Optical neural unit device and construction method of deep learning optical neural network

Abstract

The application discloses a kind of optical nerve unit device, the construction method of deep learning optical nerve network, it is related to neural network technical field, including contact metal electrode, graphical graphene array, dielectric layer film and substrate, the graphical graphene array is contacted with dielectric layer film formed on substrate, the contact metal electrode is contacted with graphical graphene array and is used to generate electric field with substrate, regulate device output optical signal as the input weight of synapse, the graphical graphene array includes graphene and infrared light source wave, the graphene is single-layer graphene, the graphical graphene array unit is periodic strip, the beneficial effects of the application are: compared with other materials, graphical graphene is due to high carrier density and support surface plasmon propagation, provides an incomparable advantage, i.e. in room temperature to terahertz range in the wavelength of mid-infrared can be in situ dynamically tuned by using gate voltage.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, specifically to an optical neural unit device and a method for constructing a deep learning optical neural network. Background Technology

[0002] Due to the unique extraction capabilities of the retina, the human visual system can recognize various information and objects in complex environments, greatly inspiring the development of Convolutional Neural Networks (CNNs). The retina can essentially be viewed as an adjustable convolutional filter, extracting different features of an image based on brightness, contrast, saturation, and color balance. Similarly, CNNs require numerous specialized filters to extract different information from the input image before constructing a fully connected network. In recent years, Complementary Metal-Oxide-Semiconductor (CMOS) has been proposed as a convolutional filter, but it has been significantly limited by bottlenecks. Therefore, the development of novel in-memory computing devices with high energy efficiency has attracted widespread attention. One solution is to use cross arrays of synaptic memory, such as Resistive Random Access Memory (RRAM) or Phase-Change Random Access Memory (PCRAM), to implement in-memory computing and Artificial Neural Network (ANN) algorithms. Furthermore, three-terminal devices such as Ferroelectric Random Access Memory (FeRAM) and Ion Floating Gate Memory can also overcome efficiency bottlenecks through parallel programming and readout. In these devices, their conductance (synaptic weights) can be adjusted via voltage pulses during learning using a training dataset and a backpropagation (BP) algorithm. However, these devices (including RRAM, PCRAM, and FeRAM) still have some drawbacks that significantly hinder the development of future neural networks. For example: 1. Due to sudden changes after setting or resetting, they are inevitably affected by nonlinear and asymmetric conductance, which rapidly reduces the accuracy of ANNs. Although alternatives to linear and symmetric ion floating gate memories have been proposed, their instability due to lithium-ion or hydrogen-ion injection still cannot guarantee ANN accuracy. 2. Because the relationship between conductance and voltage pulses is unclear, the synapses proposed by these schemes usually require additional control circuitry and standard memory to reset synaptic weights. The theoretical conductance calculated using the BP algorithm needs to be stored in a standard memory implemented by static random access memory (SRAM) before being written, and the decision circuit and comparator circuit need to execute an increment or decrement voltage pulse.

[0003] Optical neural units (optical synapses), whose weights can be easily adjusted via all-optical input, enable an ideal combination of photonics and electronics. Currently, various optical synapses, utilizing optical diffraction and interference, have been reported for machine learning applications. However, these devices still face the challenge of compact assembly techniques for future large-scale integration; in-situ tunable optical neural units and networks have not yet been constructed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an optical neural unit device and a method for constructing a deep learning optical neural network, thus solving the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: an optical neural unit device, comprising a contact metal electrode, a patterned graphene array, a dielectric layer film and a substrate, wherein the patterned graphene array is in contact with the dielectric layer film formed on the substrate, the contact metal electrode is in contact with the patterned graphene array and together with the substrate is used to generate an electric field, and the control device outputs an optical signal as the input weight of the synapse.

[0006] Preferably, the patterned graphene array comprises graphene and an infrared light source, wherein the graphene is a single layer of graphene, the patterned graphene array unit is a periodic strip, the period width of the patterned graphene array unit is 400 nm, the width of the graphene strip is 80 nm, 120 nm, 160 nm, 180 nm, 220 nm and 300 nm, and the wavelength of the infrared light source is 5 to 50 μm;

[0007] Preferably, the contact metal electrode is a gold, titanium, palladium, or nickel metal electrode.

[0008] Preferably, the dielectric layer film is a high-quality oxide film grown on a substrate, the contact metal electrode is preferably hafnium dioxide (HfO2) or silicon dioxide (SiO2), and the thickness of the dielectric layer film is preferably 300 nm.

