A photonic neural synapse device with double micro-ring structure and convolution operation network model

Abstract

This invention discloses a photonic neural synapse device with a dual-micro-ring structure and a convolutional computation network model. The device includes a front resonant ring, a rear resonant ring, and two coupled phase-change material (PCM) films. The two PCM films are matched and correspond to the front and rear resonant rings, respectively. Synaptic weighting is achieved by controlling the coupled PCM films. The front and rear resonant rings are used to split the input optical signal in the bus waveguide into a 50:50 ratio, and output the difference signal current through a balanced photodetector. The photonic neural synapse of this invention has the advantages of high bandwidth and high speed. Once the synaptic weights are determined, it is a passive device with theoretically zero power consumption. Furthermore, this invention designs a dual-micro-ring structure, which ensures that the modulation of light does not affect the resonant wavelength, improving modulation accuracy, reducing optical transmission loss, and broadening the range of synaptic weights. This invention can be widely applied in the field of micro-nano optoelectronics.

Description

Technical Field

[0001] This invention relates to cutting-edge research fields that intersect multiple disciplines, including optics, integrated optoelectronics, neural network computing, nanomaterials, and micro / nano fabrication, and in particular to a photonic neural synapse device with a dual micro-ring structure and a convolutional computation network model. Background Technology

[0002] In recent years, Artificial Neural Networks (ANNs) have received considerable attention due to their widespread application in machine learning. With the rapid development of applications such as artificial intelligence and big data, the scale and computational demands of neural networks are growing rapidly. Convolutional Neural Networks (CNNs) are powerful and widely used tools capable of extracting features from large datasets, holding an unparalleled position in computer vision and natural language processing. CNNs require a large number of matrix operations during training. However, while the traditional von Neumann architecture can simulate the architecture of a neural network, its separation of computation and data storage is inherently unsuitable for neural network implementation. Convolution operations consume over 80% of the computer's computation time, resulting in significant energy consumption. In contrast, photonics has unique advantages in terms of energy consumption, bandwidth, and speed, especially in matrix computation and interconnection. Optical methods have enormous potential and are naturally compatible with neural networks, supporting various large-scale neuron connections. The combination of ANNs and optics provides a novel computational architecture: Optical Neural Networks (ONNs).

[0003] ONN views neural networks as a collection of many computational units and connections between them. Photonic neurons perform two important tasks: weighted summation and nonlinear activation. Weighted summation employs two methods: microring weighting and MZI (Multi-Incremental Photodetector). Microring weighting utilizes a waveguide-based multiply-accumulate (WMD) to transmit multiple optical signals in parallel. Microrings of different radii couple optical signals of different wavelengths. By altering the coupling efficiency of the microrings, the power of the output optical signal can be adjusted, thus loading weights. These weights are then input to a pair of balanced photodetectors (PDs) to sum the multiple optical signals, achieving the multiply-accumulate (MAC) operation. Changing the coupling efficiency of the microrings requires altering their effective refractive index, using methods such as thermo-optic modulation and electro-optic modulation. However, since the effective refractive index of a material is only slightly affected by electro-optic and thermo-optic effects, achieving a significant change requires adding a large modulator, which is inconvenient for integration. Furthermore, these two methods are not non-volatile, requiring heating or current to maintain effective device operation, increasing power consumption. Phase change materials are non-volatile, small in size, have a large effective refractive index change between crystalline and amorphous states, and have a long state retention time. They are ideal materials for micro-ring weighting and reassembly to reduce the integrated size.

[0004] In addition, for a single-wavelength coupled microring structure with through and download ports, weight allocation is achieved by subtraction through sending the outputs of both ports to a balanced photodiode. This method requires direct modulation of the effective refractive index onto the ring, which alters the resonant wavelength of the microring and affects the wavelength division multiplexing (WDM) effect. Furthermore, due to coupling losses between the waveguide and the microring, as well as waveguide losses, a complete synaptic weighting from -1 to 1 cannot be achieved. Summary of the Invention

[0005] In order to at least partially solve one of the technical problems existing in the prior art, the present invention aims to provide a photonic neural synapse device with a dual micro-ring structure and a convolutional computation network model.

