A neuron device and a neural network apparatus based on a magnetic tunnel junction

Abstract

The application discloses a neuron device based on a magnetic tunnel junction and a neural network device, wherein the neuron device based on the magnetic tunnel junction comprises a synthetic antiferromagnetic structure layer, a first side of the synthetic antiferromagnetic structure is provided with a bottom electrode, a barrier layer is arranged on a second side of the synthetic antiferromagnetic structure layer, a ferromagnetic free layer is arranged on a side of the barrier layer away from the bottom electrode, a top electrode is arranged on a side of the ferromagnetic free layer away from the bottom electrode, a first boundary antiferromagnetic pinning layer and a second boundary antiferromagnetic pinning layer are both arranged on a side of the ferromagnetic free layer away from the bottom electrode and are respectively located on two sides of the top electrode; the neuron device based on the magnetic tunnel junction can better control the strength of a stray field by adjusting the synthetic antiferromagnetic structure layer, realizes adjustment of a leakage speed, guarantees a self-leakage function with high reliability, and is conducive to miniaturization and integration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial neural networks, and in particular to a neuron device based on a magnetic tunnel junction and a neural network device. BACKGROUND

[0002] With the rapid development of the field of artificial intelligence, the demand for large-scale data processing has increased dramatically. Neuromorphic networks have shown unique advantages in improving the efficiency of processing complex tasks and reducing power consumption. At present, although various algorithms have greatly developed artificial neural networks, the progress of neuromorphic computing is still limited by the lack of specialized hardware. In view of the gradual slowing down of Moore's law and the limitations of the von Neumann bottleneck, the computing speed and energy efficiency of CMOS (Complementary Metal Oxide Semiconductor) hardware are gradually approaching their theoretical limits, while emerging spintronic devices have significant advantages and application potential in neuromorphic computing due to their ultrafast dynamics, low power consumption, non-volatility, fatigue resistance, and randomness. The core of neuromorphic networks is to realize the functions of neurons and synapses. Using spintronic devices can effectively simulate synaptic weight functions, and by regulating the tunneling magnetoresistance of magnetic tunnel junctions (MTJs), multi-state modulation can be achieved. This non-volatile resistance can be used to construct cross-point arrays to realize the key "vector-matrix" multiplication function in neural networks. However, because the function of neurons is relatively complex, there is relatively little research on the hardware implementation of neurons. Research in neuroscience shows that the function of neurons includes accumulating external input excitation signals, releasing output signals to the outside world when a certain threshold is reached, and gradually leaking accumulated signals when there is no input excitation, i.e., the leaky-integrate-fire (LIF) function. Spintronic devices based on magnetic domain wall motion have great potential in the hardware implementation of neuron devices. Through spin transfer torque (STT) or spin orbit torque (SOT), the motion of magnetic domain walls can be driven. By setting the device structure and materials, the automatic retreat of the magnetic domain wall can be achieved. When the magnetic domain wall moves to the threshold position, the tunneling magnetoresistance of the magnetic tunnel junction changes, which can output a current spike signal in combination with an external circuit. Therefore, through all-electric control means, the leaky-integrate-fire function of neurons can be effectively simulated, providing a feasible solution for the hardware implementation of large-scale, high-speed parallel computing of neuromorphic networks.

[0003] At present, the existing technology uses the magnetic field generated by a hard magnet, or the energy gradient generated by a trapezoidal ferromagnetic free layer, or the energy gradient generated by a uniaxial magnetic anisotropy gradient, to realize a self-leakage function. The spin transfer torque is used to drive the movement of a magnetic domain wall, and when the magnetic domain wall moves to a threshold region, a change in the tunneling magnetoresistance of a magnetic tunnel junction is caused to output a sharp pulse signal. Thus, the leaky-integrate-fire function of a neuron is simulated. However, when the existing technology realizes the leakage function, it is difficult to control the leakage speed. SUMMARY

[0004] The embodiment of the present application provides a neuron device based on a magnetic tunnel junction and a neural network device, so that the strength of a stray field generated by a synthetic antiferromagnetic structure layer is better controlled, the adjustment of a leakage speed is realized, a high-reliability self-leakage function is ensured, and microscaling and integration are facilitated.

[0005] In a first aspect, an embodiment of the present application provides the following technical scheme:

[0006] A neuron device based on a magnetic tunnel junction, comprising:

[0007] a synthetic antiferromagnetic structure layer, a first side of the synthetic antiferromagnetic structure layer being provided with a bottom electrode; a barrier layer, the barrier layer being arranged on a second side of the synthetic antiferromagnetic structure layer; a ferromagnetic free layer, the ferromagnetic free layer being arranged on a side of the barrier layer away from the bottom electrode; wherein a stray field received by the ferromagnetic free layer is determined by a structure of the synthetic antiferromagnetic structure layer; a top electrode, the top electrode being arranged on a side of the ferromagnetic free layer away from the bottom electrode; a first boundary antiferromagnetic pinning layer and a second boundary antiferromagnetic pinning layer, both of which are arranged on a side of the ferromagnetic free layer away from the bottom electrode and are located on two sides of the top electrode, respectively; the first boundary antiferromagnetic pinning layer and the second boundary antiferromagnetic pinning layer are respectively used to determine the magnetization directions of two ends of the ferromagnetic free layer, so that a magnetic domain wall in the ferromagnetic free layer moves between the first boundary antiferromagnetic pinning layer and the second boundary antiferromagnetic pinning layer; a first boundary electrode, the first boundary electrode being arranged on a side of the first boundary antiferromagnetic pinning layer away from the bottom electrode; and a second boundary electrode, the second boundary electrode being arranged on a side of the second boundary antiferromagnetic pinning layer away from the bottom electrode.

