A tunable LIF neuron hardware unit based on electrochemical doping

Abstract

This invention discloses a tunable LIF neuron hardware unit based on electrochemical doping, comprising an electrochemical random access memory (ECRAM) as an adjustable load resistor and a three-terminal threshold switch device with neuron firing function. The ECRAM is a three-terminal memory device, and the ECRAM and the three-terminal threshold switch device are connected in series through their respective channels; that is, the source of the ECRAM is connected to the drain of the three-terminal threshold switch device. Changes in the state of both are achieved by applying a suitable electrical signal to the third terminal. Applying this tunable LIF neuron hardware unit to a spiking neural network, by changing the channel resistance of the ECRAM and the threshold voltage of the three-terminal threshold switch device channel, the membrane potential time constant τ and threshold voltage V of the LIF neuron can be controlled. th Adaptive adjustments are made. Therefore, the hardware accelerator for the spiking neural network based on this tunable LIF neuron hardware unit can adjust parameters for different application environments, thereby optimizing network performance.

Description

Technical Field

[0001] This invention belongs to the field of microelectronic devices, specifically relating to a tunable LIF (Leaky Integrate and Fire) neuron hardware unit based on electrochemical doping and its fabrication method. It can adjust various important parameters of the LIF neuron and adaptively optimize its performance for different application environments, thereby further improving the performance of the spiking neural network. Background Technology

[0002] In recent years, artificial intelligence technology has made many breakthroughs in application areas such as image recognition, autonomous driving, natural language processing, and generative dialogue chat systems. However, the hardware systems and parameters it relies on are becoming increasingly large, and the corresponding power consumption is also increasing rapidly. Currently used deep learning algorithms initially simulated the human brain's nervous system in their connection structure and processing methods, but mainly simplified them. For example, some functions related to spatiotemporal information processing were not integrated, making it impossible to achieve more advanced cognitive functions in the human brain through traditional deep learning. Secondly, most current deep learning network algorithms are deployed on traditional hardware systems, and the storage-computing separation characteristic of traditional hardware architectures severely limits the network inference speed to the memory access stage, resulting in additional latency and power consumption.

[0003] To fully simulate the characteristics of human brain computation, a spiking neural network based on pulse signal processing has been proposed. Compared with traditional artificial neural networks, its neuron model retains the core characteristics of human brain neurons, such as leakage, integration, and firing functions, enabling better processing of spatiotemporal information. Furthermore, its event-driven nature means that computation is only triggered when a spiking neuron fires a pulse, resulting in lower power consumption compared to traditional artificial neural networks based on continuous signal processing.

[0004] For hardware systems deploying networks, in-memory computing architectures based on storage-computing convergence are better suited for accelerating neural network inference, including traditional artificial neural networks and spiking neural networks. Their core mainly comprises two parts: synaptic devices based on non-volatile memory and neuron-type device units or circuits with nonlinear integration capabilities. For neuron-type hardware units, the neuron portion based on traditional CMOS circuits is generally too complex in structure, resulting in high power consumption. Meanwhile, the neuron portion based on novel devices lacks adjustable degrees of freedom, and for different tasks and application scenarios, there is a lack of adjustable dimensions for further optimization. Summary of the Invention

[0005] To fill the gap in existing technologies, this invention proposes a tunable LIF neuron hardware unit based on electrochemical doping for spiking neural networks. Based on the effect of electrochemical doping, the membrane potential time constant and firing threshold voltage of the LIF neuron can be controlled separately, and the performance can be adaptively optimized for different application environments, thereby further improving the performance of spiking neural networks.

[0006] Specifically, this invention provides a tunable LIF neuron hardware unit based on electrochemical doping and a corresponding fabrication method, while optimizing the parameters of a spiking neural network based on its tunable characteristics.

[0007] The tunable LIF neuron hardware unit of the present invention mainly includes an electrochemical random access memory (ECRAM) as an adjustable load resistor and a three-terminal threshold switch device with neuron firing function.

