Optoelectronic synaptic device based on anisotropic orientation-dependent β-ga2o3 and preparation method

Abstract

The application discloses a photoelectric synapse device based on anisotropic orientation dependence of beta-Ga2O3 and a preparation method, and is used for solving the limitations of a traditional synapse device in dynamic object recognition, time series prediction and multi-modal perception and the like tasks, and realizing a collaborative response across space-time dimensions. The synapse device adopts a sandwich structure and sequentially comprises a c-Al2O3 substrate, a beta-Ga2O3 functional layer and an electrode from bottom to top. The application firstly utilizes the inherent crystal anisotropy of the beta-Ga2O3 thin film to realize a uniaxial axon-multi synapse neuromorphic system. The device realizes a robust short-term to long-term memory (STM-LTM) conversion, a paired pulse facilitation (PPF) and an effective learning behavior. Through direction evolution synapse dynamics and space-time coding capability, the axon-multi synapse system realizes dynamic motion recognition of six in-plane trajectories under blind ultraviolet light irradiation, and the accuracy is more than 98% under a noise-free condition, and even when the input noise reaches 50%, the accuracy still remains more than 80%.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor electronic device technology, specifically relating to photoelectric synapse devices based on β-Ga2O3 anisotropic orientation-dependent and their fabrication methods. Background Technology

[0002] The rapid development of data applications has placed higher demands on computing efficiency and energy consumption. However, the traditional von Neumann architecture has fundamental limitations due to the physical separation of memory and processing units. This design not only requires frequent data transfers but also leads to excessive energy consumption and reduced overall system performance. To overcome these bottlenecks, neuromorphic computing, inspired by biological neural systems, has emerged. In biological neural systems, synapses, as the core functional connections between neurons, regulate the intensity and temporal dynamics of signal transmission through synaptic plasticity. Most existing neuromorphic computing systems employ an axon-monosynaptic connection structure, where each input signal interacts with only a single synapse. While this configuration can reproduce basic synaptic plasticity, it lacks the multi-branch, multi-channel collaborative mechanisms unique to biological neural systems, making it difficult to support highly parallel information transmission and complex task processing. The limitations of this monosynaptic structure are particularly prominent in tasks such as dynamic object recognition, time series prediction, and multimodal perception, as it cannot achieve collaborative responses across spatiotemporal dimensions.

[0003] Recent research has revealed an important yet often overlooked biological structure—multisynaptic connections. In this structure, a single axon forms multiple synapses on the same postsynaptic neuron. Each of these neurons has a unique firing threshold, and this form of synaptic organization is prevalent in the nervous systems of humans and other animals. Compared to the traditional single-axon-single-synapse structure, multisynaptic connections enable a single axon to form multiple synaptic connections with the same postsynaptic neuron, thereby achieving parallel reception of multidirectional, temporally differentiated, and multimodal inputs, and supporting more efficient and complex information processing.