[0008] Optionally, the synthetic antiferromagnetic structure layer has the same width at different length positions, so that the ferromagnetic free layer receives a uniform stray field.

[0009] Optionally, the width of the synthetic antiferromagnetic structure layer decreases from the middle to the two ends, so that the ferromagnetic free layer receives a non-uniform stray field.

[0010] Optionally, the ferromagnetic free layer is a material having perpendicular magnetic anisotropy.

[0011] Optionally, the ferromagnetic free layer is a material having tilted magnetic anisotropy.

[0012] Optionally, the synthetic antiferromagnetic structure layer comprises a ferromagnetic reference layer, a synthetic antiferromagnetic coupling layer and a bottom ferromagnetic layer; the bottom electrode is arranged on a first side of the bottom ferromagnetic layer, the synthetic antiferromagnetic coupling layer is arranged on a second side of the bottom ferromagnetic layer, and the ferromagnetic reference layer is arranged on a side of the synthetic antiferromagnetic coupling layer away from the bottom ferromagnetic layer; wherein the saturation magnetization of the ferromagnetic reference layer and the bottom ferromagnetic layer is used to determine the compensation degree of the synthetic antiferromagnetic structure layer, so as to adjust the strength of the stray field received by the ferromagnetic free layer.

[0013] Optionally, the angle between the easy axis of the magnetic moment of the ferromagnetic free layer and the plane where the ferromagnetic free layer is located is 30°-90°.

[0014] Optionally, the heavy metal layer is arranged on a side of the ferromagnetic free layer away from the bottom electrode.

[0015] In a second aspect, an embodiment of the present application provides the following technical scheme:

[0016] A neural network device comprises the magnetic tunnel junction-based neuron device of any one of claims 1-9.

[0017] Optionally, the neural network device further comprises a write control word line, a write bit line, a read control word line, a read bit line, a source line, a first switch tube and a second switch tube; the write control word line is connected to the gate of the first switch tube, the write bit line is connected to the first boundary electrode through the first switch tube, the read control word line is connected to the gate of the second switch tube, the read bit line is connected to the top electrode through the second switch tube, the second boundary electrode is grounded, and the source line is connected to the bottom electrode.

[0018] The one or more technical schemes provided in the embodiments of the present application have at least the following technical effects or advantages:

[0019] 1. The magnetic domain wall motion-based neuron device in the embodiment can simulate the LIF function of a neuron under full electrical control, and can be used in a high-energy-efficiency spiking neuron network (SNN). The stray field received by the ferromagnetic free layer is adjusted by the synthetic antiferromagnetic structure layer, the automatic retreat of the magnetic domain wall is realized, the self-leakage function with high reliability is realized, and further miniaturization and integration are facilitated.

[0020] 2、The neuron device based on magnetic domain wall motion in the embodiment can realize accurate adjustment of the leakage speed of different regions by simply adjusting the shape of the synthetic antiferromagnetic structure layer, and then simulate various linear or nonlinear LIF neuron characteristics, and has good adjustability.

[0021] 3、The neuron device based on magnetic domain wall motion in the embodiment can realize a high-speed neuron device by adjusting the thickness of the ferromagnetic free layer with a tilted magnetic anisotropy, so that the ferromagnetic free layer has a certain in-plane component, and the movement speed of the magnetic domain wall in the accumulation and leakage process is improved.

[0022] 4、The neuron device based on magnetic domain wall motion in the embodiment can realize the Sigmoid activation function function in the artificial neural morphological network by appropriately reducing the driving current, adjusting the shape of the synthetic antiferromagnetic structure layer to gradually widen from both ends to the middle, making the movement speed of the magnetic domain wall first increase and then decrease, and combining the tunneling magnetoresistance relationship of the magnetic tunnel junction, and can be used for the neural network architecture such as the commonly used convolutional neural network (CNN). BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0024] Figure 1 is an implementation structure schematic diagram of a neuron device based on a magnetic tunnel junction in an embodiment of the present application;

[0025] Figure 2 shows the magnetic domain wall position-time change image in the case of different magnetic moment easy axis orientations during the neuron leakage process in the embodiment of the present application;

[0026] Figure 3 shows the magnetic domain wall position-time change image in the case of different magnetic moment easy axis orientations during the neuron accumulation process in the embodiment of the present application;

[0027] Figure 4 shows the magnetic domain wall position-time change image under the driving of different density currents during the neuron accumulation process in the embodiment of the present application.