[0008] In some embodiments of the present invention, the ECRAM described above as an adjustable load resistor is a three-terminal memory device, mainly comprising a source, a drain, a channel layer, a gate electrolyte layer, a passivation layer, and a gate. The source and drain are located at opposite ends of the channel layer, the gate electrolyte layer is located on the channel layer, and the gate electrolyte layer contains mobile ions. The passivation layer covers the gate electrolyte layer, and the gate is located on the passivation layer. By applying a stimulation signal to the gate, the mobile ions (e.g., hydrogen ions, lithium ions, etc.) in the gate electrolyte layer move under the action of an electric field, completing the doping and dedoping of the channel layer, thereby affecting the band structure of the channel material and dynamically adjusting the conductivity of the channel. When a strong gate stimulation is applied, the ions will stably exist in the channel, and the channel conductivity state can be maintained for a long time when the stimulation is removed. The material of the channel layer may be, but is not limited to, WO3. x NbO x VO x MoO x The electrolyte layer consists of ZnO, IGZO, ITO, and MoS2, with 1 ≤ x ≤ 3, and a thickness ranging from 10 to 100 nm. The gate electrolyte layer is preferably a solid electrolyte film, and its material can be, but is not limited to, LiPON or LATP (Li 1+x Al x Ti 2-x(PO4)3, 0≤x≤0.5, lithium aluminum titanium phosphate), PSG (Phospho Silicate Glass), the thickness of the gate electrolyte layer is in the range of 50-300 nm; the material of the passivation layer can be, but is not limited to, SiO2, Al2O3, Si3N4, HfO2, Ta2O5, and its thickness is in the range of 1-20 nm; the materials of the gate, source and drain can be, but are not limited to, Ti, Cr, Sc, Al, Pd, Au, Pt, Ag, W, TiN, TaN, and the electrode thickness is in the range of 10-300 nm.

[0009] The aforementioned three-terminal threshold switching device with neuron firing function mainly includes a source, a drain, a channel layer, a gate electrolyte layer, a passivation layer, and a gate. The source and drain are located at opposite ends of the channel layer, the gate electrolyte layer is located on the channel layer and contains mobile ions, and the passivation layer covers the gate electrolyte layer, with the gate located on the passivation layer. By applying a stimulation signal to the gate, mobile ions (such as hydrogen ions, lithium ions, etc.) in the gate electrolyte move under the influence of an electric field, completing the doping and dedoping of the channel layer. This changes the initial resistance of the channel layer, correspondingly altering the threshold voltage of the channel threshold switching device. Furthermore, this threshold voltage change is non-volatile, meaning that its characteristics are retained even after the gate signal is removed. The channel layer material needs to possess insulator-metal transition (IMT) characteristics under electric field stimulation, and is generally selected from, but not limited to, VO2, V3O5, and NbO2, with a thickness ranging from 1 to 100 nm; the gate electrolyte layer is preferably a solid electrolyte film, and its material can be, but is not limited to, LiPON, LATP (Li 1+x Al x Ti 2-x (PO4)3, 0≤x≤0.5, lithium aluminum titanium phosphate), PSG (Phospho Silicate Glass), the thickness of the gate electrolyte layer is in the range of 50-300 nm; the material of the passivation layer can be, but is not limited to, SiO2, Al2O3, Si3N4, HfO2, Ta2O5, and its thickness is in the range of 1-20 nm; the materials of the gate, source and drain can be, but are not limited to, Ti, Cr, Sc, Al, Pd, Au, Pt, Ag, W, TiN, TaN, and the electrode thickness is in the range of 10-300 nm.

[0010] In the aforementioned tunable LIF neuron hardware unit, the ECRAM and the three-terminal threshold switch are connected in series via their respective channels. That is, the source of the ECRAM is connected to the drain of the three-terminal threshold switch. Changes in the state of both are achieved by applying a suitable electrical signal to the third terminal (gate). For details, please refer to [reference needed].Figure 2 .