[0004] In recent years, significant breakthroughs have been achieved in anisotropic low-dimensional materials (including black phosphorus, rhenium sulfides, and triclinic selenium nanosheets), opening new avenues for realizing crystal orientation-dependent axonal multisynaptic systems. Despite the enormous potential of these systems, inherent drawbacks such as poor environmental stability, complex fabrication processes, and limited scalability still exist, collectively hindering their practical application in large-scale neuromorphic computing. β-Ga₂O₃ materials possess an ultrawide bandgap of approximately 4.8 eV, with its intrinsic absorption edge located in the deep ultraviolet region (<280 nm), exhibiting minimal sensitivity to visible and near-ultraviolet light. This characteristic allows β-Ga₂O₃-based optoelectronic devices to achieve solar-blind detection without complex optical filtering, offering significant advantages for high signal-to-noise ratio ultraviolet detection, space communication, and security imaging applications. Furthermore, β-Ga₂O₃ readily forms oxygen vacancies (V₂O₃). O The concentration of these defect states can be effectively controlled through growth conditions, annealing atmosphere, and photoexcitation. These defect states exhibit multi-scale carrier capture and release dynamics under electrical or photostimulation, and their dynamic regulation mechanism is remarkably similar to the regulation of neurotransmitters in biological synapses, thus providing a key physical basis for achieving synaptic plasticity in artificial devices. Under deep ultraviolet light irradiation, photogenerated electron-hole pairs are generated in β-Ga₂O₃, and holes are captured by oxygen vacancy-related trap states. When the photostimulation is removed, the trap-assisted decapitation barrier in β-Ga₂O₃ prolongs the decay time of the photocurrent, thereby generating synaptic behaviors such as sustained photoconductivity. More importantly, β-Ga₂O₃ crystallizes in a low-symmetry monoclinic crystal system, where different crystallographic orientations exhibit significant differences in atomic arrangement, oxygen vacancy formation and migration behavior, and optical transition probabilities. Due to the intrinsic anisotropy of β-Ga₂O₃, devices fabricated along different crystallographic orientations exhibit heterogeneous synaptic properties. Therefore, it is hoped that orientation-dependent synaptic dynamics can be studied, and an integrated axon-multisynaptic device can be realized. This can effectively address the limitations of traditional synaptic devices in tasks such as dynamic object recognition, time series prediction, and multimodal perception, enabling collaborative responses across time and space dimensions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an anisotropic orientation-dependent photoelectric synaptic device and its fabrication method based on β-Ga2O3. The core of this invention lies in leveraging the characteristic of β-Ga2O3 crystallizing into a low-symmetry monoclinic structure, primarily the significant differences in atomic arrangement, oxygen vacancy formation and migration behavior, and optical transition probabilities exhibited by different crystallographic orientations. The anisotropic physical properties resulting from the fewer symmetry operations and reduced crystallographic symmetry of β-Ga2O3 are utilized. An orientation-dependent neural synaptic device was designed and studied for the first time, and an integrated axon-multisynaptic device was realized. β-Ga2O3 is crystallized into a low-symmetry monoclinic structure belonging to the C2 / m space group, with lattice parameters a≈6.5 Å, b≈8.0 Å, and c≈15.5 Å. The unit cell of β-Ga₂O₃ contains a mirror-symmetric (010) plane and a

[010] crystallographic direction, which exhibits both double rotational symmetry and helical rotational symmetry, with the b-axis serving as the rotation axis, indicating its intrinsic anisotropy. This crystal structure consists of two inequivalent Ga₂O₃ crystals. 3+ Cation (Ga I and Ga Ⅱ ) and three types of O 2- Anionic composition. Specifically, [Ga] Ⅱ O6 octahedrons form double-chain structures by sharing edges, and these double chains are composed of [Ga I [O4] Tetrahedral interconnections. Due to the inherent anisotropy of β-Ga2O3, devices fabricated along different crystallographic orientations exhibit heterogeneous synaptic properties. This will effectively address the limitations of single synapses in tasks such as dynamic object recognition, time series prediction, and multimodal perception. Such a multi-synaptic connection structure enables a single axon to form multiple synaptic connections with the same postsynaptic neuron, thereby achieving parallel reception of multidirectional, temporally differentiated, and multimodal inputs, and realizing collaborative responses across spatiotemporal dimensions.

[0006] A first aspect of this invention provides an anisotropic orientation-dependent photoelectric synaptic device based on β-Ga2O3. The synaptic device employs a sandwich structure, comprising, from bottom to top, a c-Al2O3 substrate, a β-Ga2O3 functional layer, and electrodes. The core innovation of this invention lies in the fact that the β-Ga2O3 functional layer is a material layer with high oxygen vacancies, regulated by growth conditions. These defect states exhibit multi-scale carrier capture and release dynamics under electrical or optical stimulation, and their dynamic regulation mechanism is remarkably similar to the regulation of neurotransmitters in biological synapses, thus providing a key physical basis for achieving synaptic plasticity in artificial devices. Furthermore, the β-Ga2O3 functional layer possesses a low-symmetry monoclinic structure, indicating its inherent anisotropic properties. Therefore, oxygen vacancy transport exhibits anisotropy, ensuring that devices fabricated along different crystallographic orientations exhibit heterogeneous synaptic characteristics, which is a crucial foundation for fabricating orientation-dependent neural synaptic devices.