[0028] Figure 5 is another implementation structure schematic diagram of a neuron device based on a magnetic tunnel junction in an embodiment of the present application;

[0029] Figure 6Fig. 1 shows a diagram of the position of a domain wall over time in a neuron leak process in an embodiment of the present application at different stray field strengths;

[0030] Figure 7 Fig. 3 shows a diagram of the spatial distribution of the stray field of a synthetic antiferromagnetic structure layer of different width in an embodiment of the present application;

[0031] Figure 8 Fig. 4 shows a diagram of another implementation structure of a neuron device based on a magnetic tunnel junction in an embodiment of the present application;

[0032] Figure 9 Fig. 5 shows a diagram of another implementation structure of a neuron device based on a magnetic tunnel junction and a diagram of the easy axis direction of the magnetic moment in an embodiment of the present application;

[0033] Figure 10 Fig. 6 shows a diagram of a structure of a neural network device in an embodiment of the present application. DETAILED DESCRIPTION

[0034] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It should be understood, however, that the description which follows is merely exemplary and is not intended to limit the scope of the present disclosure. Furthermore, in the following description, description of well-known structures and techniques is omitted to avoid obscuring the concept of the present disclosure.

[0035] In the drawings, various structural diagrams according to embodiments of the present disclosure are shown. These diagrams are not drawn to scale in which certain details are exaggerated for clarity and others are omitted. The shapes of various regions, layers, and their relative sizes and positional relationships shown in the diagrams are merely exemplary, and in actuality, they can deviate due to manufacturing tolerances or technical limitations, and regions / layers having different shapes, sizes, and relative positions can be additionally designed by those skilled in the art as needed.

[0036] In the context of the present disclosure, when a layer / element is said to be located "on" another layer / element, the layer / element can be directly located on the other layer / element, or an intervening layer / element can be present therebetween. In addition, if a layer / element is located "on" another layer / element in one orientation, it can be located "under" the other layer / element when the orientation is reversed.

[0037] Referring to Figure 1In an embodiment of the present application, a magnetic tunnel junction based neuron device 10 is provided, which can be used to simulate the neuron firing process; the magnetic tunnel junction based neuron device 10 comprises a synthetic antiferromagnetic structure layer 111, a barrier layer 107, a ferromagnetic free layer 106, a top electrode 103, a first boundary antiferromagnetic structure layer 102, a second boundary antiferromagnetic structure layer 105, a first boundary electrode 101 and a second boundary electrode 104.

[0038] The synthetic antiferromagnetic structure layer 111 (SAF structure) is used to generate a stray field and adjust the stray field received by the ferromagnetic free layer 106; the first side of the synthetic antiferromagnetic structure layer 111 is provided with a bottom electrode.

[0039] Specifically, the synthetic antiferromagnetic structure layer 111 comprises a ferromagnetic reference layer 108, a synthetic antiferromagnetic coupling layer 109 and a bottom ferromagnetic layer 110; the bottom electrode is arranged on the first side of the bottom ferromagnetic layer 110, the synthetic antiferromagnetic coupling layer 109 is arranged on the second side of the bottom ferromagnetic layer 110, and the ferromagnetic reference layer 108 is arranged on the side of the synthetic antiferromagnetic coupling layer 109 away from the bottom ferromagnetic layer 110; wherein the saturation magnetization of the ferromagnetic reference layer 108 and the bottom ferromagnetic layer 110 is used to determine the compensation degree of the synthetic antiferromagnetic structure layer 111 to generate a stray field with corresponding strength. In addition, by adjusting the saturation magnetization of the bottom ferromagnetic layer 110 and the ferromagnetic reference layer 108 in the synthetic antiferromagnetic structure layer 111, the compensation degree of the synthetic antiferromagnetic structure layer 111 can be modulated, different strength stray fields can be generated, and the overall leakage speed can be adjusted.

[0040] In order to achieve the above effects in the present embodiment, the composition materials of the bottom ferromagnetic layer 110 and the ferromagnetic reference layer 108 include any one or more of the following materials with perpendicular magnetic anisotropy: CoFeB, CoFe, Co / Pt (multilayer film of cobalt and platinum alternately), Ni / Co (multilayer film of nickel and cobalt alternately); the composition material of the synthetic antiferromagnetic coupling layer 109 includes one or more of the following metals: Ru, Ta, etc.

[0041] The barrier layer 107, i.e. the non-magnetic barrier layer, is arranged on the second side of the synthetic antiferromagnetic structure layer 111; the composition material of the barrier layer 107 can include any one or more of the following: MgO, HfO x and Al2O3.