[0006] The technical solution adopted in this invention is:

[0007] A photonic neural synapse device with a dual microring structure includes a front resonant ring, a rear resonant ring, and two coupled phase change material films. The two coupled phase change material films are matched and correspond to the front resonant ring and the rear resonant ring, respectively. Synaptic weighting is achieved by controlling the coupled phase change material films.

[0008] The front and rear resonant rings are used to split the input optical signal in the bus waveguide into 50%:50% beams, which are then output as a differential signal current via a balanced photodetector.

[0009] Furthermore, the phase change material selected for the coupled phase change material film is Ge2Sb2Te5, whose complex refractive index at 1550 nm is nc for its crystalline and amorphous states, respectively. c =7.197 + 1.27i and n a =4.433+0.119i.

[0010] Furthermore, the coupled phase change material film structure is a multi-level structure, which has n different phase change material units, and each phase change material unit can switch between crystalline and amorphous states;

[0011] The waveguide transmittance can be varied in stages from 0 to 1 through a multi-level structure.

[0012] Furthermore, the multi-level structure is divided into 5 levels, with a thickness of 30 nm, a width of 0.4 μm, and a single-level length of 0.2 μm.

[0013] Furthermore, the coupled phase change material film layer is used to achieve single-level control. The difference signal is output by inputting the pre-resonant ring and the post-resonant ring to adjust the balanced photodetector, that is, the weighted signal is output between -1, 0 and +1.

[0014] Furthermore, the photonic neural synapse, from bottom to top, consists of silicon (Si), a silicon dioxide (SiO2) buried oxide layer covering the silicon substrate, a waveguide layer, and an oxide cladding. The waveguide uses materials such as Si, Si3N4, and LiNbO3, which have the same transparent window as the silicon substrate. The waveguide structure is designed as a strip waveguide with a width of 400 nm and a thickness of 220 nm. The radius of the microrings and the coupling gap can be adjusted and selected according to the operating wavelength, bandwidth, and splitting coefficient, based on the photonic neural synapse unit, so that light pulse signals of different wavelengths operate within the corresponding microrings.

[0015] Another technical solution adopted in this invention is:

[0016] A convolutional computation network model includes a bus waveguide, a balanced photodetector, a transimpedance amplifier, and k photonic neural synaptic devices as described above; k is a positive integer greater than 1.

[0017] The front resonant ring and the corresponding coupling phase change material film in the k photonic synaptic devices constitute the upper waveguide, and the rear resonant ring and the corresponding coupling phase change material film in the k photonic synaptic devices constitute the lower waveguide.

[0018] The bus waveguide contains input optical signals of all wavelengths. After being modulated by the coupling phase change material film in the photonic nerve synapse, synaptic weighting is achieved. That is, the intensity of each different wavelength optical signal is multiplied by the corresponding synaptic weight. Finally, the incoherent summation of all product signals is completed by the balanced photodetector, realizing the function of matrix operation. The magnitude of the output current signal represents the operation result.

[0019] The transimpedance amplifier is used to amplify the signal output by the balanced photodetector.

[0020] Furthermore, the coupled phase change material film can be controlled in a single stage, and then the upper and lower waveguides are respectively input to adjust the balanced photodetector to achieve the difference signal output of the upper and lower waveguides, that is, the weighted signal is output between -1, 0 and +1.

[0021] Furthermore, when the coupling phase change material films corresponding to the upper and lower waveguides are both crystalline or both amorphous, the output of the balanced photodetector is 0, that is, the weighted value is 0.

[0022] When the coupling phase change material film corresponding to the upper waveguide is amorphous and the coupling phase change material film corresponding to the lower waveguide is crystalline, the positive output current of the balanced photodetector represents a weight of 1.

