Sb / Sb2S3 type superlattice thin film material for neuromorphic calculation and preparation method of Sb / Sb2S3 type superlattice thin film material

Abstract

The invention discloses a Sb / Sb2S3 superlattice film material for neuromorphic calculation and a preparation method thereof, the chemical structure of the Sb / Sb2S3 superlattice film material is [Sb (x nm) / Sb2S3 (y nm)] n, x is 2-4, y is 2-4, and n is a numerical value of an integer part of 100 / (x + y), the preparation method comprises the following steps: Sb and Sb2S3 are used as target materials, a double-target alternate sputtering method is adopted, high-purity argon is used as a working gas, and the Sb / Sb2S3 superlattice film material is prepared. Adopting a quartz plate or a silicon wafer as a substrate material, and performing surface deposition on the substrate material; the Sb direct current sputtering power is adjusted to be 20 W, the radio frequency power of the Sb2S3 target is adjusted to be 25 W, and the Sb / Sb2S3 type superlattice thin film material is obtained through sputtering at the room temperature. The Sb / Sb2S3 type superlattice thin film material has the advantages of being high in crystallization temperature and data retention, low in resistance drift, high in crystallization speed, good in thermal stability and high in crystalline and amorphous resistivity.

Description

Technical Field

[0001] The present invention relates to the field of phase change electronic synaptic materials, and in particular to a Sb / Sb phase change electronic synaptic material for neuromorphic computing. 2 S 3 Superlattice-like thin film material and preparation method thereof. Background Art

[0002] Among many new information devices, phase change memory (PCM) has become one of the most competitive candidates for future neuromorphic devices due to its advantages such as non-volatility, miniaturization, high switching speed, low operating voltage, low power consumption, long cycle life and compatibility with existing semiconductor processes. In recent years, PCM has been widely used in the simulation of biological synaptic characteristics due to its bidirectional controllable multi-level conductivity regulation characteristics similar to biological synaptic structures. The top electrode and bottom electrode of PCM correspond to the biological preneuron axon and postneuron dendrite respectively, and the conductance of the intermediate phase change layer corresponds to the weight of the biological synapse. When an electric pulse acts on the top electrode of PCM, the resistance of the phase change layer changes with the change of the phase change layer material, thereby generating a changing current signal in the phase change layer and transmitting it to the bottom electrode, thus completing the simulation of the biological synaptic weight regulation function. Synaptic plasticity (synaptic enhancement, synaptic inhibition) is achieved by applying a series of low-amplitude and wide pulses to the electronic synapse, which increases the conductance of the phase change layer, corresponding to the enhancement effect of the biological synapse. Conversely, when a series of pulses with slightly higher amplitudes and narrower pulse widths are applied to the electronic synapse, the conductance of the phase change layer weakens, corresponding to the inhibition of the biological synapse. Phase change layer materials GST still face challenges in practical applications. For example, the uncertainty of the nucleus during the crystallization process leads to a relatively slow setting speed (about 280 nanoseconds). The low crystallization temperature (about 150°C) and its insufficient data retention capability (about 90°C for ten years) result in unsatisfactory thermal stability. Structural relaxation causes amorphous resistance drift. In addition, Ge, Sb, and Te elements may migrate toward the cathode and anode after multiple set-to-reset cycles, resulting in large fluctuations in resistance. In addition, accurately adjusting the phase change volume to obtain incremental changes in conductivity, as well as simulating long-term potentiation (LTP) and long-term depression (LTD) of synaptic weights, are quite challenging tasks.

