Three-dimensional heterogeneous integrated optical computing chip

Abstract

The application discloses a three-dimensional heterogeneous integrated optical computing chip, and belongs to the optical computing field.The computing chip is composed of three parts: a spatial diffraction optical neural network core particle, a planar matrix reconfigurable optical neural network component and a planar switchable optical switch module; a single or multiple spatial diffraction optical neural network core particles, planar matrix reconfigurable optical neural network components and planar switchable optical switch modules are combined together in a three-dimensional heterogeneous integrated manner to realize efficient multi-task optical computing processing.The application utilizes the mutual combination of the spatial diffraction optical neural network and the planar matrix reconfigurable optical neural network, breaks through the traditional optical computing architecture, has a wide application prospect in the optical computing field, and fills the blank in the related technical field.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of optical computing, and more particularly relates to a three-dimensional heterogeneous integrated optical computing chip. BACKGROUND

[0002] With the rapid development of artificial intelligence and deep learning, the demand for processing large-scale data and complex tasks is gradually increasing. Traditional computing platforms such as CPU and GPU have performance bottlenecks and excessive energy consumption when facing these demands, which prompts people to seek new computing methods to meet the growing computing demand.

[0003] Optical computing, as a new computing method, can exhibit great advantages in processing large-scale data with the help of optical parallelism and high speed. Compared with electronic computing, optical computing has the advantages of high parallelism, low power consumption, anti-interference and large capacity transmission. Optical computing has a wide application prospect in various fields, including artificial intelligence, image processing, data mining, pattern recognition, etc.

[0004] With the development of optical computing, the current mainstream optical neural network has a spatial diffraction neural network and a planar linear neural network. The spatial diffraction neural network uses the diffraction of light between multiple phase planes to realize multi-layer full connection to realize image recognition and other tasks. The network has a large number of neurons and can realize a large neural network, but the network cannot be freely adjusted after training and has certain limitations. The planar linear neural network can be flexibly regulated to realize different matrix configurations and operations, but the operation scale is often limited by the network size. In view of this, a three-dimensional heterogeneous integrated optical computing chip is proposed, which combines the spatial diffraction neural network and the planar linear matrix network through three-dimensional heterogeneous integration to realize efficient multi-task optical computing processing. SUMMARY

[0005] For the traditional optical neural network computing architecture, the present application provides a three-dimensional heterogeneous integrated optical computing chip, which aims to provide a new type of optical neural network architecture to realize efficient multi-task optical computing processing. The input task is feature extracted and dimensionally reduced by the spatial diffraction optical neural network, and then processed and operated by the planar matrix reconfigurable optical neural network. Finally, the output calculation result is detected by the detector array. By flexibly switching the planar matrix reconfigurable optical neural network and different spatial diffraction optical neural network chips, multi-task flexible switching can be realized.

[0006] To achieve the above object, the application provides a three-dimensional heterogeneous integrated optical computing chip, which comprises a spatial diffraction optical neural network core particle, a planar matrix reconfigurable optical neural network component and a planar switchable optical switch module, the spatial diffraction optical neural network core particle, the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module are integrated in a three-dimensional heterogeneous manner, the spatial diffraction neural network core particle comprises not less than one layer of multi-phase planes, and the multi-phase planes are propagation layers; the planar matrix neural network component comprises a linear matrix network module and a detector array connected in sequence.

[0007] The spatial diffraction neural network core particle is used for feature extraction and dimension reduction of a task to be processed to obtain a feature output result, the feature output result is input into the planar switchable optical switch module in a coupling manner for state configuration, and the planar matrix neural network component is used for operation on the feature output result after state configuration.

[0008] The application has the beneficial effects that the combination of a single or multiple spatial diffraction optical neural network core particles, a planar matrix reconfigurable optical neural network component and a planar switchable optical switch module is realized in a heterogeneous integration manner, and efficient multi-task optical computing processing is realized. In the computing process of the optical computing chip, a task to be processed is input into the spatial diffraction optical neural network core particle for feature extraction and dimension reduction, then the task is input into the planar switchable optical switch module in a coupling manner for state configuration, finally the task is input into the planar matrix reconfigurable optical neural network component for operation and processing, and the computing result is detected and output by the detector array integrated on the planar matrix reconfigurable optical neural network component. In the optical computing process, a nonlinear activation function can be inserted to realize nonlinear operation, so as to improve the accuracy of task recognition. For a data set task with similar features, the matrix parameters of the planar matrix reconfigurable optical neural network component can be reconfigured; for a data set task with significantly different features, the configuration state of the planar switchable optical switch module and the matrix configuration of the planar matrix reconfigurable optical neural network component can be adjusted; meanwhile, feedback control and intelligent algorithms can be combined, the planar switchable optical switch module and the planar matrix reconfigurable optical neural network component are adaptively adjusted by the task, so as to realize efficient intelligent self-learning and self-configuration multi-task recognition.