[0011] For the aforementioned tunable LIF neuron hardware unit, refer to Figure 1 The membrane potential V of its neurons out satisfy:

[0012]

[0013] Where C M R is the membrane capacitance of this LIF neuron. E R is the channel resistance of the ECRAM. T This represents the high-resistance state of the channel of a three-terminal threshold switch, i.e., the resistance state before the threshold transition occurs. in Input from external stimuli.

[0014] When V out >V th When the neuron fires a pulse, it will do so; conversely, it will accumulate and leak, with V... th This is the threshold voltage of the channel of a three-terminal threshold device.

[0015] For the LIF neuron mathematical model commonly used in spiking neural networks:

[0016]

[0017] Among them, V m V is the membrane potential of the LIF neuron. rest Let t be the resting potential of the LIF neuron, τ be the membrane potential time constant of the LIF neuron, which controls the decay rate of the membrane potential, and I(t) be the external input stimulus at time t, which can be either a current stimulus or a voltage input stimulus.

[0018] When V m >V th When this happens, the neuron fires a pulse, in which V th This represents the threshold voltage of the LIF neuron.

[0019] Comparing the above-mentioned tunable LIF neuron hardware unit and the LIF neuron mathematical model of the spiking neural network, it can be seen that V out This is the membrane potential of the LIF neuron, whose resting potential is 0, C M (R E +R T This is the membrane potential time constant of the LIF neuron. The input stimulus is from the outside, and the threshold voltage of the channel of the three-terminal threshold switch is the threshold voltage of the LIF neuron. Therefore, by changing R... EThe threshold voltage V of the LIF neuron can be changed by altering the membrane potential time constant τ of the LIF neuron and by changing the threshold voltage of the channel of the three-terminal threshold switch. th .

[0020] Therefore, based on the above principle, by changing the channel resistance of the ECRAM and the threshold voltage of the three-terminal threshold switch, the core parameters τ and V of the LIF neuron in the spiking neural network can be adjusted. th By making adaptive adjustments, the hardware accelerator of the spiking neural network based on this tunable LIF neuron hardware unit can adjust parameters for different application environments, thereby optimizing network performance.

[0021] Another objective of this invention is to provide a method for fabricating the tunable LIF neuron hardware unit, the main process steps of which include:

[0022] 1) Prepare the substrate, including cleaning and drying the substrate;

[0023] 2) Deposit the channel layer material for the three-terminal threshold switching device on the substrate;

[0024] 3) Fabricate the source and drain contact electrodes of the three-terminal threshold switch device;

[0025] 4) Pattern the region where the electrochemical random access memory is designed and fabricated, and deposit an isolation dielectric layer, which may be selected but not limited to HfO2, Al2O3 and SiO2;

[0026] 5) Fabricate source and drain contact electrodes for ECRAM on the isolation dielectric layer deposited in step 4);

[0027] 6) Pattern and deposit the ECRAM channel layer on the isolation medium layer deposited in step 4);

[0028] 7) The gate electrolyte layer and passivation layer of electrochemical random access memory and three-terminal threshold switching device are prepared by patterning and sequentially depositing electrolyte layer and passivation layer;

[0029] 8) Fabricate source, drain, and gate connection electrodes for electrochemical random access memory and three-terminal threshold switching devices.

[0030] In the fabrication process, to simplify the process steps, the same material is used for the gate electrolyte layer of the ECRAM and the three-terminal threshold switch device in step 7), and the gate electrolyte layer and the passivation layer are fabricated simultaneously.

[0031] This invention provides a tunable LIF neuron hardware unit based on electrochemical doping and a corresponding fabrication method, which has the following advantages:

[0032] The tunable LIF neuron hardware unit in this invention, based on electrochemical doping, allows for the independent tuning of the membrane potential time constant and firing threshold voltage of the LIF neuron. This enables adaptive optimization of performance in different application environments, thereby further improving the performance of the spiking neural network. Furthermore, the electrical equations of the hardware unit correspond to the mathematical model of the LIF neuron in the spiking neural network, reducing network performance degradation caused by errors in hardware simulation. The proposed fabrication method also enables the reuse of multiple material types, reducing processing difficulty and cost. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the tunable LIF neuron hardware unit based on electrochemical doping proposed in this embodiment of the invention. It includes an ECRAM as an adjustable load resistor and a three-terminal threshold switch device with neuron firing function. The gates of these two devices are denoted as the EG terminal and the TG terminal, respectively, and their corresponding gate signals are VEG and VTG, respectively. EG and V TG The node voltage between the two is taken as the membrane potential of the neuron, C in the figure. M This refers to the parasitic capacitance of the entire unit or the external capacitance.