[0023] When the coupling phase change material film corresponding to the upper waveguide is crystalline and the coupling phase change material film corresponding to the lower waveguide is amorphous, the reverse current of the output of the balanced photodetector represents a weight of -1.

[0024] Furthermore, the coupled phase change material film structure is a multi-level structure, which has n different phase change material units, and multi-level control is achieved based on multiple phase change material units;

[0025] When all phase change material units are in the amorphous state, the optical signal transmittance in the waveguide is close to 1; when all phase change material units are in the crystalline state, the optical signal transmittance in the waveguide is close to 0; when some phase change material units are in the crystalline state and others are in the amorphous state, the optical signal transmittance in the waveguide is between 0 and 1. The optical transmittance value can be designed and adjusted according to the number, structure and size of the phase change material film units.

[0026] Furthermore, in the multi-level modulation mode of photonic neural synapse devices, when the light transmittance of the upper waveguide is T + The light transmittance of the lower waveguide is T - The balanced photodetector converts the optical signals from the upper and lower waveguides into electrical signals and subtracts them. The output current of the balanced photodetector is i. PD With (T) + -T - Linear correlation; with T + Maximum value, T - The minimum value is normalized as a synaptic weight of 1 to achieve quasi-continuous control of multiple state values ​​between -1 and 1.

[0027] The photonic neural synapse device of the present invention has the following features and advantages:

[0028] Compared to electronic devices operating in traditional computers, the photonic neural synapse of this invention has the advantages of high bandwidth and high speed. Once the synaptic weights are determined, it is a passive device with theoretically zero power consumption. Compared to other photonic neural synapse devices, this invention designs a dual-micro-ring structure. This method does not affect the resonant wavelength when modulating light, thus improving modulation accuracy, reducing light transmission loss, and broadening the range of synaptic weights.

[0029] The convolutional neural network architecture of this invention has the following features and advantages:

[0030] Compared to the electronic neural networks built on traditional von Neumann architecture computers, the optical convolutional neural network of this invention has the advantages of high bandwidth, high speed, and low power consumption. It utilizes wavelength division multiplexing (WDM) technology to fully leverage the high bandwidth advantage of photonics, enabling rapid parallel data processing and the formation of large-scale data processing networks, with broad application prospects in data centers. This lays the foundation for the application of optical computing chips for convolutional neural network operations. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the photonic neural synapse device in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of a phase change material with multi-level regulation in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the transmittance of the front resonant ring in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the transmittance of the rear resonant ring in an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the transmittance of phase change materials with different weighting levels in an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram of the convolutional neural network architecture in an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of a single microring resonator in an embodiment of the present invention;

[0039] Figure 8 This is a schematic diagram of the convolution operation in an embodiment of the present invention.

[0040] Figure 1 The attached figures are labeled as follows: 1-front resonant ring; 2-rear resonant ring; 3-first coupling phase change material film layer; 4-second coupling phase change material film layer; 5-balanced photodetector; 6-transimpedance amplifier; 7-bus waveguide.

[0041] Figure 6 The attached diagrams are labeled as follows: 8-bus waveguide; 21-upper waveguide; 22-lower waveguide; 23-balanced photodetector; 24-transimpedance amplifier. Detailed Implementation

[0042] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0043] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0044] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0045] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0046] Example 1

[0047] like Figure 1 The present embodiment provides a photonic neural synapse device with a dual microring structure, including a front resonant ring 1, a rear resonant ring 2 and two coupled phase change material films. The two coupled phase change material films are respectively matched and correspond to the front resonant ring 1 and the rear resonant ring 2. Synaptic weighting is achieved by controlling the coupled phase change material films.

[0048] The front resonant ring 1 and the rear resonant ring 2 are used to split the input optical signal in the bus waveguide 7 by 50%:50%, and output the differential signal current through the balanced photodetector 5.