[0009] Preferably, the spatial diffraction optical neural network core particle is based on the spatial diffraction optical principle, the multi-layer phase planes are optimized and designed, the unique structure of the spatial diffraction optical neural network is combined, the pixel points of the multi-layer phase planes are used as neurons, the layers are connected by optical diffraction to establish approximate full connection, and large-scale linear operation is realized. By utilizing the diffraction characteristics of light waves in the propagation process, the distribution of the phase planes is adjusted, the propagation and superposition of the light field are accurately controlled, and the dimension reduction feature extraction of high-pixel image information is realized.

[0010] Preferably, the spatial diffractive optical neural network core particle is stacked by three-dimensional integration to form a spatial diffractive optical neural network core particle, the number of layers is not less than 1, and there is a propagation layer composed of a low refractive index material between the layers. The three-dimensional integration process can be realized by transfer printing, direct bonding or femtosecond laser 3D direct writing technology, and the multi-layer phase surface is integrated into a spatial diffractive optical neural network core particle. The phase surface can be realized based on resonance phase, transmission phase or geometric phase principle, and the composition unit of each phase surface can be a metasurface unit or a diffractive optical unit.

[0011] Preferably, the planar matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, linearly process the signal output by the planar switchable optical switch module, and then output the calculation result through the detector array.

[0012] Preferably, the linear matrix network includes a network composed of a Mach-Zehnder interferometer, a micro-ring array or a multi-mode interference waveguide. Through the reconfigurable units on the linear matrix network module, any linear optical matrix configuration can be realized to adapt to the needs of different tasks.

[0013] Preferably, the planar switchable optical switch module includes a high-efficiency coupler array and a switchable optical switch. The high-efficiency coupler array is used to receive feature output results from different spatial diffractive optical neural network core particles and convert spatial light modes into waveguide modes; the switchable optical switch is used to realize the working state configuration of the planar switchable optical switch module, and selectively output the feature output results from different spatial diffractive optical neural network core particles to the planar matrix reconfigurable optical neural network component.

[0014] Preferably, the processing materials of the spatial diffractive optical neural network core particle include silicon, silicon nitride, metal, silicon dioxide and polymer; the processing materials of the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module include silicon, germanium, silicon dioxide, indium phosphide, gallium arsenide, thin film lithium niobate, polymer or phase change material, or a mixture of the above materials.

[0015] Preferably, the non-linear activation function can be realized by inserting a material with non-linear response (such as graphene, phase change material, anti-saturation absorber material, etc.) between the layers of the spatial diffractive optical neural network core particle; the non-linear activation function in the planar matrix neural network component can be composed of micro-rings, germanium-silicon materials, SOA materials, etc.; or after completing the optical operation, the output results are operated in the electrical domain to realize the non-linear activation function, thereby improving the recognition accuracy.

[0016] Preferably, the three-dimensional heterogeneous integration method can be realized by multi-core particle integration, and the spatial diffraction optical neural network core particles are directly bonded on a plane chip composed of a plane matrix reconfigurable optical neural network component and a plane switchable optical switch module; or different spatial diffraction optical neural network core particles, plane matrix reconfigurable optical neural network components and plane switchable optical switch modules are bonded in specific areas of an optical adapter board, and the connection between the spatial diffraction optical neural network core particles and the plane matrix reconfigurable optical neural network components and the plane switchable optical switch modules is realized through the optical adapter board.

[0017] Preferably, the three-dimensional heterogeneous integrated optical computing chip task configuration method dynamically switches the routing connection between different spatial diffraction optical neural network core particles and plane matrix reconfigurable optical neural network components through the plane switchable optical switch module according to the task type, ensures the flexibility and high efficiency of the task processing path, dynamically adjusts the matrix weight configuration of the plane matrix reconfigurable optical neural network component to adapt to the computing requirements of different tasks, realizes efficient and accurate feature extraction and result output, and further combines feedback control and intelligent algorithms to adaptively optimize and adjust the routing state of the plane switchable optical switch module and the matrix parameters of the plane matrix reconfigurable optical neural network component, realizes task self-driven self-learning and self-configuration multi-task configuration, and improves the efficiency and intelligent level in the switching and identification of different tasks. This configuration method not only can meet the rapid identification of similar feature tasks, but also can realize accurate processing of differentiated tasks through deep optimization, and provide self-driven, self-learning and self-configuration capabilities, and finally realize efficient and intelligent identification and calculation in a multi-task scenario.

