An artificial photoelectric synapse device based on topological insulator and application thereof

Abstract

The application relates to the technical field of artificial synapses, and specifically discloses an artificial optoelectric synapse device based on a topological insulator, which comprises an Sb2Te3 film and two Au electrodes arranged at two ends of the Sb2Te3 film; compared with a conventional three-terminal transistor synapse which needs an additional stimulation gate voltage, the optoelectric synapse device has a simple preparation process, the artificial optoelectric synapse device based on the topological insulator has a simple structure, a simple preparation process, stable performance in air, and excellent photoelectric current response in a visible light range; the photo response of the device can be modulated by adjusting parameters of light stimulation. In addition, the artificial optoelectric synapse device is combined with a recurrent neural network to process image information, which helps to improve the recognition accuracy in the recurrent neural network; a series of synapse signals capable of representing images are taken as inputs of the RNN, the image recognition accuracy of the RNN is improved by effectively utilizing the synapse time sequence data, and the accuracy can reach 100%.

Description

Technical Field

[0001] This application relates to the field of artificial synapse technology, and more specifically to an artificial opto-synapse device based on a topological insulator and its application. Background Technology

[0002] With the advent of the information age, the amount of data that human society needs to process has increased dramatically. Existing computers, based on the von Neumann architecture, separate computation and storage, resulting in significant energy consumption and computational redundancy. Because biological synapses in the human brain's nervous system possess both computational and storage functions, the human brain exhibits a significant advantage in simultaneously processing multiple tasks while maintaining low energy consumption. Therefore, there is a strong demand for developing novel devices that integrate data sensing, data storage, and computational functions.

[0003] Artificial photosynapses have attracted significant attention in neuromorphic computing due to their unique integration of optical sensing and synaptic functions. However, traditional three-terminal photosynapses require electrical stimulation of the gate terminals to improve charge separation efficiency, which greatly limits the device's processing speed and bandwidth. Complex device manufacturing processes hinder their integration and application, while also posing significant challenges to large-scale production and stable performance. Therefore, it is necessary to develop two-terminal artificial photosynapses with simple device architecture, high photoresponse, and suitability for mass production.

[0004] Furthermore, as an important field within artificial intelligence, machine learning, and deep learning, image information processing has been widely adopted and applied in various scenarios. It also plays a crucial cognitive role in neuromorphic optical systems and is attracting increasing attention due to its potential applications in medicine, security, and autonomous driving. A significant current direction in image information processing is image recognition, which utilizes the characteristics of different neural networks to process images. These neural networks are constantly evolving to develop unique algorithmic models to improve the accuracy of image recognition. Using neural network systems in image recognition is a relatively new technology. By combining image recognition with neural network algorithms, the grayscale features of the image are typically extracted first, and then these features are mapped onto the neural network for image recognition and classification.

[0005] Traditional optoelectronic devices exhibit no correlation or synaptic behavior in their photodetector signals. This means that applying a light pulse generates a current signal, which disappears immediately upon removal of the pulse, and the preceding signal has no effect on the following one; each photocurrent signal is independent and unaffected. By converting photoelectric signals into electrical signals, image information is transformed into electrical signals that are then fed into a neural network for training, ultimately enabling the network to perform image recognition and classification. However, due to the "independence" of image signal conversion in traditional optoelectronic devices, they cannot effectively identify image features, thus affecting the accuracy of the neural network's image recognition. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial photoelectric synapse device based on topological insulators and its application. This artificial photoelectric synapse device has a simple fabrication process, good device performance, stable output, and can exhibit excellent response under visible light stimulation. This artificial photoelectric synapse device can be combined with RNN for image recognition and classification, thereby improving the image recognition accuracy of RNN.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] The present invention provides an artificial photoelectric synapse device based on a topological insulator, comprising an Sb2Te3 thin film and two Au electrodes disposed at both ends of the Sb2Te3 thin film.

[0009] Preferably, the wavelength of the light that generates the photocurrent response of the artificial photoelectric synapse device is 400–700 nm.