[0034] Figure 2 This is an optical microscopic image (top view) of the tunable LIF neuron hardware unit prepared in an embodiment of the present invention, wherein 2-VO2 film layer, 9-source and drain contact electrodes of the three-terminal threshold switch device, 3-isolation dielectric layer, 4-source and drain contact electrodes of ECRAM, 5-channel layer of ECRAM, 7-passivation layer, and 8-connection electrode.

[0035] Figure 3 This is a schematic diagram of the fabrication process of the adjustable LIF neuron hardware unit in an embodiment of the present invention. Its final structure includes an ECRAM with adjustable load resistance and a cascaded unit of three-terminal threshold switching devices.

[0036] Figure 4 This is a schematic diagram of the structure of the adjustable LIF neuron hardware unit in an embodiment of the present invention, wherein: 1-Al2O3 substrate, 2-VO2 film layer, 3-isolation medium layer, 4-source and drain contact electrodes of ECRAM, 5-channel layer of ECRAM, 6-electrolyte layer, 7-passivation layer, 8-connection electrode, 9-source and drain contact electrodes of three-terminal threshold switching device.

[0037] Figure 5 The diagram shows the conductivity regulation characteristics (a) and retention characteristics (b) of ECRAM in different states in an embodiment of the present invention.

[0038] Figure 6This is a threshold adjustable characteristic data diagram of a three-terminal threshold switch device in an embodiment of the present invention. The initial value of the channel is dynamically doped by electrochemical action to change its initial resistance state, thereby changing its threshold value at the same time.

[0039] Figure 7 The images show the accumulation, leakage, and discharge characteristics of the adjustable LIF neuron hardware unit under different load resistances in this embodiment of the invention, where the corresponding loads are 6kΩ, 8kΩ, and 10kΩ (from left to right).

[0040] Figure 8 The images show the accumulation, leakage, and firing characteristics of the adjustable LIF neuron hardware unit under different firing threshold voltages in this embodiment of the invention, where the corresponding threshold voltages are 3.6V, 2.8V, and 1.9V (from left to right).

[0041] Figure 9 This is a schematic diagram of the spiking neural network structure in an embodiment of the present invention. A 784×10 fully connected structure is used, employing the MNIST dataset and Poisson coding. The neurons in the last layer use... Figure 1 The adjustable LIF neuron hardware unit in the process.

[0042] Figure 10 In this embodiment of the invention, network performance is improved by selecting an appropriate decay time constant.

[0043] Figure 11 In this embodiment of the invention, network performance is improved by selecting an appropriate threshold voltage. Detailed Implementation

[0044] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0045] See Figure 3 The tunable LIF neuron hardware unit is prepared according to the following steps:

[0046] (1) Prepare an alumina substrate, including cleaning and drying the substrate to ensure that the cleanliness of the subsequent coating is not affected;

[0047] (2) 20 nm of VO2 is deposited on the substrate prepared in step (1) by pulsed laser deposition, which is used as the channel layer material of the three-terminal threshold switching device.

[0048] (3) Patterning is performed on the VO2 film prepared in step (2) by electron beam exposure. 5nm Ti and 30nm Au are deposited by electron beam evaporation as source and drain contact electrodes of the three-terminal threshold switch device. Then, the excess material is removed by stripping.

[0049] (4) The VO2 film prepared in step (2) is patterned by electron beam exposure, and a 30 nm HfO2 layer is deposited as an isolation medium layer by atomic layer deposition. Then, the excess material is removed by stripping.

