Random nerve component structure without external magnetic field and application thereof

Abstract

The invention discloses a random nerve component structure without an external magnetic field and application thereof. The device structure comprises an antiferromagnetic lamination layer, a magnetic reference layer, a barrier layer and a magnetic tunnel junction free layer. The magnetic moment direction of the magnetic reference layer is in-plane or vertical orientation, and the magnetic moment direction of the magnetic tunnel junction free layer is inclined orientation; an antiferromagnetic pinning structure fixes a magnetic moment maintaining in-plane orientation of the magnetic reference layer. The single magnetic tunnel junction device provided by the invention can realize corresponding functions, is small in size, and is favorable for realizing high-density integration; corresponding functions can be realized without a magnetic field, a magnetic field generating device does not need to be designed, and low power consumption and high expansibility are realized; the probability of random neuron flipping is modulated by external voltage, and the neuron has adjustability and can be used for simulating the working characteristics of random neuron, the preparation process can be compatible with the CMOS process, and the neuron is expected to serve as a basic unit of a neuromorphic chip to construct a hardware neural network to realize large-scale application.

Description

Technical Field

[0001] The invention relates to the technical field of spin electronics, and in particular to a random neuron device structure without the need for an external magnetic field and its application. Background Art

[0002] With the development of artificial intelligence and big data science, in order to complete more complex classification tasks, the number of neurons and layers of deep neural networks has greatly increased, and the computing power requirements for computing platforms such as GPUs have gradually increased, and the hardware power consumption has increased to the order of 100W. In recent years, small-volume, low-power hardware computing platforms such as ASIC and FPGA have been used for the hardware implementation of neural network models. However, these hardware platforms are based on simple digital-analog circuit logic calculations, with low computing power, and it is difficult to complete high-precision floating-point calculations of deep neural networks. Therefore, a model of random computing neural network was created, which binarized high-precision floating-point data into logical numbers composed of 0 / 1, and replaced high-precision floating-point operations with simple logical bit operations, reducing hardware requirements and system power consumption.

[0003] The core of stochastic computational neural network is stochastic computational neurons. Neurons can realize different nonlinear activation functions to enable the neural network built by them to realize various functions in a targeted manner according to different needs. The current research on stochastic computational neural network mainly uses the finite state machine method to realize stochastic computational neurons based on CMOS circuits, but it requires the use of multiple transistors to simulate the function of a single nonlinear neuron, which is not conducive to improving the system integration.

[0004] Random magnetic tunnel junctions have the characteristics of low power consumption, small size and rich dynamic characteristics. Only a single device is needed to realize the function of random computing neurons, and rich research results have been achieved in the fields of neuromorphic computing and random number generators. At present, in the research on neurons based on random tunnel junctions, magnetic tunnel junctions driven by spin-orbit current can realize random neuron devices that encode information with current, but a certain external magnetic field is still required to achieve symmetry breaking of the ferromagnetic / non-magnetic material interface. The external magnetic field not only brings additional burdens in device design and circuit design, but also interferes with the device itself due to the existence of the magnetic field; in the in-plane magnetized tunnel junction device, the current-controlled zero-field random flip is realized by the spin transfer torque effect, but there is a long flip relaxation time (up to 20ms) and low information encoding efficiency, which is not conducive to the realization of efficient neuromorphic computing.

[0005] Therefore, there is an urgent need in the field to develop a high-efficiency random tunnel junction device that does not require an external magnetic field. Summary of the invention

[0006] In view of the deficiencies of the prior art, the object of the present invention is to provide a random neuron device structure and application thereof that does not require an external magnetic field.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention includes:

[0008] In a first aspect, the present invention provides a random neuron device structure that does not require an external magnetic field, which includes an antiferromagnetic stack, a magnetic reference layer, a barrier layer, and a magnetic tunnel junction free layer stacked in a specified direction;

[0009] When no deflection voltage is applied, the magnetic moment direction of the magnetic reference layer is fixed to an in-plane orientation or a vertical orientation, and the magnetic moment direction of the magnetic tunnel junction free layer is an inclined orientation; the antiferromagnetic stack and the magnetic reference layer form an antiferromagnetic pinning structure to fix the magnetic moment direction of the magnetic reference layer, and when the deflection voltage is applied, no significant change in the magnetic moment direction occurs, and the in-plane orientation is still maintained.

