Photodetector based on autodoping effect and its application as dynamic neural network accelerator

Abstract

The present invention discloses a photodetector based on the self-doping effect and its application as a dynamic neural network accelerator, belonging to the technical field of high-performance sensing and sensing-storage-computing integration. The photodetector uses a layered two-dimensional semiconductor bismuth oxide selenium as a channel layer, and regulates the position of selenium vacancies through an interfacial acid etching method to separate it from the electron transport layer, thereby realizing a self-modulation effect of the photoresponsivity, and realizing a wide range of photoresponsivity amplitude modulation and photoresponsivity change time constant modulation. The use of this photodetector as a dynamic on-chip optical neural network accelerator can effectively perform dynamic weight adjustment, and is used for continuously changing object feature recognition, including the recognition and detection tasks of objects under rotation, movement, and degradation conditions.

Description

Technical Field

[0001] The present invention relates to a dynamic neural network accelerator, and in particular to a method for realizing a self-doping modulated photoconductance and dynamic neural network accelerator based on a self-doping effect, and belongs to the technical field of high-performance sensing and sensing-storage-computing integration. Background Art

[0002] The state of objects in the natural environment presents a wide range of Gaussian distribution. Due to the limited training samples, traditional static neural networks often face performance degradation when performing tasks with correlation and time-series changes. Dynamic neural networks are widely used in the fields of meta-learning and adaptive learning due to their structural adjustability, environmental adaptability, and operational robustness. Traditional research on dynamic neural networks remains at the software algorithm level. This is because the network structure of hardware-based neural network accelerators is fixed after the weights are written, and structural adaptability cannot be achieved in subsequent operations. In order to realize the functions of dynamic neural networks at the hardware level, additional functional circuit structures such as feedback circuits, compensation circuits, and correction circuits need to be configured in the hardware on-chip neural network, which greatly reduces the space utilization of the chip and increases additional power consumption and delay overhead. In contrast, the basic device unit with self-modulation effect can spontaneously regulate parameters inside the device to act as a dynamic convolution kernel. The integrated device array directly constructs an on-chip neural network accelerator, which completes the preprocessing and recognition tasks of the dynamically changing environment by performing simulated in-situ multiplication and addition operations. This highly parallel way of capturing and processing information can effectively reduce the transmission of redundant data, thereby reducing energy consumption. However, existing research has not reported on such a device unit with self-modulation effect, which greatly limits the development of existing sensing, storage and computing device systems in the field of dynamic neural networks. Therefore, there is an urgent need for an electronic device unit with self-modulation effect. Summary of the invention

[0003] In order to realize a dynamic neural network accelerator unit with self-modulation effect, the present invention proposes a photodetector based on the self-doping effect of high-mobility bismuth oxygen selenium, which realizes a continuously dynamically changing light response (the light response range is 10 A / W ~ 10 5 A / W, relaxation time is 10 ms ~ 10 6 ms, and the corresponding light weight accuracy reaches 93.2%). Using this photodetector as a dynamic on-chip optical neural network accelerator can effectively perform dynamic weight adjustment for continuously changing object feature recognition, including recognition and detection tasks of objects under rotation, movement, and degradation conditions.

[0004] In a first aspect of the present invention, a photodetector is provided, comprising a substrate and a gate dielectric layer, and a layered two-dimensional bismuth oxygen selenide channel layer located on the gate dielectric layer, wherein both ends of the channel layer are source / drain (S / D) electrodes, wherein the lower portion of the layered two-dimensional bismuth oxygen selenide channel layer is an interface defect layer having selenium vacancies, and the upper portion is a channel transmission layer.

[0005] Preferably, the interface defect layer is 1-2 layers of two-dimensional bismuth oxide selenium with selenium vacancies, and its thickness is 0.7 nm-1.4 nm. The selenium vacancy defect concentration in the interface defect layer is 4.2×10 12 / cm 2 ~1.6×10 13 / cm 2 .

[0006] The thickness of the channel transfer layer is preferably 8 nm to 12 nm.

[0007] The substrate may be a silicon substrate, which is a back gate. The gate dielectric layer may be made of dielectric materials such as hafnium oxide, aluminum oxide, silicon oxide, etc. The source and drain electrodes may be metal electrodes such as titanium gold, palladium gold, titanium palladium gold, etc.