[0018] Compared with the prior art, the above technical scheme conceived by the present application has the following beneficial effects:

[0019] 1. The three-dimensional heterogeneous integrated optical computing chip disclosed by the present application makes full use of the advantages of spatial diffraction neural networks and plane matrix neural networks, and performs efficient multi-task parallel computing, thereby providing a new idea for improving computing power compared with the serial computing mode of electronic computers.

[0020] 2. The present application utilizes the strong processing capability of spatial diffraction optical neural networks to perform dimension reduction processing on input tasks, realizes feature extraction, reduces the processing difficulty of plane matrix reconfigurable optical neural networks, and can realize efficient operation processing.

[0021] 3. The present application utilizes the flexibility of plane matrix reconfigurable optical neural networks to flexibly configure input tasks, performs linear operation on the features extracted by spatial diffraction optical neural networks, and can realize configuration under different tasks.

[0022] 4、The application utilizes the flexibility of the multi-core particle scheme to flexibly configure different tasks with large differences, and connects different spatial diffractive optical neural network core particles to a planar matrix reconfigurable optical neural network to process different data tasks.

[0023] 5、The application has universality by using the idea of the computing architecture, and can also be applied to other neural network information processing. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 It is a three-dimensional heterogeneous integrated optical computing chip architecture.

[0025] Figure 2 It is a spatial diffractive optical neural network core particle working unit.

[0026] Figure 3 It is a spatial diffractive optical neural network core particle working principle diagram.

[0027] Figure 4 It is a planar switchable optical switch module principle diagram.

[0028] Figure 5 It is a planar matrix reconfigurable optical neural network component schematic diagram.

[0029] Figure 6 It is a planar matrix reconfigurable optical neural network component working principle diagram.

[0030] Figure 7 It is a planar chip schematic diagram.

[0031] Figure 8 It is a three-dimensional heterogeneous integrated optical computing chip schematic diagram.

[0032] Figure 9 It is a three-dimensional heterogeneous integrated optical computing chip system working schematic diagram. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0034] The application provides a three-dimensional heterogeneous integrated optical computing chip, which is composed of three parts: a spatial diffraction optical neural network core particle, a planar matrix reconfigurable optical neural network component, and a planar switchable optical switch module. Through heterogeneous integration, the combination of a single or multiple spatial diffraction optical neural network core particles, planar matrix reconfigurable optical neural network components, and planar switchable optical switch modules is realized, thereby realizing efficient multi-task optical computing processing. During the computing process, the spatial diffraction optical neural network core particle is input with the task to be processed for feature extraction and dimension reduction; then, the planar switchable optical switch module is input through coupling for state configuration; finally, the planar matrix reconfigurable optical neural network component is input with the task for operation processing, and the detector array integrated on the planar matrix reconfigurable optical neural network component is used to detect and output the computing result. In the optical computing process, a nonlinear activation function can be inserted to realize nonlinear operation, thereby improving the accuracy of task recognition. For data set tasks with similar features, the matrix parameters of the planar matrix reconfigurable optical neural network component can be reconfigured; for data set tasks with significantly different features, the configuration state of the planar switchable optical switch module and the matrix configuration of the planar matrix reconfigurable optical neural network component can be adjusted; at the same time, feedback control and intelligent algorithms can be combined to adaptively adjust the planar switchable optical switch module and the planar matrix reconfigurable optical neural network component driven by the task, thereby realizing efficient intelligent self-learning and self-configuration multi-task recognition.

[0035] Specifically, the spatial diffraction optical neural network core particle is based on the principle of spatial diffraction optics. Through the optimization design of multiple phase planes, combined with the unique structure of the spatial diffraction optical neural network, the pixel points of the multiple phase planes serve as neurons, and the layers are connected through optical diffraction to form an approximate full connection, thereby realizing large-scale linear operation. By adjusting the phase plane distribution, the propagation and superposition of the light field are accurately controlled, thereby realizing the dimension reduction feature extraction of high-pixel image information.