[0007] As a preferred technical solution, the thickness of the β-Ga2O3 functional layer of the neural synaptic device is 100 nm to 400 nm. The electrode is preferably a circular silver electrode with a diameter of 1 mm. Furthermore, the β-Ga2O3 functional layer is grown in a low-oxygen environment, which ensures that the prepared β-Ga2O3 functional layer has high oxygen vacancies, a crucial prerequisite for the synaptic phenomenon exhibited by the β-Ga2O3 material.

[0008] A second aspect of the present invention provides a method for preparing the above-mentioned orientation-dependent neural synapses. The key to this method lies in ensuring high oxygen vacancies while achieving thin film crystallization through a plasma-enhanced chemical vapor deposition process. The method includes the following steps:

[0009] S1. Clean the c-Al2O3 substrate;

[0010] S2. Preparation of the β-Ga2O3 functional layer: A certain amount of metallic gallium (99.99%) was placed in a ceramic boat. Then, a cleaned c-Al2O3 substrate was inverted at the other end of the ceramic boat, 5 cm away from the metallic gallium. The ceramic boat was placed in the center of the quartz tube furnace of the PECVD apparatus. Once the pressure inside the quartz tube dropped to approximately 1 Pa, argon gas (99.999%) was introduced at a suitable flow rate to maintain a constant pressure inside the tube. In the CNC instrument of the PECVD apparatus, the temperature was set and the heater was turned on. The temperature inside the tube furnace was slowly increased to a suitable temperature at a rate of 10 °C / min. Then, oxygen was introduced and radio frequency was activated for material growth. After growth was complete, the material was cooled to room temperature and removed.

[0011] S3. A circular silver electrode with a diameter of 1 mm is formed on the surface of the β-Ga2O3 thin film by evaporation through a mask, thus completing the device fabrication.

[0012] Based on the aforementioned orientation-dependent synaptic characteristics, the memristor of this invention exhibits three outstanding application functions, further demonstrating its innovative comprehensive advantages:

[0013] 1. Continuously Tunable Photosynaptic Plasticity: Under stimulation with 254 nm ultraviolet light pulses, the device exhibits weighting behavior similar to that of biological synapses. By systematically changing the width, number, and intensity of the light pulses, the device's conductivity state can transition from short-term memory to long-term memory, and it can completely simulate the advanced neural activity process of "learning → forgetting → relearning," demonstrating its enormous application potential in biomimetic neural network hardware.

[0014] 2. High-Accuracy MNIST Handwritten Digit Recognition: To mitigate the impact of non-ideal synaptic features, an optimized weight update strategy was employed during network training, where weight values ​​were stored and updated as continuous values. A three-layer artificial neural network (ANN) consisting of 784 input neurons, 300 hidden neurons, and 10 output neurons was constructed for training and classification of the MNIST handwritten digit dataset. MNIST images representing the digits "0" to "9" were preprocessed with vertical and horizontal convolutional kernels, and the extracted features were then input into the input, hidden, and output layers of the ANN. A non-linear activation function was used in the hidden layers to enhance the network's mapping ability. Results showed that after 50 training epochs, the classification accuracy for the three channels on the MNIST dataset stabilized at approximately 95%, 90%, and 85%, respectively.

[0015] 3. Spatiotemporal Object Trajectory Recognition: Leveraging the anisotropic nature of oxygen vacancy transmission in the device, different motion directions of objects are encoded through different synaptic channels. By connecting adjacent frames, a pulse sequence is constructed for each pixel, thereby encoding spatiotemporal visual information into a temporal pulse pattern, which is then injected into a sensor-memory array. A convolutional neural network (CNN) with hidden layers is used as the readout layer, trained through backpropagation, to classify motion trajectories based on the temporally encoded input. This achieves the spatiotemporal encoding capability of the β-Ga2O3 axon-multisynaptic system.

[0016] In summary, this invention creatively constructs an orientation-dependent opto-synaptic device by leveraging the low-symmetry monoclinic structure of β-Ga₂O₃ and utilizing the anisotropic physical properties resulting from fewer symmetry operations and reduced crystallographic symmetry in β-Ga₂O₃. This structure not only possesses the performance of traditional synaptic devices but also fundamentally overcomes the limitations of traditional single-axon-single-synapse structures in parallel reception of multi-directional, temporally differentiated, and multimodal inputs. It enables a single axon to form multiple synaptic connections with the same postsynaptic neuron, thereby supporting more efficient and complex information processing. Therefore, this invention provides a core device-level solution for developing next-generation heterosynaptic, sensor-memory-computing integrated brain-like systems.

