Skymin-based simulated neuron device and simulated neural network

Abstract

This invention discloses a skyrmion-based simulated neuron device and a simulated neural network. The skyrmion-based simulated neuron device includes: a magnetic skyrmion track; a voltage control region is located at one end of the magnetic skyrmion track near a first current injection port; the voltage control region controls the magnetic anisotropy constant of its region through voltage control; skyrmions generated at the first current injection port stop moving after reaching the voltage control region, and only move towards a second current injection port after the accumulated skyrmions in the voltage control region exceed a preset threshold; a skyrmion detection device is located at one end of a ferromagnetic layer near the second current injection port; the skyrmion detection device detects the state of the skyrmions and outputs a trigger signal when a change in the skyrmion state is detected. The technical solution of this invention provides hardware and physical principle support for large-scale deep neural networks.

Description

Technical Field

[0001] The present invention relates to electromagnetic technology, and more particularly to a skyrmion-based analog neuron device and analog neural network. Background Technology

[0002] Neuronal computing refers to the computational methods that mimic the human brain. It has broad applications in fields such as Artificial Intelligence (AI). Given the high computational demands of neurons in the Blue Group, deep neural networks cannot be limited to simply increasing network depth. Finding more effective propagation media and transmission mechanisms between neurons has become a new direction of exploration in the field of neuronal computing.

[0003] Currently, neural network simulation devices based on Complementary Metal Oxide Semiconductor (CMOS) dominate the field of neural network computing. However, when faced with neural computations with extremely high time complexity, CMOS-based neural network simulation devices are not only not small enough in size and consume a lot of power, but also have relatively low computing speeds, making it difficult to perform complex neural network computations well. Summary of the Invention

[0004] This invention provides a skyminzi-based simulated neuron device and simulated neural network, providing hardware and physical principle support for large-scale deep neural networks.

[0005] In a first aspect, embodiments of the present invention provide a simulated neuron device based on skyrmions, comprising:

[0006] A magnetic skyrmion track, wherein the skyrmion track comprises a substrate layer, a heavy metal layer, and a ferromagnetic layer arranged sequentially from bottom to top;

[0007] The skyrmion track is provided with a first current injection port and a second current injection port at both ends. The first current injection port and the second current injection port inject current in a direction perpendicular to the skyrmion track to generate skyrmions. The skyrmions flow from the first current injection port to the second current injection port.

[0008] A voltage control region is provided on one end of the ferromagnetic layer near the first current injection port. The voltage control region controls the magnetic anisotropy constant of the region by voltage control. Skyrmions generated by the first current injection port stop moving after moving to the voltage control region. They continue to move to the second current injection port until the accumulated skyrmions in the voltage control region exceed a preset threshold.

[0009] A skyrmion detection device is provided on one end of the ferromagnetic layer near the second current injection port. The skyrmion detection device is used to detect the state of skyrmions and outputs a trigger signal when a change in the state of skyrmions is detected.

[0010] In one possible implementation of the first aspect, the voltage of the voltage control region is adjustable, and the preset threshold is adjusted by adjusting the voltage of the voltage control region.

[0011] In one possible implementation of the first aspect, the skyrmion detection device includes a detection MTJ structure and a reference MTJ structure arranged from bottom to top. The detection MTJ structure is used to detect the state of skyrmions in the skyrmion track. When the detection MTJ structure detects an increase in the number of skyrmions in the skyrmion track, its resistivity changes. When the resistivity of the detection MTJ structure changes, the trigger current signal is generated between the reference MTJ structure and the detection MTJ structure.

[0012] In one possible implementation of the first aspect, the detection MTJ structure and the reference MTJ structure are sandwich structures.

[0013] In one possible implementation of the first aspect, the current injected at the second current injection port is used to drive the skyrmion arriving at the second current injection port back to its initial state.

[0014] In one possible implementation of the first aspect, the first current injection port is used to simulate the presynapse of a neuron, the second current injection port is used to simulate the postsynapse of a neuron, and the preset threshold of the voltage control region is used to simulate the synaptic weights of a neuron.

