A bionic polarization vision chip device and its preparation method, and convolutional neural network

Abstract

The present invention relates to the field of semiconductor chip technology, and in particular to a bionic polarization vision chip device and a preparation method thereof, and a convolutional neural network. The device comprises: a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure prepared by two-dimensional materials, the structure comprising a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer stacked in sequence; the channel layer has anisotropy, and the ferroelectric layer controls the conductivity in the channel layer by switching the polarization switch under the induction of polarized light, which is used for polarization image recognition. In the present invention, a bionic polarization vision chip device is prepared by using two-dimensional materials, and the atomic layers of the two-dimensional materials are combined by weak van der Waals forces, and can be stacked on each other to form a van der Waals heterostructure without considering the problem of lattice mismatch. The anisotropy of the channel layer gives it polarization sensitivity, and the polarization effect of the ferroelectric layer can change the conductivity in the channel layer, thereby realizing the recognition of polarized images.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor chips, and in particular to a bionic polarization vision chip device and a preparation method thereof, and a convolutional neural network. Background Art

[0002] With the increasing popularity of diverse intelligent applications such as robots, unmanned driving and visual recognition, the current polarization image sensing architecture faces huge challenges. Traditional polarization image sensing systems mainly use complementary metal oxide semiconductor polarization image sensors and von Neumann architecture for polarization image sensing and processing respectively. Due to the physical separation of storage and computing in the von Neumann architecture, the energy efficiency problem caused by the frequent transmission of data is becoming more and more obvious. The polarization vision chip industry must make urgent innovations at the level of the underlying computing architecture. Inspired by the human retina, the neuromorphic vision sensor architecture with multiple functions such as integrated image perception, memory and computing helps to avoid the power consumption bottleneck of data processing in the von Neumann architecture, improve overall efficiency and reduce computational delays from detection to computing and redundant data storage in image sensors. Compared with the human visual system, the animal visual system has more advanced features in some aspects. Taking the desert locust as an example, its excellent polarization information sensing, storage and computing capabilities provide more dimensional degrees of freedom for the design of bionic visual information systems. This feature allows the system to obtain visual information more comprehensively, enhances the ability to detect and identify targets, and has unique applications in polarization navigation, encrypted communications, geological remote sensing and machine vision. Therefore, the bionic polarization vision chip device inspired by biological vision integrates polarization information perception, storage and computing capabilities, which is expected to overcome the limitations of traditional polarization image sensing systems based on the von Neumann architecture and provide a full hardware solution for realizing more compact, efficient and multi-dimensional bionic vision chips.

[0003] However, the material system used in current bionic polarization vision chip devices is traditional organic / inorganic materials, which often leads to challenges of thermal expansion coefficient differences and lattice mismatch when constructing multilayer ferroelectric heterostructures, which has become an important obstacle for the bionic vision system to integrate more functions (such as polarization-sensitive properties). In addition, due to the limitations of its defect state charge capture mechanism, it is difficult to ensure non-volatility and high-speed storage at the same time under the constraints of voltage-time parameters, and the high voltage required for writing and erasing (about 10 V to 20 V) limits the storage speed and increases the energy consumption burden. Furthermore, technical challenges such as interface compatibility, defect control, and effective integration of multilayer ferroelectric heterojunctions are still the main obstacles faced by bionic polarization vision chips for high-dimensional information processing and high-speed data storage in the process of energy-saving full hardware implementation. Summary of the invention

[0004] In view of this, the present invention provides a bionic polarization vision chip device and a preparation method thereof, and a convolutional neural network to solve the technical problem that the bionic polarization vision chip device prepared in the prior art is difficult to integrate polarization-sensitive characteristics.

[0005] The technical solution provided by the present invention is as follows:

[0006] The first aspect of the present invention provides a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, the device comprising: a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure prepared using a two-dimensional material, the structure comprising a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer stacked in sequence; the channel layer has anisotropy, and the ferroelectric layer controls the conductivity in the channel layer by switching the polarization switch under the induction of polarized light, so as to be used for polarized image recognition.

[0007] In the present invention, a bionic polarization vision chip device is prepared using a two-dimensional material, and the atomic layers of the two-dimensional material are combined by weak van der Waals forces, and can be stacked on each other to form a van der Waals heterostructure without considering the problem of lattice mismatch. Therefore, the bionic polarization vision chip device of the present invention avoids the problem of lattice mismatch existing in the multilayer ferroelectric heterostructure constructed by traditional organic / inorganic materials. At the same time, the anisotropy of the channel layer gives it polarization sensitivity, and the ferroelectric layer uses the property of the ferroelectric material to change the polarization direction under an applied electric field, which can change the conductivity in the channel layer, thereby realizing the recognition of polarized images.