[0008] In some embodiments of the present invention, the substrate is a silicon substrate; the gate dielectric layer is made of hafnium oxide (HfO2) with a thickness in the range of 10-20 nm; the source and drain electrodes are Ti / Au metal stacks with a thickness not exceeding 60 nm; the thickness of the two-dimensional bismuth oxygen selenium channel layer is in the range of 5-20 nm; the interface defect layer can induce selenium vacancies through interface modification engineering caused by wet etching, plasma bombardment, element doping or other methods.

[0009] The photodetector of the present invention has gate voltage dependence: when the gate voltage is 0, the interaction between channel electrons and defects reaches equilibrium, and the light response intensity does not change with time; under negative gate voltage, the Fermi level decreases, defect electrons are released, the defect sites have local positive charge, channel electron transmission is promoted by self-modulation doping, and the light response intensity increases with time; under positive gate voltage, the Fermi level increases, electrons fill defects, defect sites have local negative charge, channel electron transmission is suppressed by self-modulation doping, and the light response intensity decreases with time.

[0010] The above-mentioned characteristics of the photodetector can be used as an optical neural network accelerator to realize dynamic optical network weights with adjustable gate voltage and bidirectional changes. The time-varying increase and decrease light response of the photodetector that depends on the gate voltage corresponds to the weight in the optical neural network calculation, realizing the integration of sensing, storage and computing.

[0011] In the second aspect of the present invention, a dynamic neural network is provided, whose weights can be adjusted over time. When external information changes, the weights can also change accordingly to make certain compensations, thereby improving the recognition accuracy of dynamic targets and unrecognized targets.

[0012] The hardware structure of the dynamic neural network proposed in the present invention is an array of the above-mentioned photodetectors. The light response current of the device can be expressed as I ph =R·P, where R is the photoresponsivity of the device, acting as the network weight stored in the device, and P corresponds to the input external information. The external image is input to the device array in the form of optical power P, and the output photocurrent is obtained by the above formula inside the device. Then, by constructing the on-chip device array (see Fig.15 ), the drain of each device is connected and controlled separately, and the gate bias of each device is also controlled independently. The gate voltage of each device can be provided through the peripheral PCB or FPGA connection to achieve the modulation of dynamic light responsiveness, and the source of the devices on each column is connected. According to Kirchhoff's current law, the output current of the nth column can be expressed as:

[0013]

[0014] Wherein, m and n are positive integers, n is the column coordinate corresponding to the device, and m is the horizontal coordinate corresponding to the device; n (t) is the current output by the device in the nth column at time t; R mn (t) is the photoresponsivity of the device in the nth column and the mth row at time t, P mn (t) is the optical power input to the device in the nth column and the mth row at time t.

[0015] Since P changes with time, the dynamic R is constructed to achieve information compensation tracking and improve the recognition accuracy of dynamic targets. Each column of current corresponds to a type of output label, and the label corresponding to the column with the largest output current is used as the output result of the array. The exponential function is fitted to the change of the device's light response (i.e. weight information) over time, and the fitting equation is:

[0016]

[0017] Where, t represents time; R mn (t) is the photoresponsivity of the device in the nth column and the mth row at time t; R mn,0 , R mn,1 is the fitting parameter, and its value can be achieved by adjusting the source-drain voltage and gate bias. The specific correspondence between the fitting parameter value and the gate voltage and source-drain voltage can be referred to Fig.12 As shown, R initial For R mn(t) The light response value at time 0 in the function can be expressed as R initial = R mn,0 + R mn,1 , R final For R mn The light response value at the time when the time is infinite in the function (t) can be expressed as R final = R mn,0 Therefore, in order to achieve the setting of fitting parameters and the regulation of time constants, the corresponding gate voltage and source-drain voltage combination can be designed to achieve effective regulation; τ mn is the time constant of the photoresponsivity of the device in the nth column and the mth row over time, which can be achieved by adjusting the gate bias voltage. The specific adjustment method is as follows Fig.12 shown.