[0015] In one possible implementation of the first aspect, the ferromagnetic layer is made of iron, cobalt, nickel or other magnetic metals or alloys, the heavy metal layer is made of platinum, and the substrate layer is made of silicon.

[0016] In one possible implementation of the first aspect, the skyrmion track has a length of 600 nanometers, a width of 100 nanometers, and a thickness of 1 nanometer.

[0017] In one possible implementation of the first aspect, the control voltage of the voltage control region is between -2 volts and 2 volts.

[0018] In a second aspect, embodiments of the present invention provide a skyminton-based simulated neural network, characterized in that it includes: a plurality of skyminton-based simulated neuron devices;

[0019] Each skyminton-based analog neuron device includes:

[0020] Skyrmion track, wherein the skyrmion track comprises a substrate layer, a heavy metal layer, and a ferromagnetic layer arranged sequentially from bottom to top;

[0021] The skyrmion track is provided with a first current injection port and a second current injection port at both ends. The first current injection port and the second current injection port inject current in a direction perpendicular to the skyrmion track to generate skyrmions. The skyrmions flow from the first current injection port to the second current injection port.

[0022] A voltage control region is provided on one end of the ferromagnetic layer near the first current injection port. The voltage control region controls the magnetic anisotropy constant of the region by voltage control. Skyrmions generated by the first current injection port stop moving after moving to the voltage control region. They continue to move to the second current injection port until the accumulated skyrmions in the voltage control region exceed a preset threshold.

[0023] A skyrmion detection device is provided on one end of the ferromagnetic layer near the second current injection port. The skyrmion detection device is used to detect the state of skyrmions and outputs a trigger signal when a change in the state of skyrmions is detected.

[0024] The plurality of skyrmion-based analog neuron devices are connected according to a preset network structure, wherein the trigger signal output by the previous skyrmion-based analog neuron device is used to control the voltage value of the voltage control region of the next skyrmion-based analog neuron device.

[0025] The skyrmion-based simulated neuron device and simulated neural network provided in this invention utilize the properties of skyrmions to demonstrate the plasticity of neurons, including their enhancement and inhibition behaviors. This demonstrates great potential for future use in complex, interconnected, and dense neural networks, significantly reducing transmission energy consumption between components and providing hardware and physical principle support for large-scale deep neural networks. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a skyminzi-based simulated neuron device provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0028] To better simulate the neuronal mechanisms of the human brain, skyrmions, as small particles, have become the preferred propagation medium in neuronal computation due to their unique properties. Skyrmions possess extremely high stability and very low energy consumption during transmission, making them highly suitable as a transmission medium between neurons to demonstrate the advantages of neural conduction. Skyrmions excel in reducing time complexity and also exhibit superior performance compared to traditional components in terms of memory usage and device size.

[0029] However, in existing skyrmion devices, thresholding devices have not been able to control the aggregation and transmission of skyrmions with very high precision and efficiency. Therefore, this invention employs voltage-controlled magnetic anisotropy, utilizing an external electric field to precisely and effectively control the number of skyrmions, thereby achieving effective pulses. This significantly increases the accuracy of skyrmion conduction while reducing the power and time complexity of the device.

[0030] Figure 1 This is a schematic diagram of a skyrmion-based simulated neuron device provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the skyminton-based simulated neuron device provided in this embodiment includes:

[0031] Skyrmion track 11, first current injection terminal 12, second current injection terminal 13, voltage control region 14, and skyrmion detection device 15.

[0032] The skyrmion track 11 comprises a substrate layer 21, a heavy metal layer 22, and a ferromagnetic layer 23 arranged sequentially from bottom to top. A first current injection port 12 and a second current injection port 13 are respectively provided at both ends of the skyrmion track 11. Current is injected into the first current injection port 12 and the second current injection port 13 in a direction perpendicular to the skyrmion track 11 to generate skyrmions. The skyrmions flow from the first current injection port 12 to the second current injection port 13. The substrate layer 21, the heavy metal layer 22, and the ferromagnetic layer 23 form a sandwich-structured track. The skyrmion track 11 can be a nanoscale skyrmion generation structure. Both the first current injection port 12 and the second current injection port 13 are disposed on a nanofilm.