[0010] In a second aspect, the present invention further provides a random signal generation method based on the above-mentioned random neuron device structure, which comprises:

[0011] Applying a deflection voltage to the random neuron device structure, and detecting a change in the tunneling resistance of the random neuron device structure to distinguish whether the random neuron device structure is in a low resistance state or a high resistance state, and using the low resistance state and the high resistance state to represent a random signal;

[0012] The deflection voltage is adjusted to regulate the probability of random switching of the random neuron device structure between the low resistance state and the high resistance state.

[0013] In a third aspect, the present invention also provides an application of the above-mentioned random neuron device structure in a random computing neural network.

[0014] Based on the above technical solution, compared with the prior art, the beneficial effects of the present invention at least include:

[0015] The advantages of the random neuron device structure provided by the present invention are that a single magnetic tunnel junction device can realize the corresponding function, which greatly saves area compared to other architectures that require multiple devices to realize the corresponding function, and the size of the magnetic tunnel junction device can be as low as a few nanometers, which is conducive to achieving high-density integration; the neuron device in the present invention does not respond when receiving an input below the threshold, and has sparsity; it randomly flips when receiving an input above the threshold, and can realize the corresponding function without a magnetic field, and does not need to design a magnetic field generating device in the circuit, thereby avoiding the additional power consumption caused by the generation of the magnetic field and the limitations and influences on the device and circuit design. Therefore, the pulse neuron device in the present invention has low power consumption and high scalability; the probability of random flipping of neurons is modulated by an external voltage, is adjustable, and can be used to simulate the working characteristics of random neurons, and the preparation process is compatible with CMOS technology, and is expected to be used as the basic unit of neuromorphic chips to construct hardware neural networks and realize large-scale applications.

[0016] The above description is only an overview of the technical solution of the present invention. In order to enable those skilled in the art to more clearly understand the technical means of the present application and implement them according to the contents of the specification, the following is a description of the preferred embodiments of the present invention in conjunction with detailed drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the structure and principle of a random signal generation system with a random neuron device structure provided by a typical implementation case of the present invention;

[0018] Figure 2 It is a schematic diagram of the random flipping control principle of a random neuron device structure provided by a typical implementation case of the present invention.

[0019] Description of reference numerals:

[0020] 10. Top electrode; 20. Bottom electrode; 30. Magnetic tunnel junction;

[0021] 31. Antiferromagnetic stack; 32. Magnetic reference layer; 33. Barrier layer; 34. Magnetic tunnel junction free layer. DETAILED DESCRIPTION

[0022] In view of the deficiencies in the prior art, the inventor of this case has proposed the technical solution of the present invention after long-term research and extensive practice. The technical solution, its implementation process and principle will be further explained as follows.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0024] Before introducing the technical solutions and examples provided by the present invention, in order to facilitate those skilled in the art to understand the technical solutions of the present invention, some terms and concepts involved in the present invention are introduced as follows:

[0025] Spin electrons: a concept in quantum mechanics, which refers to the spin property of electrons. In quantum mechanics, electrons not only have electric charge and orbital motion (i.e., motion around the nucleus), but also have an intrinsic angular momentum, which is called spin. Spin is a basic quantum property of electrons, which determines the behavior of electrons in a magnetic field. In particular, under the action of a magnetic field, the spin of electrons can present two states: spin up (or spin +1 / 2) and spin down (or spin -1 / 2).