[0042] The ferromagnetic free layer 106 is disposed on the barrier layer 107 away from the bottom electrode. The ferromagnetic free layer 106 can be composed of one or more of Co-Ni, Co, and the like. In this embodiment, the stray field can be adjusted by incompletely compensating the magnetization of the bottom ferromagnetic layer 110 of the synthetic antiferromagnetic layer 111 and the ferromagnetic reference layer 108, so that the magnetic domain wall in the ferromagnetic free layer 106 has a tendency to move in the opposite direction of the current driving direction, simulating the leak function of a neuron. In this embodiment, the easy axis orientation of the magnetic moment of the ferromagnetic free layer 106 can also be adjusted by adjusting the thickness of the ferromagnetic free layer 106, so that the magnetic moment has a certain in-plane component, and the movement speed of the magnetic domain wall in the accumulation and leak processes is increased, thereby realizing a high-speed neuron device. Specifically, the angle between the easy axis of the magnetic moment of the ferromagnetic free layer 106 and the plane in which the ferromagnetic free layer 106 is disposed (in this embodiment, the horizontal direction is taken as an example) is 30° to 90°.

[0043] Specifically, the orientation of the easy axis of the magnetic moment is the result of the mutual competition of multiple magnetic anisotropies. When the ferromagnetic thin film is very thin, the easy axis of the magnetic moment tends to be oriented in the direction perpendicular to the thin film due to the effect of surface anisotropy. When the deposition incident angle of the deposition of some ferromagnetic material thin film is greater than 60°, the easy axis of the magnetic moment is limited in the incident plane. Experiments show that when the deposition is performed by tilting, the thickness of the ferromagnetic free layer 106 can be adjusted to adjust the easy axis orientation of the magnetic moment of the free layer, so that the easy axis of the magnetic moment has a certain in-plane component, and the movement speed of the magnetic domain wall in the accumulation and leak processes is increased, thereby realizing a high-speed neuron device.

[0044] Referring to Figure 2 , Figure 2 Fig. 6 shows the magnetic domain wall position-time variation images in the neuron leak process with different orientations of the easy axis (EA) of the magnetic moment. The neuron leak is realized by using a stray field of 2 mT. When the angle between the easy axis of the magnetic moment and the horizontal direction is 90°, the speed of the neuron leak is very slow. When the angle between the easy axis of the magnetic moment and the horizontal direction is reduced to 60°, the magnetic moment has a certain in-plane component, and the speed of the neuron leak is increased. When the angle between the easy axis of the magnetic moment and the horizontal direction is further reduced to 45°, the speed of the neuron leak is reduced, but is still faster than that without the in-plane component. When the angle between the easy axis of the magnetic moment and the horizontal direction is further reduced to 30°, the speed of the neuron leak is similar to that without the in-plane component. Therefore, the in-plane component of the magnetic moment can be modulated by appropriately adjusting the easy axis orientation of the free layer, thereby increasing the speed of the neuron leak.

[0045] Referring to Figure 3 , Figure 3 Fig. 7 shows the magnetic domain wall position-time variation images in the neuron accumulation process with different orientations of the easy axis (EA) of the magnetic moment. The neuron accumulation is realized by using a current density of 1 x 10 8 A / cm2 When the angle between the easy axis of the magnetic moment and the horizontal direction is less than 90°, the magnetic moment has a certain in-plane component, and the accumulation speed of the neuron is faster than that without the in-plane component. When the angle between the easy axis of the magnetic moment and the horizontal direction changes between 30°, 45° and 60°, the overall speed of the neuron accumulation changes little, only the speed at different moments of the accumulation process is different. When the angle between the easy axis of the magnetic moment and the horizontal direction is 0°, the spin-orbit torque of the driving current has weak flipping ability on the horizontal magnetic moment, and the accumulation speed of the domain wall is the slowest. Therefore, the accumulation speed of the neuron can be accelerated by properly adjusting the easy axis direction of the magnetic moment of the ferromagnetic free layer 106.

[0046] Further, the accumulation speed of the neuron can also be accelerated by adjusting the size of the injected current density, as shown in Figure 4 . Figure 4 When the easy axis direction of the magnetic moment of the ferromagnetic free layer 106 is 45° and the stray field strength is constant, the image of the position of the magnetic domain wall changing with time under the driving of the current with different densities during the accumulation of the neuron is shown. With the increase of the current density, the accumulation speed of the magnetic domain wall is accelerated, and the time required for the magnetic domain wall to move from the left end to the right end is significantly reduced. When the current density is small, the driving effect of the spin transfer torque of the spin current on the magnetic domain wall is not enough to overcome the reverse inhibition of the stray field, and the magnetic domain wall cannot be accumulated to the right end. Therefore, in the presence of the stray field, the accumulation speed of the neuron can be adjusted by adjusting the size of the injected current density.

[0047] The top electrode 103 is arranged on the side of the ferromagnetic free layer 106 away from the bottom electrode.

[0048] The first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer are both arranged on the side of the ferromagnetic free layer 106 away from the bottom electrode, and are respectively located on the two sides of the top electrode 103. When the first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer are made, the thickness of the end of the ferromagnetic free layer 106 can be increased to achieve the same.