[0008] In an optional embodiment, the bottom electrode layer and the input polarized light are used as a presynaptic membrane, the ferroelectric layer, the intermediate metal layer and the insulating layer are used as presynaptic neurons, the channel layer is used as postsynaptic neurons, and the weight between the presynaptic neuron and the postsynaptic neuron is used to be adjusted under the action of electrical stimulation and / or light stimulation to achieve learning and memory of polarized visual information.

[0009] In the present invention, the unique polarization switch switching mechanism of the ferroelectric layer is used to enable it to have ultra-low power consumption reading and writing capabilities and high-speed non-volatile storage characteristics. Furthermore, by utilizing the characteristics of no dangling bonds on the surface of two-dimensional ferroelectric materials and the ability of seamless integration, a multi-layer two-dimensional ferroelectric heterojunction can be constructed, and combined with the high sensitivity of the channel layer to polarized light, the size of the synaptic weight can be adjusted through electrical stimulation and optical stimulation, thereby realizing the construction of a bionic polarization vision chip device with more efficient high-dimensional information storage and processing capabilities.

[0010] In an optional embodiment, the channel layer is a palladium diselenide channel layer, a rhenium disulfide channel layer, a black phosphorus channel layer, a germanium selenide channel layer or a tellurium channel layer.

[0011] In the present invention, palladium diselenide, rhenium disulfide, black phosphorus, germanium selenide or tellurium and other anisotropic materials are used as the channel layer, so that the conductivity of the channel layer can be changed by electrical stimulation and light stimulation to simulate the weight change in biological synapses.

[0012] In an optional embodiment, the ferroelectric layer is an indium-doped copper sulfide phosphide ferroelectric layer, a two-dimensional barium titanate ferroelectric layer, a two-dimensional bismuth ferrite ferroelectric layer, or a twisted double-layer nitride ferroelectric layer.

[0013] In the present invention, indium-doped copper sulfide phosphide, two-dimensional barium titanate, two-dimensional bismuth ferrite or twisted double-layer nitride and other two-dimensional materials with out-of-plane polarization and insulation are used as the ferroelectric layer, which can make the ferroelectric layer have a polarization effect and maintain the conductivity in the channel layer.

[0014] In an optional implementation, the insulating layer is a boron nitride insulating layer, a hafnium oxide insulating layer, an aluminum oxide insulating layer, a titanium oxide insulating layer or a cobalt oxide insulating layer.

[0015] In the present invention, a boron nitride insulating layer, a hafnium oxide insulating layer, an aluminum oxide insulating layer, a titanium oxide insulating layer or a cobalt oxide insulating layer is provided between the intermediate metal layer and the channel layer, which can play an insulating role, reduce leakage current and improve the stability of the device.

[0016] In an optional embodiment, the bottom electrode layer is a multi-layer graphene bottom electrode layer, and the intermediate metal layer is a few-layer graphene intermediate metal layer.

[0017] In the present invention, a multi-layer graphene bottom electrode layer is used to enable the device to have better gate control capability. A few-layer graphene intermediate metal layer is used to compensate for the residual polarization in the ferroelectric layer and prolong the retention time of the device conductivity.

[0018] The second aspect of the present invention provides a convolutional neural network, comprising: a bionic polarization vision chip device array composed of the bionic polarization vision chip devices based on two-dimensional ferroelectric heterojunction described in the first aspect of the present invention and any item of the first aspect, wherein the bionic polarization vision chip device array is used to perform convolution operations on an input polarization image.

[0019] In an optional embodiment, the convolutional neural network is a U-net convolutional neural network, the first layer of the U-net convolutional neural network includes a bionic polarization vision chip device array and a first convolutional layer, and the input voltage of the bionic polarization vision chip device array is determined based on the parameters of the first convolutional layer.

[0020] In the present invention, a bionic polarization vision chip device array composed of bionic polarization vision chip devices is arranged in a convolutional neural network, thereby combining the image feature enhancement and refinement of the U-net convolutional neural network, de-mosaicing and classifying the polarization image, and at the same time improving the polarization information learning ability of the convolutional neural network.

[0021] The third aspect of the present invention provides a method for preparing a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, the method comprising: using a mechanical stripping method to strip off the materials for preparing a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer to obtain a two-dimensional material sample, wherein the two-dimensional material sample comprises a bottom electrode layer sample, a ferroelectric layer sample, an intermediate metal layer sample, an insulating layer sample and a channel layer sample respectively formed on a glass slide; transferring the bottom electrode layer sample, the ferroelectric layer sample, the intermediate metal layer sample, the insulating layer sample and the channel layer sample formed on the glass slide to a silicon wafer to obtain a stacked bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer, wherein the stacked bottom electrode layer, the ferroelectric layer, the intermediate metal layer, the insulating layer and the channel layer constitute a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure.

[0022] In an optional embodiment, the method further comprises: forming a source-drain electrode layer on a surface of the channel layer that is away from the insulating layer.