[0018] The network is trained by software, and the training process is as follows: Taking the recognition training process of rotating animal heads as an example, the selected training data set is continuously rotating animal heads, where each frame image is a 200 pixel × 200 pixel image, and each animal contains ten frames of images for training. These images are input into Python, and by constructing a layer of fully connected neural network, firstly, each node in the fully connected neural network in Python is assigned random initial weight information, that is, the initial fitting parameters R0 and R1 and the value of the time constant τ. Here, the neurons at the software level correspond to a device in the subsequent hardware structure. According to the aforementioned method of summing the current of each column based on Kirchhoff's current law, read the column with the largest output value. During the training process, we assign a target label value to the recognition result, that is, the object category we hope to identify; the actual output result of the software is the output value, we determine whether its output label is the same as the target label, iteratively feedback the difference between the output label and the target label, and update the network weight until all ten frames of images of the same animal are trained, and the network algorithm can converge to achieve accurate recognition; then repeat the above process, train the images of other animals, and accurately distinguish the animal categories in the training stage to obtain the weight information of each network node, that is, the R0, R1 and τ values ​​of each network node. After that, the trained weight information is written into the hardware neural network by controlling the gate bias and source-drain voltage of the device. Therefore, the present invention designs a dynamic neural network whose weight can change over time, utilizes the conductivity relaxation characteristics of the device, uses the conductivity to store the weight of the network, and when training the network, trains the R0, τ and R1 values ​​of the network, and studies the relationship between the relaxation characteristics of the device and the external bias, and then realizes the spontaneous change of the electrical weight by writing the gate voltage and the source-drain voltage, thereby realizing the recognition and detection of dynamic targets.

[0019] The present invention further provides a method for preparing the above-mentioned photoelectric detector, comprising the following steps:

[0020] 1) Growth of layered two-dimensional bismuth oxygen selenide materials;

[0021] 2) Modify the interface of layered two-dimensional bismuth oxide selenium materials to introduce and regulate the concentration of interfacial selenium vacancy defects;

[0022] 3) Transferring the layered two-dimensional bismuth oxide selenide material onto the gate dielectric layer on the substrate with the interface-modified side facing downward;

[0023] 4) Position the transferred layered two-dimensional bismuth oxide selenide material and define the source and drain electrode regions by electron beam exposure;

[0024] 5) Prepare source and drain electrodes.

[0025] The above step 1) can grow a layered two-dimensional bismuth oxide selenide material on a mica sheet by chemical vapor deposition (CVD); step 2) spin-coat a polymer such as PMMA (polymethyl methacrylate) on the layered two-dimensional bismuth oxide selenide material as a protective layer, and then etch the mica sheet interface with dilute hydrofluoric acid, and adjust the interface selenium vacancy defect concentration by adjusting the etching time to achieve an interface defect layer; step 3) transfer the layered two-dimensional bismuth oxide selenide material together with the protective layer to the gate dielectric layer, and then remove the protective layer by immersion in acetone solution or the like.

[0026] The advantages of the present invention are as follows:

[0027] 1) High-mobility and high-stability ionic layered two-dimensional semiconductor bismuth oxide selenide is used as the channel layer. For the first time, the position of selenium vacancies is regulated by interfacial acid etching to separate it from the electron transport layer, thereby achieving a self-modulation effect of photoresponsivity, and realizing a wide range of photoresponsivity amplitude modulation and photoresponsivity change time constant modulation.

[0028] 2) The time-varying photoresponsivity that depends on gate voltage is used as the network weight, and in-situ multiplication and addition operations are performed with the incident light power inside the device to realize a hardware-level optical neural network accelerator. The average power consumption for single-pixel operations is as low as 1 pW, greatly reducing the power consumption overhead of hardware neural networks.