[0033] A basic skyrmion generation structure consists of a substrate layer 21, a heavy metal layer 22, and a ferromagnetic layer 23. The substrate layer 21 is a non-metallic substrate material. The heavy metal layer 22 and the ferromagnetic layer 23 generate skyrmions after current injection, which then move within a skyrmion track 11 under the influence of a driving current. Current injection terminals, a first current injection terminal 12 and a second current injection terminal 13, are located at opposite ends of the skyrmion track 11 along its length. When current is injected into the skyrmion track 11 through the first current injection terminal 12, this current generates a magnetic field within the track, thus producing skyrmions. Driven by the current injected at the first current injection terminal 12, the skyrmions move towards the second current injection terminal 13. When a skyrmion reaches the second current injection terminal 13, the second current injection terminal 13 also injects current into the skyrmion track 11, causing the skyrmion to leave the area and return to its initial state.

[0034] In one embodiment, the ferromagnetic layer 23 is made of iron, cobalt, nickel or other magnetic metals or alloys, the heavy metal layer 22 is made of platinum, and the substrate layer 21 is made of silicon.

[0035] The above structure is the basic structure for generating skyrmions. However, in order for skyrmions to achieve the effect of simulating neurons, it is necessary to accurately control the generation of skyrmions. In this embodiment, the generation and transmission of skyrmions are controlled by the principle of pressure-controlled magnetic anisotropy.

[0036] A voltage control region 14 is provided on one end of the ferromagnetic layer 23 near the first current injection port 12. The voltage control region 14 controls the magnetic anisotropy constant of the region by voltage control. Skyrmions generated by the first current injection port 12 stop moving after moving to the voltage control region 14, until the accumulated skyrmions in the voltage control region 14 exceed a preset threshold, and then move to the second current injection port 13.

[0037] The voltage control region 14 is located on the side of the skyrmion track 11, which is adjacent to the first current injection terminal 12 that generates skyrmions. When the voltage control region 14 applies a voltage to the skyrmion track 11, this voltage controls the perpendicular magnetic anisotropy constant of the ferromagnetic layer 23. Controlling the perpendicular magnetic anisotropy constant can pin or unpin the skyrmions. In other words, by controlling the perpendicular magnetic anisotropy constant with voltage, the movement or stop of the skyrmions can be controlled. Changes in the voltage applied by the voltage control region 14 will cause changes in the perpendicular magnetic anisotropy constant of the ferromagnetic layer 23. Therefore, by setting a magnetic anisotropy region in the voltage control region 14, the voltage applied by the voltage control region 14 becomes the threshold affecting whether the skyrmions move. When current is injected into the first current injection terminal 12, the generated skyrmions move towards the second current injection terminal 13. When they reach the voltage control region 14, due to the influence of the voltage applied by the voltage control region 14, the skyrmions will be pinned and unable to move. When the number of skyrmions accumulated in the voltage control region 14 reaches a preset threshold, the skyrmions will break through the magnetic anisotropy region controlled by the voltage control region 14 and continue to move towards the second current injection terminal 13. By setting the voltage control region 14, the movement of skyrmions can be controlled.

[0038] The voltage of the voltage control region 14 is adjustable, and the preset threshold is adjusted by adjusting the voltage of the voltage control region 14. The skyrmion-based simulated neuron device provided in this embodiment simulates a single neuron. However, to achieve complex neural network calculations, the simulated neural network needs numerous neurons. Therefore, the skyrmion-based simulated neuron device provided in this embodiment can be used as the basic neuron structure. Multiple skyrmion-based simulated neuron devices can be deployed and connected together according to a preset method to achieve the structure of the entire neural network. The trigger signal output by the previous skyrmion-based simulated neuron device is used to control the voltage of the voltage control region 14 of the next skyrmion-based simulated neuron device, thus simulating the synaptic weights of biological neurons and realizing the plasticity of neuronal synapses.