[0049] The front resonant ring 1 of the photonic neural synapse device couples 50% of the light of a single wavelength in the bus waveguide 7, and the rear resonant ring 2 couples the remaining 50% of the light to 100%, thereby achieving 50%:50 splitting of the light of a single wavelength.

[0050] As an optional implementation, the two coupled phase change material (PCM) films include a first coupled PCM film 3 and a second coupled PCM film 4. The first coupled PCM film 3 is matched with the front resonant ring 1, and the second coupled PCM film 4 is matched with the rear resonant ring 2. The structure of the coupled PCM film is as follows: Figure 2 The multi-level structure shown demonstrates how phase change materials (PCMs) can significantly alter the transmittance of the optical waveguide. The multi-level structure contains n different PCM units, each capable of switching between crystalline and amorphous states. While a single PCM unit only allows for two states, a multi-level structure with n PCM units allows for two states to be switched. n These are different states. After normalization, the multi-level structure achieves hierarchical control of waveguide transmittance between 0 and 1.

[0051] As an optional implementation, the phase change material is a non-volatile reversible material. In this embodiment, the phase change material is GST, i.e., crystalline Ge2Sb2Te5. The phase change material is divided into 5 stages, with a thickness of 30 nm, a width of 0.4 μm, and a single stage length of 0.2 μm. The planar waveguide is 220 nm high and 400 nm wide. The resonant ring has a width of 400 nm and a radius of 5 μm. The distance between the front resonant ring 1 and the bus waveguide 7 is 100 nm, and the distance between the front resonant ring 1 and the output branch waveguide is 100 nm. The distance between the rear resonant ring 2 and the bus waveguide 7 is 100 nm, and the distance between the rear resonant ring 2 and the output branch waveguide is 200 nm. The bus waveguide 7 of the convolution kernel unit can be directly connected to the bus waveguide 7 of the next convolution kernel structure unit with a different resonant wavelength, realizing wavelength division multiplexing.

[0052] As an optional implementation, the photonic neural synapse device, from bottom to top, consists of silicon (Si), a silicon dioxide (SiO2) buried oxide layer covering the silicon substrate, a waveguide layer, and an oxide cladding. The waveguide uses materials such as Si, Si3N4, or LiNbO3, which have the same transparent window as the silicon substrate. The waveguide structure is designed as a strip waveguide with a width of 400 nm and a thickness of 220 nm. The radius of the microrings and the coupling gap can be adjusted and selected according to the operating wavelength, bandwidth, and splitting coefficient, based on the photonic neural synapse unit, so that optical pulse signals of different wavelengths operate within the corresponding microrings.

[0053] See Figure 6As an optional implementation, a convolutional computation network model can be further constructed based on the aforementioned photonic synaptic devices, including k photonic synaptic devices, a bus waveguide 7, an upper waveguide 21, a lower waveguide 22, a balanced photodetector 23, and a transimpedance amplifier 24. The bus waveguide 8 contains input optical signals of all wavelengths, which are modulated by a phase-change material to achieve synaptic weighting, i.e., the intensity of each different wavelength optical signal is multiplied by its corresponding synaptic weight. Finally, the balanced photodetector 23 performs incoherent summation of all product signals, realizing matrix operations. The magnitude of the output current signal represents the computation result. Figure 6 In the diagram, the front resonant ring 9 represents the front resonant ring of the first photonic synaptic device, the front resonant ring 13 represents the front resonant ring of the second photonic synaptic device, and the front resonant ring 17 represents the front resonant ring of the k-th photonic synaptic device; the coupling phase change material film 10 represents the first coupling phase change material film of the first photonic synaptic device, the coupling phase change material film 14 represents the first coupling phase change material film of the second photonic synaptic device, and the coupling phase change material film 18 represents the first coupling phase change material film of the k-th photonic synaptic device; the aforementioned front resonant rings and the first coupling phase change material film constitute the upper waveguide 21. Post-resonant ring 11 represents the post-resonant ring of the first photonic synaptic device, post-resonant ring 15 represents the post-resonant ring of the second photonic synaptic device, and post-resonant ring 19 represents the post-resonant ring of the k-th photonic synaptic device; coupling phase change material film 12 represents the second coupling phase change material film of the first photonic synaptic device, coupling phase change material film 16 represents the second coupling phase change material film of the second photonic synaptic device, and coupling phase change material film 20 represents the second coupling phase change material film of the k-th photonic synaptic device; the aforementioned post-resonant rings and second coupling phase change material films constitute the lower waveguide 22.

