Artificial neuron device based on magnetic tunnel junction and application

Abstract

This application discloses an artificial neuron device and its application based on a magnetic tunnel junction. The artificial neuron device includes a magnetic tunnel junction for transmitting radio frequency (RF) signals in response to a driving electrical signal. The magnetic tunnel junction includes a capping layer, a magnetic free layer, a barrier layer, a pinning layer, and an antiferromagnetic layer. The magnetic free layer has magnetic anisotropy along the direction from the barrier layer to the capping layer. The capping layer is used to adjust the magnetic anisotropy of the magnetic free layer so that the transmission frequency of the RF signal is independent of the driving electrical signal. A first electrode layer is disposed on one side of the magnetic tunnel junction; a second electrode layer is disposed on the other side of the magnetic tunnel junction. The second electrode layer is used to cooperate with the first electrode layer to input the driving electrical signal into the magnetic tunnel junction. The artificial neuron device of this application can achieve a fixed transmission frequency of the RF signal by changing the thickness of the capping layer; it can simulate the working characteristics of artificial neurons in artificial neural networks and can be used as a spin oscillator in RF-based artificial neural networks.

Description

Technical Field

[0001] This application belongs to the field of artificial neuron device technology, specifically relating to an artificial neuron device based on a magnetic tunnel junction and its application. Background Technology

[0002] Currently, radio frequency (RF) signals are widely used in fields such as wireless communication, transforming human work and lifestyles. With the new industrial revolution represented by the Internet of Things (IoT), the scope of RF applications is further expanding, and the demand for accurate and rapid analysis is becoming increasingly strong. In recent years, artificial neural networks have demonstrated advantages in RF-related applications, such as medicine, fingerprint recognition, finger sensing, radar applications, and aircraft detection and identification. Furthermore, when applied in real-world environments, artificial neural networks often outperform traditional algorithms that rely on complex analysis tools.

[0003] However, current methods for processing radio frequency (RF) signals using artificial neural networks first require digitizing the antenna-sensed signals, followed by processing on CMOS-based hardware such as CPUs, GPUs, FPGAs, or ASICs. Both stages of this process require significant computation, resulting in high latency and high energy consumption, which limits the further application and development of RF technology. Summary of the Invention

[0004] The purpose of this application is to provide an artificial neuron device and its application based on a magnetic tunnel junction, in order to solve the technical problems that existing artificial neural networks have high latency and high power consumption, which limit the further application and development of radio frequency technology.

[0005] To achieve the above objectives, one technical solution adopted in this application is:

[0006] An artificial neuron device with a magnetic tunnel junction is provided, comprising:

[0007] A magnetic tunnel junction for transmitting radio frequency (RF) signals in response to a driving electrical signal. The magnetic tunnel junction includes a cap layer, a magnetic free layer, a barrier layer, a pinning layer, and an antiferromagnetic layer stacked sequentially. The magnetic free layer has magnetic anisotropy along the direction from the barrier layer to the cap layer. The cap layer is used to adjust the magnetic anisotropy of the magnetic free layer, and the thickness of the cap layer is configured such that the transmission frequency of the RF signal is independent of the driving electrical signal.

[0008] The first electrode layer is disposed on one side of the magnetic tunnel junction;

[0009] A second electrode layer is disposed on the other side of the magnetic tunnel junction. The second electrode layer is used to cooperate with the first electrode layer to input the driving electrical signal to the magnetic tunnel junction.

[0010] In one or more embodiments, the driving electrical signal is a direct current to drive the magnetic moment oscillation of the magnetic current layer.

[0011] In one or more embodiments, the absolute value of the voltage of the driving electrical signal is greater than 0V and less than 10V.

[0012] In one or more embodiments, the material of the magnetic free layer is one or more combinations of CoFeB, Py, Co, CoFe, FePt, Co / Ni, and FeB.