[0026] Magnetic Tunnel Junction (MTJ): A device consisting of two layers of magnetic film and an insulating barrier film in the middle. When the magnetization directions of the two magnetic layers are the same, the device will be in a state of low resistance; when the magnetization directions of the two magnetic layers are different, the device will be in a state of high resistance. The resistance of the magnetic film in the device can be changed by applying an external magnetic field or a DC bias current.

[0027] Tunneling magnetoresistance effect: When current is injected into the tunnel junction, the tunneling behavior of electrons in the tunnel junction satisfies the following rule: electrons only tunnel between the spin subbands with the same spin orientation in the two layers of ferromagnetic materials. Therefore, when the magnetization directions of the ferromagnetic layers on both sides of the barrier layer are parallel to each other, the majority and minority spin subbands corresponding to the two have the same spin orientation, and the probability of electrons tunneling between the majority subbands and the minority subbands is relatively high, and the tunnel junction exhibits a low resistance state; when the magnetization directions of the two ferromagnetic layers are arranged in antiparallel, the probability of electrons tunneling between the majority and minority spin subbands becomes smaller, resulting in a higher tunneling resistance, making the tunnel junction exhibit a high resistance state.

[0028] Spin transfer torque effect: When current passes through a ferromagnetic film, the magnetization direction of the ferromagnetic layer will change due to the interaction between the spin electrons in the current and the magnetic moment in the ferromagnetic layer.

[0029] Voltage Control of Magnetic Anisotropy: The VCMA effect refers to the regulation of the magnetic state by applying an external voltage to change the magnetic anisotropy of the ferromagnetic layer in a ferroelectric / ferromagnetic composite material. The physical mechanism of this effect mainly involves the rearrangement of magnetic domains and rotation of magnetic moments in the ferromagnetic layer by the electric field.

[0030] Random flipping characteristics: When the magnetic moments of the two ferromagnetic layers of the magnetic tunnel junction are arranged in parallel or anti-parallel, the device is in the lowest energy state, and the two potential wells are separated by a MgO potential barrier. When the device potential barrier is high, the free layer magnetic moment is bound in a stable state, and the electron energy is low, far less than the barrier height, and cannot jump to the other potential well; when the device potential barrier is lowered to approach room temperature thermal disturbance, the free layer magnetic moment absorbs the energy of the thermal disturbance. Due to the quantum tunneling mechanism, the electrons in the two potential wells have a certain probability of passing through the potential barrier, causing the free layer magnetic moment to continuously jump between the two potential wells, and the device magnetization state continuously switches between the parallel and anti-parallel states, resulting in a continuous random jump resistance state. This phenomenon is called random flipping.

[0031] Stochastic Computing Neural Networks (SCNNs): A neural network algorithm that uses random numbers and probabilities to perform calculations. Stochastic computing neural networks use random bit streams to represent weights and activation values, and the forward propagation and backpropagation of the network can be efficiently implemented through stochastic computing. This approach can simplify hardware implementation in some aspects and improve computational efficiency, especially in low-power and embedded systems.

[0032] Random neurons: Nonlinear neurons in random computing neural networks are usually implemented by introducing nonlinear activation functions. These activation functions (such as Sigmoid, Tanh, ReLU, etc.) can map input signals to output signals and enable neural networks to learn complex nonlinear relationships. For random neurons, the output signal generally corresponds to the probability of a random event.

[0033] Combined with the above basic concepts and definitions, the purpose of the present invention is to make a magnetic device that can realize the function of nonlinear neurons in random computational neural networks. This device can realize the flip probability control under zero magnetic field by voltage regulating the potential barrier of random flipping. This device does not need to be combined with other CMOS devices, and can independently realize the function of random neurons, significantly saving area and reducing power consumption; in addition, the purely electrically regulated random neurons that do not require magnetic fields are conducive to being used to build a scalable random computational neural network; finally, this device can also be integrated into the CMOS system to reduce the production cost of the neuromorphic chip built using this device.