[0049] The first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer are used to respectively determine the magnetization direction of the two ends of the ferromagnetic free layer 106, so that the magnetic domain wall in the ferromagnetic free layer 106 moves between the first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer. Specifically, the magnetization directions of the first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer are opposite to each other to realize the injection and pinning of the magnetic domain wall.

[0050] That is, the first boundary anti-ferromagnetic pinning layer and the second boundary anti-ferromagnetic pinning layer can pin the magnetic moments at the two ends of the ferromagnetic free layer 106 in the +z and -z directions, respectively, as the magnetic domain wall nucleation region; wherein the +z direction is the direction perpendicular to the ferromagnetic free layer 106 and away from the bottom electrode, and the -z direction is opposite to the +z direction. The spin transfer torque generated by the spin-polarized current can drive the magnetic domain wall to move in the ferromagnetic free layer 106, simulating the accumulation process of the neuron. The stray field generated by the synthetic anti-ferromagnetic structure layer 111 can make the magnetic moment in the ferromagnetic free layer 106 flip, realizing the automatic retreat of the magnetic domain wall in the ferromagnetic free layer 106, simulating the leakage process of the neuron. When the magnetic domain wall in the free layer moves to the threshold position, the magnetization direction of the ferromagnetic free layer 106 at this position flips, the magnetic moments of the ferromagnetic free layer 106 at the two ends of the magnetic tunnel junction switch from the anti-parallel state to the parallel state, the tunneling magnetoresistance decreases, and a current spike signal can be output in combination with the external circuit, simulating the neuron firing process.

[0051] Further, the magnetic domain wall movement in the ferromagnetic free layer 106 is driven by the amplitude, pulse width and number of the current pulse from the synapse; when there is no current pulse or the current pulse is small, the magnetic domain wall will move in the opposite direction under the action of the stray field of the SAF structure; when the magnetic domain wall moves to the region where the top electrode 103 is located, i.e. the region where the read current passes through, the magnetic moments of the ferromagnetic free layer 106 at the two ends of the MTJ switch from the anti-parallel state to the parallel state, and the MTJ outputs a current spike signal in combination with the peripheral circuit, thereby simulating the complete leakage-accumulation-release characteristics of the neuron.

[0052] Please refer to Figure 5 In some implementations, the neuron device 100 of the magnetic tunnel junction can also drive the magnetic domain wall by depositing a heavy metal layer 401 above the ferromagnetic free layer 106, injecting spin-polarized current by spin Hall effect, and driving the magnetic domain wall by spin-orbit torque, in addition to driving the magnetic domain wall by spin transfer torque through spin-polarized current generated in the local pinning region at the boundary of the ferromagnetic free layer 106.

[0053] In this embodiment, the materials of the top electrode 103, the first boundary electrode 101 and the second boundary electrode 104 can include one or more of Cu, Au and other metals.

[0054] The first boundary electrode 101 is arranged on the side of the first boundary anti-ferromagnetic pinning layer away from the bottom electrode, and the second boundary electrode 104 is arranged on the side of the second boundary anti-ferromagnetic pinning layer away from the bottom electrode.

[0055] In this embodiment, the width of the synthetic anti-ferromagnetic structure layer 111 at different lengths can be designed to regulate the stray field strength of the ferromagnetic free layer 106 in different regions, thereby regulating the leakage speed in different regions, so as to realize various linear / non-linear neuron characteristics. Please refer toFigure 6 , Figure 6 Fig. 6 shows the images of the position of the domain wall of the ferromagnetic free layer 106 with the easy axis direction of 45° under different stray field intensities during the neuron leaking process. Without the injection of the current, the speed of the domain wall leaking is accelerated with the increase of the stray field intensity, and the time required for the domain wall to return from the right end to the left end is significantly reduced. Therefore, the leaking process of the neuron can be effectively realized by using the stray field, and the leaking speed of the neuron can be adjusted by adjusting the intensity of the stray field. Please refer to Figure 7 Fig. 7 shows the images of the position of the domain wall of the ferromagnetic free layer 106 with the easy axis direction of 45° under different stray field intensities during the neuron leaking process. Without the injection of the current, the speed of the domain wall leaking is accelerated with the increase of the stray field intensity, and the time required for the domain wall to return from the right end to the left end is significantly reduced. Therefore, the leaking process of the neuron can be effectively realized by using the stray field, and the leaking speed of the neuron can be adjusted by adjusting the intensity of the stray field. Please refer to Figure 7 Fig. 8 shows the images of the stray field spatial distribution of the SAF structure with different size widths. The smaller the size width of the SAF structure, the smaller the distance between the ferromagnetic free layer 106 and the edge of the SAF structure, and the greater the stray field acting on the ferromagnetic free layer 106. Therefore, the size width of the SAF structure can be adjusted to effectively control the size of the stray field acting on the ferromagnetic free layer 106.

[0056] For example, in some implementations, the width of the synthetic antiferromagnetic structure layer 111 at different length positions can be set to be the same, so that the synthetic antiferromagnetic structure layer 111 can generate a uniform stray field, and the ferromagnetic free layer 106 can be subjected to a uniform stray field, thereby realizing the linear neuron characteristics, as shown in Figure 1 .

