A neurosynaptic biomimetic device and a biomimetic control method thereof

Abstract

The application discloses a kind of nerve synapse biomimetic devices and its biomimetic control method, belong to microelectronic device technical field, including from bottom to top sequentially distributed lower electrode, resistance change layer, barrier layer and upper electrode;Among them, resistance change layer is resistance change layer with energy level defect, can spontaneously capture and release carrier, and without preforming operation to produce soft breakdown can have stable resistance change characteristics, and will not cause serious damage to device internal structure, can more accurately simulate the function of nerve synapse.Meanwhile it can also inhibit the interference between nerve signals, reduce the influence of non-ideal factors of memristor device in the process of neural network training and inference operation in memristor array.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of microelectronic devices, and more particularly, to a neural synapse bionic device and a bionic control method thereof. BACKGROUND

[0002] The development of information technology towards big data requires highly dense storage arrays and faster electronic devices. However, the traditional von Neumann computing architecture has performance limitations and high power consumption due to the separation of memory and processor, so a new computing architecture that integrates computing and storage is urgently needed. Inspired by the human brain, data storage and processing are both completed in synapses and neurons, which perform complex functions such as perception, learning, and memory. The neuron is the basic structural unit of the neural network, and the synapse is the basis for the connection between neurons. The change in the morphology, function, and efficiency of the synapse, i.e. synaptic plasticity, plays a crucial role in the information transmission, processing, and storage between neurons in the neural network, and it is also an important basis for the learning ability of the human brain. Therefore, it is of great significance to study a synaptic bionic device and the corresponding bionic control method.

[0003] Traditional electronic devices used for neuromorphic computing have large area and power overheads. In recent years, the study of memristors has promoted the application in this regard. The memristor is a simple two-terminal device, which has the characteristics of low power consumption, small feature size 4F 2 , fast read-write speed, and compatibility with CMOS, compared with traditional electronic devices. Since the resistance value of the memristor is determined by the charge flowing through it, it has memory properties, which are very similar to biological synapses. For example, the endocytosis and exocytosis of biological calcium ions are similar to the diffusion dynamics of metal cations or oxygen vacancies, or the carrier trapping and de-trapping processes of interface-type memristors; therefore, in the prior art, the gradual I-V characteristic curve of the memristor is often adjusted to obtain multiple conductance values, thereby realizing synaptic bionics.

[0004] However, most of the existing memristors used for synaptic bionics need to be formed to produce soft breakdown and form a pre-conductive channel. Due to the influence of the internal structure of the material and the external operating conditions on the filament formation morphology, the device characteristics often exhibit inconsistent characteristics, which can seriously affect the accuracy of neuromorphic computing and reduce the bionic accuracy. In addition, one of the important requirements for building artificial neural networks is to realize a high-density synaptic array, and the memristor devices used for synaptic bionics often adopt a crossbar array structure. However, the crossbar array has the problem of electrical crosstalk, which often causes hidden currents in the process of training and reasoning the synapses in the synaptic array, leading to incorrect recognition results and greatly limiting its development. SUMMARY

[0005] In order to solve the above problems, the present application provides a neural synapse biomimetic device and a biomimetic control method thereof, which can solve the problem of low biomimetic precision caused by soft breakdown of the memristor due to pre-forming.

[0006] In order to achieve the above-mentioned purpose, in a first aspect, the present application provides a neural synapse biomimetic device, comprising: a lower electrode, a resistive switching layer, a blocking layer and an upper electrode distributed in sequence from bottom to top.

[0007] Preferably, the resistive switching layer is a resistive switching layer with energy level defects; the blocking layer is used for self-limiting current to prevent the resistive switching layer from being broken down.

[0008] Further preferably, the activation process of the neural synapse biomimetic device is realized by an activation module.

[0009] Preferably, the activation module is used for activating the neural synapse biomimetic device to make the neural synapse biomimetic device have stable resistive switching characteristics, comprising:

[0010] The lower electrode is grounded, and a limiting current is set; a first direct current scan and a second direct current scan are sequentially applied to the upper electrode, and after repeating a preset number of times, the direct current scan is stopped; wherein the voltage polarity of the first direct current scan and the second direct current scan is opposite.

[0011] Further preferably, the resistive switching layer is prepared by a magnetron sputtering method, an atomic deposition method, a pulsed laser deposition method, a chemical vapor deposition method or a thermal oxidation method.

[0012] Further preferably, the thickness of the resistive switching layer is 1-100 nm. The thickness of the blocking layer is 1-100 nm. Based on the following formula:

[0013]

[0014]

[0015] Wherein, And are the bias voltages distributed to the resistive switching layer and the blocking layer, respectively. And are the dielectric constants of the resistive switching layer and the blocking layer, respectively. is the voltage applied to the neural synapse biomimetic device.