[0017] This invention provides an anisotropic orientation-dependent photoelectric synapse device based on β-Ga2O3 and its fabrication method. It has the following beneficial effects:

[0018] 1. This invention is the first to utilize the inherent crystal anisotropy of β-Ga2O3 thin films to realize an integrated axon-multi-synaptic system on a single device. This structure enables a single input to form multiple synaptic connections with the same output unit, each with different characteristics, thereby supporting parallel reception and coordinated response to multi-directional, time-differenced information, fundamentally overcoming the limitations of traditional axon-single-synaptic architectures.

[0019] 2. This invention utilizes the anisotropy of oxygen vacancy transport in β-Ga2O3, enabling the device to encode the motion trajectory of an object in different directions into different synaptic channels. Based on this, the system successfully achieves dynamic recognition of motion directions in six planes, with an accuracy exceeding 98% under noise-free conditions. Even with up to 50% Gaussian noise added to the input signal, the system maintains a recognition accuracy of over 80%, demonstrating strong environmental adaptability and robustness.

[0020] 3. Based on the synaptic characteristics of this device, an artificial neural network was constructed. When classifying and recognizing the classic MNIST dataset, the synaptic channels with three different crystal orientations achieved accuracies of approximately 95%, 90%, and 85%, respectively. This fully demonstrates that the device can effectively support complex pattern recognition tasks and provides differentiated computational performance due to its anisotropy, verifying its feasibility in practical applications.

[0021] 4. This invention seamlessly integrates solar-blind ultraviolet light sensing, synaptic weight information storage, and neuromorphic computing functions into a single optoelectronic synaptic device. This opens up new avenues for solving the memory wall problem in the traditional von Neumann architecture and developing efficient, low-power neuromorphic computing hardware, providing crucial device-level support for next-generation brain-inspired intelligent systems.

[0022] In summary, this invention not only solves the core problem of the limitations of traditional synaptic devices, but also realizes the function of recognizing the motion trajectory of orientation-dependent objects on a single device, with significant overall performance. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a photoelectric synapse device based on the anisotropic orientation dependence of β-Ga2O3.

[0024] Figure 2 The oxygen atom configuration is along the

[010] and

[102] directions;

[0025] Figure 3 The results show the oxygen vacancy concentration of the β-Ga2O3 thin film.

[0026] Figure 4 The angle-resolved Raman intensity of the β-Ga2O3 thin film;

[0027] Figure 5 An angle-resolved Raman image of a β-Ga2O3 thin film;

[0028] Figure 6 The image shows the dependence of EPSC on optical pulse intensity and pulse number for channels 1-3;

[0029] Figure 7 The learning-forgetting-relearning curves are for channels 1-3;

[0030] Figure 8 The PPF characteristics of channels 1-3;

[0031] Figure 9 The confusion matrix for handwritten digit recognition in channels 1-3;

[0032] Figure 10 Clustering mapping for handwritten digit recognition in channels 1-3;

[0033] Figure 11 The functional relationship between the classification accuracy of handwritten digits in MNIST on channels 1-3 and the number of training epochs;

[0034] Figure 12 A schematic diagram for recognizing the dynamic motion of an object;

[0035] Figure 13 The photocurrent response diagrams for channels 1-3 under spatiotemporally encoded optical pulse input are shown.

[0036] Figure 14 The motion recognition accuracy of channels 1-3 under different Gaussian noise levels. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1

[0039] The fabrication method of β-Ga2O3 anisotropic orientation-dependent photoelectric synaptic device includes the following steps:

[0040] S1. The c-Al2O3 substrate is pretreated by ultrasonic cleaning with acetone, anhydrous ethanol and deionized water for 10 minutes in sequence. After cleaning, it is dried with high-purity nitrogen and set aside for use.