[0013] In one or more embodiments, the capping layer is made of MgO; and / or the thickness of the capping layer is 0.5 to 1 nm.

[0014] In one or more embodiments, the thickness of the magnetic free layer is 0.1 to 10 nm.

[0015] In one or more embodiments, the thickness of the barrier layer is 1 to 5 nm.

[0016] In one or more embodiments, a magnetoelectric coupling device is further included, the magnetoelectric coupling device being used to generate stress in response to a control electrical signal to modulate the magnetic anisotropy of the magnetic free layer.

[0017] Preferably, the magnetoelectric coupling device includes a piezoelectric thin film disposed on the side of the first electrode layer opposite to the magnetic tunnel junction or disposed on the side of the second electrode layer opposite to the magnetic tunnel junction.

[0018] In one or more embodiments, a magnetic metal layer is further included, the magnetic metal layer being disposed on a side of the cap layer opposite to the magnetic free layer, and the magnetic metal layer being used to provide an external magnetic field to control the magnetization direction of the magnetic free layer.

[0019] To achieve the above objectives, another technical solution adopted in this application is:

[0020] The application of the magnetic tunnel junction-based artificial neuron device described in any of the above embodiments in radio frequency-based artificial neural networks is provided.

[0021] The advantages of this application, which differ from existing technologies, are:

[0022] The artificial neuron device proposed in this application adjusts the magnetic anisotropy of the magnetic free layer through the capping layer. By changing the thickness of the capping layer, the transmission frequency of the radio frequency signal can be made independent of the driving signal, thereby achieving a fixed transmission frequency of the radio frequency signal.

[0023] The magnetic free layer of the artificial neuron device proposed in this application has a low flip barrier, thus resulting in low power consumption and effectively reducing the energy consumption of the artificial neural network when processing radio frequency signals.

[0024] The overall size of the artificial neuron device proposed by this applicant can be made very small, enabling high integration.

[0025] The artificial neuron device proposed in this application can simulate the working characteristics of artificial neurons in artificial neural networks. It can be used as a spin oscillator in radio frequency-based artificial neural networks to realize microwave detection of artificial neuron devices. As a unit device for performing multiplication calculations, it can realize accumulation calculations through device interconnection. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of one embodiment of the artificial neuron device based on a magnetic tunnel junction according to this application;

[0027] Figure 2 The ratio I of the power P to the DC current and AC current in the magnetic tunnel junction of this application is... dc / I th Relationship diagram;

[0028] Figure 3 This is a schematic diagram of another embodiment of the artificial neuron device based on a magnetic tunnel junction in this application;

[0029] Figure 4 This is a schematic diagram of another embodiment of the artificial neuron device based on a magnetic tunnel junction in this application. Detailed Implementation

[0030] The present application will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0031] Currently, when using artificial neural networks to process radio frequency (RF) signals, the signals sensed by the antenna must first be digitized, and then processed and executed on CMOS-based hardware (such as central processing units (CPUs), graphics processing units (GPUs), field-programmable arrays (FPGAs), or application-specific integrated circuits (ASICs). Both stages of this process require a large amount of computation, resulting in high latency and high energy consumption, which limits the further application and development of RF technology.

[0032] Today, spintronic devices are gaining increasing attention due to their low power consumption and ability to transmit information over long distances via spin currents and microwave signals. However, when a spin oscillator is used as a neural device corresponding to a spin diode synapse, its information is represented solely by the injected DC current. Different information corresponds to different DC currents. Therefore, it is necessary to ensure that the transmission frequency of the radio frequency signal emitted by the spin oscillator does not change with the input current, thereby enabling microwave detection in artificial neuron devices. As a unit device performing multiplication calculations, it achieves accumulation calculations through device interconnection.

[0033] To this end, the applicant has developed an artificial neuron device based on a magnetic tunnel junction. This artificial neuron device has low power consumption and high integration, and the transmission shielding of its output radio frequency electrical signal is not affected by its input electrical signal. Thus, it can be used as a spin oscillator to provide a fixed transmission frequency radio frequency electrical signal when applied to radio frequency-based artificial neural networks.