[0054] As an optional implementation, the photonic synapses used in the convolutional neural network architecture operate in non-overlapping wavelengths, and the resonant wavelengths of the photonic synapses correspond one-to-one with the wavelengths of the optical pulses transmitted in the main waveguide. Optionally, the convolutional neural network architecture contains at least two photonic synapses. The number of photonic synapses and their corresponding resonant wavelengths can be adjusted according to actual application requirements.

[0055] As an optional implementation, the balanced photodetector 23 converts the optical signals from the upper and lower waveguides into electrical signals and subtracts them, then amplifies the differential signal via a transimpedance amplifier 24. Through the hierarchical modulation of the aforementioned phase change material, graded control of the synaptic weights between -1 and 1 can be achieved.

[0056] Example 2

[0057] like Figure 1As shown, the photonic neural synapse device with a dual-microring structure consists of a front resonant ring 1, a rear resonant ring 2, a phase change material (i.e., a coupling phase change material film layer), a balanced photodetector 5, a transimpedance amplifier 6, and a bus waveguide 7 for wavelength division multiplexing. The phase change material includes a first coupling phase change material film layer 3 corresponding to the front resonant ring 1, and a second coupling phase change material film layer 4 corresponding to the rear resonant ring 2.

[0058] The front resonant ring 1 of the photonic neural synapse device achieves 50% coupling of a single wavelength of light in the main waveguide, and the rear resonant ring 2 achieves 100% coupling of the remaining 50% of the light. Under this effect, a 50%:50 split of light of a single wavelength is achieved.

[0059] The structure of the phase change material is as follows: Figure 2 The multi-level structure shown demonstrates how the phase change material significantly alters the transmittance of the optical waveguide, achieving graded variations in transmittance between 0 and 1. The phase change material is a non-volatile, reversible material; in this embodiment, it is GST, i.e., crystalline Ge2Sb2Te5. The phase change material is divided into 5 levels, with a thickness of 30 nm, a width of 0.4 μm, and a single-level length of 0.2 μm. The planar waveguide has a height of 220 nm and a width of 400 nm. The number of levels and dimensions of the phase change material 34 can be set according to actual application requirements. In this embodiment, the phase change material is divided into 5 levels, with a thickness of 30 nm, a width of 0.4 μm, and a single-level length of 0.2 μm.

[0060] The microring radius, coupling gap, and number of the microrings (i.e., front resonant ring 1 and rear resonant ring 2) are adjusted according to the requirements of the operating wavelength, bandwidth, and splitting coefficient. In this embodiment, the resonant ring width is 400 nm, the resonant ring radius is 5 μm, the distance between the front resonant ring and the bus waveguide 7 is 100 nm, the distance between the front resonant ring and the output branch waveguide is 100 nm, the distance between the rear resonant ring and the bus waveguide 7 is 100 nm, and the distance between the rear resonant ring and the output branch waveguide is 200 nm.

[0061] The balanced photodetector 5 converts the optical signals from the upper and lower waveguides into electrical signals, subtracts them, and then amplifies them via the transimpedance amplifier 6. Through the hierarchical modulation of the aforementioned phase change material, hierarchical modulation of the synaptic weights between -1 and 1 can be achieved.