[0016] Further preferably, the Schottky barrier between the lower electrode and the resistance change layer is lower than the Schottky barrier between the upper electrode and the resistance change layer; the material of the lower electrode is a low work function metal material or a single crystal material, including Al, Mg, Ti, TiN, Zn, In, Li, Ta or Zr; the material of the upper electrode is a high work function metal material or a single crystal material, including Cr, Ni, C, Au, Pt, Co, Pd or Si.

[0017] Further preferably, the resistance change layer is also used to block carrier tunneling.

[0018] In a second aspect, the present application provides a biomimetic control method based on the neural synapse biomimetic device provided in the first aspect of the present application, comprising:

[0019] After the neural synapse biomimetic device is activated, the lower electrode is grounded, and a set pulse is continuously applied to the upper electrode within a first preset time period to stimulate the neural synapse biomimetic device, thereby realizing the biomimicry of the learning function of the neural synapse.

[0020] The method for activating the neural synapse biomimetic device comprises grounding the lower electrode, setting a limit current, and sequentially applying a first direct current scan and a second direct current scan to the upper electrode, and after repeating a preset number of times, stopping applying the direct current scan; wherein the voltage polarity of the first direct current scan and the second direct current scan is opposite.

[0021] Further preferably, after realizing the learning function of the neural synapse, the set pulse is stopped being applied to the upper electrode within a preset second preset time period to biomimic the forgetting function of the neural synapse; wherein the second preset time period is less than the first preset time period.

[0022] Further preferably, the learning degree of the neural synapse biomimetic device is regulated by changing the size of the first preset time period.

[0023] The forgetting degree of the neural synapse biomimetic device is regulated by changing the size of the second preset time period.

[0024] The forgetting rate of the neural synapse biomimetic device is regulated by changing the parameters of the set pulse; wherein the parameters of the set pulse include the duty cycle, the amplitude and the pulse width of the set pulse.

[0025] In a third aspect, the present application provides a biomimetic control method based on the neural synapse biomimetic device provided in the first aspect of the present application, comprising:

[0026] After the neural synapse biomimetic device is activated and is in a high resistance state, the lower electrode is grounded, and a forward direct current scan voltage or a set pulse is continuously applied to the upper electrode to change the state of the neural synapse biomimetic device to a low resistance state, thereby biomimic the strengthening function of the neural synapse.

[0027] The method for activating the nerve synapse biomimetic device comprises: grounding the lower electrode, setting a limit current; sequentially applying a first direct current scan and a second direct current scan on the upper electrode, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarities of the first direct current scan and the second direct current scan are opposite.

[0028] In a fourth aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, comprising:

[0029] For the nerve synapse biomimetic device in a low resistance state after activation, the lower electrode is grounded, and a negative direct current scan voltage or a reset pulse is continuously applied on the upper electrode to change the state of the nerve synapse biomimetic device to a high resistance state, so as to mimic the inhibitory function of the nerve synapse.

[0030] The method for activating the nerve synapse biomimetic device comprises: grounding the lower electrode, setting a limit current; sequentially applying a first direct current scan and a second direct current scan on the upper electrode, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarities of the first direct current scan and the second direct current scan are opposite.

[0031] In a fifth aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, comprising:

[0032] S11, for the activated nerve synapse biomimetic device, the initial resistance value is adjusted to an intermediate resistance value;

[0033] S12, a first continuous pulse is applied on the upper electrode of the nerve synapse biomimetic device, and a second continuous pulse same as the first continuous pulse is applied on the lower electrode; wherein the interval time of applying the first continuous pulse and the second continuous pulse is ΔT;

[0034] S13, the interval time ΔT is changed;

[0035] S14, steps S12-S13 are repeated to mimic the STDP function of the nerve synapse.

[0036] The method for activating the nerve synapse biomimetic device comprises: grounding the lower electrode, setting a limit current; sequentially applying a first direct current scan and a second direct current scan on the upper electrode, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarities of the first direct current scan and the second direct current scan are opposite.

[0037] In a sixth aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, comprising:

[0038] The following steps are performed on the activated and high resistance state neural synapse biomimetic device:

[0039] S21, a first pulse is applied on the upper electrode of the neural synapse biomimetic device, and after a time interval T, a second pulse identical to the first pulse is applied again;

[0040] S22, a negative direct current scan is applied on the upper electrode, so that the neural synapse biomimetic device returns to the high resistance state;

[0041] S23, change the time interval T;

[0042] S24, repeat steps S21-S23 to simulate the double-peak pulse facilitation function of the neural synapse;

[0043] The method for activating the neural synapse biomimetic device comprises grounding the lower electrode thereof and setting a limit current; a first direct current scan and a second direct current scan are sequentially applied on the upper electrode thereof, and after being repeated for a preset number of times, the direct current scan is stopped; wherein the voltage polarities of the first direct current scan and the second direct current scan are opposite.

[0044] In a sixth aspect, the present application provides an artificial neural network comprising a plurality of neurons connected to each other through a synaptic device array; wherein the synaptic device array is a cross array structure; and the synaptic device in the synaptic device array is the neural synapse biomimetic device provided in the first aspect of the present application.