[0010] Preferably, the light intensity of the photocurrent response generated by the artificial photosynaptic device is 0.35–1.8 W / cm². 2 .

[0011] Preferably, the frequency at which the artificial photoelectric synapse device generates photocurrent response is 0.1 to 5 Hz.

[0012] Preferably, the number of pulses that generate a photocurrent response in the artificial photoelectric synapse device is 1 to 50.

[0013] Preferably, the thickness of the Sb2Te3 thin film is 10–100 nm.

[0014] To achieve the above objectives, another aspect of the present invention provides an application of an artificial photoelectric synapse device in image information processing, which combines the artificial photoelectric synapse device with a recurrent neural network to recognize handwritten digit images.

[0015] Preferably, the application of the artificial photoelectric synapse device in image information processing includes: extracting pixel features of a handwritten digit image from two sequences along the diagonal direction; calculating the duty cycle and frequency of the input light signal based on the extracted pixel features; converting the duty cycle and frequency into corresponding light pulse signals; applying the light pulse signals and reading voltage in the photoelectric test of the artificial photoelectric synapse device; and using the measured photocurrent response as the input of a recurrent neural network for image training and classification.

[0016] Preferably, the photoelectric test is performed using a semiconductor parameter analyzer.

[0017] Preferably, the photoelectric test of the artificial photoelectric synapse device includes: irradiating the artificial photoelectric synapse device with light of wavelength 400-700nm, and applying a reading voltage to the electrodes at both ends of the artificial photoelectric synapse device.

[0018] Preferably, the reading voltage is 0.05 to 5V.

[0019] The beneficial effects of this invention are:

[0020] (1) Compared to traditional three-terminal transistor synapses that require additional stimulation gate voltage, the fabrication process is simple. The artificial opto-synapse device based on topological insulators of this invention has a simple architecture, a simple fabrication process, stable performance in air, and exhibits excellent photocurrent response in the visible light range. The photoresponse of the device can be modulated by adjusting the parameters of the photostimulation, especially at a light power intensity of 1.4 W / cm². 2 Under stimulation by light with a wavelength of 450nm, the photocurrent can reach up to 0.24mA;

[0021] (2) The artificial opto-synaptic device based on topological insulators of the present invention exhibits typical characteristics of synapses, including short-term memory, long-term memory, and the learning process that simulates the human brain;

[0022] (3) This invention combines an artificial photoelectric synaptic device based on topological insulator with a recurrent neural network for image information processing. The photocurrent response of the synaptic device exhibits time-dependent characteristics, which helps to improve the recognition accuracy in the recurrent neural network (RNN). A series of synaptic signals that can characterize the image are used as inputs to the RNN. By effectively utilizing synaptic time series data, the image recognition accuracy of the RNN is improved, and its accuracy can reach 100%. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the structure of the artificial photoelectric synapse device in this application, where A represents the Sb2Te3 thin film and B represents the Au electrode;

[0024] Figure 2 (a) shows the typical photocurrent response of the artificial photoelectric synapse device in Example 1;

[0025] Figure 2 (b) shows the change in synaptic current generated by the artificial photoelectric synapse device in Example 1 under different light intensities and exposure times at 450nm;

[0026] Figure 2 (c) The change in synaptic current generated by applying different numbers of light pulses to the artificial photoelectric synapse device in Example 1;

[0027] Figure 2(d) shows the change in synaptic current generated by applying light pulses of different frequencies to the artificial photoelectric synapse device in Example 1;

[0028] Figure 2 (e) The artificial optical synaptic device in Example 1 simulates the learning process of the human brain.

[0029] Figure 3 This is a schematic diagram of the recurrent neural network recognition of handwritten digit images using the artificial photoelectric synapse device in Example 2;

[0030] Figure 4 This is a comparison of the recognition accuracy of the RNN with synaptic signals, the FNN with synaptic signals, and the conventional FNN with image pixel grayscale values ​​during 50 training cycles in Example 2. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. It should be understood that the embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0032] The purpose of this invention is to provide an artificial photoelectric synapse device based on topological insulators and its application. This artificial photoelectric synapse device has a simple fabrication process, good device performance, stable output, and can exhibit excellent response under visible light stimulation. This artificial photoelectric synapse device can be combined with RNN for image recognition and classification, thereby improving the image recognition accuracy of RNN.