[0050] (5) Patterning is performed on the isolation dielectric layer prepared in step (4) by electron beam exposure. 5nm Ti and 20nm Au are deposited as source and drain contact electrodes of ECRAM by electron beam evaporation. Then, the excess material is removed by stripping.

[0051] (6) Patterning is performed on the isolation dielectric layer prepared in step (4) by electron beam exposure, and 30 nm WO3 is deposited as the channel layer of ECRAM by magnetron sputtering. Then, the excess material is removed by stripping.

[0052] (7) Patterning is performed by electron beam exposure, and 100nm LiPON and 25nm SiO2 are deposited sequentially by magnetron sputtering to serve as the electrolyte layer and passivation layer of ECRAM and three-terminal threshold switching device, respectively. Then, the excess material is removed by stripping.

[0053] (8) Patterning is performed by electron beam exposure, and 20 nm Ti and 200 nm Au are deposited by electron beam evaporation as connecting electrodes for each part. Then, the excess material is removed by stripping.

[0054] Figure 4 This is a schematic diagram of the device structure after all process steps are completed, including: Al2O3 substrate 1, VO2 film layer 2, isolation dielectric layer 3, source and drain contact electrodes of ECRAM 4, channel layer of ECRAM 5, electrolyte layer 5, passivation layer 7, connection electrode 8, and source and drain contact electrodes of three-terminal threshold switch device 9.

[0055] corresponding Figure 2 This is an optical microscopic image of the tunable LIF neuron hardware unit prepared in an embodiment of the present invention.

[0056] Its corresponding Figure 1 This is a schematic diagram of the hardware unit for the tunable LIF neuron ultimately applied in a spiking neural network, which includes a series structure of ECRAM and a three-terminal threshold switch.

[0057] for Figure 1 The electrical characteristics of ECRAM and three-terminal threshold switching devices are as follows: Figure 5 and Figure 6 As shown: Figure 5The diagram shows the conductance regulation characteristics of ECRAM and the retention characteristics of different states in this embodiment of the invention. The channel resistance of ECRAM is changed by applying stimulation to the gate, thereby changing the membrane potential time constant of the LIF neuron. Figure 6 To address the volatile resistive switching characteristics of the three-terminal threshold switching device in this embodiment of the invention, the initial resistive state of the device is altered through electrochemical doping, thereby changing its threshold voltage. Correspondingly, this alters the threshold voltage of the LIF neuron.

[0058] Figure 7 The accumulation, leakage, and firing characteristics of the tunable LIF neuron hardware unit under different load resistances are shown, with corresponding loads of 6kΩ, 8kΩ, and 10kΩ (from left to right). It can be seen that as the load resistance increases, the corresponding membrane potential time constant also increases, and the leakage speed becomes slower.

[0059] Figure 8 The accumulation, leakage, and firing characteristics of the adjustable LIF neuron hardware unit under different firing threshold voltages are shown, with the corresponding threshold voltages being 3.6V, 2.8V, and 1.9V (from left to right). As the threshold voltage decreases, the firing frequency also increases.

[0060] In the network model validation using this tunable LIF neuron hardware unit, the following methods were employed: Figure 9 The spiking neural network shown is a 784×10 fully connected model, using the MNIST dataset and Poisson encoding. The neurons in the last layer employ... Figure 1 The tunable LIF neuron hardware unit in the network can further improve network performance by changing the decay time constant and threshold voltage, respectively, as shown below. Figure 10 and Figure 11 As shown.