[0045] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0046] 1. The present application provides a neural synapse biomimetic device, comprising a lower electrode, a resistance variable layer, a barrier layer and an upper electrode distributed in order from bottom to top; wherein the resistance variable layer is a resistance variable layer with energy level defects, and there are a large number of energy level defects that can act as defect sites for electron capture and de-trapping to capture and release carriers, which is an attribute of the neural synapse biomimetic device itself, and does not need to be formed into a conductive channel to have stable resistance variable characteristics, and will not cause serious damage to the internal structure of the device, and can more accurately simulate the function of neural synapse.

[0047] 2. The neural synapse biomimetic device provided by the present application has a lower potential barrier between the lower electrode and the resistance variable layer than the potential barrier between the upper electrode and the barrier layer, and can achieve a rectification ratio of 10 4 The neural synapse biomimetic device has a self-rectifying characteristic, and when used to construct an artificial neural network, it can effectively reduce the electrical crosstalk problem caused by other current paths passing through the target unit when inferring and training the weights of the synaptic array.

[0048] 3、The nerve synapse biomimetic device provided by the application can accurately realize various nerve synapse functions, including nerve synapse excitation and inhibition, double-peak pulse facilitation, STDP and complex learning and forgetting processes, and the above-mentioned synapse functions are realized based on the capture and de-capture processes of the electronic carriers under deep and shallow energy levels. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A structural schematic diagram of the nerve synapse biomimetic device provided by the application;

[0050] Figure 2 A resistance change curve diagram of the nerve synapse biomimetic device provided by the application;

[0051] Figure 3 Resistance change curve diagrams of HfO x / Al2O3 devices with different thicknesses provided by the application;

[0052] Figure 4 A voltage / current-time relationship diagram when simulating the learning and forgetting functions of a synapse through 10 stimulation stages and intermittent stages provided by the application;

[0053] Figure 5 Energy band diagrams of the nerve synapse biomimetic device under Set operation and the nerve synapse biomimetic device under 0 bias voltage provided by the application; wherein (a) is an energy band diagram of the nerve synapse biomimetic device under Set operation; (b) is an energy band diagram of the nerve synapse biomimetic device under 0 bias voltage;

[0054] Figure 6 A current-pulse number relationship diagram when simulating the potentiation and inhibition functions of a synapse by applying positive and negative pulse stimulation to the nerve synapse biomimetic device provided by the application. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0056] In order to achieve the above-mentioned purpose, in a first aspect, the present application provides a nerve synapse biomimetic device, as shown in Figure 1 including, from bottom to top, a lower electrode 1, a resistance change layer 2, a barrier layer 3 and an upper electrode 4;

[0057] The resistance change layer is a resistance change layer with energy level defects, which can be prepared by a magnetron sputtering method, an atomic deposition method, a pulsed laser deposition method, a chemical vapor deposition method, a thermal oxidation method or the like, has a large number of defect energy levels including deep defect energy levels and shallow defect energy levels, and serves as a carrier trapping layer and a de-trapping layer, and the resistance change is based on the trapping and de-trapping processes of electrons. The material of the resistance change layer can be HfO x , TaO x , TiO x , ZnO x , etc.

[0058] The barrier layer is used for self-limiting, acts as a series resistor, prevents the device from being broken down (hard breakdown or soft breakdown) due to excessive current, avoids the forming process, and greatly improves the durability of the device. The barrier layer can be prepared by an atomic deposition method, a plasma enhanced atomic layer deposition (PEALD), a PECVD or the like, has a small number of defect energy levels, and in addition to the self-limiting effect, can also be used for blocking the migration of oxygen vacancies to improve the retention characteristics of the device (because the drift of oxygen vacancies will cause the resistance to drift), and blocking the tunneling of carriers to form a self-rectifying characteristic (because if the reverse tunneling occurs, the reverse current will be large, and the rectification effect will fail). Further, the material of the barrier layer can be Al2O3, NbO2, TiO2, ZrO2, SiO2 or the like.

[0059] In order to redistribute and partially generate the energy level defects and achieve a stable resistance change state, the neural synapse biomimetic device needs to be activated before use. The neural synapse biomimetic device provided by the application has a forming-free characteristic, and the activation process is realized by using an activation module. Specifically, the activation module is used to activate the neural synapse biomimetic device to make the neural synapse biomimetic device have stable resistance change characteristics, and specifically includes: grounding the lower electrode and setting a limit current; sequentially applying a first direct current scan and a second direct current scan on the upper electrode, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarity of the first direct current scan and the second direct current scan is opposite. Specifically, as shown in Figure 2 , in an optional embodiment, +7V direct current voltage is applied to the upper electrode, and then -7V direct current voltage is applied, and 10-20 cycles of circulation are performed to achieve a stable rectification ratio of about 10 4 . It should be noted that the neural synapse biomimetic device provided by the application has low operating current, and the corresponding power consumption is also low. Specifically, according to the I-V curve, the operating current of the application is in the order of μA, and the power consumption is 6 orders of magnitude lower than that of most existing neural synapse biomimetic devices based on memristors. Figure 2