[0034] Based on the above purpose, an embodiment of the present invention first provides a random neuron device structure that does not require an external magnetic field, which includes an antiferromagnetic stack, a magnetic reference layer, a barrier layer and a magnetic tunnel junction free layer stacked along a specified direction; the magnetic moment direction of the magnetic reference layer is fixed to an in-plane orientation or a vertical orientation, and the magnetic moment direction of the magnetic tunnel junction free layer is an inclined orientation; the antiferromagnetic stack and the magnetic reference layer form an antiferromagnetic pinning structure to fix the magnetic moment direction of the magnetic reference layer.

[0035] Some existing technologies have proposed a technical solution that uses a magnetic free layer, a barrier layer, a pinned layer and an antiferromagnetic layer stacked in sequence, wherein the magnetic free layer introduces a magnetic tunnel junction with perpendicular magnetic anisotropy in conjunction with two upper and lower electrodes to realize a random neuron device. However, in this technical solution, a magnetic moment adjustment element needs to be set to realize magnetic moment reversal.

[0036] Although the present invention also contains a magnetic tunnel junction structure of an antiferromagnetic stack, a magnetic reference layer, a barrier layer and a magnetic tunnel junction free layer, the magnetic moment direction of the film layer on both sides of the barrier layer is significantly different from the above-mentioned prior art. This is the key difference that enables the present invention to achieve no external magnetic field and low relaxation time (high efficiency).

[0037] In principle, the biggest difference between the zero magnetic field random flipping neuron device proposed in the present invention and the neuron device provided by the above-mentioned prior art is that the way they achieve random flipping, or the mechanism, is significantly different. The zero magnetic field random flipping neuron proposed in the present invention uses the VCMA effect to reduce the flipping barrier, so that the free layer undergoes continuous random flipping under the action of room temperature thermal disturbance. During the process, a DC voltage is applied, and the voltage is also the effect. The neuron device mentioned in the prior art uses the STT effect to drive the free layer to flip from one end to the other. It is a one-way, one-time flipping behavior, and the free layer does not flip back during this process. This process is achieved by applying a pulse voltage. Since the device resistance is small, the essence of this pulse voltage is to provide a current that can produce an STT effect, so it is the current that produces the effect.

[0038] Therefore, the random flipping process of the random flipping neurons proposed in the present invention is continuous and can be observed in real time and continuously during the process of applying voltage. There is no need to measure whether a flipping occurs after each voltage pulse is applied as in the prior art, and after the measurement, a reverse voltage or magnetic field needs to be applied to reset the free layer.

[0039] Therefore, the free layer used in the zero magnetic field random flip neuron device proposed in the present invention is tilted magnetized. The purpose of tilted magnetization is to make the flip barrier of the free layer relatively lower, which is more conducive to lowering the flip barrier to a position where random flipping can occur through the VCMA effect.

[0040] Regarding the specific tilted magnetization angle, in some embodiments, when the magnetic moment direction of the magnetic reference layer is in-plane oriented, the angle formed by the magnetic moment direction of the magnetic tunnel junction free layer and the plane where the magnetic tunnel junction free layer is located is 5-30°. In practical applications, the angle between the free layer and the reference layer under zero magnetic field can be preferably 5-30°, and the value of the applied voltage when flipping occurs at different angles will change accordingly. As the angle increases, the voltage during flipping is relatively reduced, and the relaxation time is also relatively shorter.

[0041] Regarding other components of the device structure, in some embodiments, the random neuron device structure may further include a first electrode layer disposed on one side of the antiferromagnetic stack and a second electrode layer disposed on one side of the magnetic tunnel junction free layer.

[0042] In some embodiments, the material of the first electrode layer includes any one or a combination of two or more of Ta, Ru, Cr, Mo, Al, Ni, W, Pt, Pb, Cu, Ti, Au, TMDCs, SiC, MoO2, ITO, Ti3C2Tx, and graphene, with a thickness of 5 to 200 nm.

