neuromorphic member 100. Further, the reserve tank element 12 according to embodiment 3 achieves the same effects as those of the reserve tank element 10 according to embodiment 1. In addition, the reservoir element 12 of embodiment 3 can efficiently generate spin current from spins injected into the spin-conductive layer 2. Thus, interference of spin flow in the spin conduction layer 2 is promoted, and the reservoir element 12 can exhibit a more complicated phenomenon.
"embodiment 4
Fig. 13 is a sectional view of the reservoir member of embodiment 4. The reservoir element 13 of embodiment 4 is different from the reservoir element 12 of embodiment 3 in that it has the 2 nd tunnel barrier layer 5. The other structures are the same as those of the reservoir member 12 of embodiment 3, and description thereof is omitted. In fig. 13, the same components as those in fig. 12 are denoted by the same reference numerals.
The 2 nd tunnel barrier layer 5 is located between the spin conduction layer 2 and the communication wiring 3. The 2 nd tunnel barrier layer 5 extends continuously in the xy plane, for example. The 2 nd tunnel barrier layer 5 may be dispersed in the xy plane only at a position between the ferromagnetic layer 1 and the spin conduction layer 2.
The 2 nd tunnel barrier layer 5 is made of a non-magnetic insulator. The 2 nd tunnel barrier layer 5 is made of the same material as the 1 st tunnel barrier layer 4. The thickness of the 2 nd tunnel barrier layer 5 is equal to the thickness of the 1 st tunnel barrier layer 4.
The 2 nd tunnel barrier layer 5 has a larger spin resistance than the spin conduction layer 2. The 2 nd tunnel barrier layer 5 suppresses the flow of spins that have reached the spin conduction layer 2 to the communication wiring 3.
The reservoir member 13 of embodiment 4 can be applied to the neuromorphic member 100. Further, the reserve tank element 13 according to embodiment 4 achieves the same effects as those of the reserve tank element 10 according to embodiment 1. In addition, the reservoir element 13 of embodiment 4 can efficiently generate spin current from spins injected into the spin-conductive layer 2. Therefore, interference of spin flow in the spin conduction layer 2 is promoted, and the reservoir element 13 can exhibit a more complicated phenomenon.
Fig. 18 is a plan view of the reservoir member of the 5 th embodiment. The reservoir element 14 of embodiment 5 differs from the reservoir element 10 of embodiment 1 in that the spin-conductive layer 2 'is annular, and a plurality of ferromagnetic layers 1 are dispersed along the annular spin-conductive layer 2'. The other structures are the same as those of the reservoir member 10 of embodiment 1, and description thereof is omitted. In fig. 18, the same components as those in fig. 1 are denoted by the same reference numerals.
The input signal is divided, for example, for each time domain, and is input to the pool element 14 as a plurality of signals S1, S2, S3, and S4 …. For example, a signal S1 is input to the ferromagnetic layer 1A, a signal S2 is input to the ferromagnetic layer 1B, and a signal S3 is input to the ferromagnetic layer 1C. Signals S1, S2, S3, and S4 … are converted into spin currents, which are input from the ferromagnetic layers 1A, 1B, and 1C to the spin-conductive layer 2', respectively. The spin currents diffused from the respective ferromagnetic layers 1A, 1B, 1C propagate along the circumferential direction of the spin conduction layer 2' and interfere with each other.
The interference of spin currents in the spin conduction layer 2' is output as a potential difference from the ferromagnetic layers 1E, 1F, 1G, for example. In addition, the input position of the signal to the ferromagnetic layer and the output position of the signal from the ferromagnetic layer may be sequentially changed.
Although a preferred embodiment of the present invention has been described in detail, the present invention is not limited to the embodiment, and various modifications and changes can be made within the scope of the gist of the present invention described in the scope of the claims.
For example, the features of the reserve cell elements 10 to 14 according to embodiments 1 to 5 may be combined.