[0061] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A hardware unit for a tunable LIF neuron based on electrochemical doping, comprising an electrochemical random access memory as an adjustable load resistor and a three-terminal threshold switch with neuron firing function, wherein: The electrochemical random access memory is a three-terminal memory device including a source, a drain, a channel layer, a gate electrolyte layer, a passivation layer, and a gate. The source and drain are located at the two ends of the channel layer, the gate electrolyte layer is located on the channel layer, and the gate electrolyte layer contains mobile ions; the passivation layer covers the gate electrolyte layer, and the gate is located on the passivation layer. The conductivity of the channel is dynamically adjusted by applying a stimulation signal to the gate. The three-terminal threshold switch device with ion emission function includes a source, a drain, a channel layer, a gate electrolyte layer, a passivation layer, and a gate. The source and drain are located at opposite ends of the channel layer, and the gate electrolyte layer is located on the channel layer, containing mobile ions. The passivation layer covers the gate electrolyte layer, and the gate is located on the passivation layer. The material of the channel layer has an insulator-metal phase transition characteristic under electric field stimulation. By applying a stimulation signal to the gate to change the initial resistance of the channel layer, the threshold voltage of the channel threshold switch device is changed accordingly. Moreover, this threshold voltage change is non-volatile, that is, the characteristic is retained even after the gate signal is removed. The source of the electrochemical random access memory is connected to the drain of the three-terminal threshold switch, so that the two are connected in series through their respective channels.

2. The tunable LIF neuron hardware unit as described in claim 1, characterized in that, In the electrochemical random access memory that serves as an adjustable load resistor, the channel layer is made of tungsten oxide, niobium oxide, vanadium oxide, molybdenum oxide, zinc oxide, molybdenum disulfide, IGZO, or ITO; the gate electrolyte layer is a solid electrolyte membrane made of LiPON, LATP, or PSG; the passivation layer is made of SiO2, Al2O3, Si3N4, HfO2, or Ta2O5; and the source, drain, and gate are made of Ti, Cr, Sc, Al, Pd, Au, Pt, W, TiN, or TaN.

3. The tunable LIF neuron hardware unit as described in claim 1, characterized in that, In the three-terminal threshold switching device, the channel layer is made of VO2, V3O5, or NbO2; the gate electrolyte layer is a solid electrolyte membrane made of LiPON, LATP, or PSG; the passivation layer is made of SiO2, Al2O3, Si3N4, HfO2, or Ta2O5; and the source, drain, and gate are made of Ti, Cr, Sc, Al, Pd, Au, Pt, W, TiN, or TaN.

4. A method for fabricating the tunable LIF neuron hardware unit according to any one of claims 1 to 3, comprising the following steps: 1) Prepare the substrate, including cleaning and drying the substrate; 2) Deposit the channel layer material for the three-terminal threshold switching device on the substrate; 3) Fabricate the source and drain contact electrodes of the three-terminal threshold switch device; 4) Pattern the region where the electrochemical random access memory is designed and fabricated, and deposit an isolation dielectric layer; 5) Fabricate source and drain contact electrodes for the electrochemical random access memory on the isolation dielectric layer deposited in step 4); 6) Pattern and deposit the channel layer of the electrochemical random access memory on the isolation dielectric layer deposited in step 4); 7) Fabrication of the gate electrolyte layer and passivation layer for electrochemical random access memory and three-terminal threshold switching devices; 8) Fabricate source, drain, and gate connection electrodes for electrochemical random access memory and three-terminal threshold switching devices.

5. The preparation method according to claim 4, characterized in that, In step 7), the gate electrolyte layer of the electrochemical random access memory and the three-terminal threshold switch is made of the same material and is fabricated simultaneously with the passivation layer.

6. The application of the tunable LIF neuron hardware unit according to any one of claims 1 to 3 as a LIF neuron in a spiking neural network, characterized in that, The membrane potential of the LIF neurons satisfy: in, The membrane capacitance of LIF neurons; The channel resistance of the electrochemical random access memory; This refers to the high-resistance state of the channel of the three-terminal threshold switch device, i.e., the resistance state before the threshold transition occurs. The input is an external stimulus; the resting potential of this LIF neuron is 0. This is the membrane potential time constant of LIF neurons. The threshold voltage of the channel of the three-terminal threshold switch is the threshold voltage of the LIF neuron. .

7. The application as described in claim 6, characterized in that, By changing the channel resistance of the electrochemical random access memory and the channel threshold voltage of the three-terminal threshold switch, the membrane potential time constant of the LIF neuron is affected. and threshold voltage Make adaptive adjustments.

8. The application as described in claim 6, characterized in that, The tunable LIF neuron hardware unit serves as the neuron in the last layer of a fully connected spiking neural network.

Images