[0043] In some embodiments, the material of the second electrode layer includes any one or a combination of two or more of Ta, Ru, Mo, Pt, Al, W, Cu, Au, Cr, Ti, Ni, Pb, CuN, TMDCs, SiC, MoO2, ITO, Ti3C2Tx, and graphene, with a thickness of 5 to 200 nm.

[0044] Specifically regarding the structure and material of the antiferromagnetic stack, in some embodiments, the antiferromagnetic stack includes an antiferromagnetic layer, a ferromagnetic layer and a non-magnetic layer, the material of the antiferromagnetic layer includes any one of PtMn, FeMn, CrSBr, RuO2, and IrMn, or a combination of two or more, the material of the ferromagnetic layer includes any one of CoFe, FeB, Py, Co, FePt, and CoNi, or a combination of two or more, and the material of the non-magnetic layer includes any one of Ru, Al2O3, and MgO, or a combination of two or more.

[0045] In some embodiments, the thickness of the antiferromagnetic layer, the ferromagnetic layer, or the non-magnetic layer is 0.1-10 nm.

[0046] In some embodiments, the material of the magnetic reference layer includes any one or a combination of two or more of CoFeB, FeB, Py, Co, CoFe, FePt, and CoNi, and the thickness is 0.1 to 10 nm.

[0047] In some embodiments, the material of the barrier layer includes any one of MgO, Al2O3, MgAl2O4, TiO2 or a combination of two or more thereof, and the thickness is 0.5 to 10 nm.

[0048] In some embodiments, the material of the magnetic tunnel junction free layer includes any one or a combination of two or more of CoFeB, FeB, Py, Co, CoFe, FePt, and CoNi, and the thickness is 0.1 to 10 nm.

[0049] As a further application of the above technical solution, a second aspect of an embodiment of the present invention further provides a random signal generation method based on the random neuron device structure provided in any of the above embodiments, which comprises the following steps:

[0050] Applying a deflection voltage to the random neuron device structure, and detecting a change in the tunneling resistance of the random neuron device structure to distinguish whether the random neuron device structure is in a low resistance state or a high resistance state, and using the low resistance state and the high resistance state to represent a random signal;

[0051] The deflection voltage is adjusted to regulate the probability of random switching of the random neuron device structure between the low resistance state and the high resistance state.

[0052] In some embodiments, the above signal generation method may specifically include:

[0053] The random neuron device is placed in a constant current modulation state, a voltage change at both ends of the random neuron device is recorded, and the low resistance state and the high resistance state are distinguished according to the voltage change.

[0054] Furthermore, the third aspect of the embodiment of the present invention also provides the application of the above-mentioned random neuron device structure in a random computing neural network.

[0055] The technical solution of the present invention is further described in detail below through several embodiments and in conjunction with the accompanying drawings. However, the selected embodiments are only used to illustrate the present invention, and do not limit the scope of the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment shows a random neuron device and its random probability control and signal generation system.

[0058] The top electrode 10 adopts a Ta / Ru / Cu / Ta composite metal structure; the magnetic tunnel junction 30 includes:

[0059] Artificial antiferromagnetic stack 31 (multi-layer material is PtMn / Co 70 Fe 30 / Ru);

[0060] The magnetic reference layer 32 (Co 40 Fe 40 B 20 ), the magnetization direction is in-plane oriented, and the artificial antiferromagnetic pinning structure composed of the artificial antiferromagnetic stack 31 and the magnetic reference layer 32 can fix the magnetic moment direction of the magnetic reference layer 32;

[0061] The barrier layer 33 is MgO);

[0062] The magnetic tunnel junction free layer 34 (Co20Fe60B20) has a tilted magnetization magnetic moment;

[0063] Bottom electrode 20 (made of Ta / Ru / Ta / CuN / Ta / CuN / Ta).