[0060] ​It should be noted that, since the existing soft-breakdown required memristor needs to apply a large voltage (forming operation) to make the device generate oxygen vacancy filament channel to obtain stable resistance change characteristics, and the external operation is often uncontrollable, the filament formation morphology often shows inconsistent characteristics in the device characteristics under the influence of the internal structure of the material and the external operation conditions. In the nerve synapse biomimetic device proposed in the present application, taking hafnium oxide resistance layer as an example, a large number of oxygen vacancies exist in the resistance layer obtained by the above preparation process (for example, the number of oxygen vacancies is controlled by the flow ratio of argon and oxygen gas by magnetron sputtering method, generally, the less the oxygen content, the more the oxygen defects); these oxygen defects can be used as electron capture and de-trapping defect sites, which are the properties of the nerve synapse biomimetic device itself, and do not need to form a conductive channel, so there is no need to perform forming operation to obtain stable resistance change characteristics as in the existing above-mentioned memristor, and will not cause serious damage to the internal structure of the device.

[0061] Further, in order to suppress the electrical crosstalk problem existing when the above-mentioned nerve synapse biomimetic device is used to construct an artificial neural network, in an optional embodiment, the above-mentioned nerve synapse biomimetic device is designed as: the Schottky barrier of the lower electrode and the resistance layer is lower than the Schottky barrier of the upper electrode and the barrier layer, so that the current flows asymmetrically, so that the nerve synapse biomimetic device has a self-rectifying characteristic, and further suppresses the crosstalk between the nerve synapse biomimetic devices. Preferably, the material of the lower electrode is a low work function metal material or a single crystal material, including: Al, Mg, Ti, TiN, Zn, In, Li, Ta or Zr; the material of the upper electrode is a high work function metal material or a single crystal material, including: Cr, Ni, C, Au, Pt, Co, Pd or Si. The barrier layer is also used to block carrier tunneling.

[0062] Further, in order to make the resistance change effect (forming-free characteristic) of the nerve synapse biomimetic device more stable, and the self-rectifying characteristic more obvious, it is necessary to reasonably distribute the appropriate bias voltage to the resistance layer and the barrier layer. If the bias voltage distributed to the barrier layer is too small, the carrier is easy to tunnel, which reduces the self-rectifying performance of the nerve synapse biomimetic device. If the bias voltage distributed to the resistance layer is too small, on the one hand, the electrons may not have enough energy to overcome the TiN / HfO xThe barrier enters the defect-filling energy level, but there is not enough energy for oxygen vacancy defects to be generated and redistributed to penetrate the channel between the upper and lower electrodes, thus preventing significant resistive switching. If the bias voltage allocated to the resistive switching layer is too large, the resistive switching layer is prone to forming conductive filaments, causing breakdown. Therefore, when designing the barrier layer and the resistive switching layer, it is necessary to consider that they can be allocated an appropriate bias voltage. This invention derives the relationship between the bias voltage of the barrier layer and the resistive switching layer and their thickness by treating the barrier layer and the resistive switching layer as equivalent to a series capacitor structure and dividing the voltage. Thus, by designing the resistive switching layer and the barrier layer with appropriate thickness matching, they can be allocated an appropriate bias voltage. This not only achieves more obvious self-rectification but also allows an appropriate bias voltage to be allocated to the resistive switching layer, making the resistive switching characteristics of the neural synapse bionic device more stable. Specifically, in one optional embodiment, the thickness of the resistive switching layer is... and the thickness of the barrier layer Calculated based on the following formula:

[0063]

[0064]

[0065] in, and These are the bias voltages allocated to the resistive switching layer and the blocking layer, respectively; and These are the dielectric constants of the resistive switching layer and the barrier layer, respectively. This refers to the voltage applied to the bionic device at the neural synapse. This is determined by... and The range of values ​​is used to determine the thickness of the resistive switching layer. and the thickness of the barrier layer The range.

[0066] It should be noted that for the resistive switching effect, it is necessary to ensure that the bias voltage allocated to the resistive switching layer is appropriate to avoid breakdown or insufficient current drive. This is because areas with concentrated electric fields generally have a large number of defects, and the current transport mechanism may involve electrons jumping and conducting within the defects. A large drive current can cause the defects to redistribute and generate. For the self-rectification characteristic, a thin barrier layer results in a large reverse tunneling current and an insignificant rectification effect; a thick barrier layer leads to a small bias voltage allocated to the resistive switching layer, resulting in an insufficient electric field for significant resistive switching behavior (the voltage may be too low to cause defect generation and redistribution, and these defects may penetrate the upper and lower electrodes, causing electrons to jump and conduct within them), a small forward current, and an insignificant resistive switching effect. Based on the above principles, this invention determines the appropriate bias voltage values ​​allocated to the resistive switching layer and the barrier layer, and then determines the appropriate thickness of the resistive switching layer and the barrier layer based on the above relationship, so as to ensure the stability of the resistive switching effect and the significant self-rectification characteristic of the neural synapse bionic device.