[0023] In the present invention, a mechanical stripping method and a sample layer transfer method are used to form the desired device structure. Not only is the preparation method simple, does not require complex instruments and equipment, and is low-cost, but the two-dimensional material sample formed also has many advantages such as few defects, a smooth surface, and high mobility, thereby further improving the performance of the prepared device. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 A schematic diagram of the structure of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0026] Figure 2 It is a flow chart of a method for preparing a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0027] Figure 3aA schematic diagram of the structure of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0028] Figure 3b This is a light microscope image of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0029] Figure 4a Schematic diagram of leakage characteristics of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0030] Figure 4b Schematic diagram of transfer characteristics of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0031] Figure 5a Schematic diagram of the functional relationship between the EPSC peak value (ΔEPSC) and the polarization angle of the bionic polarization vision chip device based on the two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0032] Figure 5b It is a schematic diagram of the change of the PPF index with Δt of the bionic polarization vision chip device based on the two-dimensional ferroelectric heterojunction in the embodiment of the present invention under the stimulation of light pulses with polarization angles of 0° and 90°;

[0033] Figure 5c Schematic diagram of the functional relationship between the Gmax / Gmin ratio and the polarization angle of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0034] Figure 6 A schematic diagram of polarization-related image storage and recognition of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction in an embodiment of the present invention;

[0035] Figure 7 This is a schematic diagram of a bionic polarization vision chip array in an embodiment of the present invention;

[0036] Figure 8 Schematic diagrams showing original images and demosaiced images at four polarization angles in an embodiment of the present invention;

[0037] Fig. 9 Schematic diagram of polarization-related recognition accuracy and training cycle of a convolutional neural network including a bionic polarization vision chip device array and not including a bionic polarization vision chip device array in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0040] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, it can also be the internal connection of two components, it can be a wireless connection, or it can be a wired connection. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0041] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0042] The embodiment of the present invention provides a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, the device comprising: a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure prepared by a two-dimensional material, such as Figure 1 As shown, the structure includes a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer stacked in sequence; the channel layer has anisotropy, and the ferroelectric layer controls the conductivity in the channel layer by switching the polarization switch under the induction of polarized light, which is used for polarization image recognition.

[0043] Specifically, the bionic polarization vision chip device of the present embodiment is prepared by two-dimensional materials, and the atomic layers of the two-dimensional materials are bound by weak van der Waals forces and can be stacked on each other to form a van der Waals heterostructure without considering the problem of lattice mismatch. Therefore, the bionic polarization vision chip device of the present embodiment avoids the problem of lattice mismatch existing in the multilayer ferroelectric heterostructure constructed by traditional organic / inorganic materials. At the same time, the anisotropy of the channel layer gives it polarization sensitivity, and the ferroelectric layer uses the property of the ferroelectric material to change the polarization direction under an applied electric field, which can change the conductivity in the channel layer, thereby realizing the recognition of polarized images.

[0044] Among them, when performing polarization image recognition, polarization-related image recognition can be achieved by using light signals with different polarization angles as input and single-pixel imaging technology, and simulating the polarization perception and memory functions in biological vision. Specifically, the optical mask can be illuminated by a laser outputting non-polarized light to form a pattern of preset letters and projected on the surface of the device. It should be noted that when the laser passes through a polarizer, it changes from non-polarized light to polarized light, and the change in the laser polarization angle can be controlled by a half-wave plate. Finally, a preset letter pattern image with polarization information is formed and projected on the surface of the device to complete the image recognition process.

[0045] In an optional embodiment, the bottom electrode layer and the input polarized light are used as a presynaptic membrane, the ferroelectric layer, the intermediate metal layer and the insulating layer are used as presynaptic neurons, the channel layer is used as postsynaptic neurons, and the weight between the presynaptic neuron and the postsynaptic neuron is used to be adjusted under the action of electrical stimulation and / or light stimulation to achieve learning and memory of polarized visual information.

[0046] Specifically, in addition to the recognition of polarized images, the device can also process images through changes in conductance under electrical and optical stimulation. Among them, by simulating the presynaptic neurons and postsynaptic neurons formed between the various layers of the device structure, the polarization-related long-term potentiation (LTP) and electrically adjustable long-term depression (LTD) conductance changes are induced based on programmable light pulses and electrical pulses, respectively, so that the device can simulate the multi-signal regulation mechanism in biological neural synapses.