[0062] Specifically, Figure 7 This is a schematic diagram of a single microring resonator. An optical pulse is input to the microring resonator from input port 25. If the wavelength of the input pulse is within the operating band of the microring resonator, the optical pulse is coupled through microring 27 and output from download port 28. Otherwise, it is output from through port 26. Figure 3The graph shows the relationship between the transmittance of the download and through ports of the pre-resonant ring 1 and the input light wavelength. The output light pulse of the download section of the pre-resonant ring 1 can maintain a transmittance of more than 50% in the 1574-1575nm and 1591-1592nm wavelength bands, and a transmittance of more than 40% in the 1557-1558nm wavelength band. Figure 4 The graph shows the relationship between the transmittance of the download and through ports of the rear resonant ring 2 and the input light wavelength. The output light pulse of the download section of the rear resonant ring 2 can achieve nearly 100% transmittance in the wavelength bands of 1557–1558 nm, 1574–1575 nm, and 1591–1592 nm. Therefore, by reasonably selecting the input pulse wavelength, a 50%:50% beam splitting can be achieved. The operating wavelength band and splitting coefficient of the resonant ring can be adjusted by changing the resonant ring radius and coupling gap.

[0063] Specifically, Figure 2 This is a phase change material unit with 5 levels of modulation. At level 0 weighting, phase change materials a, b, c, d, and e are set to amorphous state; at level 1 weighting, phase change materials b, c, d, and e are set to amorphous state, and a is set to crystalline state; at level 2 weighting, phase change materials c, d, and e are set to amorphous state, and a and b are set to crystalline state; at level 3 weighting, phase change materials d and e are set to amorphous state, and a, b, and c are set to crystalline state; at level 4 weighting, phase change material e is set to amorphous state, and a, b, c, and d are set to crystalline state; at level 5 weighting, phase change materials a, b, c, d, and e are set to crystalline state.

[0064] Specifically, Figure 5 The graph shows the transmittance versus wavelength for the six states of this unit. At the resonant wavelength of 1.5745 nm, the waveguide transmittances for weighted levels 0, 1, 2, 3, 4, and 5 are 0.025, 0.058, 0.103, 0.205, 0.491, and 0.888, respectively. After normalization, these are 0.0281, 0.0653, 0.116, 0.231, 0.553, and 1, representing weights of 0, 0.0625, 0.125, 0.25, 0.5, and 1, respectively. By changing the state of this phase change material unit, six levels of transmittance control can be achieved. This unit is placed at the download ports of the front resonant ring 1 and the rear resonant ring 2, enabling a total of 6 × 6 = 36 different synaptic weights. By removing duplicate weight values, 23 states can be achieved: 1, 0.9375, 0.875, 0.75, 0.5, 0.4375, 0.375, 0.25, 0.1875, 0.125, 0.0625, 0, -0.0625, -0.125, -0.1875, -0.25, -0.375, -0.4375, -0.5, -0.75, -0.875, -0.9375, and -1.

[0065] The bus waveguide 7 of the convolution kernel unit can be directly connected to the bus waveguide of the next convolution kernel structure unit with a different resonant wavelength to realize wavelength division multiplexing function.

[0066] Example 3

[0067] This embodiment provides an optical convolutional neural network architecture implemented using the aforementioned photonic synaptic device, the structure of which is as follows: Figure 6 As shown, it includes k photonic neural synapse devices, a bus waveguide 8, an upper waveguide 21, a lower waveguide 22, a balanced photodetector 23, and a transimpedance amplifier 24.

[0068] The optical convolutional neural network consists of a front resonant ring (9, 13, 17), a rear resonant ring (11, 15, 19), and phase change materials (10, 12, 14, 16, 18, 20).