[0034] Specifically, please refer to Figure 1 , Figure 1 This is a schematic diagram of one embodiment of the artificial neuron device based on a magnetic tunnel junction according to this application.

[0035] The artificial neuron device includes a magnetic tunnel junction 30 and a first electrode layer 10 and a second electrode layer 20 for cooperating to input driving electrical signals to the magnetic tunnel junction 30.

[0036] Specifically, the first electrode layer 10 is arranged on one side of the magnetic tunnel junction 30, and the second electrode layer 20 is arranged on the other side of the magnetic tunnel junction 30. The first electrode layer 10 and the second electrode layer 20 are electrically connected to a power supply 40, thereby inputting a driving electrical signal to the magnetic tunnel junction 30.

[0037] Specifically, the driving electrical signal can be a direct current, and the absolute value of its voltage can be greater than 0V and less than 10V.

[0038] The magnetic tunnel junction 30 is used to transmit radio frequency electrical signals in response to a driving electrical signal. It includes a cap layer 31, a magnetic free layer 32, a barrier layer 33, a pinning layer 34, and an antiferromagnetic layer 35 stacked sequentially.

[0039] The magnetic free layer 32 has magnetic anisotropy along the direction from the capping layer 31 to the barrier layer 33, i.e., perpendicular magnetic anisotropy. The magnetic free layer 32 can be made of CoFeB material and the thickness can be 1 nm. In other embodiments, the magnetic free layer 32 can also be made of other materials. For example, the material of the magnetic free layer 32 can be one or more combinations of CoFeB, Py, Co, CoFe, FePt, Co / Ni, and FeB. The thickness of the magnetic free layer 32 can be 0.1 to 10 nm, all of which can achieve the effect of this embodiment.

[0040] The capping layer 31 is used to adjust the magnetic anisotropy of the magnetic free layer 32 so that the transmission frequency of the radio frequency signal is independent of the driving signal.

[0041] Specifically, the capping layer 31 can be made of MgO, and its thickness can be 0.5 nm. In other embodiments, the thickness of the capping layer 31 can also be 0.5–1 nm. It is worth noting that the thickness of the capping layer 31 should be adjusted based on the vertical magnetic anisotropy of the magnetic free layer 32, thereby achieving the control of the vertical magnetic anisotropy of the magnetic free layer 32, so that the radio frequency electrical signal transmission frequency of the magnetic tunnel junction 30 is independent of the driving electrical signal.

[0042] The barrier layer 33 allows charge carriers to tunnel between the magnetic free layer 32 and the pinning layer 34. The barrier layer 33 can be made of MgO material and can have a thickness of 2 nm. In other embodiments, the thickness of the barrier layer 33 can also be 1 to 5 nm. The barrier layer 33 can also be made of other materials that can allow charge carriers to tunnel between the magnetic free layer 32 and the pinning layer 34, all of which can achieve the effect of this embodiment.

[0043] The pinning layer 34 can be made of CoFeB material with a thickness of 1 nm. In other embodiments, the thickness of the pinning layer 34 can be 0.1–10 nm. The magnetization direction of the pinning layer 34 can be in-plane or perpendicular to the in-plane orientation in other embodiments, both achieving the effects of this embodiment. Specifically, the in-plane orientation means that its magnetic moments are arranged along the plane of the pinning layer 34, and the magnetization direction is fixed by the antiferromagnetic layer 35 and does not change significantly under the operating conditions of the device. The perpendicular to the in-plane orientation means that its magnetic moments are arranged perpendicular to the plane of the pinning layer 34.

[0044] Understandably, since the magnetic free layer 32 of the artificial neuron device in this embodiment has a low flip barrier, the power consumption is low; at the same time, the overall structural size can be made very small, thus enabling high integration of the artificial neuron device.