[0047] In the case of LTP, if a series of strong pulses arrive continuously, the synaptic weight may increase; while in the case of LTD, if the pulses are fewer or weaker, the synaptic weight may decrease. Therefore, by modulating multidimensional modulation signals such as pulse intensity, pulse number, pulse width and polarization angle, the combination form of LTP / LTD multi-level conductance states can be updated, so that the device can better simulate the behavior of biological neural networks and obtain LTP / LTD behavior curves with better linearity and more conductance states. Therefore, by modulating multidimensional modulation signals, operations such as filtering and feature extraction can be directly performed on the input image, thereby realizing functions such as image preprocessing. At the same time, due to the effect of ferroelectric polarization, the device has a memory behavior similar to that of human synapses, so the input image information can be stored and erased. In addition, by outputting a photocurrent equal to the input optical power divided by the voltage, the device can also realize memory computing tasks, and the data does not need to be moved from the memory to the processor for processing during calculation, which can greatly reduce the delay and energy consumption caused by data movement.

[0048] In addition, after the device is prepared, the device can also be subjected to polarization-dependent synaptic plasticity tests, including tests of synaptic behaviors such as excitatory postsynaptic current, double-pulse facilitation and long-term plasticity. During the test, the conductivity of the channel layer can be controlled by switching the polarization switch in the ferroelectric layer induced by polarized light to simulate the polarization-dependent synaptic weight changes in biological visual neurons.

[0049] In an optional embodiment, the channel layer is a palladium diselenide channel layer, a rhenium disulfide channel layer, a black phosphorus channel layer, a germanium selenide channel layer or a tellurium channel layer. The ferroelectric layer is an indium-doped copper sulfide phosphide ferroelectric layer, a two-dimensional barium titanate ferroelectric layer, a two-dimensional bismuth ferroelectric layer or a twisted double-layer nitride ferroelectric layer. The insulating layer is a boron nitride insulating layer, a hafnium oxide insulating layer, an aluminum oxide insulating layer, a titanium oxide insulating layer or a cobalt oxide insulating layer. The bottom electrode layer is a multi-layer graphene bottom electrode layer, and the intermediate metal layer is a few-layer graphene intermediate metal layer.

[0050] Specifically, the bottom electrode layer uses multilayer graphene, which can make the device have better gate control ability. The ferroelectric layer uses a two-dimensional material with out-of-plane polarization and insulation, for example, an indium-doped copper sulfide phosphide ferroelectric layer, a two-dimensional barium titanate ferroelectric layer, a two-dimensional bismuth ferrite ferroelectric layer or a twisted double-layer nitride ferroelectric layer. When an external electric field is applied, the polarization direction in the ferroelectric layer switches, that is, an induced polarization phenomenon occurs, which helps to maintain the conductivity in the channel layer. The intermediate metal layer uses a few layers of graphene, which can compensate for the residual polarization in the ferroelectric layer and extend the retention time of the device conductivity. The insulating layer can use ultra-thin (such as less than 20nm) and insulating materials, thereby reducing the thickness of the device, and playing an insulating role, reducing leakage current, and improving the stability of the device. For example, a boron nitride insulating layer, a hafnium oxide insulating layer, an aluminum oxide insulating layer, a titanium oxide insulating layer or a cobalt oxide insulating layer can be used. The channel layer uses an anisotropic material to give the channel layer polarization sensitivity. At the same time, the channel layer can also serve as a polarized light receiving layer, and the conductance of the channel layer changes through electrical stimulation and optical stimulation to simulate the weight change in biological synapses. Specifically, the channel layer can be a palladium diselenide channel layer, a rhenium disulfide channel layer, a black phosphorus channel layer, a germanium selenide channel layer, or a tellurium channel layer.

[0051] The present embodiment provides a convolutional neural network, including: a bionic polarization vision chip device array composed of a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction of the above embodiment, wherein the bionic polarization vision chip device array is used to perform a convolution operation on an input polarization image. Specifically, 9 bionic polarization vision chip devices can be used to construct a 3×3 bionic polarization vision chip device array, and the array can use the relationship between voltage and current to implement the convolution operation on the polarization image.

[0052] In an optional embodiment, the convolutional neural network is a U-net convolutional neural network, and the first layer of the U-net convolutional neural network includes a bionic polarization vision chip device array and a first convolutional layer, and the input voltage of the bionic polarization vision chip device array is determined based on the parameters of the first convolutional layer. Specifically, the parameters of the first convolutional layer can be determined by learning the polarization image data, and for the parameters of the determined first convolutional layer, it is converted into a voltage matrix that the device can respond to, and then applied to the device array. Convolution operations are implemented in the device array by Kirchhoff's law and Ohm's law. Among them, in the device array, the flow of current simulates the sliding process of the convolution kernel on the polarization image, that is, the way in which the current flows through the device array is equivalent to the convolution kernel and the polarization image performing a dot product operation. In addition, after the first layer of the U-net convolutional neural network is processed, the output result can be input to the back part of the U-net convolutional neural network for further processing. Thus, the polarization image is feature enhanced and refined by the U-net convolutional neural network to complete the demosaicing and classification tasks of the polarization image.