[0069] Specifically, wavelength division multiplexing (WDM) signals contain optical pulse signals of different wavelengths, λ... i The intensity value of the light pulse signal represents different pixels A. λi The wavelength of the optical pulse signal is the same as the resonant wavelength of each microring, and it is transmitted through the bus waveguide 8 to each photonic synapse device (9 / 10 / 11 / 12, 13 / 14 / 15 / 16, 17 / 18 / 19 / 20). Each photonic synapse device has a different resonant wavelength due to its different ring radius. For optical pulses outside the operating band of the microring, the optical pulse signal passes normally. For optical pulse signals in the operating band, the front resonant ring located above the bus waveguide acquires 50% of the optical pulse signal, while the rear resonant ring located below the bus waveguide has the same resonant wavelength as the corresponding front resonant ring and acquires the remaining 50% of the optical pulse signal. Different order-weighted phase change materials result in different transmittances in the branch waveguides below them. The branch waveguide at the exit of the front resonant ring has a transmittance... The branch waveguide at the exit of the resonant ring has transmittance. The optical signal in the upper waveguide is the sum of the output signals of the pre-resonant loops of each neural synaptic device. The optical signal in the lower waveguide is the sum of the output signals of the post-resonant loops of each neural synaptic device. A balanced photodetector converts the optical signals from the upper and lower waveguides into electrical signals, and then subtracts them. The output is...

[0070]

[0071] Where g is the conversion coefficient, let The output is

[0072]

[0073] make but

[0074]

[0075] It implements vector operations.

[0076] Figure 8 This is a schematic diagram of convolution. The input image is represented as a numerical matrix of dimensions H×W×D, where H, W, and D are the height, width, and depth of the image, respectively. Each element A i,j The kernel F represents the intensity of a pixel at a specific spatial location. It is an R×R×D matrix where each element is defined as Fintensity. i,j The kernel is multiplied by an element of the same size in the image, resulting in one pixel of the output convolutional image. The kernel slides across the image, ultimately producing a complete output image. For example... Figure 8 As shown, the process of multiplying a convolution kernel with elements of the same size in the image can be considered as performing a vector-to-vector multiplication operation between the vector composed of elements in the image and the vector composed of elements in the convolution kernel. Assuming the input image is a 3×3×1 image and the kernel is a 2×2 convolution kernel, then one convolution operation requires...

[0077] (A 1，1 A 1，2 A 2，1 A 2，2 )·(F 1，1 ,F 1，2 ,F 2，1 ,F 2，2 )

[0078] (A 1，2 A 1，3 A 2，2 A 2，3 )·(F 1，1 ,F 1，2 ,F 2，1 ,F 2，2 )

[0079] (A 2，1 A 2，2 A 3，1 A 3，2 )·(F 1，1 ,F 1，2 ,F 2，1 ,F 2，2 )

[0080] (A 2，2 A 2，3 A 3，2 A 3，3 )·(F 1，1 ,F 1，2 ,F 2，1 ,F 2，2)

[0081] A total of four vector multiplications are performed to obtain four output pixels, ultimately resulting in a complete output image.

[0082] Therefore, convolution can be achieved through operations between vectors.

[0083] In summary, this embodiment provides a photonic neural synapse device with a dual-microring structure and a corresponding photonic convolutional neural network architecture based on this synapse. This architecture stabilizes the resonant wavelength of the convolutional units, increases the range of synaptic weights, and significantly reduces power consumption. The coupled phase-change material film is located on the branch waveguide. When subjected to thermal / electrical stimulation, it undergoes a phase transition between an amorphous and crystalline state, thus adjusting the synaptic weights, changing the transmittance of the branch waveguide, and altering the signal output energy. Ultimately, this causes the output signal of the balanced photodetector to change across multiple states. Finally, by cascading multiple of these units, combining the low power consumption and high bandwidth characteristics of the microring resonator with the stability and speed advantages of phase-change modulation, the convolutional neural network computation function is realized. This provides a framework and application basis for developing high-speed, low-power, and high-precision photonic convolutional kernels and other photonic integrated applications.