[0064] Devices used for voltage-controlled magnetic anisotropy all have relatively large resistance. According to Ohm's law, when the DC bias is constant, the voltage difference between the high and low resistance states of the device under zero magnetic field is much smaller than the voltage applied to the device. Therefore, the constant current method can be used for modulation. The input current is fixed, and the power supply is applied to both ends of the magnetic tunnel junction to generate a relatively stable voltage. The intensity of the voltage can regulate the flipping barrier of the magnetic tunnel junction, thereby achieving the regulation of the probability of neuron flipping.

[0065] The flip probability modulation principle of the above device The function of the random neuron is mainly realized by the magnetic tunnel junction device and the power supply. The magnetic tunnel junction mainly includes a magnetic free layer with variable magnetization direction, a barrier layer, a magnetic reference layer with fixed magnetization direction, and an artificial antiferromagnetic stack; the power supply is a DC power supply, which is used to regulate the perpendicular magnetic anisotropy field of the magnetic tunnel junction.

[0066] A DC voltage signal is injected into the device. The voltage applied across the magnetic tunnel junction changes the orbital hybridization between Fe and O atoms in CoFeB and MgO, thereby achieving electrical regulation of the perpendicular magnetic anisotropy field. The tunnel junction barrier is related to the anisotropy field, so the flipping probability can be regulated by voltage.

[0067] See also Figure 2 The barrier height state of the tunnel junction shown in the figure. When the barrier height is reduced, the magnetic moment of the free layer can cross the barrier under the action of room temperature thermal disturbance and randomly flip at both ends of the barrier; when the barrier height is increased, the magnetic moment tends to be stable.

[0068] Since the height of the potential barrier determines the difficulty of the magnetic moment to randomly flip over the potential barrier, the voltage controls the difficulty of random flipping by adjusting the height of the potential barrier. The difficulty of random flipping can be expressed by the probability of random flipping, which is defined as the proportion of high and low resistance states in a period of time. The high resistance state of the device is defined as the "1" state, and the low resistance state is defined as the "0" state; the time interval between two consecutive flips of the device is defined as the relaxation time. For example, the time period between the device flipping from the low resistance state to the high resistance state and the subsequent flipping from the high resistance state to the low resistance state is defined as the relaxation time τ of the high resistance state. high Similarly, the time between the device flipping from high resistance state to low resistance state and then flipping from low resistance state to high resistance state is defined as the relaxation time τ of low resistance state. low By statistically analyzing the relaxation time distribution at different time intervals, the average value of the relaxation time statistical distribution can be obtained, and the flip probability P of the high-resistance state (low-resistance state) is defined as the ratio of the high-resistance state (low-resistance state) relaxation time to the sum of the high-resistance and low-resistance state relaxation times, as shown in the following three formulas:

[0069]

[0070] Among them, <τ high >、<τ low > is the average relaxation time of the device in the high and low resistance states, respectively, Δ0 is the thermal stability factor under zero magnetic field and zero bias, Hs is the in-plane stray field caused by the magnetic moment of the reference layer, Keff is the energy density of the magnetic anisotropy field, and ξ is the VCMA coefficient. Equations 1.1 to 1.3 show that the external bias can affect the magnetic anisotropy field, thereby regulating the flip barrier and thermal stability factor.

[0071] Example 2

[0072] This embodiment illustrates random neuron devices with different materials and film thicknesses, which are substantially the same as the embodiment 1, except that:

[0073] Adopt and Figure 1 With the same thin film structure, the top electrode layer 10 can be replaced by one or a combination of Cr, Mo, Al, Ni, W, Pt, Pb, Cu, Ti, Au, TMDCs, SiC, MoO2, ITO, Ti3C2Tx, graphene, etc., and the thickness can be adjusted in the range of 5nm to 200nm;

[0074] The antiferromagnetic layer in the antiferromagnetic stack 31 can be replaced by at least one of FeMn, CrSBr, RuO2, and IrMn, the ferromagnetic layer can be replaced by at least one of FeB, Py, Co, CoFe, FePt, or Co / Ni, and the non-magnetic layer can be replaced by at least one of Al2O3 and MgO. The thickness of each of the above film layers can be adjusted in the range of 0.1 nm to 10 nm.