[0003] In recent years, to solve these problems, a superlattice phase change memory based on van der Waals heterostructure has been developed by alternately depositing ultrathin films. The low thermal conductivity and high resistivity of superlattice-like (SLL) films create electrothermal confinement through van der Waals-like interfaces, thereby improving the energy efficiency of superlattice-like phase change memory devices (SLL-PCM). Superlattice structure composed of phase change materials, with alternating layers of GeTe / Sb 2 Te 3 、TiTe 2 / Sb 2Te 3 ,GeSb 2 Te 4 / Sb 2 Te 3 and Sb 2 Te 3 / Ge 2 Sb 2 Te 5 , reduced switching current, minimal resistance drift and enhanced thermal stability have been achieved. However, the volatility of elemental tellurium is detrimental to semiconductor technology, and the presence of tellurium may lead to phase separation and insufficient data retention after periodic cycling. Using single-atom Sb as a phase change material is a viable solution to overcome the reliability or fatigue challenges in programming multi-component compounds. However, there are still some problems with the application of Sb in phase change memory, mainly manifested in: 1. The explosive crystallization mechanism hinders the realization of a stable amorphous state and poor thermal stability. 2. The contrast between the amorphous and crystalline resistances of Sb multilayers is still not high enough, making it difficult to use in emerging fields such as neuromorphic computing. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a Sb / Sb for neuromorphic computing with high crystal stability and data retention capability, fast crystallization speed and low resistance drift. 2 S 3 A superlattice-like thin film material and a preparation method thereof, the method has low cost, strong process controllability and is easy for large-scale production.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a Sb / Sb for neuromorphic computing 2 S 3 Superlattice thin film material, its chemical structure is [Sb(x nm) / Sb 2 S 3 (y nm)] n , where x is 2 to 4, y is 2 to 4, and n is the integer part of 100 / (x+y).

[0006] Preferably, the chemical structure of the thin film material is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 The data retention of the phase change film material can be stored for ten years at 126.5°C, with a low resistance drift v=0.0025.

[0007] Preferably, the chemical structure of the thin film material is [Sb(2nm) / Sb 2 S 3 (2nm)] 25, low resistance drift v reaches 0.0006.

[0008] The above Sb / Sb for neuromorphic computing 2 S 3 Preparation method of superlattice-like thin film material, using high-purity round block Sb and Sb 2 S 3 As the target material, a magnetron sputtering instrument is used, through double-target alternating sputtering, high-purity argon is used as the working gas, and silicon wafers or quartz wafers are used as substrate materials for surface deposition. The specific steps are as follows:

[0009] (1) Sb round metal target and Sb 2 S 3 The back of the round metal oxide target is completely bonded to a copper sheet with the same diameter and thickness of 1 mm to make a magnetron sputtering coating target. The Sb target is installed in the magnetron DC sputtering target. 2 S 3 Installed in magnetron RF sputtering target;

[0010] (2) placing the quartz wafer or silicon wafer substrate material in deionized water and anhydrous ethanol for ultrasonic cleaning in turn, taking it out, blowing it dry with high-purity nitrogen, and placing it in a sputtering chamber;

[0011] (3) The sputtering chamber is closed and evacuated. When the vacuum degree of the sputtering chamber reaches 2×10 -4 When the pressure reaches 2.3 Pa, high-purity argon gas is introduced at a gas flow rate of 50 ml / min, so that the gas pressure in the sputtering chamber reaches the ignition pressure of 2.3 Pa required for sputtering;

[0012] (4) Turn on the DC power supply, adjust the Sb DC sputtering power to 20W, turn on the RF power supply, and adjust the Sb 2 S 3 The target RF power is 25W. After the light is stable, the coating is started at room temperature. 2 S 3 The target material is sputtered alternately for 14 to 25 cycles to obtain Sb / Sb for neuromorphic computing. 2 S 3 Superlattice-like thin film materials.

[0013] Preferably, the Sb target and the Sb 2 S 3 The purity of the target materials is 99.99%.

[0014] Preferably, the deposited Sb / Sb obtained in step (4) 2 S 3The superlattice-like thin film material is placed in a rapid annealing furnace and rapidly heated to 180-300°C for annealing in an atmosphere of high-purity nitrogen to obtain the Sb / Sb for neuromorphic computing after heat treatment. 2 S 3 Superlattice-like thin film materials.