[0045] The specific principle of this implementation method is as follows:

[0046] When the magnetic tunnel junction 30 has no external driving force, its self-consistent dynamic equation can be expressed as:

[0047]

[0048] Where c is the oscillation amplitude, ω(P) is the resonant frequency of the resonant element, and Γ + (P) represents the damping rate of natural energy dissipation (natural damping or positive damping), Γ -(P) represents the negative damping ratio, which describes the effective function of an active device.

[0049] To find the constant solution of equation (1), it is rewritten as two equations concerning the power P = |c|. 2 The system of equations for phase φ = arg(c):

[0050]

[0051]

[0052] Equation (2) represents the dynamic equation of the oscillation power P in the magnetic tunnel junction 30, and equation (3) represents the dynamic equation of the oscillation phase φ. Equation (2) is independent of the oscillation phase φ, meaning that only the oscillation power P can be considered first.

[0053] For equation (2), there are two possible steady-state solutions (dP / dt = 0).

[0054] The first steady-state solution: P = 0, i.e., when there is no oscillation, linearizing around the zero-power solution (P→0) yields the following equation:

[0055]

[0056] Among them, when Γ + (0)>Γ - When (0), the zero-power solution is stable and physically realizable. Therefore, it is possible to obtain the first steady-state solution from Γ. + (0)=Γ - (0), determine the threshold for non-zero power oscillations in the magnetic tunnel junction 30.

[0057] The second steady-state solution: when Γ + (P0)=Γ - When (P0), a non-zero solution P = P0 is obtained. Its physical meaning is that under static conditions, due to natural energy loss (Γ)... + The energy loss caused by (P0) should be completely offset by the energy provided by the active device (Γ). - (P0)) exactly compensates, that is, the total effective damping Γ + (P0)-Γ - (P0) is exactly equal to 0.

[0058] For most devices, positive damping (Γ) + (P0) is proportional to the oscillation power P (i.e., dΓ) + (P0) / P>0), and negative damping (Γ) - (P0) is inversely proportional to the oscillation power P (i.e., dΓ) - (P0) / P<0), in this case, equation Γ +(P0)=Γ - The solution to (P0) is unique.

[0059] Furthermore, in the structure of its magnetic tunnel junction 30, the frequency ω(P) and damping ratio Γ ± The relationship between (P0) and power P is expressed using a weakly nonlinear (P << 1) Taylor expansion:

[0060] ω(P)≈ω0+N P (5a)

[0061] Γ + (P)≈Γ G (1+Q P (5b)

[0062] Γ - (P)≈σI(1-P) (5c)

[0063] Where ω0 is the ferromagnetic resonance frequency, N P σ is the nonlinear frequency shift coefficient caused by amplitude-phase coupling, I is the spin transmission efficiency, and I is the DC current.

[0064] Furthermore, substituting the approximate value (5) into the equation, we obtain the static power P0 of the magnetic tunnel junction 30:

[0065]

[0066] Where Q is the nonlinear damping coefficient. The dimensionless supercritical parameter ζ is defined as:

[0067]

[0068] Among them, I th This is the threshold current.

[0069] Substituting equation (7) into equation (6) yields

[0070]

[0071] Among them, I dc It is direct current.

[0072] Furthermore, from equation (8), the ratio of power P to DC current and AC current in the magnetic tunnel junction 30 can be obtained. dc / I th Relationships, such as Figure 2 As shown, Figure 2 The ratio I of power P to DC current and AC current in the magnetic tunnel junction 30 of this application is... dc / I th Relationship diagram.

[0073] Depend on Figure 2It can be seen that the relationship between microwave transmission power and injected DC drive current is exactly similar to that of the activation function of a linear rectifier unit, so it can be regarded as a nonlinear activation unit.