[0029] 3) The light responsivity that changes with the increase or decrease of gate voltage corresponds to the self-adjustment of weights in the dynamic neural network, and then the multiplication and addition of light weights (light responsivity) and light intensity can be realized based on the optical dynamic neural network device array, realizing the dynamic neural network function on the chip. Based on the on-chip optical neural network accelerator, the recognition of objects that are in continuous motion and whose features gradually disappear is realized. Compared with the static neural network with fixed weights, the device's ability to recognize and detect data correlation is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1It is a schematic structural diagram of a photoelectric detector with self-modulation effect of the present invention;

[0031] Figure 2~Figure 7 A schematic diagram of the process steps for preparing a photoelectric detector with a self-modulation effect according to a specific embodiment of the present invention, wherein:

[0032] Figure 2 The method comprises the steps of growing bismuth oxide selenide material on a mica sheet;

[0033] Figure 3 The step of spin coating and etching a PMMA protective layer on the surface of the bismuth oxide selenide material using a PMMA assisted transfer method;

[0034] Figure 4 The step of immersing the bismuth oxide selenium material coated with a PMMA protective layer into a dilute acid solution to introduce interfacial selenium vacancies;

[0035] Figure 5 The step of transferring bismuth oxide selenide / PMMA onto Si / HfO2 substrate;

[0036] Figure 6 A step for removing the PMMA protective layer;

[0037] Figure 7 The step of defining the source-drain electrode region and preparing the source-drain electrode by electron beam exposure;

[0038] Figures 1 to 7 Middle: 1—gate dielectric layer; 2—substrate; 3—source electrode; 4—drain electrode; 5—bismuth oxide selenide channel layer; 6—interface defect layer; 7—mica sheet; 8—PMMA protective layer.

[0039] Figure 8 The light response performance of the photodetector with self-modulation effect in the embodiment of the present invention is characterized. In the figure, a is a schematic diagram of applying continuous light pulses to the device in the embodiment, b and c are graphs of the channel current variation with time measured when the device is applied with light pulses under gate voltage bias of -1.3 V and 1.3 V, respectively, and d is a graph of the extracted light response variation with time.

[0040] Fig. 9 : is a curve showing the variation of the modulation coefficient with the source-drain voltage at a wavelength of 600 nm for the photodetector with self-modulation effect in an embodiment of the present invention, wherein a is the curve showing the variation of the modulation coefficient with the source-drain voltage under a negative gate voltage, and b is the curve showing the variation of the modulation coefficient with the source-drain voltage under a positive gate voltage.

[0041] Fig.10 4 is a curve showing the change of the modulation coefficient of the photodetector with self-modulation effect at a wavelength of 600 nm as a function of the gate voltage in an embodiment of the present invention.

[0042] Fig.111 is an increase / decrease light response intensity curve of the photodetector with self-modulation effect in the wavelength range of 500 nm to 800 nm under a gate voltage bias of -2 V to 2 V in an embodiment of the present invention, wherein the first to fourth columns correspond to wavelengths of 500 nm, 600 nm, 700 nm, and 800 nm, respectively; the first to fifth rows correspond to gate voltages of 2 V, 1 V, 0 V, -1 V, and -2 V, respectively.

[0043] Fig.12 Graph showing the initial intensity (a) and steady-state intensity (b) of the photoresponsivity change of the photodetector in an embodiment of the present invention modulated at a gate voltage of -1.5 V to 1.5 V and a source-drain voltage of 0 to 0.7 V.

[0044] Fig.13 3 is a time-varying curve of the light response intensity extracted from the photodetector with self-modulation effect in an embodiment of the present invention under 1000 cycles of continuous light pulse stimulation, wherein a shows the repeatably enhanced light response when the gate voltage is -1.5 V, and b is a parameter distribution curve obtained by parameter fitting of the light response curve.

[0045] Fig.14 Schematic diagram of the working principle of the photoelectric detector with self-modulation effect in an embodiment of the present invention.

[0046] Fig.15 The schematic diagram is a structural diagram of realizing dynamic neural network recognition based on the light response detector array proposed in the present invention.

[0047] Fig.16 The recognition results of handwritten digits by the dynamic neural network and the static neural network using the photoelectric detector with self-modulation effect as an accelerator in the embodiment of the present invention. Wherein, a is the image data set to be trained, and b is the recognition results of handwritten digits by comparing the static neural network (SANN) and the dynamic neural network (DANN).