[0075] The magnetic reference layer 32 may be replaced by at least one of FeB, Py, Co, CoFe, FePt or Co / Ni, with a thickness of 0.1 to 10 nm;

[0076] The barrier layer 33 can be replaced by Al2O3 or MgAl2O4 or TiO2, with a thickness of 0.5nm to 10nm;

[0077] The magnetic tunnel junction free layer 34 can be replaced by FeB, Py, Co, CoFe, FePt or Co / Ni or several combinations thereof, and the thickness can be adjusted in the range of 0.1 to 10 nm;

[0078] The bottom electrode layer 20 can be replaced by one or more combinations of Mo, Pt, Al, W, Cu, Au, Cr, Ti, Ni, Pb, TMDCs, SiC, MoO2, ITO, Ti3C2Tx or graphene, with a thickness of 5 to 200 nm.

[0079] By using the above-mentioned different material and thickness combinations to meet the corresponding film layer functions and physical properties, it is possible to achieve efficient random tunnel junction functions with low relaxation time without the need for an external magnetic field.

[0080] Comparative Example 1

[0081] This comparative example is substantially the same as Example 1, except that:

[0082] The magnetic moment of the free layer of the magnetic tunnel junction is adjusted to be in-plane oriented.

[0083] If the free layer is purely in-plane oriented, it will be more difficult to achieve random flipping under VCMA, or in other words, the flipping barrier will be higher when it is purely in-plane, and a larger energy barrier needs to be lowered to make it randomly flipped. It is even possible that even after reaching the breakdown voltage of the device without the assistance of a magnetic field, the barrier has not been lowered to a position where random flipping can occur.

[0084] Comparative Example 2

[0085] This comparative example is substantially the same as Example 1, except that:

[0086] The magnetic moment of the free layer of the magnetic tunnel junction is aligned perpendicular to the in-plane orientation.

[0087] If the free layer is oriented perpendicular to the out-of-plane direction, random flipping controlled by VCMA cannot be achieved, because the component of the free layer in the magnetization direction of the reference layer is 0, and the directions parallel and antiparallel to the reference layer are no longer the lowest energy points. The overall energy state changes, which is not conducive to random flipping.

[0088] In addition to the above implementation cases, regarding the change in the direction of the magnetic moment of the reference layer, based on the film structure design principle of the magnetic tunnel junction, the reference layer is generally not designed to be tilted magnetized. If the magnetic moment of the reference layer is adjusted to be perpendicular to the in-plane orientation, as long as the angle between the orientation of the free layer and the reference layer is adjusted to a corresponding certain range, at this time the free layer uses the magnetization direction upward and downward as the lowest energy point, then in theory, random reversal of zero magnetic field can be achieved, but the corresponding angle range and conditions need to be determined separately, and the present invention will not elaborate on them.

[0089] Based on the above embodiments and comparative examples, it can be clearly seen that the advantage of the random neuron device structure provided by the embodiments of the present invention is that a single magnetic tunnel junction device can realize the corresponding function, which greatly saves area compared to other architectures that require multiple devices to realize the corresponding function, and the size of the magnetic tunnel junction device can be as low as a few nanometers, which is conducive to achieving high-density integration; the neuron device in the present invention does not respond when receiving an input below the threshold, and has sparsity; it randomly flips when receiving an input above the threshold, and can realize the corresponding function without a magnetic field, and does not need to design a magnetic field generating device in the circuit, thereby avoiding the extra power consumption caused by the generation of the magnetic field and the limitations and influences on the device and circuit design. Therefore, the pulse neuron device in the present invention has low power consumption and high scalability; the probability of random flipping of neurons is modulated by an applied voltage, is adjustable, and can be used to simulate the working characteristics of random neurons, and the preparation process is compatible with CMOS technology, and is expected to be used as the basic unit of neuromorphic chips to construct hardware neural networks and achieve large-scale applications.