[0015] Compared with the prior art, the advantages of the present invention are: the present invention is a [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Superlattice thin film material. The film limits the Sb layer to a scale below 6 nanometers and is combined with the amorphous sulfide binary material Sb 2 S 3 Alternating deposition forms Sb / Sb-like 2 S 3 Superlattice films improve the thermal stability of Sb by utilizing van der Waals (vdW)-like interfaces in superlattice structures, achieving extremely low resistance drift (v = 0.0006) and up to 10 4 The resistance switching ratio is 225℃, and the amorphous resistance is 10 5 ~10 7 Ω, crystalline resistance 10 3 ~10 4 Ω; the temperature of 10-year data retention is tested to be 126.5°C. The increase of thin film crystalline resistance will help reduce the power consumption of PCM. The present invention has low production cost, good repeatability, and strong process controllability. The prepared [Sb(2nm) / Sb 2 S 3 (4nm)] 16 The thin film material not only has the advantages of small component error, high adhesion strength, and uniform and dense film quality, but also can adjust the crystallization temperature according to the controlled components. It has good thermal stability, fast crystallization temperature, and large crystalline / amorphous resistivity. It can be used for industrial large-scale preparation of large-area phase change films, thereby meeting the application needs of future phase change electronic synapse materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Sb / Sb 2 S 3 Thermal stability diagram of superlattice-like thin film materials, where (a) is Sb / Sb 2 S 3 Schematic diagram of the superlattice-like film structure, (b) the variation of sheet resistance with temperature at a heating rate of 40°C / min, (c) the 10-year data retention curve obtained by fitting the Arrhenius equation, (d) the resistance drift coefficient under 85°C heat treatment;

[0017] Figure 2 (a) is Sb / Sb of the same composition 2 S 3 Superlattice-like films in the deposited state The relationship diagram with lv, (b) is [Sb(2nm)Sb 2 S 3 (4nm)] 16 Under different annealing conditions Relationship diagram with lv;

[0018] Figure 3 = [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Crystallization kinetics of superlattice-like films, where (a) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 In the resistance-temperature curves at heating rates from 10°C / min to 50°C / min, the inset shows the curves between 200°C and 240°C, (b) the relationship between crystallinity and temperature, (c) the crystallization kinetic index n, and (d) the fluctuation of the crystallization kinetic index n between 212°C and 226°C;

[0019] Figure 4 (a) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 X-ray diffraction patterns of the film at different heating temperatures, (b) [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Raman spectra of the films, the red line indicates the amorphous state, while the blue line indicates annealing at 280 °C for 10 min in nitrogen;

[0020] Figure 5 = [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Cross-sectional BF-TEM images of superlattice-like thin films in crystalline and amorphous states, where (a) is the deposited state, (b) is the state after annealing at 300°C for 5 minutes, (c) is the enlarged BF-TEM image obtained from the crystal in (b), and (df) are BF-TEM images of the corresponding positions in (c);

[0021] Figure 6 = [Sb(2nm) / Sb 2 S 3 (4nm)] 16STEM-EDX element distribution diagram of superlattice-like films, (ac) are deposited state, (df) are crystalline state;

[0022] Figure 7 Based on [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Electrical performance characteristics of superlattice-like thin film devices, where (a) is a photo of the fabricated PCM unit and a schematic diagram of the T-shaped unit, (b) is the current-voltage (IV) characteristic, and the inset shows the resistance drift curve of the device unit, and (c) is the SET-RESET operation using 50-2000 nanosecond pulses;

[0023] Figure 8 Based on [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Thin film electronic synaptic devices are used for handwritten digit recognition applications, where (a) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Synaptic properties of the device. The inset shows the composition of a biological synapse. (b) Schematic diagram of a neural network, including a flat layer and two multi-layer layers. (c) Evaluation of the accuracy of the MNIST dataset in capturing the optimal and suboptimal features of the proposed device. (d) Output of 10 neurons after inputting digital images 0 to 9. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below with reference to the accompanying drawings. 1. Specific embodiments

[0026] An Sb / Sb for neuromorphic computing 2 S 3 Superlattice thin film material, its chemical structure is [Sb(xnm) / Sb 2 S 3 (y nm)] 100 / (x+y) The crystallization temperature of the material is between 160 and 225°C, and the amorphous resistance is between 10 5 ~10 7 Ω, crystalline resistance 10 3 ~10 4 Ω. The preparation method is: using high purity round blocks of Sb and Sb 2 S 3 Materials are used as targets. A magnetron sputtering device is used, dual targets are sputtered alternately, high-purity argon is used as the working gas, and quartz wafers or silicon wafers are used as substrate materials for surface deposition.