[0039] Finally, a skyrmion detection device 15 is disposed on the ferromagnetic layer 23 near the second current injection port 13. The skyrmion detection device 15 is used to detect the state of skyrmions and outputs a trigger signal when a change in the skyrmion state is detected. The skyrmion detection device 15 can be any existing structure capable of detecting changes in skyrmions, as long as it can detect the change and output a trigger signal based on that change. In this embodiment, a magnetic tunnel junction (MTJ) structure is used as the skyrmion detection device 15.

[0040] The skyrmion detection device 15 may include a sensing MTJ structure 24 (sensing MTJ) and a reference MTJ structure 25 arranged from bottom to top. The sensing MTJ structure 24 is used to detect the state of skyrmions in the skyrmion track 11. When the sensing MTJ structure 24 detects an increase in the number of skyrmions in the skyrmion track 11, its resistivity changes. When the resistivity of the sensing MTJ structure 24 changes, a trigger current signal I is generated between the reference MTJ structure 24 and the sensing MTJ structure 25. In addition, a signal transmission circuit for transmitting the trigger current signal I is also provided between the reference MTJ structure 25 and the sensing MTJ structure 24.

[0041] Specifically, the detection of magnetic skyrmions occurs due to the change in resistance caused by the presence or absence of skyrmions in the magnetic tunnel junction. The movement of these skyrmions is influenced by in-plane currents flowing through the device or currents perpendicular to the plane. Both the detection MTJ structure 24 and the reference MTJ structure 25 can be sandwich structures. The detection MTJ structure 24 can be a spin valve sensor, capable of reading the number of skyrmions.

[0042] In the neural network model, the preneuron transmits weights to the postneuron via the magnetic tunneling junction effect. Both the preneuron and postneuron can utilize the skyrmion-based simulated neuron device provided in this embodiment. For example, given a certain operating voltage in the voltage control region, the magnitude of the current flowing through this region will be modulated by the conductivity. Therefore, the skyrmion can be induced by the free layer under the magnetic tunneling junction and further adjust the synaptic weights of the next neuron. The synaptic weights are reflected in the magnitude of the control current, because a larger current is needed for the skyrmion to break through the pincers and move smoothly.

[0043] The skyrmion-based simulated neuron device provided in this embodiment simulates neurons based on a skyrmion generation device, a voltage control region, and a skyrmion detection device. Specifically, the basic function of a biological neuron is "leakage-integrate-fire," which means that when a neuron has only one input, the membrane potential will automatically leak and gradually fall back to the resting state due to ion exchange across the cell membrane. Only when all the pulses received by a neuron from the axon terminals connected to that neuron exceed a certain threshold will the neuron continue to transmit pulses. At the same time, after the neuron has sent all the pulses, it will enter a hyperpolarized state. During this refractory period, the neuron will no longer receive stimulation and will maintain a resting potential.

[0044] In this embodiment, the first current injection terminal 12 and the second current injection terminal 13 simulate the presynapse and postsynapse of a neuron, respectively. The preset threshold of the voltage control region 14 is used to simulate the synaptic weights of the neuron, which is the same as the basic functional principle of a biological neuron, thus effectively simulating the neuron. The membrane potential of the simulated neuron increases as the current increases when the accumulated neurons in the presynaptic region reach a certain threshold and break through the restriction to enter the postsynaptic region. This also indicates a correlation between the number of skyrmions and the driving current density. The "integration" of neurons is manifested as the aggregation of skyrmions in the presynaptic region regulated by piezo-controlled magnetic anisotropy. After release, these skyrmions reach the skyrmion detection device 15 located in the postsynaptic region. After the skyrmions are detected, the resistivity changes after recording the number of skyrmions, meaning that the presynaptic region releases several "neurotransmitters" into the postsynaptic region and is then "excited." Finally, the second current injection port 13 at the end will drive the skyrmion under the magnetic tunnel junction away from the region and back to the initial state, completing the "reset" process.

[0045] The skyrmion-based simulated neuron device provided in this embodiment utilizes the properties of skyrmions to demonstrate the plasticity of neurons, including their enhancement and inhibition behaviors. It shows great potential for future use in complex and dense neural networks, greatly reduces the transmission energy consumption between components, and provides hardware and physical principle support for large-scale deep neural networks.