[0090] It should be understood that the above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A random neuron device structure that does not require an external magnetic field, characterized in that: It includes an antiferromagnetic stack, a magnetic reference layer, a barrier layer and a magnetic tunnel junction free layer stacked in a specified direction; When no deflection voltage is applied, the magnetic moment direction of the magnetic reference layer is in-plane oriented or vertically oriented, and the magnetic moment direction of the magnetic tunnel junction free layer is tilted; the antiferromagnetic stack and the magnetic reference layer form an antiferromagnetic pinning structure to fix the magnetic moment direction of the magnetic reference layer.

2. The random neuron device structure according to claim 1, characterized in that: When the magnetic moment direction of the magnetic reference layer is in-plane oriented, the angle formed by the magnetic moment direction of the magnetic tunnel junction free layer and the plane where the magnetic tunnel junction free layer is located is 5-30°.

3. The random neuron device structure according to claim 1, characterized in that: It also includes a first electrode layer arranged on one side of the antiferromagnetic stack and a second electrode layer arranged on one side of the magnetic tunnel junction free layer.

4. The random neuron device structure according to claim 3, characterized in that: The material of the first electrode layer includes any one or a combination of two or more of Ta, Ru, Cr, Mo, Al, Ni, W, Pt, Pb, Cu, Ti, Au, TMDCs, SiC, MoO2, ITO, Ti3C2Tx, and graphene, and the thickness is 5N200 nm; And / or, the material of the second electrode layer includes any one or a combination of two or more of Ta, Ru, Mo, Pt, Al, W, Cu, Au, Cr, Ti, Ni, Pb, CuN, TMDCs, SiC, MoO2, ITO, Ti3C2Tx, and graphene, with a thickness of 5 to 200 nm.

5. The random neuron device structure according to claim 1, characterized in that: The antiferromagnetic stack includes an antiferromagnetic layer, a ferromagnetic layer and a non-magnetic layer. The material of the antiferromagnetic layer includes any one of PtMn, FeMn, CrSBr, RuO2, and IrMn, or a combination of two or more thereof. The material of the ferromagnetic layer includes any one of CoFe, FeB, Py, Co, FePt, and CoNi, or a combination of two or more thereof. The material of the non-magnetic layer includes any one of Ru, Al2O3, and MgO, or a combination of two or more thereof.

6. The random neuron device structure according to claim 5, characterized in that: The thickness of the antiferromagnetic layer, ferromagnetic layer or non-magnetic layer is 0.1-10 nm.

7. The random neuron device structure according to claim 1, characterized in that: The material of the magnetic reference layer includes any one or a combination of two or more of CoFeB, FeB, Py, Co, CoFe, FePt, and CoNi, and the thickness is 0.1 to 10 nm; And / or, the material of the barrier layer includes any one of MgO, Al2O3, MgAl2O4, TiO2 or a combination of two or more thereof, and the thickness is 0.5 to 10 nm; And / or, the material of the magnetic tunnel junction free layer includes any one or a combination of two or more of CoFeB, FeB, Py, Co, CoFe, FePt, and CoNi, and has a thickness of 0.1 to 10 nm.

8. A random signal generation method based on the random neuron device structure according to any one of claims 1 to 7, characterized in that: include: Applying a deflection voltage to the random neuron device structure, and detecting a change in the tunneling resistance of the random neuron device structure to distinguish whether the random neuron device structure is in a low resistance state or a high resistance state, and using the low resistance state and the high resistance state to represent a random signal; The deflection voltage is adjusted to regulate the probability of random switching of the random neuron device structure between the low resistance state and the high resistance state.

9. The random signal generation method according to claim 8, characterized in that: Specifically include: The random neuron device is placed in a constant current modulation state, a voltage change at both ends of the random neuron device is recorded, and the low resistance state and the high resistance state are distinguished according to the voltage change.

10. Application of the random neuron device structure described in any one of claims 1 to 7 in a random computational neural network.

Images