[0067] Specifically, the resistive switching layer is hafnium oxide (HfO).x For example, the nerve synapse biomimetic device with the barrier layer of Al2O3, the resistive switching curves of HfO2 / Al2O3 devices with different thicknesses are shown in FIG. 1, preferably, the resistive switching layer is about 8-12 nm, and the barrier layer is about 4-6 nm, so that the resistive switching effect and self-rectifying characteristics of the nerve synapse biomimetic device are more stable. x Figure 3

[0068] Further, in an optional embodiment, the hafnium oxide resistive switching layer with a thickness of 8-12 nm is prepared by a magnetron sputtering method; specifically, the cavity is pumped to 1×10 -3 ~3×10 -3 Pa, argon gas with a flow rate of 50 sccm and oxygen gas with a flow rate of 10 sccm are introduced into the cavity, then the pressure is fixed at 0.67 Pa, the Hf target sputtering power is controlled to be 90-110 W, the pre-sputtering time is 2 min, and the sputtering time is 8 min.

[0069] In an optional embodiment, the aluminum oxide barrier layer with a thickness of 4-6 nm is prepared by an atomic deposition method; specifically, the first precursor water vapor is introduced at a deposition temperature of 300°C, then argon gas is introduced for purging, the second precursor trimethylaluminum is introduced, and then argon gas is introduced for purging, which is one cycle, and a total of 38 to 57 cycles are performed.

[0070] In an optional embodiment, the TiN lower electrode is prepared by a magnetron sputtering method; specifically, the magnetron sputtering cavity is pumped to 8×10 -5 ~1×10 -4 Pa, argon gas with a flow rate of 40 sccm is introduced into the cavity, the butterfly valve is adjusted to control the cavity vacuum to be 0.4-0.6 Pa, the TiN target sputtering power is controlled to be 40 W, and the sputtering time is 10-20 min.

[0071] In an optional embodiment, the Pt upper electrode is prepared by a magnetron sputtering method; specifically, the magnetron sputtering cavity is pumped to 8×10 -5 ~1×10 -4 Pa, argon gas with a flow rate of 40 sccm is introduced into the cavity, the butterfly valve is adjusted to control the cavity vacuum to be 0.4-0.6 Pa, the Pt target sputtering power is controlled to be 30 W, and the sputtering time is 10-20 min.

[0072] In an optional embodiment, the substrate of the nerve synapse biomimetic device described above is a Si substrate with a SiO2 thickness of about 1 μm.

[0073] ​​In summary, the memory resistor with forming-free characteristics is constructed as a neural synapse biomimetic device, and the inherent defects of the memory resistor make the difference between device and device, cell and cell, and cycle and cycle small. In addition, the neural synapse biomimetic device has the advantages of gradual I-V curve, self-limiting current characteristic, and low operating current.

[0074] It should be noted that, since the memory resistor with forming-free characteristics does not need to be pre-formed to produce soft breakdown, the electrical characteristics of the memory resistor can be used to more accurately simulate the functions of synapse enhancement and inhibition, spike timing-dependent plasticity STDP, double-peak pulse facilitation PPF, learning and forgetting, and can also suppress the influence of other synapse devices on the target synapse device (without using a multi-integrated gating tube such as a transistor to solve the problem of crosstalk).

[0075] In a second aspect, the application provides a biomimetic control method based on the neural synapse biomimetic device provided in the first aspect of the application, which is used to biomimic the learning and forgetting functions of the neural synapse, and specifically includes:

[0076] After the neural synapse biomimetic device is activated, the lower electrode is grounded, and a set pulse is continuously applied to the upper electrode within a first preset time period to stimulate the neural synapse biomimetic device, thereby realizing the biomimic of the learning function of the neural synapse.

[0077] After realizing the learning function of the neural synapse, the set pulse is stopped being applied to the upper electrode within a preset second preset time period to biomimic the forgetting function of the neural synapse; wherein the second preset time period is less than the first preset time period.

[0078] The method for activating the neural synapse biomimetic device includes grounding the lower electrode, setting a limiting current, and sequentially applying a first direct current scan and a second direct current scan to the upper electrode, and after repeating a preset number of times, stopping applying the direct current scan; wherein the voltage polarity of the first direct current scan and the second direct current scan is opposite.

[0079] It should be noted that the learning degree of the neural synapse biomimetic device is regulated by changing the size of the first preset time period; the forgetting degree of the neural synapse biomimetic device is regulated by changing the size of the second preset time period; and the forgetting rate of the neural synapse biomimetic device is regulated by changing the parameters of the set pulse; wherein the parameters of the set pulse include the duty cycle, amplitude, and pulse width of the set pulse.