[0036] Specifically, the spatial diffraction optical neural network core particle is formed by stacking multiple phase planes through three-dimensional integration, and the number of layers is not less than 1, and there is a propagation layer composed of a low refractive index material between the layers. The three-dimensional integration process can be realized through transfer printing, direct bonding, or femtosecond laser 3D direct writing technology to integrate multiple phase planes into a spatial diffraction optical neural network core particle. The phase plane can be realized based on resonance phase, transmission phase, or geometric phase principles, and the composition unit of each phase plane can be a metasurface unit or a diffractive optical unit.

[0037] Specifically, the planar matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize arbitrary reconfigurable optical matrix calculation, linearly process the signals output by the planar switchable optical switch module, and then output the calculation results through the detector array.

[0038] Specifically, the linear matrix network includes a network composed of Mach-Zehnder interferometers, micro-ring arrays, or multimode interference waveguides. Through the control of the reconfigurable units on the linear matrix network module, an arbitrary linear optical matrix configuration is realized to adapt to the requirements of different tasks.

[0039] Specifically, the planar switchable optical switch module includes a high-efficiency coupler array and a switchable optical switch. The high-efficiency coupler array is used to receive feature output results from different spatial diffraction optical neural network cores and convert spatial light modes into waveguide modes; the switchable optical switch is used to realize the working state configuration of the planar switchable optical switch module and selectively output the feature output results from different spatial diffraction optical neural network cores to the planar matrix reconfigurable optical neural network component.

[0040] Specifically, the processing materials of the spatial diffraction optical neural network core include silicon, silicon nitride, metal, silicon dioxide, and polymer; the processing materials of the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module include silicon, germanium, silicon dioxide, indium phosphide, gallium arsenide, thin-film lithium niobate, polymer, or phase-change material, or a mixture of the above materials.

[0041] Specifically, the nonlinear activation function can be realized by inserting materials with nonlinear response (such as graphene, phase-change materials, and anti-saturation absorber materials) between layers of the spatial diffraction optical neural network core; the nonlinear activation function in the planar matrix neural network component can be composed of micro-rings, germanium-silicon materials, SOA materials, etc.; or after completing optical operation, the output results are subjected to nonlinear activation function operation in the electrical domain, thereby improving the recognition accuracy.

[0042] Specifically, the three-dimensional heterogeneous integration method can be realized by multi-core integration, directly bonding the spatial diffraction optical neural network core on the planar chip composed of the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module; or bonding different spatial diffraction optical neural network cores, planar matrix reconfigurable optical neural network components, and planar switchable optical switch modules in specific areas of an optical adapter board, and realizing the connection between the spatial diffraction optical neural network core and the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module through the optical adapter board.

[0043] Specifically, the three-dimensional heterogeneous integrated optical computing chip task configuration method dynamically switches the routing connection between different spatial diffraction optical neural network cores and the planar matrix reconfigurable optical neural network components through the planar switchable optical switch module according to the task type, ensuring the flexibility and efficiency of the task processing path. At the same time, by dynamically adjusting the matrix weight configuration of the planar matrix reconfigurable optical neural network component, the computing requirements of different tasks are adapted to realize efficient and accurate feature extraction and result output. In addition, the method can also combine feedback control and intelligent algorithms to adaptively optimize and adjust the routing state of the planar switchable optical switch module and the matrix parameters of the planar matrix reconfigurable optical neural network component, realizing task self-driven self-learning and self-configuration multi-task configuration, thereby improving the efficiency and intelligent level in the switching and identification of different tasks. This configuration method not only can meet the rapid identification of similar feature tasks, but also can realize accurate processing of differentiated tasks through deep optimization, and provide self-driven, self-learning and self-configuration capabilities, and finally realize efficient and intelligent identification and calculation in multi-task scenarios.

[0044] As shown in Figure 1 The present application provides a three-dimensional heterogeneous integrated optical computing chip architecture. The chip includes two core modules: spatial diffraction optical neural network core, planar switchable optical switch module and planar matrix reconfigurable optical neural network component. Through three-dimensional heterogeneous integration, the efficient feature extraction capability of spatial diffraction optical neural network and the flexible linear operation capability of planar matrix reconfigurable optical neural network are combined to realize efficient and multi-task optical computing processing.

[0045] Specific working principle:

[0046] 1. Input processing:

[0047] When the task input from different data sets, the corresponding spatial diffraction optical neural network core is selected through the external control circuit. The task input is usually in the form of optical image, and the initial light field distribution is generated by modulation.