[0084] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0085] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0086] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A photonic neural synapse device with a dual microring structure, characterized in that, It includes a front resonant ring, a rear resonant ring, and two coupled phase change material films. The two coupled phase change material films are matched and correspond to the front resonant ring and the rear resonant ring, respectively. Synaptic weighting is achieved by controlling the coupled phase change material films. The front and rear resonant rings are used to split the input optical signal in the bus waveguide into 50%:50% beams, which are then output as a differential signal current through a balanced photodetector. The coupled phase change material film structure is a multi-level structure, which has n different phase change material units, and each phase change material unit can switch between crystalline and amorphous states. The waveguide transmittance can be varied in stages from 0 to 1 through a multi-level structure; The ring radius, coupling gap, and number of the front and rear resonant rings are adjusted according to the requirements of the operating wavelength, bandwidth, and splitting coefficient.

2. The photonic neural synapse device with a dual microring structure according to claim 1, characterized in that, The phase change material selected for the coupling phase change material film layer is... .

3. The photonic neural synapse device with a dual microring structure according to claim 1, characterized in that, The coupled phase change material film is used to achieve single-stage control.

4. A convolutional operation network model, characterized in that, It includes a bus waveguide, a balanced photodetector, a transimpedance amplifier, and k photonic neural synapse devices as described in any one of claims 1-3; where k is a positive integer greater than 1. The front resonant ring and the corresponding coupling phase change material film in the k photonic synaptic devices constitute the upper waveguide, and the rear resonant ring and the corresponding coupling phase change material film in the k photonic synaptic devices constitute the lower waveguide. The bus waveguide contains input optical signals of all wavelengths. After being modulated by the coupling phase change material film in the photonic nerve synapse, synaptic weighting is achieved. That is, the intensity of each different wavelength optical signal is multiplied by the corresponding synaptic weight. Finally, the incoherent summation of all product signals is completed by the balanced photodetector, realizing the function of matrix operation. The magnitude of the output current signal represents the operation result. The transimpedance amplifier is used to amplify the signal output by the balanced photodetector.

5. A convolutional operation network model according to claim 4, characterized in that, The coupled phase change material film can be controlled in a single stage, and then the upper and lower waveguides are used to input and adjust the balanced photodetector to achieve the output of the difference signal between the upper and lower waveguides, that is, the weighted signal is output between -1, 0 and +1.

6. A convolutional operation network model according to claim 5, characterized in that, When the coupling phase change material films corresponding to the upper and lower waveguides are both crystalline or both amorphous, the output of the balanced photodetector is 0, that is, the weighted value is 0. When the coupling phase change material film corresponding to the upper waveguide is amorphous and the coupling phase change material film corresponding to the lower waveguide is crystalline, the positive output current of the balanced photodetector represents a weight of 1. When the coupling phase change material film corresponding to the upper waveguide is crystalline and the coupling phase change material film corresponding to the lower waveguide is amorphous, the reverse current of the output of the balanced photodetector represents a weight of -1.

7. A convolutional operation network model according to claim 4, characterized in that, The coupled phase change material film structure is a multi-level structure with n different phase change material units, and multi-level control is achieved based on multiple phase change material units. When all phase change material units are in the amorphous state, the optical signal transmittance in the waveguide is close to 1; when all phase change material units are in the crystalline state, the optical signal transmittance in the waveguide is close to 0; when some phase change material units are in the crystalline state and others are in the amorphous state, the optical signal transmittance in the waveguide is between 0 and 1. The optical transmittance value can be designed and adjusted according to the number, structure and size of the phase change material film units.

8. A convolutional operation network model according to claim 7, characterized in that, In the multi-level modulation mode of photonic neural synapse devices, when the light transmittance of the upper waveguide is... The light transmittance of the lower waveguide is The balanced photodetector converts the optical signals from the upper and lower waveguides into electrical signals and subtracts them. The output current of the balanced photodetector is... and Linear correlation; with Maximum value The minimum value is normalized as a synaptic weight of 1 to achieve quasi-continuous control of multiple state values ​​between -1 and 1.

Images