[0053] The bionic polarization vision chip device based on the two-dimensional ferroelectric heterojunction provided in this embodiment integrates polarization information perception, storage and computing capabilities, effectively overcoming the data processing bottleneck inherent in the traditional von Neumann architecture. The device utilizes the unique polarization switch switching mechanism of ferroelectric materials to enable it to have ultra-low power reading and writing capabilities and high-speed non-volatile storage characteristics. Furthermore, by utilizing the characteristics of no dangling bonds on the surface of two-dimensional ferroelectric materials and the ability of seamless integration, a multi-layer two-dimensional ferroelectric heterojunction can be constructed, and combined with the high sensitivity of the anisotropic channel layer to polarized light, a bionic polarization vision chip device with more efficient high-dimensional information storage and processing capabilities can be constructed.

[0054] At the same time, thanks to its superior ferroelectric polarization switching mechanism and bionic interface engineering, the device achieves a high 6 The device has an on / off ratio of 1.28kJ / 1.2kJ and an extremely low energy consumption of only 0.26 nJ / 124 aJ (photoexcitation / electric inhibition per event). The anisotropy of the channel layer gives it polarization sensitivity. The device successfully simulates a variety of polarization-related synaptic plasticity features and achieves memory imaging at different polarization angles. Based on a fully two-dimensional 3×3 bionic polarization vision unit array and combined with a U-net convolutional neural network, de-mosaicing of polarization images is achieved, and a high recognition accuracy of 99.7% is achieved in classification tasks.

[0055] This embodiment also provides a method for preparing a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, such as Figure 2 As shown, the method comprises the following steps:

[0056] Step S101, using a mechanical stripping method to strip off the materials used to prepare the bottom electrode layer, the ferroelectric layer, the intermediate metal layer, the insulating layer and the channel layer to obtain a two-dimensional material sample, wherein the two-dimensional material sample includes a bottom electrode layer sample, a ferroelectric layer sample, an intermediate metal layer sample, an insulating layer sample and a channel layer sample respectively formed on a glass slide.

[0057] Specifically, the mechanical exfoliation method can also be called micromechanical exfoliation, tape method or tape tearing method. At the same time, since the layers of two-dimensional materials are combined together by weak van der Waals forces, this method can peel out a single layer or a few atomic layers in the two-dimensional material.

[0058] The following uses the preparation of devices using two-dimensional materials such as multilayer graphene (for preparing the bottom electrode layer), indium-doped copper sulfide phosphide (for preparing the ferroelectric layer), few-layer graphene (for preparing the intermediate metal layer), boron nitride (for preparing the insulating layer) and palladium diselenide (for preparing the channel layer) as an example to illustrate the preparation method:

[0059] When peeling off the two-dimensional materials, put the block graphene, indium-doped copper sulfide phosphide, boron nitride and palladium diselenide on the blue tape, press the surface of the material gently with the blade of the scalpel, and then separate the material from the blue tape with tweezers, leaving a relatively thin and uniform two-dimensional block material on the blue tape. Then, fold the blue tape along the edge of the material 4 to 6 times until the material is evenly distributed on the tape and a metallic luster appears on the surface of the material.

[0060] When forming a sample, you can first cut the PDMS into small pieces and stick them on a glass slide, and stick the blue transparent tape with the material on the glass slide on the side with the PDMS, and use a cotton swab to gently press the tape to make the tape completely fit the PDMS. If necessary, you can also use a heating table to heat it to make more materials detach from the tape onto the PDMS. Finally, tear off the tape from the PDMS, and the required material will be pasted on the PDMS. Select the two-dimensional material of the required thickness and size under an optical microscope, and try to select samples with uniform thickness and no other overly thick materials around to obtain the required two-dimensional material samples. Among them, it should be noted that for two-dimensional materials such as multilayer graphene, indium-doped copper sulfide phosphide, few-layer graphene, boron nitride and palladium diselenide, each material can obtain multiple samples by the above-mentioned mechanical stripping method, so as to select samples that meet the requirements from multiple samples.

[0061] Step S102, transferring the bottom electrode layer sample, ferroelectric layer sample, intermediate metal layer sample, insulating layer sample and channel layer sample formed on the glass slide to the silicon wafer to obtain a stacked bottom electrode layer, ferroelectric layer, intermediate metal layer, insulating layer and channel layer. The stacked bottom electrode layer, ferroelectric layer, intermediate metal layer, insulating layer and channel layer constitute a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure.

[0062] Specifically, when transferring to form each layer of the device, the transfer platform can be used to accurately transfer the prepared two-dimensional material samples layer by layer. First, transfer the bottom electrode layer. When transferring, fix the glass slide and silicon wafer with the bottom electrode layer sample to the transfer platform, adjust the left and right front and back knobs to align the material to the target area, adjust the lifting knob to slowly lower the glass slide until the target material on the PDMS is completely attached to the silicon wafer, then turn on the heating button to heat at 60 degrees for 3 minutes. After heating is completed, slowly lift the glass slide to separate the PDMS from the silicon wafer, and the target material is transferred to the silicon wafer.