[0057] For another example, in some implementations, please refer to Figure 8 , the neuron device 20 of the magnetic tunnel junction includes a synthetic antiferromagnetic structure layer 211, a barrier layer 207, a ferromagnetic free layer 206, a top electrode 203, a first boundary antiferromagnetic structure layer 202, a second boundary antiferromagnetic structure layer 205, a first boundary electrode 201 and a second boundary electrode 204; wherein the synthetic antiferromagnetic structure layer 211 includes a ferromagnetic reference layer 208, a synthetic antiferromagnetic coupling layer 209 and a bottom ferromagnetic layer 110. The width of the synthetic antiferromagnetic structure layer 211 can be set to decrease from the middle to both ends, so that the synthetic antiferromagnetic structure layer 211 can generate a non-uniform stray field, and the ferromagnetic free layer 206 can be subjected to a non-uniform stray field, thereby realizing the nonlinear neuron characteristics. Specifically, the shape of the synthetic antiferromagnetic structure layer 211 can be set to gradually widen from both ends to the middle, so that the reverse inhibition of the stray field acting on the ferromagnetic free layer 206 gradually weakens from both ends to the middle, the movement speed of the domain wall first accelerates and then slows down, the resistance of the magnetic tunnel junction is linearly related to the movement distance of the domain wall in the ferromagnetic free layer 206, and the nonlinear Sigmoid function relationship between the pulse number and the tunneling current of the magnetic tunnel junction can be realized.

[0058] The principles of the present application will be further described and explained with reference to actual examples and drawings:

[0059] Please continue to refer to Figure 1 , inFigure 1 In the illustrated magnetic tunnel junction based neuron device 10, the left first antiferromagnetic pinning layer pins the magnetization direction of the left end region of the ferromagnetic free layer 106 in the -z direction, and the right second antiferromagnetic pinning layer pins the magnetization direction of the right end region of the ferromagnetic free layer 106 in the +z direction. The magnetic domain wall in the ferromagnetic free layer 106 moves between the two end pinning regions without annihilation. The magnetization direction of the bottom ferromagnetic layer 110 is along the +z direction, and the magnetization direction of the ferromagnetic reference layer 108 is along the -z direction. The saturation magnetization of the bottom ferromagnetic layer 110 is greater than that of the ferromagnetic reference layer 108. The magnetic fields generated by the two layers are not completely compensated, and the synthetic antiferromagnetic coupling layer 109, the bottom ferromagnetic layer 110 and the ferromagnetic reference layer 108 on both sides thereof form a SAF structure. The SAF structure can adjust the stray field generated thereby, so that the magnetization direction of the ferromagnetic free layer 106 above the SAF structure tends to be in the +z direction. Even if the magnetic domain wall moves in the -x direction (in the length direction of the ferromagnetic free layer 106 and towards the direction in which the first boundary electrode 101 is located), the leakage process of the neuron is simulated. When the width of the SAF structure is consistent (i.e., as illustrated in Figure 1 The structure illustrated), the ferromagnetic free layer 106 can be subjected to a uniform stray field, linear back driving of the magnetic domain wall of the ferromagnetic free layer 106 is achieved, and thus a uniform leakage speed is achieved.

[0060] When the width of the SAF layer varies at different length positions, as illustrated in Figure 8 The ferromagnetic free layer 206 can be subjected to a non-uniform stray field, non-linear back driving of the magnetic domain wall of the ferromagnetic free layer 206 is achieved, and thus a non-uniform leakage speed is achieved. In the initial state, the magnetization direction of the non-magnetic domain pinning region in the ferromagnetic free layer 206 is along the +z direction, i.e., the magnetic domain wall is located at the boundary of the left end pinning region. When a current is excited, the magnetization direction of the ferromagnetic free layer 206 is flipped to the -z direction by the spin transfer torque of the spin-polarized current, i.e., the magnetic domain wall is driven to move in the +x direction (the opposite direction of the -x direction), and the accumulation process of the neuron is simulated. After a series of accumulation and leakage processes, when the magnetic domain wall moves to the region below the top electrode 203, the magnetization directions of the ferromagnetic free layer 206 and the ferromagnetic reference layer 208 on both ends of the region are switched from anti-parallel to parallel by the read current between the top electrode and the bottom electrode, the tunneling magnetoresistance is reduced, a current spike signal is output by the peripheral circuit, and the discharge process of the simulated neuron is achieved.