[0027] Example 1: Sb / Sb for neuromorphic computing 2 S 3 Superlattice thin film material, its chemical structure is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 The specific steps of its preparation method are as follows:

[0028] (1) Sb round target and Sb 2 S 3 The back of the round target is completely attached with a round copper sheet with the same diameter as the target. The thickness of the copper sheet is about 1 mm, and the magnetron sputtering coating target is obtained. The Sb target is installed in the DC sputtering target. 2 S 3 Installed in a radio frequency sputtering target;

[0029] (2) placing a quartz wafer or silicon wafer in deionized water for ultrasonic cleaning for 15 minutes, and then placing it in anhydrous ethanol for ultrasonic cleaning for 15 minutes, taking it out and drying it with high-purity nitrogen, and placing it in a sputtering chamber as a substrate;

[0030] (3) The sputtering chamber is sealed and evacuated. When the vacuum degree of the sputtering chamber reaches 2×10 -4 When the pressure reaches 3 Pa, high-purity argon gas is introduced at a rate of 50 ml / min, so that the gas pressure in the sputtering chamber reaches the ignition pressure of 3 Pa required for sputtering;

[0031] (4) Turn on the DC power supply and adjust the Sb DC sputtering power to 20W; turn on the RF power supply and adjust the Sb 2 S 3 The target RF power is 25W. After the light is stable, the coating is started at room temperature. 2 S 3 The target material alternate sputtering time was 14.285s, 30s, and after 16 cycles of alternate sputtering, a deposited [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Superlattice-like films;

[0032] (5) The deposited state [Sb(2nm) / Sb 2 S 3 (4nm)] 16 The film material is placed in a rapid annealing furnace and annealed at 180-300°C under the condition of nitrogen flow, and the heat-treated [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Superlattice-like thin film materials.

[0033] The magnetron sputtering device used in the above is a Kurt Lesker PVD75 magnetron sputtering system. The purity of the sputtering target used is 99.99%, and the size is Φ50×3mm. 2 S 3 A 1mm thick copper sheet with the same diameter is pasted on the back to solve the heat dissipation problem of the glass target during sputtering.

[0034] Example 2 is the same as Example 1, except that the Sb layer is changed to 4 nm. 2 S 3 The layer was changed to 3 nm, and then 14 cycles of alternating deposition were performed to obtain a [Sb(4 nm) / Sb 2 S 3 (3nm)] 14 Superlattice-like thin film materials.

[0035] Example 3, same as Example 1, except that the Sb layer is changed to 3nm, Sb 2 S 3 The layer was changed to 3 nm, and then 16 cycles of alternating deposition were performed to obtain a [Sb(3 nm) / Sb 2 S 3 (3nm)] 16 Superlattice-like thin film materials.

[0036] Example 4 is the same as Example 1, except that the Sb layer is changed to 2nm. 2 S 3 The layer was changed to 3 nm, and then 20 cycles of alternating deposition were performed to obtain a 100 nm thick [Sb(2 nm) / Sb 2 S 3 (3nm)] 20 Superlattice-like thin film materials.

[0037] Example 5, same as Example 1, except that the Sb layer is changed to 2nm, Sb 2 S 3 The layer was changed to 2 nm, and then 25 cycles of alternating deposition were performed to obtain a 100 nm thick [Sb(2 nm) / Sb 2 S 3 (2nm)] 25 Superlattice-like thin film materials.

[0038] 2. Experimental Results Analysis

[0039] The Sb / Sb prepared in the above example 2 S 3 The film was tested for performance. Figure 1(a) is a schematic diagram of the film structure. It can be seen that this type of superlattice film consists of Sb layer and Sb 2 S 3 The layers are formed alternately. Figure 1 (b) shows the relationship between the block resistance and temperature of the thin film materials with different component ratios at a heating rate of 40°C / min. 2 S 3 The crystallization temperature of the superlattice-like film increases with the increase of the thickness ratio and the number of interfaces, which indicates that the Sb / Sb 2 S 3 The film exhibits strong thermal stability, which exceeds that of pure Sb film and traditional GST film, thereby improving performance and reducing power consumption when applied to electronic synapses. Figure 1 (c) It can be seen that the data retention also increases with Sb 2 S 3 The thickness ratio and the number of interfaces change accordingly. In [Sb(2nm) / Sb 2 S 3 (2nm)] 25 In the current case, the data retention is 137.6℃, which is much higher than GST's 82.1℃. Figure 1 (d) is the resistance drift coefficient of films with different compositions after heat treatment at 85℃, which further illustrates the superlattice-like Sb 2 S 3 The increase in the thickness ratio and the number of interfaces will increase the Sb / Sb 2 S 3 The thermal stability of the film is increased.