[0048] Fig.17 The following are the recognition results of the dynamic neural network (DANN) and the static neural network (SANN) for continuously rotating small animals, which are composed of the photodetector with self-modulation effect as the accelerator in the embodiment of the present invention. Among them, a is a schematic diagram of the continuously changing image information and the corresponding weights changing over time; b is the recognition results of the dynamic neural network and the static neural network during the training process, as well as the changing image information and weights; c shows the change of the loss function of the static neural network and the dynamic neural network during the training process with the training nodes; d shows the change of the recognition accuracy of the static neural network and the dynamic neural network during the training process with the training nodes. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below through embodiments in conjunction with the accompanying drawings.

[0050] like Figure 1 As shown, the photodetector with self-modulation effect of the present invention comprises a substrate 2 and a gate dielectric layer 1, an interface defect layer 6 and a bismuth oxygen selenium channel layer 5 sequentially located on the gate dielectric layer 1, and two ends of the bismuth oxygen selenium channel layer 5 are a source electrode 3 and a drain electrode 4. Based on the photodetector array, an on-chip sensing, storage and computing integrated structure is formed, and the dynamic change of light weights of different pixel points is realized by adjusting the voltage bias of a single device, which is applied to the on-chip dynamic neural network accelerator, further improving the error tolerance and environmental adaptability of the dynamic neural network.

[0051] The above-mentioned photodetector is a light-sensing and computing integrated device based on interface selenium vacancy self-doping modulation, and its preparation steps are as follows:

[0052] 1) Using Bi2O3 and Bi2Se3 as growth sources, a layered two-dimensional bismuth oxide selenium material is grown on a mica sheet 7, such as Figure 2 As shown;

[0053] 2) A layer of PMMA (polymethyl methacrylate) is spin-coated on the surface of the bismuth oxide selenium material on the mica sheet 7 as a protective layer. The spin-coating process conditions in this embodiment are: spin coating at a rotation speed of 4000 rpm for 1 min, and then curing on a heating plate at 170° C. for 3 min. Figure 3 As shown;

[0054] 3) The bismuth oxide selenium material protected by the PMMA protective layer 8 is etched at the interface of the mica sheet 7 with a diluted acid solution of HF:H2O = 1:50 (volume ratio), and the interface defect layer 6 is regulated by adjusting the etching time to control the concentration of selenium vacancy defects on the interface. In this embodiment, the etching time is 4 to 12 h. Figure 4 As shown;

[0055] 4) After washing with deionized water, the bismuth oxide selenium / PMMA was transferred to the Si / HfO2 composite substrate and then cured at 80°C for 3 min. Figure 5 As shown;

[0056] 5) The bismuth oxide selenium / PMMA transferred to the Si / HfO2 composite substrate is immersed in an acetone solution for 30 min to remove the PMMA protective layer 8, as shown in FIG. Figure 6 As shown;

[0057] 6) For the two-dimensional bismuth oxide selenide channel region (the thickness of the two-dimensional bismuth oxide selenide material in this embodiment is in the range of 5-20 nm), the source and drain (S / D) regions are defined by electron beam exposure (EBL), and the metal is evaporated by electron beam (the metal used in this embodiment is Ti / Au=5 / 40nm) and peeled off to form source and drain electrodes, such as Figure 7 shown.

[0058] The schematic diagram of the self-modulated dynamic light response of the photodetector prepared in this embodiment is as follows: Figure 8 As shown, the photodetector is under negative gate voltage (V g = -1.3 V), the photocurrent amplitude obtained by continuous application of light pulses increases with time, showing an increasing response; under positive gate voltage (V g = 1.3 V), the photocurrent amplitude obtained by continuous application of light pulses decreases with time, showing a decreasing response; according to I ph = I light -I dark , R = I ph / (P in ×S), the light response obtained is as follows Figure 8 As shown in d. Where: I ph represents the photocurrent, I light Represents the channel current in the presence of light, I dark represents the channel current in the dark, R represents the light response, and P in represents the incident light power, and S represents the illuminated area.

[0059] Combination Fig. 9 , Fig.10 As shown, we define the modulation coefficient a = log(R final / R initial ), where R final , R initial The light responsivity of the photodetector described in the example at the initial moment and the end moment are respectively shown in the table; under a negative gate voltage, as the source-drain voltage increases, the a value gradually decreases, indicating that the enhancement effect gradually decreases; under a positive gate voltage, as the source-drain voltage increases, the a value gradually increases, indicating that the inhibition effect gradually becomes stronger.

