Metasurface-based diffraction neural network and three-dimensional diffraction neural network
By arraying reconstructible micro-nano structural units at high density and large scale on the metasurface, building high-density and large-scale neurons, the problems of low neuron density and small scale in the existing technology are solved, and efficient Information calculation processing.
Patent Information
- Application Number
- CN202421249620.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Utility models(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-03
- Publication Date
- 2025-02-18
- Estimated Expiration
- 2034-06-03
AI Technical Summary
The existing reconfigurable diffraction neural networks have low neuron density and small scale, making it difficult to meet the needs of high-density large-scale neurons.
Using a diffraction neural network based on the metasurface, high-density and large-scale reconstructible micro-nano structural units are constructed by high-density and large-scale arraying on the metasurface.
The construction of high-density large-scale neurons has been realized, the intelligence and plasticity of the diffraction neural network have been improved, and the ability to effectively handle complex information computing tasks.
Smart Images
Figure CN222506923U_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of optical chips, and specifically to a diffraction neural network and a three-dimensional diffraction neural network based on a metasurface. Background Art
[0002] Diffractive Optical Neural Networks (DONN), also known as diffractive neural networks, use light as an information medium and utilize the diffraction of light to realize information processing, thereby executing the tasks matched during training; such as image recognition tasks, semantic recognition tasks, etc. Compared with digital neural networks, diffractive neural networks can greatly save computing resources and significantly improve computing speed; theoretically, the computing speed of diffractive neural networks for information can reach the speed of light.
[0003] The reconfigurable diffractive neural network has the ability of dynamic training, which is conducive to improving the intelligence and plasticity of the diffractive neural network. However, the reconfigurable diffractive neural network provided by the related art has a low density of neurons and a small scale. Utility Model Content
[0004] One purpose of the present application is to propose a diffractive neural network and a three-dimensional diffractive neural network based on a metasurface, which can have high-density and large-scale neurons.
[0005] According to one aspect of an embodiment of the present application, a diffractive neural network based on a metasurface is disclosed, wherein the diffractive neural network comprises: an optical signal input module; an optical signal modulation module; an optical signal output module;
[0006] Wherein, the optical signal input module is used to guide the optical signal to the optical signal modulation module;
[0007] The optical signal modulation module includes a metasurface; the metasurface is used to modulate the received optical signal; the metasurface includes at least two reconfigurable micro-nano structure units; the reconfigurable micro-nano structure unit includes a micro-nano structure and a spacing structure arranged around the micro-nano structure;
[0008] The optical signal output module is used to receive the optical signal from the optical signal modulation module and output the optical signal to a detector used for detecting signals.
[0009] In an exemplary embodiment of the present application, an arrangement period of the reconfigurable micro-nano structure units is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 2 times the central wavelength.
[0010] In an exemplary embodiment of the present application, the length of the reconfigurable micro-nanostructure unit in a direction perpendicular to the metasurface is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 5 times the central wavelength.
[0011] In an exemplary embodiment of the present application, the optical signal modulation module includes at least two layers of metasurfaces; in the direction from the optical signal input module to the optical signal output module, the metasurfaces are hierarchically arranged in sequence; and the optical signal modulated by the metasurface propagates along the hierarchical arrangement direction of the metasurface.
[0012] In an exemplary embodiment of the present application, the spacing between adjacent levels of metasurfaces is greater than or equal to 0.5 times the central wavelength of the optical signal and less than or equal to 1000 times the central wavelength.
[0013] In an exemplary embodiment of the present application, in each layer of the metasurface, one of the reconfigurable micro-nanostructure units constitutes a neuron, or at least two adjacent reconfigurable micro-nanostructure units constitute a neuron.
[0014] In an exemplary embodiment of the present application, in each layer of the metasurface, the number of neurons J is less than 2M, where M is the number of signal input ports in the optical signal input module.
[0015] In an exemplary embodiment of the present application, the material of the micro-nano structure is a phase change material; the reconfigurable micro-nano structure unit further includes a phase change excitation element for applying excitation to the micro-nano structure.
[0016] In an exemplary embodiment of the present application, the reconfigurable micro-nano structure unit further includes a refractive index matching structure, and the refractive index matching structure is used to match the refractive index of the micro-nano structure and the refractive index of the spacing structure to control the transmittance of the reconfigurable micro-nano structure unit.