[0074] Furthermore, since information is represented by injected direct current, the microwave transmission frequency must remain constant with the current. Therefore, to truly realize a spin oscillator artificial neuron, developing an oscillator with a fixed transmission frequency is essential. Theoretically, the relationship between microwave transmission frequency and current is as follows:

[0075]

[0076] Among them, f FMR denoted as the intrinsic ferromagnetic resonance frequency of the free layer of the tunnel junction, N is the nonlinear factor, and P is the normalized transmit power.

[0077] From equations (6) and (9), we can obtain

[0078]

[0079] From equation (10), we can see that the nonlinear factor N determines the frequency tunability. Therefore, to achieve the goal of making the transmission frequency independent of the current (i.e., df / dζ = 0), N needs to be 0.

[0080] Furthermore, based on the experimental conditions, the expression for the nonlinear factor N is derived:

[0081]

[0082]

[0083] Where γ is the electron gyrometry.

[0084] From equations (11) and (12), we can see that N is related to factors such as magnetic anisotropy and magnetic field. Magnetic anisotropy includes shape anisotropy, magnetocrystalline anisotropy, and perpendicular magnetic anisotropy. That is, N can be 0 by changing the structure to control the magnetic anisotropy characteristics, so that the microwave transmission frequency is independent of the current, that is, the transmission frequency of the radio frequency signal is independent of the driving signal.

[0085] Because the MgO / CoFeB formed by the capping layer 31 and the magnetic free layer 32 exhibits interfacial anisotropy within its perpendicular magnetic anisotropy, which mainly originates from the hybridization of the d orbitals of Co / Fe and the p orbitals of O, adding a MgO capping layer 31 on the magnetic free layer 32 allows for the control of the interfacial anisotropy by varying the MgO thickness, thereby ensuring that the microwave transmission frequency remains constant regardless of the current.

[0086] In one application scenario, a series of artificial neuron devices with different cap layer 31 thicknesses can be fabricated and power density map tests can be performed. Through data analysis, devices whose output frequency does not change with the input current can be found, thereby determining the optimal thickness of the cap layer 31.

[0087] To prevent inconsistencies in the manufacturing process of the artificial neuron device from affecting the value of the nonlinear factor N, in another embodiment, the artificial neuron device further includes a magnetoelectric coupling device. The magnetoelectric coupling device is used to generate stress in response to a control electrical signal to regulate the magnetic anisotropy of the magnetic free layer 32, thereby ensuring that the magnetic anisotropy of the magnetic free layer 32 is consistent with the design. This allows the cap to regulate the magnetic anisotropy of the magnetic free layer 32 as designed, so that the transmission frequency of the radio frequency signal is independent of the driving electrical signal.

[0088] Specifically, please refer to Figure 3 , Figure 3 This is a schematic diagram of another embodiment of the artificial neuron device based on the magnetic tunnel junction 30 of this application. The magnetoelectric coupling device may include a piezoelectric thin film 50 disposed on the side of the second electrode layer 20 opposite to the magnetic tunnel junction 30. The piezoelectric thin film 50 is used to generate stress according to the input control electrical signal, and the material of the piezoelectric thin film 50 may be a piezoelectric material or a ferroelectric material.

[0089] In this embodiment, the piezoelectric thin film 50 mainly utilizes a strain transfer mechanism to achieve voltage-driven magnetic anisotropy control. By applying a voltage, an electric field is generated and acts on the piezoelectric thin film 50. Based on the inverse piezoelectric effect of piezoelectric or ferroelectric materials, the piezoelectric thin film 50 generates a piezoelectric force, which is then transferred to the magnetic free layer 32 of the magnetic tunnel junction 30 through the interface. Due to the magnetostrictive effect, the magnetic anisotropy of the magnetic free layer 32 is further altered, thereby ensuring that the magnetic anisotropy of the magnetic free layer 32 is the same as the original design.