[0033] To achieve the above objectives, the technical solution of the present invention is as follows:

[0034] The present invention provides an artificial photoelectric synapse device based on a topological insulator, comprising an Sb2Te3 thin film and two Au electrodes disposed at both ends of the Sb2Te3 thin film.

[0035] This artificial photoelectric synapse device is simple to fabricate, has good performance, stable output, and exhibits excellent response under visible light stimulation.

[0036] This invention does not have special requirements for the size of the gold electrodes and the width of the Sb2Te3 film, but the spacing between the two gold electrodes and the width of the Sb2Te3 film will affect the performance of the device. The smaller the spacing between the two gold electrodes, the lower the energy consumption and the lower the resistance, which can be adjusted according to the requirements.

[0037] In some embodiments, the wavelength of light that generates the photocurrent response of the artificial photoelectric synapse device is 400–700 nm.

[0038] Preferably, the wavelength of the light corresponding to the photocurrent generated by the artificial photoelectric synapse device is 450–638 nm.

[0039] Within the aforementioned wavelength range, similar to biological synapses, the artificial photoelectric synapse device in this application will trigger a significant increase in current under illumination, generating a photocurrent response.

[0040] In some embodiments, the light intensity of the photocurrent response generated by the artificial photosynaptic device is 0.35–1.8 W / cm². 2 .

[0041] By adjusting the light intensity within the above range, different degrees of current increase can be achieved, thereby realizing different photocurrent responses.

[0042] In some embodiments, the frequency at which the artificial photoelectric synapse device generates a photocurrent response is 0.1 to 5 Hz.

[0043] Preferably, the frequency at which the artificial photoelectric synapse device generates photocurrent response is 0.1 to 1 Hz.

[0044] In some embodiments, the number of pulses that generate a photocurrent response from the artificial photosynaptic device is 1 to 50.

[0045] By adjusting the frequency and number of light pulses within the above range, the magnitude and attenuation period of the photocurrent generated by the artificial photoelectric synapse device can be further adjusted, thus forming memories of different lengths.

[0046] In some embodiments, the thickness of the Sb2Te3 film is 10–100 nm.

[0047] In the artificial photoelectric synapse device of this application, if the thickness of the Sb2Te3 film is too small or too large, it will affect the crystal structure and optical properties of the Sb2Te3 film and its photocurrent response. By controlling the thickness of the Sb2Te3 film to be 10-100 nm, the artificial photoelectric synapse device can have certain photon-modified surface states and light-induced surface photocurrent under illumination. When the illumination is removed, due to the defects in the Sb2Te3 film, these defects introduce trapped energy levels and slow down the recombination of charge carriers, thereby causing the photocurrent to gradually decrease, ensuring that the artificial photoelectric synapse device has a typical photocurrent response.

[0048] By matching the above parameters, the artificial photoelectric synapse device can be linearly adjusted to achieve photocurrent response, further enabling the artificial photoelectric synapse device to be tuned in both optical and electrical aspects. At the same time, the artificial photoelectric synapse device of this topological insulator exhibits typical synaptic characteristics, such as forming impulse facilitation, short-term memory, long-term memory, and the learning process of the human brain.

[0049] In some embodiments, the fabrication method of the artificial photoelectric synapse device includes: growing a thin film of topological insulating material Sb2Te3 on a substrate by deposition, and forming gold electrodes at both ends by magnetron sputtering to obtain an Au / Sb2Te3 / Au dual-ended photoelectric synapse device.

[0050] This application does not impose any special limitation on the type of substrate, including but not limited to glass substrates and quartz substrates.

[0051] This application does not specifically limit the deposition method, including but not limited to one of atomic layer deposition, chemical vapor deposition, pulsed laser deposition, and electrochemical deposition.