[0060] Electrical measurements were performed using a probe station and semiconductor parameter analyzer 4200, and light of different wavelengths and intensities were applied to the photodetector in combination with a light source, a monochromator, and other devices. The test results show that the spectral response range of the photodetector is 500 nm-800 nm, and the increase and decrease polarity of the device's light response is modulated by the gate voltage. Under positive gate voltage, the light response decreases over time, and under negative gate voltage, the light response increases over time. Fig.11 shown.

[0061] In order to study the control range of the photoresponsivity of the bismuth oxygen selenium photodetector, a single exponential fitting was performed on the photoresponsivity curve over time under the conditions of -1.5V~1.5V gate voltage bias and 0~0.7V source-drain voltage bias. The initial weight and steady-state weight distribution were extracted as shown in Figure 2. Fig.12As shown in Figure 2, under a fixed external voltage bias, the enhanced mode photoresponsivity can be improved by 10 5 times, the photoresponsivity can be attenuated by 10 in the suppression mode 2 times.

[0062] In order to verify the stability of the photodetector, periodic light pulses were applied to the photodetector. 20 continuous light pulses were applied in each test cycle. The photodetector remained stable after 1000 cycles. The three parameters (R0, τ, R1) obtained by fitting maintained a narrow Gaussian distribution, as shown in Fig.13 shown.

[0063] The working principle of the photoelectric detector with self-modulation effect of the present invention is as follows: Fig.14 As shown, in the initial state, the Fermi level (E F ) between the Se vacancy defect energy level (E Se ) and the bottom of the conduction band, the selenium vacancy is partially filled, and the interaction between the channel electrons and the defects is balanced; Fig.14 As shown in a, under positive gate voltage, the Fermi level jumps above the selenium vacancy defect level, and the electron crosses [BiO] 2+ Layer entry [Se] 2- layer and is captured by defects, [Se] 2- The local negative charge of the layer inhibits the electron transport in the channel, and the photoresponsivity decreases over time; Fig.14 As shown in middle b, under negative gate voltage, the Fermi level is lowered to the band gap, the electrons captured by the defect energy level are released, the Se vacancy acts as a local positive center, promoting channel electron transport, and the photoresponsivity increases with time, thereby achieving a time-varying photoresponsivity based on the self-modulation effect.

[0064] In order to further explore the recognition effect of dynamic weights on continuously changing environments, we trained the weights of the neural network through the Python platform. Unlike conventional neural networks, we trained the R0, R1 parameter values ​​and time constant τ of each node during the training process. By applying gate voltage and source-drain voltage combination to each device unit of the hardware neural network accelerator composed of a 3×3 photodetector array, we regulated the starting amplitude and steady-state amplitude of the light response intensity as the convolution kernel of the dynamic neural network. The handwritten digit task and the animal rotation recognition task were executed as examples. This process and principle can be used Fig.15 After the optical signal is input into the array, the output current is measured. The input information enters the device channel in the form of input optical power P, and is multiplied by the device light responsivity R (i.e., network weight) inside the device, and then the output result of each column current is obtained by Kirchhoff's current law, which is I n, each current channel is used as a decision result, and the recognition output is judged according to the output current value, that is, the recognition of the image with feature changes is realized. Compared with the static neural network (SANN) with fixed weights, the dynamic neural network (DANN) significantly improves the recognition accuracy, such as Fig.16 , Fig.17 As shown in the figure, the corresponding neural weight distribution is as follows Fig.17 The light responsivity training matrix is ​​shown in a.

[0065] The photoelectric detection device array is used to process continuously changing image information, effectively realizing the recognition of objects with dynamic changes and gradually degraded features. The results show that for handwritten digit tasks, the recognition accuracy of this embodiment is improved from 65% to 90% (see Fig.16 b); for the rotating animal recognition task, the recognition accuracy of this embodiment is improved from 20% to 99% (see Fig.17 (d)

[0066] The embodiments described above are not intended to limit the present invention. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention is defined by the scope of the claims.