[0080] Specifically, in an optional embodiment, the upper electrode is made of Pt material with high metal work function, the lower electrode is made of TiN material with low metal work function, the barrier layer is made of Al2O3 material, and the resistive switching layer is made of HfO xThe nerve synapse biomimetic device of the material is taken as an example, such as Figure 4 Fig. 10 shows a schematic diagram of the learning and forgetting process of the simulated synapse composed of 10 stimulation stages and intermittent stages; wherein the stimulation stage is composed of 20 pulses with a pulse width of 4 ms and a pulse amplitude of 8 V, and the intermittent time is about 100 ms. Specifically, the learning and forgetting process of the simulated synapse performs the following steps:

[0081] (a) applying a direct current scanning voltage of +7 V and -7 V on the upper electrode of the nerve synapse biomimetic device, and grounding the lower electrode, performing several cycles to achieve stable resistance change characteristics, so as to activate the nerve synapse biomimetic device;

[0082] (b) continuously applying 50 set pulses with a pulse width of 4 ms and a pulse amplitude of 8 V on the upper electrode, and the interval between two pulses is 6 ms, which is the stimulation stage;

[0083] (c) no voltage is applied for about 100 ms, which is the intermittent stage;

[0084] (d) repeating steps (b) - (c) 10 times;

[0085] Wherein, the stimulation stage is the learning process, and the intermittent stage is the forgetting process. Specifically, the energy band diagram of the nerve synapse biomimetic device under set operation and the energy band diagram of the nerve synapse biomimetic device under 0 bias are shown in Figure 5 Fig. (a) is the energy band diagram of the nerve synapse biomimetic device under set operation, corresponding to the stimulation stage, that is, the electrons cross the low TiN / HfO x potential barrier and enter the deep and shallow defect energy levels; Fig. (b) is the energy band diagram of the nerve synapse biomimetic device under 0 bias, corresponding to the intermittent stage, that is, the carrier de-trapping process. Specifically, as shown in Figure 5 In the stimulation stage, the carriers fill the deep and shallow energy levels, and at this time, it is in a low resistance state; in the intermittent stage, the Fermi level is flattened with the electrode potential, and the carriers in the shallow defect energy level higher than the electrode potential or the defect energy level excited by heat flow away from the lower electrode side, resulting in a spontaneous de-trapping process, and the deep energy level can firmly trap the carriers. If a negative direct current voltage or a pulse voltage is applied, the high metal work function between the upper electrode and the barrier layer blocks the flow of carriers, but the defect-trapped carriers of the resistance change layer can flow away from the lower electrode side, and the defect is not filled with carriers, at this time, it is in a high resistance state. In the pulse test, by controlling the change of the stimulation frequency or the pulse parameters, the control of the forgetting rate of the nerve synapse is realized.

[0086] In a third aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, which is used to biomimic the enhanced function of the nerve synapse, and specifically comprises:

[0087] For the activated nerve synapse biomimetic device in the high resistance state, the lower electrode is grounded, and the positive direct current scanning voltage or set pulse is continuously applied on the upper electrode to change the state to the low resistance state, so as to simulate the strengthening function of the nerve synapse.

[0088] The method for activating the nerve synapse biomimetic device comprises grounding the lower electrode, setting a limit current, and sequentially applying a first direct current scanning and a second direct current scanning on the upper electrode, and stopping applying the direct current scanning after repeating a preset number of times.

[0089] In the fourth aspect, the application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the application, which is used to simulate the inhibitory function of the nerve synapse, and specifically comprises:

[0090] For the activated nerve synapse biomimetic device in the low resistance state, the lower electrode is grounded, and the negative direct current scanning voltage or reset pulse is continuously applied on the upper electrode to change the state to the high resistance state, so as to simulate the inhibitory function of the nerve synapse.

[0091] The method for activating the nerve synapse biomimetic device comprises grounding the lower electrode, setting a limit current, and sequentially applying a first direct current scanning and a second direct current scanning on the upper electrode, and stopping applying the direct current scanning after repeating a preset number of times.

[0092] For the biomimetic control method of the nerve synapse biomimetic device provided in the third aspect and the fourth aspect, in an optional embodiment, the strengthening and inhibiting processes of the synapse are simulated by applying the direct current scanning voltage. Specifically, in the direct current test, 50 +6V direct current scanning voltages are continuously applied on the upper electrode, and the conductance value continuously increases, and then 50-6V negative scanning voltages are continuously applied, and the conductance value continuously decreases, and the conductance can be mapped as the synaptic weight to simulate the strengthening and inhibiting processes of the synapse.