[0052] Another aspect of the present invention provides an application of an artificial photoelectric synapse device in image information processing, which combines the artificial photoelectric synapse with a recurrent neural network to recognize handwritten digit images, thereby improving the image recognition accuracy of the RNN.

[0053] In some embodiments, the application of artificial photoelectric synaptic devices in image information processing includes: extracting pixel features of a handwritten digit image from two sequences along the diagonal direction; calculating the duty cycle and frequency of an input light signal based on the extracted pixel features; converting the duty cycle and frequency into a corresponding light pulse signal; applying the light pulse signal and reading voltage in the photoelectric test of the artificial photoelectric synaptic device; and using the measured photocurrent response as the input of a recurrent neural network (RNN) for image training and classification.

[0054] The above scheme can effectively combine artificial photoelectric synaptic devices and image information processing. This scheme is simple, efficient, and has high image recognition accuracy, which plays a vital role in various image recognition applications such as robotics, autonomous driving, and medical disease diagnosis.

[0055] In some implementations, the method of extracting pixel features of the handwritten digit image along two diagonal sequences is not particularly limited, including but not limited to Photoshop.

[0056] Specifically, the pixels filled in the digital image represent synapses where light pulses and readout voltages have been applied, while the remaining empty pixels in the image represent those where only readout voltages have been applied without light pulses. The duty cycle and frequency of the input light signal are calculated based on the ratio of filled to empty pixels in the diagonal direction.

[0057] In this application, the duty cycle is the ratio of the number of filled pixels to the total number of pixels along the diagonal of the image, and the frequency of the light signal is the reciprocal of the sum of the pixels in that period.

[0058] In some embodiments, the photoelectric testing equipment is a semiconductor parameter analyzer, which is used to apply light stimulation to the artificial photoelectric synapse device and to adjust different parameters to achieve nonlinear or linear adjustment of the photocurrent response of the device.

[0059] In some embodiments, the photoelectric testing of the artificial photoelectric synapse device includes: irradiating the artificial photoelectric synapse device with light of wavelength 400-700nm, applying a readout voltage to the electrodes at both ends of the artificial photoelectric synapse device, and measuring the photocurrent response corresponding to the diagonal pixel features that can represent handwritten digit images in the RNN.

[0060] In some embodiments, the read voltage is 0.05 to 5V, and the optical response can be linearly adjusted by adjusting the read voltage over a wide response range of 0.05 to 5V.

[0061] Example 1

[0062] Please see Figure 1 An artificial photoelectric synapse device based on a topological insulator includes an Sb2Te3 thin film A and two Au electrodes B disposed at both ends of the Sb2Te3 thin film A.

[0063] The fabrication method of this artificial photoelectric synapse device is as follows: a 50 nm thick Sb2Te3 thin film is deposited on a glass substrate by atomic layer deposition, and then two Au electrodes are fabricated at both ends of the Sb2Te3 thin film by magnetron sputtering, thus obtaining an Au / Sb2Te3 / Au dual-ended photoelectric synapse device.

[0064] The artificial photoelectric synapse device generates photocurrents with wavelengths ranging from 450 to 638 nm and light intensities ranging from 0.35 to 1.8 W / cm². 2 The frequency is 0.1 to 1 Hz, and the number of pulses is 1 to 50.

[0065] Performance testing

[0066] 1. Photocurrent response test

[0067] The artificial photoelectric synapse device in Example 1 was analyzed using a Keithley 4200 semiconductor parameter analyzer at a wavelength of 450 nm and a light intensity of 1.4 W / cm². 2 Irradiate the device under a voltage of 5V and observe the changes in current generated by the artificial photoelectric synapse.