[0040] Figure 2 (a) is Sb / Sb 2 S 3 The transmission spectrum of the superlattice-like film in the deposited state shows that the optical band gap increases with the increase of Sb 2 S 3 The increase in the bandgap width increases with the thickness ratio and the total number of periods in the film. The enlargement of the bandgap width enhances the formation of trap states and raises the barrier for charge carrier migration, reducing the number of charge particles, which directly increases the resistance of the film. Figure 2 (b) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Through the spectrum, the purpose is to explore the resistance change between the amorphous and crystalline states in terms of the optical band gap. It can be seen that with the increase of annealing temperature, the band gap energy (Eg) gradually decreases from 0.58eV to 0.46eV. When the temperature reaches 250°C, Eg suddenly drops to 0.46eV. The decrease in Eg is related to the increase in carrier concentration and conductivity, indicating that the transition of carriers to the conduction band becomes easier.

[0041] Figure 3 To study [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Crystallization kinetics of thin films, Figure 3 (a) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Resistance-temperature curves at heating rates from 10°C / min to 50°C / min, with the inset showing the curves between 200°C and 240°C. After differential integration and normalization, we obtain Figure 3 (b) shows the curve of crystallinity fraction (x) fluctuation, which reflects the crystallinity. By linear regression of the crystallinity fraction curves at different heating rates in the range of 212℃ to 226℃, we determined the crystallization kinetic index n, which characterizes the crystallization behavior, such as Figure 3 (c) is shown. The crystallization kinetics index is a widely accepted method for specifying the crystallization process. If the average crystallization kinetics index is lower than 1.5, the crystallization mechanism of the material is mainly controlled by the growth mode, resulting in rapid grain growth during the crystallization process. On the other hand, when the crystallization kinetics index is greater than 1.5, it indicates that the nucleation mode dominates the crystallization process. During the crystallization process, the increase in grain size may be relatively slow. Figure 3 As shown in (c), the average kinetic index is about 0.481, which indicates that [Sb(2nm) / Sb 2 S 3 (4nm)] 16 The crystallization mechanism of superlattice-like films is mainly growth-based, which is conducive to ultrafast phase transitions.

[0042] Figure 3 (d) is the fluctuation of the crystallization kinetic index n between 212℃ and 226℃. Figure 3 As shown in (d), during the continuous crystallization of the film, the n value first increases and then decreases with the increase of annealing temperature, indicating that the crystallization process is not uniform. The interfacial tension in the superlattice-like film hinders the formation of Sb crystals.

[0043] Figure 4 (a) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 X-ray diffraction patterns after annealing at 150℃, 200℃, 250℃ and 300℃. It can be seen that at the annealing temperature of 200℃, no obvious diffraction peaks appear, indicating that the film is in an amorphous state. When the annealing temperature is increased to 250℃, [Sb(2nm) / Sb 2 S 3 (4nm)] 16The film shows diffraction peaks of rhombohedral Sb phase with crystal planes (003), (110), (006), and Sb 2 S 3 The diffraction of the Sb phase is (020), (120), (130), and (240). Since no other crystal planes of the Sb phase appear, this is attributed to the interface inhibition effect. 2 S 3 The peak intensity increases and becomes more distinct and clear, indicating the growth of grain size. Figure 4 (b) is [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Raman diffraction patterns of the deposited state and annealed at 250°C. It can be seen that the Raman spectrum of the deposited state shows an isolated and broad peak (called peak A) at about 144 cm -1 , which is related to the resonance of disordered Sb-Sb bonds. After annealing at 250℃, [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Peak A in the film begins to decompose into two different peaks, peak B and peak C, located at 118 cm -1 and 152cm -1 , indicating that a phase transition has occurred within the film. Peak B is associated with the Eg mode of the Sb phase with A7 crystal structure, while peak C represents the A1g mode of the same Sb phase. As the annealing temperature increases, the Raman peaks become sharper, indicating enhanced formation of Sb grains. Higher annealing temperatures produce compressive stresses in the grains, reduce bond lengths, increase bond energies, enhance atomic contacts, and improve stability.