[0046] In a specific example of simulating biological neurons using the skyrmion-based simulated neuron device provided in this application, the skyrmion track 11 has a length of 600 nanometers and a width of 100 nanometers, and the ferromagnetic layer 23 has a thickness of 1 nanometer. The initial perpendicular anisotropy constant K of the ferromagnetic layer 23 is... u =0.8MJ / m 3 The interface Dzyaloshinskii-Moriya interaction (D = 3.5 mJ / m) 2 The voltage control region 14 can be a metal electrode disposed on the ferromagnetic layer 23, with an electrode size of 100 nm × 50 nm. The control voltage of the voltage control region 14 is between -2 volts and 2 volts, and this voltage can control the variation range of the vertical anisotropy constant in the range of K. u =0.8MJ / m 3 This is the median value, which can vary by 10%. This voltage reflects the weight of neural synapses. Voltage control K u This can cause skyrmions to pin or unpin. For example, when K... u =0.8MJ / m 3Drive current density greater than 10 10 A / m 2 At that time, Skomin could successfully navigate the track structure. When K is added... u 10% to 0.88 MJ / m 3 The scribbler was pinned down and unable to pass through. K was controlled by voltage. u The value of [value] can realize the switching of skyrmions. Skyrmions can only flow freely after the current reaches a threshold. This achieves the goal of neurons firing upon reaching the threshold. This device, based on magnetic skyrmions for simulating biological neurons, has a generated current density lower than the driving current density (10). 10 A / m 2 )

[0047] This application also provides a skyminton-based simulated neural network, including multiple skyminton-based simulated neuron devices, each of which can... Figure 1 The skyminton-based simulated neuron device in the illustrated embodiment.

[0048] The plurality of skyrmion-based analog neuron devices are connected according to a preset network structure, wherein the trigger signal output by the previous skyrmion-based analog neuron device is used to control the voltage value of the voltage control region of the next skyrmion-based analog neuron device.

[0049] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A device of a simulating neuron based on a sologenoid, characterized in that, The application relates to a magnetic skyrmion racetrack, which comprises, from bottom to top, a substrate layer, a heavy metal layer and a ferromagnetic layer. Two ends of the skyrmion racetrack are respectively provided with a first current injection port and a second current injection port, the first current injection port and the second current injection port inject current in a direction perpendicular to the skyrmion racetrack to generate skyrmions, the skyrmions flow from the first current injection port to the second current injection port. One end of the ferromagnetic layer close to the first current injection port is provided with a voltage control area, the voltage control area controls the magnetic anisotropy constant of the area through voltage, the skyrmions generated by the first current injection port stop moving after moving to the voltage control area, and then move to the second current injection port when the skyrmions accumulated in the voltage control area exceed a preset threshold. One end of the ferromagnetic layer close to the second current injection port is provided with a skyrmion detection device, the skyrmion detection device is used for detecting the state of skyrmions and outputs a trigger signal when the state of skyrmions is detected to change. The voltage of the voltage control area is adjustable, and the preset threshold is adjusted by adjusting the voltage of the voltage control area.

2. The device of claim 1, wherein the device is a Sine-Gordon based neuromorphic device. The skyrmion detection device comprises a detection magnetic tunnel junction (MTJ) structure and a reference MTJ structure arranged from bottom to top, the detection MTJ structure is used for detecting the state of skyrmions in the skyrmion racetrack, the resistivity of the detection MTJ structure changes when the detection MTJ structure detects that the number of skyrmions in the skyrmion racetrack increases, and a trigger current signal is generated between the reference MTJ structure and the detection MTJ structure when the resistivity of the detection MTJ structure changes.

3. The device of claim 1, wherein the device is a Sine-Gordon based neuromorphic device. The detection MTJ structure and the reference MTJ structure are sandwich structures.

4. The device of claim 3, wherein the SGM mode is a SGM soliton. The current injected by the second current injection port is used for driving the skyrmions reaching the second current injection port to return to an initial state.

5. The device of claim 1, wherein the device is a Sine-Gordon based neuromorphic device. The first current injection port is used for simulating a presynapse of a neuron, the second current injection port is used for simulating a postsynapse of the neuron, and the preset threshold of the voltage control area is used for simulating a synaptic weight of the neuron.