[0063] The ferroelectric layer sample is then transferred. During the transfer, the glass slide with the ferroelectric layer sample and the silicon wafer with the bottom electrode layer are fixed on a transfer platform. Then, the ferroelectric layer sample on the glass slide is transferred to the surface of the bottom electrode layer away from the silicon wafer in the same way.

[0064] Then, the intermediate metal layer sample is transferred, and the transfer method is the same as the transfer method of the above-mentioned bottom electrode layer, except that during the transfer, the intermediate metal layer sample on the glass slide can be aligned with the bottom electrode layer on the silicon wafer, so that the intermediate metal layer transferred to the surface of the ferroelectric layer on the side away from the bottom electrode layer is located in the middle part of the bottom electrode layer. Then, the insulating layer sample is transferred, and the insulating layer transferred to the surface of the intermediate metal layer on the side away from the ferroelectric layer is located in the middle part of the bottom electrode layer. Finally, the channel layer sample is transferred, and the channel layer after the transfer is also located in the middle part of the bottom electrode layer. Thus, by transferring layer by layer, a multilayer two-dimensional van der Waals vertical ferroelectric heterostructure including a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer stacked in sequence is obtained.

[0065] Step S103, forming a source-drain electrode layer on the surface of the channel layer away from the insulating layer. The source-drain electrode layer can be formed by photolithography, fixing and developing. Specifically, the structure obtained in the above step S102 is subjected to spin coating of photoresist, photolithography, development and fixing of the source-drain electrode pattern to obtain a sample with a source-drain electrode pattern. Then, Ti / Au source-drain electrode evaporation of 20 nm-40 nm Ti metal and 80 nm-100 nm Au metal is performed. After the evaporation is completed, the gold removal operation is performed to obtain the final desired device.

[0066] The preparation method provided in this embodiment adopts mechanical stripping and sample layer transfer to form the desired device structure. Not only is the preparation method simple, does not require complex instruments and equipment, and is low-cost, but the two-dimensional material sample formed has many advantages such as few defects, smooth surface and high mobility, thereby further improving the performance of the prepared device.

[0067] As a specific application embodiment of the embodiment of the present invention, the structure of the bionic polarization vision chip device based on the two-dimensional ferroelectric heterojunction includes a multi-layer graphene (Gr) bottom electrode layer, an indium-doped copper sulfide phosphide (CIPS) ferroelectric layer, a few-layer graphene intermediate metal layer, a boron nitride (h-BN) insulating layer and a palladium diselenide (PdSe2) channel layer stacked in sequence. At the same time, the structure also includes a source (Source), a drain (Drain) and a gate (Gate), such as Figure 3a As shown in FIG. 1 , it is a schematic diagram of the structure of the device. Figure 3b Shown is a light microscope image of the device.

[0068] For the prepared device, its basic electrical properties can be evaluated by electrical measurements. Figure 4a The leakage current curve of the device is shown in Figure 4b The transfer curve of the device is shown in Figure 2. Specifically, the leakage current curve characterizes the leakage characteristics of the device. Figure 4a The device is shown in Figure 2 at a source-drain bias of 0.1 V (V ds ), a low leakage current Igs within the range of ±3 gate voltage Vgs is essential for realizing reliable bionic polarization vision chip devices. The transfer curve characterizes the transfer characteristics of the device. Figure 4b The device is shown in Figure 2 at a source-drain bias of 0.1 V (V ds ) at different gate voltages (V gs ) to obtain the transfer curve (the gate voltage V gs and source-drain current I ds The corresponding curve of V gs When sweeping from -3 V to 3 V, 10 6 High on / off ratio and 0.6 V memory window.

[0069] At the same time, the device can also be tested for polarization-dependent synaptic plasticity. Figure 5a The typical optical excitatory post-synaptic current (EPSC) of the device under light pulses at different polarization angles is shown, where the small balls represent peaks and the curve represents the best fit to the data. It has significant polarization dependence and presents a polarization direction characteristic similar to a four-leaf clover shape. Figure 5b It shows that under two different polarization states, the decay trend of the paired-pulse facilitation (PPF) index is significantly affected by the polarization angle. Specifically, when the polarization angle is 0°, the PPF index shows a gradual downward trend, from the initial 400% to the final 100%. In contrast, within the same (pulse interval) Δt variation range, when the polarization angle increases to 90°, the decay process of the PPF index is steeper, from the initial 800% to 100%. Figure 5c The relationship between the maximum / minimum conductance ratio (Gmax / Gmin) and the polarization angle (Polarization angle) shows a polarization direction characteristic similar to the shape of the letter "M", indicating that the device has a long-term plastic behavior with obvious polarization dependence. The above results show that the polarization angle has a regulatory effect on the PPF characteristics of the device, which is of great significance for simulating polarization-sensitive neuromorphic behavior in biological vision in bionic vision systems.