[0061] Please continue to refer to Figure 8 , in Figure 8A magnetic tunnel junction based neuron device implementing a nonlinear Sigmoid function by a non-uniform stray field is shown in FIG. 1. In this example, the SAF structure has a non-uniform width, i.e., the width of the SAF structure gradually increases from both ends to the middle. When the current pulse is small, the spin transfer torque drives the domain wall in the ferromagnetic free layer 206 to move in the +x direction, while the stray field generated by the SAF structure drives the domain wall in the ferromagnetic free layer 206 to move in the -x direction, and the two compete with each other. Since the ferromagnetic free layer 206 is farther away from the edge of the synthetic antiferromagnetic structure layer 211, it is less affected by the stray field. The reverse inhibition of the stray field on the domain wall during movement gradually weakens from both ends to the middle, and the movement speed of the domain wall first increases and then decreases. Since the tunneling magnetoresistance of the magnetic tunnel junction is linearly related to the movement distance of the domain wall in the ferromagnetic free layer 206, a nonlinear Sigmoid function relationship between the number of current pulses and the tunneling current of the magnetic tunnel junction can be achieved.

[0062] Please refer to Figure 9 In Figure 9 The magnetic tunnel junction based neuron device 30 shown in FIG. 2 is a schematic diagram of the easy axis direction of the magnetic moment when using the in-plane component to implement a high-speed neuron device. Figure 1 The corresponding example is different in that the easy axis direction of the magnetic moment of the ferromagnetic free layer 306 is different. Figure 1 The ferromagnetic free layer 106 in FIG. 1 has perpendicular magnetic anisotropy, Figure 3 The ferromagnetic free layer 306 in FIG. 2 has tilted magnetic anisotropy, and the easy axis of the magnetic moment has a certain angle with the y axis, i.e., the easy axis of the magnetic moment has a certain in-plane component. Using the in-plane component of the magnetic moment can improve the efficiency of the magnetic moment flip and the movement speed of the domain wall during accumulation and leakage, thereby implementing a high-speed neuron device.

[0063] It should also be noted that:

[0064] 1. The size of the structure of each layer of the magnetic tunnel junction based neuron device 10 provided in this embodiment can be micronized according to the process.

[0065] 2. The synthetic antiferromagnetic structure layer 111 and the shape of the ferromagnetic free layer 106 used to adjust the strength of the stray field can be replaced by various linear or nonlinear geometric shapes, such as a square, a rectangle, a rhombus, an ellipse, a circle, or other irregular shapes in the xy plane, etc., to simulate various linear or nonlinear characteristics of neurons.

[0066] 3. The in-plane component of the magnetic moment in the ferromagnetic free layer 106 can also be adjusted by annealing under a certain magnetic field.

[0067] In summary, the magnetic tunnel junction based neuron device 10 provided in this embodiment has at least the following beneficial effects:

[0068] 1. The neuronal device based on magnetic domain wall motion in this embodiment can simulate the LIF function of neurons under full electronic control and can be used in high-efficiency spiking neuron networks (SNNs). By adjusting the stray field of the ferromagnetic free layer 106 through the SAF structure, the automatic retraction of the magnetic domain wall is realized, achieving a highly reliable self-leaking function, which is conducive to further miniaturization and integration.

[0069] 2. The neuron device based on domain wall motion in this embodiment can achieve precise adjustment of leakage velocity in different regions by simply adjusting the shape of the SAF structure, thereby simulating the characteristics of various linear or nonlinear LIF neurons, and has good controllability.

[0070] 3. The neuron device based on magnetic domain wall motion in this embodiment can realize a high-speed neuron device by adjusting the thickness of the ferromagnetic free layer 106 with tilted magnetic anisotropy, so that the magnetic moment of the ferromagnetic free layer 106 has a certain in-plane component, and at the same time increasing the motion speed of the magnetic domain wall during the accumulation and leakage process.

[0071] 4. The neuronal device based on magnetic domain wall motion in this embodiment only needs to reduce the driving current appropriately and adjust the shape of the SAF structure to gradually widen from both ends to the middle, so that the motion speed of the magnetic domain wall first increases and then decreases. Combined with the tunneling magnetoresistance relationship of the magnetic tunnel junction, the Sigmoid activation function in artificial neuromorphic networks can be realized. It can be used in common neural network architectures such as Convolutional Neuron Networks (CNN).

[0072] Based on the same inventive concept, another embodiment of the present invention provides a neural network device, including any of the neuron devices based on magnetic tunneling junctions described in the foregoing embodiments. For details, please refer to... Figure 10 , Figure 10 The diagram shows a schematic of the read / write unit structure of a neural network device. This read / write unit circuit adopts an MTJ structure, including a write control word line WWL, a write bit line WBL, a read control word line RWL, a read bit line RBL, a source line SL, a first switch S1, and a second switch S2. The write control word line WWL is connected to the gate of the first switch S1 and is used to control the on / off state of the first switch S1. The write bit line WBL is connected to the first boundary electrode 101 through the first switch S1. The read control word line RWL is connected to the gate of the second switch S2 and is used to control the on / off state of the second switch S2. The read bit line RBL is connected to the top electrode 103 through the second switch S2. The second boundary electrode 104 is grounded, and the source line SL is connected to the bottom electrode.