[0048] 2. Feature extraction and dimensionality reduction of spatial diffraction optical neural network:

[0049] The spatial diffraction optical neural network core is stacked by multiple layers of phase planes. Each layer of phase plane is optimized and trained through previous simulation calculation. The trained phase plane can realize feature extraction and dimensionality reduction of input data according to task characteristics. The phase plane modulates the wavefront information of the input light field, thereby extracting the key features related to the task. Through the cascade effect of multiple layers of phase planes, the high-dimensional input data is projected into a low-dimensional feature space, so as to realize efficient calculation.

[0050] 3. State configuration of planar switchable optical switch module:

[0051] The light signal (i.e., task feature) output from the spatial diffractive optical neural network enters the planar switchable optical switch module through the high-efficiency coupling array. The planar switchable optical switch selects the appropriate routing for the current spatial diffractive optical neural network core according to external control, ensuring that the task feature is correctly transmitted to the corresponding matrix network module.

[0052] 4. Matrix operation of the planar matrix reconfigurable optical neural network component:

[0053] For different data sets and tasks, the external control circuit is responsible for the working state of the planar matrix reconfigurable optical neural network. By adjusting the weight configuration of the linear matrix network module through the control circuit, efficient support for multiple tasks is achieved.

[0054] As shown in Figure 2 The spatial diffractive optical neural network core is stacked by multiple phase planes in a three-dimensional integrated manner, and there is a propagation layer composed of a low refractive index material between layers. The phase plane can be realized based on geometric phase, transmission phase or resonance phase, etc.

[0055] As shown in Figure 3 The working principle of the spatial diffractive optical neural network core is based on the principle of spatial diffraction. Through the optimization design of multiple phase planes, combined with the unique structure of the spatial diffractive optical neural network, large-scale linear operation is realized. The diffraction characteristics of light waves in the propagation process are used to accurately control the propagation and superposition of light fields by adjusting the phase plane distribution, thereby realizing the dimensionality reduction feature extraction of high-pixel image information. As shown in the figure, the digital image in the figure is dimensionality reduction feature extracted into a combination of several light points, which is provided for the processing of the following planar chip.

[0056] As shown in Figure 4 The planar switchable optical switch module is composed of a high-efficiency coupler array and a switchable optical switch. The high-efficiency coupler array is used to receive the feature output results from different spatial diffractive optical neural network cores, and converts the spatial light mode into a waveguide mode through the grating coupler in array form, realizing the conversion from space to plane. The switchable optical switch is used to realize the working state configuration of the planar switchable optical switch module, and selectively outputs the feature output results from different spatial diffractive optical neural network cores to the planar matrix reconfigurable optical neural network component, such as the signals of different colors shown in the figure.

[0057] As shown in Figure 5As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array.

[0058] As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 6 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 6 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 6 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array.

[0059] As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 7 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array.

[0060] As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 8 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array.

[0061] As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array. Figure 9 As shown in the plane matrix reconfigurable optical neural network component schematic diagram, the plane matrix reconfigurable optical neural network component is composed of a linear matrix network module and a detector array. The linear matrix network module is used to realize any reconfigurable optical matrix calculation, which can be composed of a network of Mach-Zehnder interferometers, micro-ring arrays or multi-mode interference waveguides, linearly processes the signals output by the plane switchable optical switch module, and then outputs the calculation results through the detector array.

[0062] 1、Data input

[0063] Input data type: The input data such as handwritten digit images (e.g. MNIST dataset) is converted into optical signals by external devices (e.g. light source or image projection device). The input optical signals are preliminarily modulated and encoded to adapt to the subsequent calculation requirements of the spatial diffractive optical neural network.

[0064] 2. Feature extraction of spatial diffractive optical neural network core

[0065] The input optical signals are transmitted to the spatial diffractive optical neural network core, which is composed of multiple layers of phase planes. Through the action of these phase planes, feature extraction and dimension reduction processing are performed on the input signals. Different datasets for different tasks correspond to different cores, and the phase design of each core is optimized and trained to ensure that the key features related to the task are extracted.

[0066] 3. State configuration of planar switchable optical switch module:

[0067] The optical signals output from the spatial diffractive optical neural network (i.e. task features) enter the planar switchable optical switch module through the high-efficiency coupling array. The planar switchable optical switch selects the appropriate route for the current spatial diffractive optical neural network core according to external control, ensuring that the task features are correctly transmitted to the corresponding matrix network module.