6. The device of any one of claims 1-5, wherein the device is a skyrmion-based neuromorphic device. The ferromagnetic layer is made of iron, cobalt, nickel or other metal or alloy materials with magnetism, the heavy metal layer is made of platinum, and the substrate layer is made of silicon.

7. The device of any one of claims 1-5, wherein the device is a skyrmion-based neuromorphic device. The length of the skyrmion racetrack is 600 nanometers, and the width is 100 nanometers.

8. The device of any one of claims 1-5, wherein the device is a skyrmion-based neuromorphic device. The control voltage of the voltage control area is between-2 volts and 2 volts.

9. The device of claim 1, wherein the device is a Sine-Gordon based neuromorphic device. The application relates to a magnetic skyrmion racetrack, which comprises, from bottom to top, a substrate layer, a heavy metal layer and a ferromagnetic layer.

10. A neural network based on solitons, characterized in that Two ends of the skyrmion racetrack are respectively provided with a first current injection port and a second current injection port, the first current injection port and the second current injection port inject current in a direction perpendicular to the skyrmion racetrack to generate skyrmions, the skyrmions flow from the first current injection port to the second current injection port. One end of the ferromagnetic layer close to the first current injection port is provided with a voltage control area, the voltage control area controls the magnetic anisotropy constant of the area through voltage, the skyrmions generated by the first current injection port stop moving after moving to the voltage control area, and then move to the second current injection port when the skyrmions accumulated in the voltage control area exceed a preset threshold. One end of the ferromagnetic layer close to the second current injection port is provided with a skyrmion detection device, the skyrmion detection device is used for detecting the state of skyrmions and outputs a trigger signal when the state of skyrmions is detected to change. The voltage of the voltage control area is adjustable, and the preset threshold is adjusted by adjusting the voltage of the voltage control area. The skyrmion detection device comprises a detection magnetic tunnel junction (MTJ) structure and a reference MTJ structure arranged from bottom to top, the detection MTJ structure is used for detecting the state of skyrmions in the skyrmion racetrack, the resistivity of the detection MTJ structure changes when the detection MTJ structure detects that the number of skyrmions in the skyrmion racetrack increases, and a trigger current signal is generated between the reference MTJ structure and the detection MTJ structure when the resistivity of the detection MTJ structure changes. The detection MTJ structure and the reference MTJ structure are sandwich structures. The current injected by the second current injection port is used for driving the skyrmions reaching the second current injection port to return to an initial state. The first current injection port is used for simulating a presynapse of a neuron, the second current injection port is used for simulating a postsynapse of the neuron, and the preset threshold of the voltage control area is used for simulating a synaptic weight of the neuron. The ferromagnetic layer is made of iron, cobalt, nickel or other metal or alloy materials with magnetism, the heavy metal layer is made of platinum, and the substrate layer is made of silicon. The length of the skyrmion racetrack is 600 nanometers, and the width is 100 nanometers. The control voltage of the voltage control area is between-2 volts and 2 volts. Two ends of the skyrmion racetrack are respectively provided with a first current injection port and a second current injection port, the first current injection port and the second current injection port inject current in a direction perpendicular to the skyrmion racetrack to generate skyrmions, the skyrmions flow from the first current injection port to the second current injection port; One end of the ferromagnetic layer close to the first current injection port is provided with a voltage control area, the voltage control area controls the magnetic anisotropy constant of the area by voltage, the skyrmions generated by the first current injection port stop moving after moving to the voltage control area, and then move to the second current injection port until the accumulated skyrmions of the voltage control area exceed a preset threshold value; One end of the ferromagnetic layer close to the second current injection port is provided with a skyrmion detection device, the skyrmion detection device is used to detect the state of the skyrmions, and a trigger signal is output when a change in the state of the skyrmions is detected; The plurality of skyrmion-based analog neuron devices are connected according to a preset network structure, wherein the trigger signal output by a previous skyrmion-based analog neuron device is used to control the voltage value of the voltage control area of a next skyrmion-based analog neuron device.

Images