[0073] When the write control word line WWL is open, a current path is formed from the write bit line WBL - the domain wall transport free layer - ground. The injection current is controlled by the electrical signal on the write bit line WBL, and the domain wall is driven to move. When the read control word line RWL is open, a current path is formed from the read bit line RBL - the magnetic tunnel junction (MTJ) - the source line SL - ground. At this time, the electrical signal released by the neuron device can be read out. When the domain wall moves to the corresponding area below the top electrode 103, the tunneling magnetoresistance of the MTJ decreases, and a current spike signal can be output through the read bit line RBL. The signal can also be output after being amplified by a comparison amplifier.

[0074] It should be noted that the neural network device provided in the embodiment adopts the neuron device based on the magnetic tunnel junction in the foregoing embodiments as a basic component unit; therefore, the neural network device also has the same beneficial effects as the neuron device based on the magnetic tunnel junction in the foregoing embodiments, and details can be referred to the foregoing embodiments, which will not be described herein again.

[0075] In the foregoing description, the technical details such as the patterning, etching and the like of each layer are not described in detail. However, it should be understood by those skilled in the art that the layers, regions and the like with desired shapes can be formed by various technical means. In addition, those skilled in the art can also design methods different from the methods described above to form the same structure. In addition, although each embodiment is described separately above, this does not mean that the measures in each embodiment cannot be used advantageously in combination.

[0076] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0077] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, it is intended that the present application embrace all such modifications and changes and include these within the scope of the appended claims and their equivalents.

Claims

1. A magnetic tunnel junction based neuron device, comprising: include: A synthetic antiferromagnetic structure layer is provided with a bottom electrode on the first side of the synthetic antiferromagnetic structure; A barrier layer is disposed on the second side of the synthetic antiferromagnetic structure layer; A ferromagnetic free layer is disposed on the side of the barrier layer away from the bottom electrode; wherein the stray field experienced by the ferromagnetic free layer is determined by the structure of the synthetic antiferromagnetic structure layer; The top electrode is disposed on the side of the ferromagnetic free layer away from the bottom electrode; The first boundary antiferromagnetic nailing layer and the second boundary antiferromagnetic nailing layer are both disposed on the side of the ferromagnetic free layer away from the bottom electrode, and are respectively located on both sides of the top electrode; the first boundary antiferromagnetic nailing layer and the second boundary antiferromagnetic nailing layer are respectively used to determine the magnetization direction at both ends of the ferromagnetic free layer, so that the domain walls in the ferromagnetic free layer can move between the first boundary antiferromagnetic nailing layer and the second boundary antiferromagnetic nailing layer; A first boundary electrode is disposed on the side of the first boundary antiferromagnetic nailing layer away from the bottom electrode; and a second boundary electrode is disposed on the side of the second boundary antiferromagnetic nailing layer away from the bottom electrode; The synthetic antiferromagnetic structure layer includes a ferromagnetic reference layer, a synthetic antiferromagnetic coupling layer, and a bottom ferromagnetic layer; the bottom electrode is disposed on a first side of the bottom ferromagnetic layer, the synthetic antiferromagnetic coupling layer is disposed on a second side of the bottom ferromagnetic layer, and the ferromagnetic reference layer is disposed on the side of the synthetic antiferromagnetic coupling layer away from the bottom ferromagnetic layer; The saturation magnetization of the ferromagnetic reference layer and the bottom ferromagnetic layer is used to determine the compensation degree of the synthesized antiferromagnetic structure layer, so as to adjust the intensity of the stray field experienced by the ferromagnetic free layer.

2. The magnetic tunnel junction based neuron device of claim 1, wherein, The width of the synthetic antiferromagnetic structure layer is the same at different length positions, so that the ferromagnetic free layer is subjected to a uniform stray field.

3. The magnetic tunnel junction based neuron device of claim 1, wherein, The width of the synthetic antiferromagnetic structure layer decreases from the middle to both ends, so that the ferromagnetic free layer is subjected to a non-uniform stray field.

4. The magnetic tunnel junction based neuron device of claim 1, wherein, The ferromagnetic free layer is a material with perpendicular magnetic anisotropy.

5. The magnetic tunnel junction based neuron device of claim 1, wherein, The ferromagnetic free layer is a material with tilted magnetic anisotropy.

6. The magnetic tunnel junction based neuron device of claim 1, wherein, The angle between the easy axis of the magnetic moment of the ferromagnetic free layer and the plane in which the ferromagnetic free layer is located is 30° to 90°.

7. The magnetic tunnel junction based neuron device of claim 1, wherein, It also includes a heavy metal layer disposed on the side of the ferromagnetic free layer away from the bottom electrode.

8. A neural network apparatus, characterized by comprising: Includes the neuronal device based on magnetic tunneling junctions as described in any one of claims 1-7.

9. The neural network device of claim 8, wherein, Also includes: Write control word line, write bit line, read control word line, read bit line, source line, first switch transistor and second switch transistor; The write control word line is connected to the gate of the first switching transistor, the write bit line is connected to the first boundary electrode through the first switching transistor, the read control word line is connected to the gate of the second switching transistor, the read bit line is connected to the top electrode through the second switching transistor, the second boundary electrode is grounded, and the source line is connected to the bottom electrode.

Images