[0093] In another optional embodiment, the strengthening and inhibiting processes of the synapse are simulated by applying the pulse. Specifically, in the pulse test, the conductance size is controlled by controlling the number of pulses and the pulse parameters. Figure 6Fig. 6 shows a schematic diagram of applying positive and negative pulse stimuli to the nerve synapse biomimetic device to simulate the potentiation and depression characteristics of the synapse; wherein the pulse width is 30 ms, the pulse amplitude is 5.5 V and 6 V respectively, and the read voltage is 3 V. Specifically, with fixed pulse parameters, 50 positive programming pulses with a pulse width of 30 ms and a pulse amplitude of 5.5 V are continuously applied, and 50 conductance values are read by a 3 V read pulse, followed by 50 negative programming pulses with a pulse width of 30 ms and a pulse amplitude of -5.5 V, and 50 conductance values are also read by a 3 V read voltage, to achieve pulse conductance modulation and simulate the potentiation and depression processes of the synapse. Similarly, the pulse amplitude is changed to 6 V, and the above steps are re-executed.

[0094] In a fifth aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, for simulating the STDP function of the nerve synapse, specifically comprising:

[0095] S11, adjusting the initial resistance of the activated nerve synapse biomimetic device to an intermediate resistance value;

[0096] S12, applying a first continuous pulse to the upper electrode of the nerve synapse biomimetic device and applying a second continuous pulse identical to the first continuous pulse to the lower electrode; wherein the interval time of applying the first continuous pulse and the second continuous pulse is ΔT;

[0097] S13, changing the interval time ΔT;

[0098] S14, repeating steps S12-S13 to simulate the STDP function of the nerve synapse;

[0099] Wherein, the method for activating the nerve synapse biomimetic device comprises: grounding the lower electrode and setting a limit current; sequentially applying a first direct current scan and a second direct current scan to the upper electrode, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarity of the first direct current scan and the second direct current scan is opposite.

[0100] In an optional embodiment, in the pulse test, the initial resistance of the memristor is adjusted to an intermediate resistance state G0, a pre-pulse signal is applied to the upper electrode, a post-pulse signal is applied to the lower electrode, the time interval ΔT between the pre-pulse signal and the pulse signal is changed to adjust the conductance value Gi, the relationship between the conductance change (Gi-G0) and the time interval ΔT is obtained, and the STDP function of the synapse is simulated.

[0101] In a sixth aspect, the present application provides a biomimetic control method based on the nerve synapse biomimetic device provided in the first aspect of the present application, for simulating the PPF function of the nerve synapse, specifically comprising:

[0102] The following steps are performed on the activated and high resistance state neural synapse biomimetic device:

[0103] S21, a first pulse is applied on the upper electrode of the neural synapse biomimetic device, and after a time interval T, a second pulse identical to the first pulse is applied again;

[0104] S22, a negative direct current scan is applied on the upper electrode, so that the neural synapse biomimetic device returns to the high resistance state;

[0105] S23, the time interval T is changed;

[0106] S24, steps S21-S23 are repeated to simulate the double-peak pulse facilitation function of the neural synapse;

[0107] The method for activating the neural synapse biomimetic device comprises: grounding the lower electrode thereof, setting a limit current, and sequentially applying a first direct current scan and a second direct current scan on the upper electrode thereof, and stopping applying the direct current scan after repeating a preset number of times; wherein the voltage polarities of the first direct current scan and the second direct current scan are opposite.

[0108] In an optional embodiment, in the pulse test, the pulse parameters are fixed, the pulse width is 5 ms, the pulse amplitude is 8 V, the pulse interval between the two pulses is changed from 10 μs to 10 ms, the stimulation intensity of the second pulse relative to the first pulse is adjusted, and the double-peak pulse facilitation is realized.

[0109] It should be noted that the learning and forgetting functions of the neural synapse biomimetic device can be used to realize associative memory, wherein the associative memory needs a learning stage, an associative memory stage and a forgetting stage. Further, the spike time-dependent plasticity STDP belongs to a kind of long-range plasticity; when training based on the STDP learning rule of the neural synapse biomimetic device in the spiking neural network SNN, the area overhead can be significantly reduced compared with a transistor. The double-peak pulse facilitation belongs to a kind of short-range plasticity, which is reflected in the influence of the time interval between the first stimulation and the second stimulation on the increase of the second stimulation amplitude. It can be seen that the short-range plasticity and long-range plasticity functions of the neural synapse biomimetic device are the key to processing sound information and images with time information in the nervous system.

[0110] In summary, the neural synapse biomimetic device provided by the present application has the forming-free characteristic, the self-limiting current characteristic, and the low operating current to reduce power consumption, and can realize the associative memory with 10 4The rectification ratio can accurately simulate long-term plasticity, short-term plasticity and complex learning and forgetting characteristics of synapses, and can inhibit interference of nerve signals in inference and training process of synapse weight in the memristor array, and reduce influence of non-ideal factors of the memristor device in the neural network training and inference operation process of the memristor array.

[0111] In a sixth aspect, the present application provides an artificial neural network, comprising a plurality of neurons, the plurality of neurons being connected to each other through a synapse device array; wherein the synapse device array is a cross array structure; and the synapse device in the synapse device array is the neural synapse bionic device provided in the first aspect of the present application.