[0068] 4. Task recognition of planar matrix reconfigurable optical neural network component

[0069] The planar matrix reconfigurable optical neural network component performs linear operation on the input feature signals through the dynamically configured linear matrix network. The system intelligently controls the planar switchable optical switch module and the planar matrix reconfigurable optical neural network component through external circuits, dynamically selects the network configuration that adapts to the current task, and thus realizes the flexibility of multi-task processing.

[0070] 5. Task output

[0071] The operation results are output by the planar matrix neural network component and enter the detector array for photoelectric signal conversion. The converted electrical signals are output as task processing results to external devices for further processing or direct display.

[0072] It is easily understood by those skilled in the art that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A three-dimensional heterogeneous integrated photonic computing chip, comprising: The application relates to a spatial diffractive optical neural network core particle, a planar matrix reconfigurable optical neural network component and a planar switchable optical switch module, which are integrated by three-dimensional heterogeneous integration, the spatial diffractive optical neural network core particle comprises a plurality of phase surfaces, and propagation layers are arranged between the phase surfaces; the planar matrix reconfigurable optical neural network component comprises a linear matrix network module and a detector array which are connected in sequence. The spatial diffractive optical neural network core particle is used for feature extraction and dimension reduction of a task to be processed, and a feature output result is obtained; the feature output result is input into the planar switchable optical switch module in a coupling mode for state configuration; and the planar matrix reconfigurable optical neural network component is used for operation on the feature output result after state configuration. The planar switchable optical switch module comprises a coupler array and an optical switch; the coupler array is used for receiving feature optical signals from different spatial diffractive optical neural network core particles and converting spatial light modes into waveguide modes; and the optical switch is used for realizing working state configuration of the planar switchable optical switch module and selectively outputting the feature output results from the different spatial diffractive optical neural network core particles to the planar matrix reconfigurable optical neural network component. The planar switchable optical switch module dynamically switches the routing connection between different spatial diffractive optical neural network core particles and the planar matrix reconfigurable optical neural network component according to a task type; meanwhile, the matrix weight configuration of the planar matrix reconfigurable optical neural network component is dynamically adjusted to adapt to the calculation requirement of different tasks; and the routing state of the planar switchable optical switch module and the matrix parameter of the planar matrix reconfigurable optical neural network component are adaptively optimized and adjusted in combination with feedback control and intelligent algorithms, so that task self-driving self-learning and self-configuration multi-task configuration are realized.

2. The three-dimensional heterogeneous integrated optical computing chip of claim 1, wherein, Pixel points of the phase surfaces serve as neurons, and approximate full connection is established between layers through optical diffraction, so that linear operation is realized.

3. The three-dimensional heterogeneous integrated optical computing chip of claim 1, wherein, The three-dimensional heterogeneous integration mode comprises transfer printing, direct bonding or femtosecond laser 3D direct writing.

4. The three-dimensional heterogeneous integrated optical computing chip of claim 2, wherein, The phase surfaces are metasurface units or diffractive optical units.

5. The three-dimensional heterogeneous integrated optical computing chip of claim 1, wherein, The linear matrix network module comprises a Mach-Zehnder interferometer, a microring array or a multimode interference waveguide.

6. The three-dimensional heterogeneous integrated optical computing chip of claim 1, wherein, Processing materials of the spatial diffractive optical neural network core particle comprise silicon, silicon nitride, metal, silicon dioxide and polymer; and processing materials of the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module comprise silicon, germanium, silicon dioxide, indium phosphide, gallium arsenide, thin-film lithium niobate, polymer, phase change material or a mixture of the above materials.

7. The three-dimensional heterogeneous integrated optical computing chip of claim 1, wherein, The spatial diffraction optical neural network core particle further comprises a first nonlinear activation function layer located between layers of the multi-phase plane, the first nonlinear activation function layer being used to implement a nonlinear operation, and a material of the nonlinear activation function layer being graphene, a phase change material or a reverse saturable absorber material; and the planar matrix reconfigurable optical neural network component further comprises a second nonlinear activation function layer located between the linear matrix network module and the detector array, and a material of the second nonlinear activation function layer being germanium-silicon material or SOA material.

8. The three-dimensional heterogeneous integrated optical computing chip of claim 3, wherein, The three-dimensional heterogeneous integration is achieved by multi-core particle integration, and the spatial diffraction optical neural network core particle is directly bonded on a planar chip composed of the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module; or different spatial diffraction optical neural network core particles, the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module are bonded in a preset region of an optical adapter board, and the connection between the spatial diffraction optical neural network core particle and the planar matrix reconfigurable optical neural network component and the planar switchable optical switch module is achieved through the optical adapter board.

Images