[0068] Depend on Figure 2 The test results in (a) show that, under the above conditions, the artificial photoelectric synapse device produced a typical photocurrent effect. Similar to biological synapses, illumination triggered a significant increase in current ΔI. This increase gradually decayed within tens of seconds after the stimulus was removed. The applicant believes the main reason is the presence of photon-modified surface states and photoinduced surface photocurrents in the topological insulator Sb₂Te₃ film. The photocurrent increased significantly with exposure, and gradually decreased when the light stimulus was turned off due to electron-hole recombination. The gradual decrease in photocurrent is due to defects in Sb₂Te₃, which introduced trapping levels and slowed down charge carrier recombination. ΔI initially increased with increasing exposure time, then gradually saturated due to the depletion of available electron-hole pairs, and finally reached a stable 0.24 mA due to the recombination of photogenerated and charge carriers.

[0069] Furthermore, when different light intensities and exposure times are applied to the artificial photoelectric synapse device in Example 1, the resulting changes in photocurrent are as follows: Figure 2 As shown in (b), it can be seen that increasing the light intensity and exposure time for artificial photosynapses can lead to a larger ΔI.

[0070] 2. Repetitive training

[0071] To simulate the short-term and long-term synaptic memory phenomena resulting from repeated stimulation, light pulses of different numbers and frequencies were applied to the artificial photoelectric synapse device in Example 1 at a wavelength of 450 nm. The resulting changes in photocurrent are as follows: Figure 2 (c) and Figure 2 As shown in (d), it can be seen that when the number of pulses increases from 1 to 50, or the frequency increases from 0.1Hz to 1Hz, ΔI increases significantly. This indicates that by adjusting the frequency and number of light pulses, the magnitude and decay period of the synaptic photocurrent can be adjusted, which is the memory of the artificial photoelectric synaptic device. In other words, the artificial synaptic device can achieve the transformation from short-term memory to long-term memory through repeated training.

[0072] 3. Simulate the human brain's learning process

[0073] The artificial photoelectric synapse device of Example 1 simulates the human brain's learning process at a light wavelength of 450 nm, including the first learning, first forgetting, relearning, and second forgetting processes. The light pulse parameters used are: light intensity 1.4 W / cm². 2Exposure time 0.5s, frequency 1Hz; by Figure 3 (e) It can be concluded that during the first learning process, the photocurrent of the artificial photosynaptic device increased by about 15 μA under the excitation of eight light pulses, and then the current dropped back to the initial state in 5.6 s, which is considered to be the first forgetting process. During the second learning process, only six light pulses were needed to achieve the same current increase. After that, in the same 5.6 s, the current decayed less than in the first forgetting process. This shows that the photosynaptic device can simulate human learning of learned things, and relearning is faster and forgetting is less.

[0074] Example 2

[0075] An application of an artificial photoelectric synapse device in image information processing is proposed, which combines an artificial photoelectric synapse device based on topological insulators with a recurrent neural network to recognize handwritten digit images.

[0076] Specifically, the application of artificial photoelectric synaptic devices in image information processing includes: extracting pixel features of a handwritten digit image from two sequences along the diagonal direction; calculating the duty cycle and frequency of the input light signal based on the extracted pixel features; converting the duty cycle and frequency into corresponding light pulse signals; applying the light pulse signals and reading voltage in the photoelectric test of the artificial photoelectric synaptic device; and using the measured photocurrent response as the input of a recurrent neural network (RNN) for image training and classification.

[0077] This embodiment uses images of handwritten digits "0" and "1" as examples, which respectively contain 20 rows and 20 columns of pixels, such as... Figure 3 As shown, the digital image is first extracted by two sequences along the diagonal direction, where pixel features are extracted using Photoshop to identify filled and empty pixels along the diagonal. Filled pixels in the handwritten digit image represent pixels where light pulses and readout voltages were applied during the synaptic process, while empty pixels represent pixels where only readout voltage was applied without light pulses. The duty cycle and frequency of the input light signal are calculated based on the ratio of filled to empty pixels along the diagonal direction. The duty cycle and frequency are then converted into corresponding light pulse signals. For the artificial photoelectric synapse device in Example 1, photoelectric testing is performed using a Keithley 4200 semiconductor parameter analyzer with a wavelength of 450nm and a light intensity of 1.4W / cm². 2 The artificial photosynapse device is illuminated by a light source, and a 5V reading voltage is applied across its terminals. The measured photocurrent response is used as the input to a recurrent neural network for image training and classification.