[0009] Preferably, the substrate is made of silicon.

[0010] A method for constructing a deep learning optical neural network includes the following steps:

[0011] S1. Encode the target image: Use grayscale values, and the image is preferably encoded as 256 pixels × 256 pixels;

[0012] S2. Initialize synaptic weights: Use the light output power of the optical neural unit, the magnitude of which is dynamically adjusted by voltage.

[0013] S3, Output total photosynaptic value: Output by weighted summation, with the weighting method being P = (η1 - η2)P0;

[0014] S4. Convolutional Deep Learning: Deep learning is performed using fully connected convolutional neural networks.

[0015] S5. Target Image Recognition: Image recognition is determined by accuracy.

[0016] Preferably, in S3 and the total value of output photosynapses, P and P0 represent the input and output optical power, respectively, and η1 and η2 represent the transmittance of the two photoneural units, respectively.

[0017] This invention provides an optical neural unit device and a method for constructing a deep learning optical neural network, which have the following beneficial effects:

[0018] 1. Compared with other materials, patterned graphene offers an unparalleled advantage in constructing this opto-neural unit device and deep learning optical neural network due to its high carrier density and support for surface plasmon resonance (SPP) propagation. Specifically, wavelengths in the mid-infrared to terahertz range can be dynamically tuned in situ using gate voltage at room temperature. Graphene SPP filters can easily and stably process input optical signals, and the output optical power can be easily adjusted by changing the Fermi level of graphene. The transmittance of the output light can be considered as synaptic weights, meaning it can be continuously and precisely controlled in situ using an applied voltage. Optical morphological neural networks are theoretically faster than electronic morphological synaptic devices, allowing for rapid updates of new weights calculated by the BP algorithm. Systems based on graphene opto-neurons are simpler than traditional RRAMs, eliminating the need for standard SRAM or comparator circuits. When in operation, similar to the human retina, patterned graphene arrays can preprocess and extract important features of input images. Graphene's high compatibility with mature silicon-based devices makes it possible to mass-produce optoelectronic networks and multi-pixel CMOS readout circuits at low cost. Attached Figure Description

[0019] Figure 1 This is the front view of the present invention;

[0020] Figure 2 This is a top view of the present invention;

[0021] Figure 3 This is a schematic diagram of the construction process of the present invention;

[0022] Figure 4 This is a graph showing the relationship between the width of the graphene strip and the position of the resonance peak in the photoresponse of the present invention.

[0023] Figure 5 This is a diagram showing the relationship between the applied voltage and the position of the resonance peak in the photoresponse of this invention.

[0024] Figure 6 This is a schematic diagram of the structure of the present invention.

[0025] In the figure: 1. Contact metal electrode; 2. Patterned graphene array; 3. Dielectric thin film; 4. Substrate. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Example 1:

[0028] Please see Figures 1 to 6 The present invention provides a technical solution: an optical neural unit device, comprising a contact metal electrode 1, a patterned graphene array 2, a dielectric thin film 3, and a substrate 4. The patterned graphene array 2 is in contact with the dielectric thin film 3 formed on the substrate 4 to receive input optical signals. The contact metal electrode 1 is in contact with the patterned graphene array 2 and together with the substrate 4 is used to generate an electric field. The output optical signal of the control device is used as the input weight of the synapse.

[0029] Furthermore, the patterned graphene array 2 includes graphene and an infrared light source. The graphene is a single layer of graphene, and the units of the patterned graphene array 2 are periodic strips. The period width of the units of the patterned graphene array 2 is 400 nm, and the width of the graphene strips is 80 nm, 120 nm, 160 nm, 180 nm, 220 nm and 300 nm. The wavelength of the infrared light source is 5 to 50 μm.

[0030] Furthermore, the contact metal electrode 1 is preferably a gold, titanium, palladium, or nickel metal electrode.

[0031] Furthermore, the dielectric layer film 3 is a high-quality oxide film grown on the substrate 4, the contact metal electrode 1 is preferably hafnium dioxide (HfO2) or silicon dioxide (SiO2), and the thickness of the dielectric layer film 3 is preferably 300 nm.

[0032] Furthermore, the substrate 4 is preferably made of silicon.