[0017] In an exemplary embodiment of the present application, the optical signal modulation module further includes: a nonlinear excitation layer, wherein the nonlinear excitation layer is used to apply a nonlinear response to the light intensity of the optical signal.
[0018] In an exemplary embodiment of the present application, the nonlinear excitation layer includes a nonlinear optical waveguide.
[0019] In an exemplary embodiment of the present application, the nonlinear optical waveguide is integrated on the light emitting surface of at least one layer of the metasurface using semiconductor technology.
[0020] In an exemplary embodiment of the present application, the optical signal input module includes: at least two signal input ports; at least two first optical waveguides;
[0021] The first optical waveguides are matched one-to-one with the signal input ports, and the first optical waveguides are used to couple the optical signal received by the corresponding signal input port to the optical signal modulation module.
[0022] In an exemplary embodiment of the present application, the end of the first optical waveguide close to the optical signal modulation module is tapered, and the tip of the tapered waveguide faces the same direction as the propagation direction of the optical signal.
[0023] In an exemplary embodiment of the present application, the optical signal output module comprises: at least two signal output ports; at least two second optical waveguides;
[0024] The second optical waveguides are matched one-to-one with the signal output ports, and the second optical waveguides are used to couple the optical signal from the optical signal modulation module to the corresponding signal output port.
[0025] In an exemplary embodiment of the present application, the head end of the second optical waveguide close to the optical signal modulation module is tapered, and the tip of the tapered waveguide faces in the opposite direction to the propagation direction of the optical signal.
[0026] According to one aspect of an embodiment of the present application, a three-dimensional diffractive neural network is disclosed, wherein the three-dimensional diffractive neural network includes at least two diffractive neural networks provided by any one of the above embodiments;
[0027] The diffractive neural networks are stacked and arranged perpendicular to the propagation direction of the optical signal.
[0028] In an exemplary embodiment of the present application, the three-dimensional diffraction neural network is packaged using a stacking packaging process.
[0029] The diffraction neural network provided by the present application includes: an optical signal input module; an optical signal modulation module; and an optical signal output module. Among them, the optical signal modulation module includes a metasurface; the metasurface is used to modulate the received optical signal; and the metasurface includes at least two reconfigurable micro-nano structural units. By applying semiconductor technology, the reconfigurable micro-nano structural units in the metasurface can reach nanometer-level sizes, and the spacing between adjacent reconfigurable micro-nano structural units can also reach nanometer-level spacing. Therefore, the reconfigurable micro-nano structural units in the metasurface can be arrayed in a high-density and large-scale manner, so that the diffraction neural network based on the metasurface provided by the present application can have high-density and large-scale neurons.
[0030] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.
[0031] It should be understood that the foregoing general description and the following detailed description are exemplary only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The above and other objects, features and advantages of the present application will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings.
[0033] Figure 1 A schematic diagram of the architecture of the metasurface-based diffraction neural network provided in this application is shown.
[0034] Figure 2 A schematic structural diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application is shown.
[0035] Figure 3 A schematic diagram of the architecture of a diffractive neural network including at least two layers of metasurfaces provided in one embodiment of the present application is shown.
[0036] Figure 4 A schematic structural diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application is shown.
[0037] Figure 5 A schematic structural diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application is shown.
[0038] Figure 6 A schematic structural diagram of a nonlinear optical waveguide in an embodiment of the present application is shown.
[0039] Figure 7 A schematic diagram of the architecture of a diffractive neural network in one embodiment of the present application is shown.
[0040] Figure 8 A schematic structural diagram of an optical signal input module in an embodiment of the present application is shown.
[0041] Fig. 9 A schematic structural diagram of an optical signal output module in an embodiment of the present application is shown.
[0042] Fig.10 A schematic diagram of the architecture of the three-dimensional diffraction neural network provided in this application is shown.
[0043] Fig.11 The figure shows the confusion matrix of the trained diffractive neural network in one embodiment of the present application.
[0044] Fig.12 The figure shows the confusion matrix of the trained diffractive neural network in one embodiment of the present application.
[0045] Reference numerals:
[0046] 1-supersurface; 10-substrate; 11-micro-nanostructure; 12-spacer structure; 13-phase change excitation element; 14-target structure; 2-nonlinear excitation layer; 20-substrate; 21-germanium waveguide; 22-silicon waveguide. DETAILED DESCRIPTION
[0047] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present application will be more comprehensive and complete and the concepts of the example embodiments will be fully conveyed to those skilled in the art. The accompanying drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.