Claims

1. A photodetector comprising a substrate and a gate dielectric layer, and a layered two-dimensional bismuth oxygen selenium channel layer located on the gate dielectric layer, wherein both ends of the channel layer are source and drain electrodes, characterized in that: The lower part of the layered two-dimensional bismuth oxygen selenium channel layer is an interface defect layer with selenium vacancies, and the upper part is a channel transmission layer; the interface defect layer is 1-2 layers of two-dimensional bismuth oxygen selenium with selenium vacancies, wherein the selenium vacancy defect concentration is 4.2×10 12 / cm 2 ~1.6×10 13 / cm 2 ; The thickness of the channel transmission layer is 8 nm ~12 nm.

2. The photodetector according to claim 1, wherein: The substrate is a silicon substrate, and the gate dielectric layer is hafnium oxide, aluminum oxide or silicon oxide.

3. The method for preparing the photodetector according to claim 1 or 2, characterized in that: The following steps are involved: 1) Growth of layered two-dimensional bismuth oxygen selenide materials; 2) Modify the interface of layered two-dimensional bismuth oxide selenium materials to introduce and regulate the concentration of interfacial selenium vacancy defects; 3) transferring the layered two-dimensional bismuth oxide selenide material onto a gate dielectric layer on a substrate with the interface-modified side facing downward; 4) Positioning the transferred layered two-dimensional bismuth oxide selenide material and defining the source and drain electrode regions by electron beam exposure; 5) Prepare source and drain electrodes.

4. The preparation method according to claim 3, characterized in that: Step 1) growing a layered two-dimensional bismuth oxide selenide material on a mica sheet by chemical vapor deposition; Step 2) spin coating a polymer as a protective layer on the layered two-dimensional bismuth oxide selenide material, then etching the mica sheet interface with dilute hydrofluoric acid, and regulating the interface selenium vacancy defect concentration by adjusting the etching time; Step 3) transferring the layered two-dimensional bismuth oxide selenide material together with the protective layer to a gate dielectric layer, and then removing the protective layer.

5. Use of the photodetector described in claim 1 or 2 as an optical neural network accelerator.

6. The use according to claim 5, characterized in that The photodetectors are used to form an array, the time-varying light responsivity of the photodetectors that depends on the gate voltage is used as a network weight, and the light responsivity and the incident light power are multiplied and added in situ to realize an optical neural network accelerator at the hardware level.

7. A dynamic neural network whose weights are adjusted over time, characterized in that The dynamic neural network comprises an array of photodetectors as claimed in claim 1 or 2, wherein the time-varying photoresponsivity of the photodetectors, which is gate voltage-dependent, corresponds to the self-regulation of weights in the dynamic neural network.

8. The dynamic neural network according to claim 7, characterized in that The drain of each device in the array is individually connected and controlled, the gate bias of each device is independently controlled, and the sources of the devices on each column are connected. The output current of the nth column is: ； Among them, I n (t) is the current output by the nth column device at time t, P mn (t) is the optical power input to the device in the nth column and the mth row at time t; R mn (t) is the photoresponsivity of the device in the nth column and the mth row at time t, and its fitting equation for changes with time is: ； Among them, R mn,0 , R mn,1 is the fitting parameter, and its value is achieved by adjusting the source-drain voltage and the gate bias; τ mn is the time constant of the photoresponsivity of the device in the nth column and the mth row as it changes with time, and its value is achieved by adjusting the gate bias voltage; By writing gate bias and source-drain voltage into the array, the electrical weights are changed spontaneously. Each column of current in the array corresponds to a type of output label, and the label corresponding to the column with the largest output current is used as the output result of the array.

9. The dynamic neural network according to claim 8, characterized in that The fitting parameters R0, R1 and time constant τ of each device are obtained by software training. Specifically, a layer of fully connected neural network is constructed by software, and each node in the fully connected neural network corresponds to a device in the hardware neural network; the fully connected neural network is trained for image recognition: first, each node is assigned random initial weight information, that is, the initial R0, R1 and τ values, and the column with the largest output value is read to determine whether the corresponding output label is the same as the target label, and the difference between the output label and the target label is iteratively fed back to update the network weight until the training is completed to achieve accurate recognition; thereby, the weight information of each node, that is, the R0, R1 and τ values ​​of each node, is obtained; and then the trained weight information is written into the hardware neural network by controlling the gate bias and source-drain voltage of the device.

Images