[0041] S2. Place a certain amount of gallium metal (99.99%) in a ceramic boat, then invert the cleaned sapphire at the other end of the ceramic boat, 5 cm away from the gallium metal. Place the ceramic boat in the center of the quartz tube furnace of the PECVD apparatus. Once the pressure inside the quartz tube drops to approximately 1 Pa, introduce argon gas (99.999%) at a flow rate of 80 sccm to maintain the pressure inside the tube at a constant 40 Pa. In the CNC instrument of the PECVD apparatus, set the temperature and turn on the heater, slowly raising the temperature inside the tube furnace to 850℃ at a rate of 10 ℃ / min. Then, introduce oxygen at a flow rate of 2 sccm and turn on 150W RF for material growth for 2 hours. After growth is complete, cool to room temperature and remove the material.

[0042] S3. Cover with a mask and fabricate an Ag electrode with a diameter of 1 mm on the β-Ga2O3 functional layer by vapor deposition, thus obtaining the aforementioned photoelectric synapse device based on the anisotropic orientation dependence of β-Ga2O3.

[0043] Example 2

[0044] The aforementioned β-Ga2O3-based anisotropic orientation-dependent photoelectric synapse device comprises a c-Al2O3 substrate, a β-Ga2O3 functional layer, and a Ti / Au electrode. The β-Ga2O3 functional layer is located on the c-Al2O3 substrate, and the Ti / Au electrode is located on the β-Ga2O3 functional layer, thus completing the device fabrication. The β-Ga2O3 functional layer possesses high oxygen vacancies and a low symmetry structure, and is fabricated using plasma-enhanced chemical vapor deposition.

[0045] The fabrication method of β-Ga2O3 anisotropic orientation-dependent photoelectric synaptic device includes the following steps:

[0046] S1. The c-Al2O3 substrate is pretreated by ultrasonic cleaning with acetone, anhydrous ethanol and deionized water for 10 minutes in sequence. After cleaning, it is dried with high-purity nitrogen and set aside for use.

[0047] S2. Place a certain amount of gallium metal (99.99%) in a ceramic boat, then invert the cleaned sapphire at the other end of the ceramic boat, 5 cm away from the gallium metal. Place the ceramic boat in the center of the quartz tube furnace of the PECVD apparatus. Once the pressure inside the quartz tube drops to approximately 1 Pa, introduce argon gas (99.999%) at a flow rate of 80 sccm to maintain the pressure inside the tube at a constant 40 Pa. In the CNC instrument of the PECVD apparatus, set the temperature and turn on the heater, slowly raising the temperature inside the tube furnace to 830℃ at a rate of 10℃ / min. Then, introduce oxygen at a flow rate of 1 sccm and turn on 120W RF for material growth for 1 hour. After growth is complete, cool to room temperature and remove the material.

[0048] S3. Cover with a mask and fabricate a Ti / Au electrode with a diameter of 1.5 mm on the β-Ga2O3 functional layer by vapor deposition. This yields the aforementioned photoelectric synapse device based on the anisotropic orientation dependence of β-Ga2O3.

[0049] The above embodiments provide an anisotropic orientation-dependent opto-synaptic device based on β-Ga2O3. Through the anisotropy of the crystal structure, it realizes the difference in oxygen vacancy transport, which solves the limitation of the traditional single axon-single synapse structure in parallel reception of multi-directional, time-difference and multi-modal inputs. It enables a single axon to form multiple synaptic connections with the same postsynaptic neuron, thereby supporting more efficient and complex information processing and providing ideas for the design optimization of heterosynaptic devices.

[0050] Figure 1 This is a schematic diagram of a photoelectric synapse device based on the anisotropic orientation dependence of β-Ga2O3, which shows the specific structure of the device.