[0070] In addition, this embodiment also implements the device's memory imaging test of polarization images. Figure 6 The three pictures in the upper part show that when the polarization angle increases from 0° to 30°, the memory image exhibits enhanced color intensity, while at a polarization angle of 60°, the color intensity weakens. This change trend is consistent with the polarization direction characteristics of the four-leaf clover shape, indicating that the device has polarization-dependent image recognition capabilities. Figure 6 The three images in the lower part show the initial image (Initial) at the time of light stimulation and 30 s and 60 s after the light stimulation stopped. During continuous light stimulation, the normalized EPSC (Normalized △EPSC) of the device was observed to increase and a clear image was presented. Over time, the image gradually became blurred, and the memory effect lasted at least 60 s.

[0071] When 9 bionic polarization vision chip devices are used to form a 3×3 bionic polarization vision chip device array, its structure is as follows: Figure 7 As shown, Figure 7 The left half of the middle is a full view of the array structure, and the surrounding leads are the electrode parts of the device, including source, drain electrodes and gate electrodes; Figure 7The right half of the figure is an enlarged part of 9 biomimetic polarization vision chip devices. For the U-net convolutional neural network containing the biomimetic polarization vision chip device array, the U-Net convolutional neural network can be used for polarization image denoising. During denoising, the input image (including color polarization images with four different polarization angles of 0°, 45°, 90° and 135°) is quantized into a conductivity matrix through the array, and the first convolution layer of the network is mapped to the input voltage of the array, thereby realizing the convolution operation in the sensor. Among them, in addition to the biomimetic polarization vision chip device array, the U-net convolutional neural network also includes a skip connection structure, which includes a convolution layer (Conv 3×3 stribe 2) with a convolution kernel size of 3×3 and a stride of 2. The structure uses a rectified linear unit (Rectified Linear Unit, ReLU) function Leaky ReLU as an activation function, and the structure also includes a 2×2 maximum pooling layer (Maxpooling) and an upsampling layer (Upsampling). In addition, for the U-net convolutional neural network, an unsupervised learning strategy can be implemented using a color polarization image dataset. The dataset consists of color polarization images of 40 different scenes, each of which contains four different polarization angles (0°, 45°, 90°, and 135°), and each image has a resolution of 1024×768 pixels and a color depth of 10 bits. Figure 8 It shows that when a 3×3 bionic polarization vision chip device array is used as the input carrier of the U-net convolutional neural network, the denoised polarization image is highly consistent with the original image and can effectively restore the polarization information.

[0072] In addition, in order to further verify the effectiveness of the bionic polarization vision chip device array in the color polarization image classification and recognition task, a comparative test can be conducted, in which the demosaicing test is performed using a convolutional neural network model containing a bionic polarization vision chip device array (witharray) and a convolutional neural network model without a bionic polarization vision chip device array (witharray). For each training of the convolutional neural network, 110 sets of color polarization images are loaded from the data set to execute a training algorithm including feedforward and backpropagation. Subsequently, the recognition accuracy of the output image is evaluated on a separate test data set of 70 sets of images. Fig. 9 The test results of recognition accuracy (Accuracy) are shown as the number of iterations (Training Eopch) increases. Fig. 9The three curves in the middle represent the demosaicing test of the convolutional neural network model without the biomimetic polarization vision chip device array (Demosaicking without array), the demosaicing test of the convolutional neural network model with the biomimetic polarization vision chip device array (Demosaicking with array), and the demosaicing test of the mosaic image by the traditional neural network without the polarization image denoising function (Mosaicking). After 20 trainings, the network with the array showed a performance level similar to that of the network without the array, achieving a recognition accuracy of 99.7%, showing a strong ability to learn polarization information. In contrast, the traditional neural network without the polarization image denoising function showed a weaker ability to learn polarization information, with a recognition accuracy of less than 90%. This result emphasizes the significant advantages of the U-net convolutional neural network based on the 3×3 biomimetic polarization vision chip device array in handling polarization-dependent image denoising, classification, and recognition tasks.

[0073] Although the exemplary embodiments and their advantages have been described in detail, those skilled in the art may make various changes, substitutions and modifications to these embodiments without departing from the spirit of the present invention and the scope of protection defined by the appended claims, and such modifications and variations all fall within the scope defined by the appended claims. For other examples, those of ordinary skill in the art should readily understand that the order of the process steps may be changed while maintaining the scope of protection of the present invention.