[0048] In addition, the described features, structures or characteristics may be combined in one or more example embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present application. However, those skilled in the art will appreciate that the technical solution of the present application may be practiced while omitting one or more of the specific details, or other methods, components, steps, etc. may be adopted. In other cases, known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring the present application and making the various aspects of the present application obscure.
[0049] Some of the blocks shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.
[0050] In order to make the diffraction neural network reconfigurable and thus capable of dynamic training, the neurons in the diffraction neural network that provide the phase of the optical signal need to have the ability to be reconfigured. In the related art, for an on-chip diffraction neural network built based on a Mach-Zehnder Interferometer (MZI) array, it is proposed to combine a Mach-Zehnder interferometer and titanium nitride materials to construct a neuron with reconfigurable capabilities. Specifically, the diffraction neural network provided by the related art changes the dielectric constant of the neuron through the electrothermal effect of titanium nitride, thereby changing the phase provided by the neuron, thereby making the neuron reconfigurable.
[0051] However, since the reconfigurable neurons in the related art rely on Mach-Zehnder interferometers, which are difficult to array in high density and on a large scale, the reconfigurable diffraction neural network provided by the related art has a low neuron density and a small scale.
[0052] In order to overcome the above-mentioned defects of the related art, the present application provides a diffraction neural network based on a metasurface, which can have high-density and large-scale neurons.
[0053] Figure 1 Schematic diagram of the architecture of the diffractive neural network based on the metasurface provided by the present application is shown. Figure 1 The diffraction neural network provided in this application includes: an optical signal input module; an optical signal modulation module; and an optical signal output module.
[0054] Specifically, the optical signal input module is used to guide the optical signal to the optical signal modulation module. The optical signal input module may include a light source or may not include a light source. When a light source is included, the optical signal input module is used to generate an optical signal, and then guide the generated optical signal to the optical signal modulation module. When a light source is not included, the optical signal input module is used to collect optical signals generated by an external light source, and then guide the collected optical signals to the optical signal modulation module.
[0055] Optionally, the optical signal generated by the light source is in the terahertz band. In this way, the requirements for the size of the micro-nano structure included in the metasurface in the optical signal modulation module can be reduced, so that even if the size of the micro-nano structure is larger, it can still meet the constraint of the maximum semi-cone diffraction angle.
[0056] Optionally, the light signal generated by the light source is in the visible light band. This method can effectively reduce the cost of the light source, improve the convenience of light source deployment and the safety of light source operation, and is also easy to combine with the mature machine vision technology at this stage.
[0057] Preferably, the optical signal generated by the light source is in the near-infrared band. In this way, the cost of the light source can be effectively reduced; and in the near-infrared band, the optical device involved in the diffractive neural network system can be made of silicon and can be processed using mature processes, so the cost can be effectively reduced; and in the near-infrared band, there is a high-speed photoelectric conversion device, so the optical signal processed and output by the diffractive neural network can be converted into an electrical signal by the photoelectric conversion device at a high speed according to the demand.
[0058] Preferably, when training the diffractive neural network, the light source generates an incoherent multi-wavelength optical signal. In this way, when training the diffractive neural network, the interference caused by coherent noise can be effectively suppressed, and the system information dimension can be effectively expanded to realize high-dimensional multi-channel information computing.
[0059] The optical signal modulation module includes a metasurface. In the embodiment of the present application, the metasurface is used to modulate the received optical signal; that is, in the embodiment of the present application, the metasurface is a physical carrier of neurons in the diffractive neural network. In detail, the metasurface includes at least two reconfigurable micro-nano structural units; the reconfigurable micro-nano structural unit serves as a functional unit for optical signal modulation, further constituting a reconfigurable neuron.
[0060] Figure 2 FIG. 1 shows a schematic diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application. Figure 2 In the embodiment of the present application, the reconfigurable micro-nano structure unit includes: a substrate 10; a micro-nano structure 11; and a spacing structure 12. The substrate 10 mainly provides support for the micro-nano structure 11 and the spacing structure 12; the spacing structure 12 is arranged around the micro-nano structure 11, and the spacing structure 12 may be air or other materials that are transparent in the wavelength band of the optical signal. As the refractive index of the micro-nano structure 11 changes, or as the distance between the micro-nano structures 11 in adjacent reconfigurable micro-nano structure units changes, the equivalent refractive index of the reconfigurable micro-nano structure unit changes accordingly, and thus the phase provided by the reconfigurable micro-nano structure unit also changes accordingly. If the phase provided by the reconfigurable micro-nano structure unit changes, the modulation effect provided by it for the optical signal changes accordingly, and thus the computational processing performed by it on the optical signal changes accordingly.