[0044] Figure 5 (a) shows that [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Cross-sectional TEM image of the film as deposited, showing a periodic layered structure. Figure 5 (b) reveals the TEM image after annealing at 250 °C for 5 min, where many lattice fringes are clearly visible, marking the crystallization of the film. After annealing, the interface structure of the multilayer film becomes discernible, highlighting the superior thermal stability of the prepared film. Figure 5 (c) shows Figure 5 (b) Enlarged portion to enhance the distinction between crystal planes. Figure 5 (df) shows the change of [Sb(2nm) / Sb after annealing at 250℃ 2 S 3 (4nm)] 16TEM images of different regions of the film. These images magnify Figure 5 (c) The lattice structure of the selected area. Figure 5 (d)- Figure 5 (f) Apply fast Fourier transform and its inverse transform to distinguish Sb (marked with red and blue lines) from Sb 2 S 3 The lattice fringes of the Sb layer are shown in yellow. The lattice spacing of the Sb layer is determined to be 0.216 nm for the (110) crystal plane and 0.225 nm for the (104) crystal plane. These values ​​are consistent with the rhombohedral structure of Sb and are consistent with the XRD results.

[0045] Figure 6 (a)-(c) and Figure 6 (d)-(f) show the deposited and crystallized [Sb(2nm) / Sb obtained using STEM mode and energy dispersive X-ray spectroscopy. 2 S 3 (4nm)] 16 Elemental distribution diagram of the film. In the deposited state, the difference between Sb and S elements is significant, indicating that the film is of high quality. After annealing, the multilayer interface remains basically unchanged, and the Sb and S elements maintain a clear boundary. Therefore, the film still maintains a clear multilayer structure at high temperature, showing excellent stability.

[0046] Figure 7 (a) Shows the physical arrangement of the PCM memory array and the T-type PCM device unit. Figure 7 (b) shows the current-voltage (IV) curve. The early steep slope of the curve indicates that the material is in a high-resistance amorphous state. When the applied voltage exceeds the threshold voltage, the negative resistance phenomenon becomes obvious, indicating that the film transforms from the amorphous state to the crystalline state. Figure 7 The inset of (b) shows the resistance variation over time when the device is configured in a high resistance state (HRS) and a low resistance state (LRS). The low resistance state (LRS) and intermediate resistance state (IRS) of the device show consistent low resistance drift v values ​​of 0.004 and 0.002, respectively. This indicates that the structural performance of the SLL device is stable. Figure 7 (c) shows the resistance-voltage (RV) curves at different pulse widths. The results show that even under a short pulse voltage of 50 nanoseconds, the resistance ratio between the reset state and the set state is still significantly higher than two orders of magnitude. This indicates that the thin layer [Sb(2nm) / Sb 2 S 3 (4nm)] 16The superlattice-like PCM cell can efficiently perform the set and reset processes with a short pulse voltage of 50 nanoseconds, surpassing the GST that requires hundreds of nanoseconds in the same technology process. The voltages for the set and reset operations are defined as 1.9V and 3V, respectively. According to the Joule heating equation, the energy consumption of the device is about 1.58×10 -13 Joule, about GST unit (about 10 -10 Joules) is three orders of magnitude lower.