[0033] The dielectric film 3 is made of silicon dioxide (SiO2), the substrate 4 is made of silicon, the contact metal electrode 1 is made of gold, and the graphene is a single layer of graphene. The SiO2 thickness is 300 nm, the Au thickness is 40 nm, and the silicon thickness is 500 μm. The single layer of graphene is periodically patterned into periodic stripes with a unit period width of 400 nm and graphene strip widths of 80, 120, 160, 180, 220, and 300 nm. An applied voltage of -60 V can achieve a thickness of 420 to 972 cm⁻¹. -1 Light transmittance regulation.

[0034] Example 2:

[0035] Please see Figures 1 to 6Unlike Example 1, the dielectric film 3 is made of silicon dioxide (SiO2), the substrate 4 is made of silicon, the contact metal electrode 1 is made of gold, and the graphene is a single layer of graphene. The SiO2 thickness is 300 nm, the Au thickness is 40 nm, and the silicon thickness is 500 μm. The single layer of graphene is periodically patterned into a periodic strip pattern with a unit period width of 400 nm and a graphene strip width of 180 nm. The applied voltages are -90, -60, -40, -30, and -20 V, which can achieve a thickness of 400 to 600 cm⁻¹. -1 In-situ dynamic regulation of light transmittance.

[0036] A method for constructing a deep learning optical neural network includes the following steps:

[0037] S1. Encode the target image: Use grayscale values, and the image is preferably encoded as 256 pixels × 256 pixels;

[0038] S2. Initialize synaptic weights: Use the light output power of the optical neural unit, the magnitude of which is dynamically adjusted by voltage.

[0039] S3, Output total photosynaptic value: Output by weighted summation, with the weighting method being P = (η1 - η2)P0;

[0040] S4. Convolutional Deep Learning: Deep learning is performed using fully connected convolutional neural networks.

[0041] S5. Target Image Recognition: Image recognition is determined by accuracy.

[0042] Furthermore, in S3 and the total output photosynaptic value, P and P0 represent the input and output optical power, respectively, and η1 and η2 represent the transmittance of the two photoneurons, respectively.

[0043] The mandrill can be selected as the target image, and grayscale encoding is used. The pixel array is 256 pixels × 256 pixels. The grayscale value is used as the initial input weight. For example, if the grayscale value is 1, the initial weight is 1, and if the grayscale value is 255, the initial weight is 255. The light transmittance, input light power, and output light power of the optical neural unit are calculated. A fully connected convolutional neural network is constructed for deep learning. The target image recognition accuracy can reach 98.4%.

[0044] In summary, the optical neural unit device and the method for constructing a deep learning optical neural network, when in use, achieve in-situ tunable light output within the range of 5 to 50 μm by adjusting the direction and magnitude of the voltage applied between the bottom electrode and the contact electrode. Preferably, the voltage applied between the bottom electrode and the contact electrode is -90V to +90V.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An optical neuron device comprising a contact metal electrode (1), a patterned graphene array (2), a dielectric layer film (3) and a substrate (4), characterized in that: The patterned graphene array (2) is in contact with a dielectric layer film (3) formed on a substrate (4), the contact metal electrode (1) is in contact with the patterned graphene array (2) and is used together with the substrate (4) to generate an electric field, and the regulating device outputs an optical signal as an input weight of a synapse. The patterned graphene array (2) comprises graphene and infrared light source waves, the graphene is single-layer graphene, the unit of the patterned graphene array (2) is a periodic strip, the periodic width of the unit of the patterned graphene array (2) is 400 nm, the strip width of the graphene is 80 nm, 120 nm, 160 nm, 180 nm, 220 nm and 300 nm, and the wavelength of the infrared light source is 5-50 μm.

2. The optical neuron device of claim 1, wherein: The contact metal electrode (1) is a gold, titanium, palladium or nickel metal electrode.

3. The optical neuron device of claim 1, wherein: The dielectric layer film (3) is an oxide film of high quality grown on the substrate (4), and the thickness of the dielectric layer film (3) is 300 nm.

4. The optical neuron device of claim 1, wherein: The material of the substrate (4) is silicon.

5. The optical neuron device constructing a deep learning optical neural network according to any one of claims 1-4, characterized in that, The method comprises the following steps: S1, encoding a target image: using a gray value, the image is encoded as 256 pixels x 256 pixels; S2, initializing a synapse weight: using the light output power of an optical neuron, the size of which is dynamically regulated by voltage; S3, outputting a total value of an optical synapse: outputting by means of weighted summation; S4, convolutional deep learning: using a fully connected convolutional neural network for deep learning; S5, target image recognition: determining image recognition by accuracy.

Images