[0051] Figure 2 The oxygen atom configuration is shown along the

[010] and

[102] directions. As the oxidation potential decreases from an oxygen-rich environment to an oxygen-poor environment, the formation energy further decreases, leading to an increase in the oxygen vacancy concentration. Due to the inherent structural anisotropy of β-Ga2O3, in In oriented thin films, O II The local atomic arrangement at the sites differs significantly along the

[010] and

[102] directions, as shown in the figure. Along the

[010] direction, O II The atoms are arranged more closely, forming a more continuous atomic chain; and along the

[102] direction, O IIThe sites exhibit a relatively dispersed configuration with large interatomic spacing. This anisotropic local coordination environment leads to a direction-dependent tendency for oxygen vacancy formation, resulting in an anisotropic distribution of oxygen vacancy. Therefore, the

[010] channel accommodates a higher density of oxygen vacancy-related trap states, enabling more efficient capture of photogenerated holes under ultraviolet irradiation. Under repeated optical stimulation, the accumulation and slow release of trapped carriers produce the PPC effect, exhibiting strong synaptic plasticity. In contrast, the relatively low density of oxygen vacancy in the

[102] direction limits the availability of trap centers, resulting in weakened hole-trapping ability and a significant reduction in PPC response under repeated irradiation.

[0052] Figure 3 These are the results of oxygen vacancy concentration measurements on the β-Ga2O3 thin film. It can be seen that O... II It accounts for 23.26% of the total oxygen vacancies, and high oxygen vacancies are the basis for the device to achieve synaptic performance.

[0053] Figure 4 The angle-resolved Raman intensity of the β-Ga2O3 thin film is given.

[0054] Figure 5 This is an angle-resolved Raman image of the β-Ga₂O₃ thin film. The angle-resolved polarized Raman spectroscopy further reveals the significant structural anisotropy of the β-Ga₂O₃ thin film in the 200–1000 cm⁻¹ range. -1 The characteristic Raman modes observed in the band are consistent with previous findings. Under parallel polarization configuration, the polarization angle-dependent Raman intensity and mapping image exhibit clear periodic changes with rotation angle, confirming the strong in-plane anisotropy of the thin film.

[0055] Figure 6 The figure shows the dependence of EPSC for channels 1-3 on light pulse intensity and pulse number. As the light stimulation intensity and pulse repetition number increase, all three channels exhibit a clear transition from short-term memory (STM) to long-term memory (LTM), indicating that each channel can function as an effective photoelectronic synaptic device with reliable learning capabilities. Under relatively low light power density and few pulses, the EPSC values ​​of the three channels are similar, indicating that weak stimulation activates only a limited number of oxygen vacancy-related trapping sites. However, with further increases in light power density and pulse number, the EPSC of channel 1 is significantly greater than that of channels 2 and 3.

[0056] Figure 7The "learning-forgetting-relearning" curves for channels 1-3 are shown. The learning capacity of the three synaptic channels was assessed using a repetitive light pulse stimulation system. These three channels exhibited similar learning-forgetting-relearning behavioral patterns. In the initial learning phase, 14 consecutive light pulses were required to reach a stable EPSC state. After a 14-second no-stimulation interval, partial forgetting occurred. Notably, during relearning, only 5 light pulses were needed to restore the EPSC to a similar level, indicating a significant enhancement in learning capacity.

[0057] Figure 8 The figure shows the PPF characteristics of channels 1-3. As can be seen from the figure, the PPF values ​​of channels 1, 2 and 3 are approximately 148%, 121% and 106%, respectively. The significantly higher PPF value of channel 1 indicates that it has a stronger short-term facilitation effect and slower carrier relaxation dynamics, further confirming its excellent synaptic performance.

[0058] Figure 9 This is the confusion matrix for handwritten digit recognition using channels 1-3. During neural network training, the confusion matrix is ​​used to evaluate classification performance in detail. The results show that the ANN implemented based on the synaptic features of channel 1 has higher recognition accuracy and lower misclassification rate, demonstrating superior learning performance.

[0059] Figure 10 Clustering maps for handwritten digit recognition in channels 1-3 are shown. To visually demonstrate the learning ability of synaptic devices, clustering maps of MNIST handwritten digits were generated based on the feature representations extracted after training. The clustering map corresponding to channel 1 clearly presents independent clusters of all 10 digit categories, indicating its excellent feature discrimination and representation capabilities. In contrast, the clustering maps of channels 2 and 3 show significant overlap between multiple digit categories, indicating a decrease in their distinguishability.