[0047] Figure 8 Based on [Sb(2nm) / Sb 2 S 3 (4nm)] 16 Thin-film electronic synaptic devices are used in handwritten digit recognition applications. Based on the heat accumulation effect in the phase-change material, changing the number of pre- and post-synaptic pulses can create different heat generation and dissipation conditions. It changes the ratio of crystalline to amorphous states in the phase-change material layer, thereby changing the conductivity of the device, which corresponds to the synaptic weight. Figure 8 (a) The characteristics of long-term potentiation (LTP) and long-term depression (LTD) are shown. The LTP characteristics are evaluated by implementing a preset pulse scheme, and the LTD characteristics are evaluated by applying a fixed reset pulse method. When continuous reset pulses of the same amplitude and form (1.45 V-500 ns) are applied, a gradual change in resistance from high conductivity to low conductivity during the depression process is observed (see Figure 8 The blue curve in (a) shows the process of conductivity gradually changing in the opposite direction when continuous setting pulses of the same amplitude (1.4V-1000ns) are applied to the same device. Figure 8 (b). The network is fully connected and consists of 784 input neurons (i), 250 hidden neurons (h), and 10 output neurons (o). The output of each neuron serves as the input of the subsequent neuron. A total of 40 training cycles were performed with a learning rate of 0.1. The optimization algorithm used was stochastic gradient descent (SGD). To simplify image classification, the activation function selected for the first and second layers was Sigmoid, while the activation function for the second and third layers was set to softmax. The back propagation (BP) algorithm calculates the gradient to optimize the parameter weights and biases to minimize the loss function and maximize the accuracy. In each iteration, the training process uses 50,000 randomly selected photos as input, and another 10,000 images are used for the testing process. The weights are updated after every 100 images are input. The updated values ​​are converted into corresponding P / D pulse counts. Figure 8 (c) shows the accuracy of recognizing MNIST handwritten digits, which reaches a baseline accuracy of 97.79% under ideal algorithm conditions. 2 S3 (4nm)] 16 When the device updated its weights, the network achieved 96.57% accuracy. Figure 8 (d) shows the output of ten output neurons when the digital image 0-9 is input, proving that [Sb(2nm) / Sb 2 S 3 (4nm)] 16 High reliability of devices as synapses.

[0048] The above description is not intended to limit the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by a person skilled in the art within the essential scope of the present invention shall also fall within the protection scope of the present invention, and the protection scope of the present invention shall be subject to the claims.

Claims

1. A Sb / Sb2S3-based superlattice thin film material for neuromorphic computing, characterized in that Its chemical structure is [Sb(x nm) / Sb2S3(y nm)] n , where x is 2 to 4, y is 2 to 4, and n is the integer part of 100 / (x+y).

2. The Sb / Sb2S3-based superlattice thin film material for neuromorphic computing according to claim 1, characterized in that Its chemical structure is [Sb(2nm) / Sb2S3(4nm)] 16 .

3. The Sb / Sb2S3-based superlattice thin film material for neuromorphic computing according to claim 1, characterized in that Its chemical structure is [Sb(2nm) / Sb2S3(2nm)] 25 .

4. A method for preparing the Sb / Sb2S3-based superlattice thin film material for neuromorphic computing according to claim 1, characterized in that: High-purity round block Sb and Sb2S3 are used as target materials, a magnetron sputtering instrument is used, double targets are co-sputtered, high-purity argon is used as the working gas, and silicon wafers or quartz wafers are used as substrate materials for surface deposition. The specific steps are as follows: (1) The backs of the Sb2S3 round block metal target and the Sb round block metal target are completely bonded to the copper sheet to prepare magnetron sputtering coating targets, the Sb target is installed in the magnetron DC sputtering target, and the Sb2S3 is installed in the magnetron RF sputtering target; (2) placing the quartz wafer or silicon wafer substrate material in deionized water and anhydrous ethanol for ultrasonic cleaning in turn, taking it out, blowing it dry with high-purity nitrogen, and placing it in a sputtering chamber; (3) The sputtering chamber is closed and evacuated. When the vacuum degree of the sputtering chamber reaches 2×10 -4 When the pressure reaches 2.3 Pa, high-purity argon gas is introduced at a gas flow rate of 50 ml / min, so that the gas pressure in the sputtering chamber reaches the ignition pressure of 2.3 Pa required for sputtering; (4) Turn on the DC power supply and adjust the Sb DC sputtering power to 20 W. Turn on the RF power supply and adjust the RF power of the Sb2S3 target to 25 W. After the glow stabilizes, start coating at room temperature. Alternately sputter the Sb and Sb2S3 targets for 14 to 25 cycles to obtain Sb / Sb2S3-type superlattice thin film materials for neuromorphic computing.

5. The method for preparing a Sb / Sb2S3-based superlattice thin film material for neuromorphic computing according to claim 4, characterized in that: The purity of the Sb target material and the Sb2S3 target material are both 99.99%.

6. The method for preparing a Sb / Sb2S3-based superlattice thin film material for neuromorphic computing according to claim 4, characterized in that: The deposited Sb / Sb2S3 superlattice thin film material obtained in step (4) is placed in a rapid annealing furnace, and rapidly heated to 180-300°C for annealing in an atmosphere of high-purity nitrogen, thereby obtaining the heat-treated Sb / Sb2S3 superlattice thin film material for neuromorphic computing.

Images