[0061] The optical signal output module is used to receive the optical signal from the optical signal modulation module and output the optical signal to a detector for detecting the signal. After the detector detects the optical signal processed by the diffraction neural network, it can output the processing result corresponding to the optical signal.
[0062] It should be noted that, by applying semiconductor technology, the reconfigurable micro-nano structural units in the metasurface can reach nanometer-level sizes, and the spacing between adjacent reconfigurable micro-nano structural units can also reach nanometer-level. Therefore, the reconfigurable micro-nano structural units in the metasurface can be arrayed in high density and large scale, so that the diffraction neural network based on the metasurface provided in this application can have high-density and large-scale neurons.
[0063] In one embodiment, the arrangement period of the reconfigurable micro-nano structure unit is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 2 times the central wavelength. By setting the lower limit of the arrangement period of the reconfigurable micro-nano structure unit to 0.3 times the central wavelength, excessive requirements on the processing technology can be avoided; by setting the upper limit of the arrangement period of the reconfigurable micro-nano structure unit to 2 times the central wavelength, high-order diffraction of the optical signal modulated by the metasurface can be avoided, thereby avoiding the waste of optical signal energy due to high-order diffraction.
[0064] In one embodiment, the length of the reconfigurable micro-nano structure unit in the direction perpendicular to the metasurface, that is, the height of the reconfigurable micro-nano structure unit, is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 5 times the central wavelength. By setting the lower limit of the height of the reconfigurable micro-nano structure unit to 0.3 times the central wavelength, it is ensured that the phase provided by the metasurface can cover the [0,2π] interval, thereby ensuring that the metasurface can provide sufficient modulation for the optical signal; by setting the upper limit of the height of the reconfigurable micro-nano structure unit to 5 times the central wavelength, it is avoided to put too high requirements on the processing technology.
[0065] In the embodiment of the present application, each layer of the metasurface in the optical signal modulation module is a diffraction layer containing multiple neurons. After passing through each layer of the metasurface, the optical signal will be diffracted under the modulation effect applied by the metasurface, thereby realizing the computational processing of the information therein.
[0066] In one embodiment, the optical signal modulation module includes only one layer of metasurface; that is, the metasurface-based diffraction neural network provided in this embodiment is a single-layer neural network.
[0067] In one embodiment, the optical signal modulation module includes at least two layers of metasurfaces; that is, the diffractive neural network based on the metasurface provided in this embodiment is a multi-layer neural network. In this embodiment, in the direction from the optical signal input module to the optical signal output module, the metasurfaces are arranged in a hierarchical manner in sequence; and the optical signal modulated by the metasurface propagates along the hierarchical arrangement direction of the metasurface.
[0068] In detail, Figure 3 FIG. 1 shows a schematic diagram of the architecture of a diffractive neural network including at least two layers of metasurfaces provided in one embodiment of the present application. Figure 3 In one embodiment, the direction from the optical signal input module to the optical signal output module is Figure 3 In this embodiment, the multi-layer metasurface in the optical signal modulation module is arranged according to Figure 3 From left to right, they are arranged in a hierarchical manner. Figure 3In the figure, the metasurfaces are arranged from left to right, and they are: the first metasurface, the second metasurface, and the last metasurface. The optical signal modulated by the first metasurface propagates to the right to the second metasurface; the optical signal modulated by the second metasurface propagates to the right to the third metasurface, and the optical signal modulated by the last metasurface propagates to the right to the optical signal output module.
[0069] In one embodiment, the optical signal modulation module includes at least two layers of metasurfaces; the spacing between adjacent layers of metasurfaces is greater than or equal to 0.5 times the central wavelength of the optical signal and less than or equal to 1000 times the central wavelength.
[0070] It should be noted that for a diffraction neural network including multiple diffraction layers, in order to use the diffraction of light to realize the operation and processing of information, the spacing between adjacent diffraction layers needs to be controlled within the Fresnel diffraction range. Specifically, if the spacing between adjacent diffraction layers is too small and does not reach the lower limit of the Fresnel diffraction range, the output light of the previous diffraction layer will be directly projected to the next diffraction layer, resulting in the failure to realize the specific diffraction function of the previous diffraction layer. If the spacing between adjacent diffraction layers is too large and exceeds the upper limit of the Fresnel diffraction range, the output light of the previous diffraction layer will enter the Fraunhofer diffraction range and undergo Fourier transform, which does not meet the information operation and processing requirements of the diffraction neural network.