[0074] In addition, the scope of application of the present invention is not limited to the processes, mechanisms, manufactures, material compositions, means, methods and steps of the specific embodiments described in the specification. From the disclosure of the present invention, it will be easily understood by those skilled in the art that the processes, mechanisms, manufactures, material compositions, means, methods or steps that currently exist or will be developed in the future, which perform substantially the same functions as the corresponding embodiments described in the present invention or obtain substantially the same results, can be applied according to the present invention. Therefore, the claims attached to the present invention are intended to include these processes, mechanisms, manufactures, material compositions, means, methods or steps within their scope of protection.

Claims

1. A bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, characterized in that: The device comprises: a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure prepared by using two-dimensional materials, the structure comprising a bottom electrode layer, a ferroelectric layer, an intermediate metal layer, an insulating layer and a channel layer stacked in sequence; The channel layer has anisotropy, and the ferroelectric layer controls the conductivity of the channel layer by switching the polarization switch under the induction of polarized light, so as to be used for polarization image recognition; The channel layer is a palladium diselenide channel layer, a rhenium disulfide channel layer, a black phosphorus channel layer, a germanium selenide channel layer or a tellurium channel layer; The bottom electrode layer and the input polarized light are used as a presynaptic membrane, the ferroelectric layer, the intermediate metal layer and the insulating layer are used as presynaptic neurons, and the channel layer is used as postsynaptic neurons. The weight between the presynaptic neuron and the postsynaptic neuron is used to be adjusted under the action of electrical stimulation and / or light stimulation to achieve learning and memory of polarized visual information.

2. The bionic polarization vision chip device based on two-dimensional ferroelectric heterojunction according to claim 1 is characterized in that: The ferroelectric layer is an indium-doped copper sulfide phosphide ferroelectric layer, a two-dimensional barium titanate ferroelectric layer, a two-dimensional bismuth ferrite ferroelectric layer or a twisted double-layer nitride ferroelectric layer.

3. The bionic polarization vision chip device based on two-dimensional ferroelectric heterojunction according to claim 1 is characterized in that: The insulating layer is a boron nitride insulating layer, a hafnium oxide insulating layer, an aluminum oxide insulating layer, a titanium oxide insulating layer or a cobalt oxide insulating layer.

4. The bionic polarization vision chip device based on two-dimensional ferroelectric heterojunction according to claim 1 is characterized in that: The bottom electrode layer is a multi-layer graphene bottom electrode layer, and the intermediate metal layer is a few-layer graphene intermediate metal layer.

5. A convolutional neural network, characterized in that: include: The bionic polarization vision chip device array composed of the bionic polarization vision chip devices based on the two-dimensional ferroelectric heterojunction as described in any one of claims 1 to 4 is used to perform convolution operation on the input polarization image.

6. The convolutional neural network according to claim 5, characterized in that The convolutional neural network is a U-net convolutional neural network, the first layer of which includes a bionic polarization vision chip device array and a first convolutional layer, and the input voltage of the bionic polarization vision chip device array is determined based on the parameters of the first convolutional layer.

7. A method for preparing a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction, characterized in that: The method comprises: The materials for preparing the bottom electrode layer, the ferroelectric layer, the intermediate metal layer, the insulating layer and the channel layer are peeled off by a mechanical peeling method to obtain a two-dimensional material sample, wherein the two-dimensional material sample comprises a bottom electrode layer sample, a ferroelectric layer sample, an intermediate metal layer sample, an insulating layer sample and a channel layer sample respectively formed on a glass slide; The bottom electrode layer sample, ferroelectric layer sample, intermediate metal layer sample, insulating layer sample and channel layer sample formed on the glass slide are transferred to the silicon wafer to obtain a stacked bottom electrode layer, ferroelectric layer, intermediate metal layer, insulating layer and channel layer, and the stacked bottom electrode layer, ferroelectric layer, intermediate metal layer, insulating layer and channel layer constitute a multi-layer two-dimensional van der Waals vertical ferroelectric heterostructure; the channel layer has anisotropy, and the ferroelectric layer controls the conductivity in the channel layer by switching the polarization switch under the induction of polarized light, which is used for polarization image recognition; The channel layer is a palladium diselenide channel layer, a rhenium disulfide channel layer, a black phosphorus channel layer, a germanium selenide channel layer or a tellurium channel layer; The bottom electrode layer and the input polarized light are used as a presynaptic membrane, the ferroelectric layer, the intermediate metal layer and the insulating layer are used as presynaptic neurons, and the channel layer is used as postsynaptic neurons. The weight between the presynaptic neuron and the postsynaptic neuron is used to be adjusted under the action of electrical stimulation and / or light stimulation to achieve learning and memory of polarized visual information.

8. The method for preparing a bionic polarization vision chip device based on a two-dimensional ferroelectric heterojunction according to claim 7, characterized in that: The method further comprises: A source-drain electrode layer is formed on a surface of the channel layer away from the insulating layer.

Images