[0060] Figure 11 This paper presents a functional relationship between the classification accuracy of MNIST handwritten digits on channels 1-3 and the number of training epochs. The device is used for training and classifying the MNIST handwritten digit dataset. MNIST images representing the digits "0" to "9" are preprocessed with vertical and horizontal convolutional kernels, and the extracted features are input into the input, hidden, and output layers of the ANN. The hidden layers employ a non-linear activation function to enhance the network's mapping ability. Simulation results show that after 50 training epochs, the classification accuracy of the three channels on the MNIST dataset stabilizes at approximately 95%, 90%, and 85%, respectively.

[0061] Figure 12This is a schematic diagram for recognizing dynamic motion of an object. To explore the application potential of axonal multi-synaptic devices in dynamic motion recognition, we constructed a physical perception scenario to identify the trajectory of a fighter jet and trained the system to recognize motion directions in six planes, including up, down, left, right, up-right, and down-left.

[0062] Figure 13 The figure shows the photocurrent response of channels 1-3 under spatiotemporally encoded optical pulse input. As shown, the photocurrent response of the three synaptic channels under optical pulse stimulation exhibits a significant dependence on the temporal sequence and spatial location of the input signal, highlighting the inherent sequence sensitivity of the device.

[0063] Figure 14 The motion recognition accuracy of channels 1-3 under different Gaussian noise levels is shown in the figure. To simulate a real sensing environment, Gaussian noise was introduced into the input signal. The classification accuracy and loss under different noise levels are summarized in the figure. As can be seen from the figure, the system achieves an accuracy of over 98% under noise-free conditions, and maintains an accuracy of over 80% even under 50% noise interference, demonstrating strong robustness to input perturbations. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A photoelectric synapse device based on the anisotropic orientation dependence of β-Ga2O3, comprising a substrate, a β-Ga2O3 functional layer, and electrodes, characterized in that: The β-Ga2O3 functional layer is composed of a β-Ga2O3 layer with high oxygen vacancies and low structural symmetry; The high oxygen vacancies in the β-Ga2O3 layer are beneficial to the synaptic performance of the device. Under repeated optical stimulation, the accumulation and slow release of trapped charge carriers will produce the PPC effect, which manifests as strong synaptic plasticity. The β-Ga2O3 layer has low structural symmetry, In the oriented thin film, the local atomic arrangement of the OII sites differs significantly along the [010] and [102] directions. Along the [010] direction, since the OII atoms are more densely packed, they can accommodate a higher density of oxygen vacancy-related trap states, which makes it more efficient to capture photogenerated holes under ultraviolet light irradiation. Conversely, the [102] direction can accommodate fewer oxygen vacancy-related trap states, thus generating oxygen vacancy differences.

2. The photoelectric synapse device based on β-Ga2O3 anisotropic orientation dependence according to claim 1, characterized in that, The β-Ga2O3 functional layer has an oxygen vacancy concentration of 10%-30%.

3. The photoelectric synapse device based on β-Ga2O3 anisotropic orientation dependence according to claim 1, characterized in that, The thickness of the β-Ga2O3 functional layer is 100 nm-500 nm.

4. The photoelectric synapse device based on β-Ga2O3 anisotropic orientation dependence according to claim 1, characterized in that, The electrode is a silver, titanium, gold, or platinum electrode.

5. A photoelectric synapse device, characterized in that, The synaptic weighting simulation unit includes a β-Ga2O3 anisotropic orientation-dependent photoelectric synaptic device as described in any one of claims 1 to 4. The conductivity state of the synaptic weight simulation unit can be continuously, reversibly, and non-discretely adjusted by applying ultraviolet light pulses.

6. The method for fabricating a β-Ga₂O₃ anisotropic orientation-dependent photoelectric synaptic device according to any one of claims 1 to 4, characterized in that, Includes the following sequential steps: S1. A β-Ga2O3 functional layer is formed on a c-Al2O3 substrate by PECVD; S2. Electrodes are prepared in the β-Ga2O3 functional layer by thermal evaporation.

7. The method for fabricating a photoelectric synaptic device based on β-Ga2O3 anisotropy orientation dependence according to claim 6, characterized in that, In step S1, the β-Ga2O3 functional layer is grown at 800 ℃-870 ℃ for 40 minutes to 120 minutes using ion-enhanced chemical vapor deposition (PECVD).

Images