[0090] In order to enable the magnetic tunnel junction 30 to output high-frequency radio frequency electrical signals, in another embodiment, the artificial neuron device further includes a magnetic metal thin film 60, which is used to provide an external magnetic field to regulate the magnetization direction of the magnetic free layer 32.

[0091] Specifically, please refer to Figure 4 , Figure 4This is a schematic diagram of another embodiment of the artificial neuron device based on the magnetic tunnel junction 30 of this application. A magnetic metal thin film 60 can be arranged between the cap layer 31 and the first electrode layer 10. The magnetic metal thin film 60 can be made of a magnetic metal material, such as Fe, Ni, or Co. The magnetic metal thin film 60 mainly provides an external magnetic field to control the magnetization direction of the magnetic free layer 32, ensuring that the magnetization direction of the magnetic free layer 32 is not collinear with the magnetization direction of the pinning layer 34, thus satisfying the conditions for microwave emission. Furthermore, based on the input DC signal, when the spin torque generated by the spin-polarized current is sufficient to overcome the intrinsic damping torque of the ferromagnetic material, the magnetic moment of the magnetic free layer 32 can undergo high-frequency steady-state precession. This magnetic moment precession will cause periodic changes in magnetoresistance, thereby converting the DC signal into a high-frequency microwave signal output. In this case, the artificial neuron device of this embodiment can be used as a spin-moment nano-oscillator.

[0092] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles applicable herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. An artificial neuron device based on a magnetic tunnel junction, characterized in that, include: A magnetic tunnel junction is used to transmit radio frequency (RF) signals in response to a driving electrical signal. The magnetic tunnel junction includes a cap layer, a magnetic free layer, a barrier layer, a pinning layer, and an antiferromagnetic layer stacked sequentially. The magnetic free layer has magnetic anisotropy along the direction from the barrier layer to the cap layer. The cap layer is used to adjust the magnetic anisotropy of the magnetic free layer, and the thickness of the cap layer is configured such that the transmission frequency of the RF signal is independent of the driving electrical signal. The material of the cap layer is MgO, and the thickness of the cap layer is 0.5–1 nm. The first electrode layer is disposed on one side of the magnetic tunnel junction; A second electrode layer is disposed on the other side of the magnetic tunnel junction. The second electrode layer is used to cooperate with the first electrode layer to input the driving electrical signal to the magnetic tunnel junction.

2. The artificial neuron device according to claim 1, characterized in that, The driving electrical signal is a direct current, which drives the magnetic moment oscillation of the magnetic free layer.

3. The artificial neuron device according to claim 2, characterized in that, The absolute value of the voltage of the driving electrical signal is greater than 0V and less than 10V.

4. The artificial neuron device according to claim 1, characterized in that, The material of the magnetic free layer is one or more combinations of CoFeB, Py, Co, CoFe, FePt, Co / Ni, and FeB.

5. The artificial neuron device according to claim 1, characterized in that, The thickness of the magnetic free layer is 0.1–10 nm.

6. The artificial neuron device according to claim 1, characterized in that, The thickness of the barrier layer is 1–5 nm.

7. The artificial neuron device according to claim 1, characterized in that, It also includes a magnetoelectric coupling device, which is used to generate stress in response to a control electrical signal to modulate the magnetic anisotropy of the magnetic free layer.

8. The artificial neuron device according to claim 7, characterized in that, The magnetoelectric coupling device includes a piezoelectric thin film disposed on the side of the first electrode layer opposite to the magnetic tunnel junction or disposed on the side of the second electrode layer opposite to the magnetic tunnel junction.

9. The artificial neuron device according to claim 1, characterized in that, It also includes a magnetic metal layer disposed on the side of the cap layer opposite to the magnetic free layer, and the magnetic metal layer is used to provide an external magnetic field to control the magnetization direction of the magnetic free layer.

10. The application of a magnetic tunnel junction-based artificial neuron device as described in any one of claims 1 to 9 in a radio frequency-based artificial neural network.

Images