[0078] In this application, the duty cycle is the ratio of the number of filled pixels to the total number of pixels along the diagonal of the image, and the frequency of the light signal is the reciprocal of the sum of the pixels in that period.

[0079] During image training, the image data is randomized in the database. The parameters in the recurrent neural network (RNN) are set as follows: the ratio of training set to test set is 80% and 20%, respectively. A single-layer perceptron with Softmax as the activation function is used, an Adam optimizer with a learning rate of 3% is used, and the cross-entropy loss is minimized by gradient descent. The hidden layer consists of 10 RNN units, the second layer contains 50 neurons, and finally, a fully connected layer is connected to have two classification outputs to classify the digits 0 and 1.

[0080] Training effect:

[0081] The recognition accuracy of this embodiment was tested during 50 training cycles. The results show that after 23 training cycles, the recognition accuracy of the RNN can instantly stabilize at 100%. Figure 4 As shown.

[0082] To compare with traditional artificial neural network image recognition, the photocurrent response output by the artificial synaptic photoelectric device in this embodiment is input into a feedforward neural network (FNN) based on a multilayer perceptron for training and learning. Figure 4 It can be seen that after two training cycles, the image recognition accuracy reaches 63% and then saturates, unable to be further improved within 50 cycles. Furthermore, when inputting 400 pixels per image into the FNN for image recognition, its image recognition accuracy remains at 92.5%, lower than the accuracy of RNNs with synaptic signals.

[0083] Therefore, it can be seen that when artificial opto-synaptic devices based on topological insulators are combined with RNNs for image recognition, the accuracy can reach 100%, effectively improving the accuracy of image recognition. This plays a crucial role in robotics, autonomous driving, and medical disease diagnosis. Thus, RNNs based on artificial synapses show great potential in next-generation neuromorphic computing, artificial sensory perception, and the hardware implementation of humanoid robots.

[0084] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An application of an artificial photoelectric synapse device based on a topological insulator in image information processing, characterized in that, The artificial opto-synaptic device based on topological insulator consists of a glass substrate, an Sb2Te3 thin film on the substrate, and two Au electrodes at both ends of the thin film. The frequency at which the artificial photoelectric synapse device generates photocurrent response is 0.1–5 Hz; The number of pulses that generate a photocurrent response in the artificial photoelectric synapse device is 1 to 50. The thickness of the Sb2Te3 thin film is 10–100 nm; The artificial photosynapse is combined with a recurrent neural network to recognize handwritten digit images.

2. The application of the artificial photoelectric synapse device according to claim 1 in image information processing, characterized in that, The wavelength of light that generates photocurrent response in the artificial photoelectric synapse device is 400–700 nm.

3. The application of the artificial photoelectric synapse device according to claim 1 in image information processing, characterized in that, The light intensity of the photocurrent response generated by the artificial photoelectric synapse device is 0.35–1.8 W / cm². 2 .

4. The application of the artificial photoelectric synapse device according to claim 1 in image information processing, characterized in that, include: The pixel features of the handwritten digit image are extracted from two sequences along the diagonal direction. The duty cycle and frequency of the input light signal are calculated based on the extracted pixel features. The duty cycle and frequency are converted into corresponding light pulse signals. The light pulse signals and reading voltage are applied in the photoelectric test of the artificial photoelectric synapse device. The measured photocurrent response is used as the input of the recurrent neural network for image training and classification.

5. The application of the artificial photoelectric synapse device according to claim 4 in image information processing, characterized in that, The photoelectric test was conducted using a semiconductor parameter analyzer.

6. The application of the artificial photoelectric synapse device according to claim 4 in image information processing, characterized in that, The photoelectric test of the artificial photoelectric synapse device includes: irradiating the artificial photoelectric synapse device with light of wavelength 400-700nm, and applying a reading voltage to the electrodes at both ends of the artificial photoelectric synapse device.

7. The application of the artificial photoelectric synapse device according to claim 6 in image information processing, characterized in that, The reading voltage is 0.05 to 5V.

Images