[0071] Therefore, in this embodiment, for a diffraction neural network including multiple layers of metasurfaces, the spacing between adjacent layers of metasurfaces is controlled to be between 0.5 times the central wavelength and 1000 times the central wavelength to ensure that the spacing between adjacent diffraction layers is controlled within the Fresnel diffraction range.
[0072] In the embodiment of the present application, the reconfigurable micro-nano structure unit is used as a functional unit for optical signal modulation, and further constitutes a reconfigurable neuron. Specifically, depending on the division of the composition of neurons in the design stage and the training stage of the diffractive neural network, in each layer of the metasurface, a neuron may be composed of only one reconfigurable micro-nano structure unit, or may be composed of at least two adjacent reconfigurable micro-nano structure units.
[0073] Neurons composed of only one reconfigurable micro-nano structural unit can cooperate and connect with each other with a high degree of freedom and complexity, thus helping to process more complex operations. Neurons composed of multiple adjacent reconfigurable micro-nano structural units can help improve the robustness of the diffractive neural network operation processing.
[0074] In one embodiment, in each layer of the hypersurface, the number of neurons J is less than 2M, where M is the number of signal input ports in the optical signal input module. By controlling J to be less than 2M, the overfitting degree of the diffractive neural network can be effectively reduced.
[0075] In the embodiments of the present application, for the micro-nano structure in the reconfigurable micro-nano structure unit, phase change materials, two-dimensional materials, flexible materials and the like can be used, and corresponding excitations can be applied, thereby making the equivalent refractive index of the reconfigurable micro-nano structure unit adjustable.
[0076] Considering that compared with other materials, phase change materials have faster response speed, more flexible excitation methods, and a wider adjustable refractive index range, in one embodiment, the material of the micro-nano structure is a phase change material, and the reconfigurable micro-nano structure unit also includes a phase change excitation element for applying excitation to the micro-nano structure; in response to the excitation applied by the phase change excitation element, the phase change material changes its phase state, thereby changing its dielectric constant, and then changing its refractive index, and then changing the equivalent refractive index of the reconfigurable micro-nano structure unit.
[0077] Specifically, in this embodiment, the phase change material can be selected from germanium antimony tellurium alloy GST, or solid vanadium dioxide VO 2 ; The excitation method applied can be electrical excitation, optical excitation, or thermal excitation.
[0078] Take GST as an example: GST has two phases, crystalline and amorphous. The dielectric constants of these two phases are quite different, and these two phases can be converted to each other. From amorphous to crystalline, electrical excitation, optical excitation, and thermal excitation can be used; from crystalline to amorphous, optical excitation and thermal excitation can be used. Whether it is converted from amorphous to crystalline or from crystalline to amorphous, after the phase conversion is completed, even if the excitation is removed and the environment returns to room temperature, GST can still maintain the phase after phase change for a long time, and has good non-volatility.
[0079] Figure 4 FIG. 1 shows a schematic diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application. Figure 4 In one embodiment, the reconfigurable micro-nano structure unit includes: a substrate 10; a micro-nano structure 11 of a GST material; a spacer structure 12 disposed around the micro-nano structure; and a phase change excitation element 13 disposed between the substrate 10 and the micro-nano structure 11. In this embodiment, the phase change excitation element 13 is an electrically controlled heating element, and the material selected for the electrically controlled heating element is transparent in the wavelength band of the optical signal. The electrically controlled heating element heats the micro-nano structure 11 of the GST material by electrically controlling, thereby controlling the micro-nano structure 11 to switch phases between the crystalline state and the amorphous state.
[0080] In one embodiment, the reconfigurable micro-nano structure unit further includes a refractive index matching structure, and the refractive index matching structure is used to match the refractive index of the micro-nano structure and the refractive index of the spacing structure to control the transmittance of the reconfigurable micro-nano structure unit.
[0081] Specifically, the spacing structure arranged around the micro-nano structure is usually also regarded as a part of the reconfigurable micro-nano structure unit; in addition, the reconfigurable micro-nano structure unit generally includes a substrate in addition to the micro-nano structure and the spacing structure, and the material of the substrate is transparent under the wavelength band of the optical signal. The micro-nano structure and the spacing structure are arranged on the surface of the substrate. That is, the reconfigurable micro-nano structure unit includes: a substrate; a micro-nano structure and a spacing structure arranged on the surface of the substrate.