[0112] Those skilled in the art will easily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A neural synapse bionic device, characterized in that, include: The bottom electrode, resistive switching layer, barrier layer, and top electrode are distributed from bottom to top. The resistive switching layer has energy level defects; The barrier layer is used for self-current limiting to prevent the resistive switching layer from being broken down; The activation process of the neural synapse bionic device is achieved using an activation module; The activation module is used to activate the neural synapse bionic device to enable the neural synapse bionic device to have stable resistive switching characteristics, including: The lower electrode is grounded, and a current limit is set; a first DC scan and a second DC scan are applied sequentially to the upper electrode, and the DC scan is stopped after repeating a preset number of times; wherein the voltage polarities of the first DC scan and the second DC scan are opposite.

2. The neural synapse bionic device according to claim 1, characterized in that, The resistive switching layer is prepared by magnetron sputtering, atomic deposition, pulsed laser deposition, chemical vapor deposition, or thermal oxidation.

3. The neural synapse bionic device according to claim 1, characterized in that, The thickness of the resistive layer and the thickness of the barrier layer Determined based on the following formula: in, and These are the bias voltages distributed to the resistive switching layer and the blocking layer, respectively; and These are the dielectric constants of the resistive switching layer and the barrier layer, respectively. The voltage applied to the neural synapse bionic device.

4. The neural synapse bionic device according to claim 1, characterized in that, The Schottky barrier between the lower electrode and the resistive switching layer is lower than the Schottky barrier between the upper electrode and the barrier layer; The barrier layer also serves to prevent charge carriers from tunneling.

5. The neural synapse bionic device according to claim 4, characterized in that, The material of the lower electrode is a low work function metallic material or single crystal material, including: Al, Mg, Ti, TiN, Zn, In, Li, Ta or Zr; The material of the upper electrode is a high work function metal or single crystal material, including: Cr, Ni, C, Au, Pt, Co, Pd or Si.

6. A biomimetic control method for the neural synapse biomimetic device according to any one of claims 1-5, characterized in that, include: Biomimetic synaptic learning function: After activation, the lower electrode of the biomimetic neural synapse device is grounded, and a set pulse is continuously applied to its upper electrode within a first preset time period to stimulate the biomimetic neural synapse device, thereby realizing the biomimetic neural synaptic learning function. Biomimetic synaptic forgetting function: After completing the biomimetic synaptic learning function, within a preset second preset time period, the application of set pulses is stopped on the upper electrode of the neural synaptic biomimetic device to biomimize the forgetting function of the neural synapse; The second preset time period is shorter than the first preset time period; The method for activating the neural synapse bionic device includes: grounding its lower electrode and setting a current limit; sequentially applying a first DC scan and a second DC scan to its upper electrode, repeating this process a preset number of times, and then stopping the application of the DC scan; the voltage polarities of the first DC scan and the second DC scan are opposite.

7. The biomimetic control method according to claim 6, characterized in that, The learning level of the neural synapse bionic device is controlled by changing the size of the first preset time period; The degree of forgetting in the neural synapse bionic device is controlled by changing the size of the second preset time period; The forgetting rate of the neural synapse bionic device is controlled by changing the parameters of the set pulse; wherein the parameters of the set pulse include: the duty cycle, amplitude and pulse width of the set pulse.

8. The biomimetic control method according to claim 6, characterized in that, Also includes: Bionic enhancement of synaptic function: For the activated neural synaptic bionic device in a high-resistivity state, its lower electrode is grounded, and a positive DC scanning voltage or set pulse is continuously applied to its upper electrode to change its state to a low-resistivity state, so as to bionic enhance the function of neural synapse. Biomimetic synaptic inhibition function: For the neural synaptic biomimetic device that is activated and in a low-resistance state, its lower electrode is grounded, and a negative DC scanning voltage or reset pulse is continuously applied to its upper electrode to change its state to a high-resistance state, so as to mimic the inhibitory function of the neural synapse. Bionics of synaptic STDP function: S11. Adjust the initial resistance of the activated neural synapse bionic device to an intermediate resistance value. S12. A first continuous pulse is applied to the upper electrode of the neural synapse bionic device, and a second continuous pulse identical to the first continuous pulse is applied to its lower electrode; wherein the time interval between applying the first continuous pulse and the second continuous pulse is ΔT; S13. Change the interval time ΔT; S14. Repeat steps S12-S13 to mimic the STDP function of a biomimetic neural synapse. Bionics of synaptic biphasic pulse facilitation function: S21. A first pulse is applied to the upper electrode of the neural synapse bionic device, and after an interval of time T, a second pulse identical to the first pulse is applied again; S22. Apply a negative DC scan to the upper electrode to return the neural synapse bionic device to a high-resistivity state. S23. Change the time interval T; S24. Repeat steps S21-S23 to mimic the bi-peak pulse facilitation function of the biomimetic neural synapse.

9. An artificial neural network, characterized in that, It includes multiple neurons, which are interconnected by an array of synaptic devices; wherein the array of synaptic devices is a cross-shaped array structure; and the synaptic devices in the array are the neural synaptic bionic devices according to any one of claims 1-5.

Images