[0082] Therefore, in this embodiment, the substrate can be used as a refractive index matching structure, and the refractive index of the micro-nano structure and the refractive index of the spacing structure can be matched by configuring the material of the substrate, thereby controlling the transmittance of the entire reconfigurable micro-nano structure unit to the optical signal.
[0083] In this embodiment, a target structure different from the micro-nano structure and the spacing structure can also be arranged on the surface of the substrate provided with the micro-nano structure, and the material of the target structure is transparent in the wavelength band of the optical signal. The target structure is used as a refractive index matching structure, and the material of the target structure is configured to match the refractive index of the micro-nano structure and the refractive index of the spacing structure, thereby controlling the transmittance of the entire reconfigurable micro-nano structure unit to the optical signal.
[0084] Figure 5 FIG. 1 shows a schematic diagram of a reconfigurable micro-nano structure unit in an embodiment of the present application. Figure 5 In one embodiment, the reconfigurable micro-nano structure unit includes: a substrate 10; a micro-nano structure 11, a spacing structure 12 and a target structure 14 arranged on the surface of the substrate 10. Among them, the spacing structure 12 is arranged around the micro-nano structure 11. In this embodiment, the substrate 10 and the target structure 14 can both be used as refractive index matching structures to match the refractive index of the micro-nano structure and the refractive index of the spacing structure, thereby controlling the transmittance of the reconfigurable micro-nano structure unit as a whole to the optical signal.
[0085] The operation and processing of the optical signal by each diffraction layer is equivalent to a linear operation in matrix form. A diffraction neural network containing only a single diffraction layer can efficiently perform linear tasks and is easy to optimize during training. Compared with a diffraction neural network containing only a single diffraction layer, a diffraction neural network containing multiple diffraction layers can perform more complex linear tasks.
[0086] It should be noted that multiple linear operations can be equivalent to one linear operation; while in the diffractive neural network, it is only the operation processing between the diffractive layers, which is still a linear operation in essence. Therefore, if the diffractive neural network only has multiple diffractive layers, then these multiple diffractive layers can actually be equivalent to one diffractive layer. In this case, although increasing the number of diffractive layers can enable the diffractive neural network to perform more complex linear tasks, it is difficult to effectively deal with nonlinear tasks, and the degree of intelligence is very limited.
[0087] To this end, in one embodiment, the optical signal modulation module further includes: a nonlinear excitation layer. The nonlinear excitation layer is used to apply a nonlinear response to the light intensity of the optical signal. Specifically, the nonlinear excitation layer is arranged after at least one diffraction layer; by applying a nonlinear response to the light intensity of the optical signal, the nonlinear excitation layer introduces a nonlinear factor into the diffraction neural network, thereby enabling the diffraction neural network to improve its task execution capability and also be suitable for processing nonlinear tasks, thereby improving the intelligence of the diffraction neural network.
[0088] In one embodiment, the nonlinear excitation layer includes a nonlinear optical waveguide; that is, the nonlinear excitation element in the nonlinear excitation layer is a nonlinear optical waveguide. In this embodiment, the nonlinear optical waveguide refers to an optical waveguide with nonlinear response properties. Specifically, the nonlinear optical waveguide includes but is not limited to: germanium silicon coupled waveguide; silicon nitride groove waveguide.
[0089] Figure 6 FIG. 2 shows a schematic diagram of the structure of a nonlinear optical waveguide in an embodiment of the present application. Figure 6 In one embodiment, the nonlinear optical waveguide used in the nonlinear excitation layer is a germanium-silicon coupled waveguide. The germanium-silicon coupled waveguide includes: a substrate 20 of silicon dioxide material; a germanium waveguide 21; and a silicon waveguide 22 having a ridge structure. The ridge of the silicon waveguide 22 makes the silicon waveguide 22 present an asymmetric shape. Specifically, the ridge of the silicon waveguide 22 divides the surface of the silicon waveguide 22 into a narrower sub-surface and a wider sub-surface; the germanium waveguide 21 is arranged on the wider sub-surface.
[0090] It should be noted that the metasurface is usually processed by semiconductor technology, and the nonlinear optical waveguide is also usually suitable for semiconductor technology. Therefore, in one embodiment, the nonlinear optical waveguide is integrated on the light emitting surface of at least one metasurface by semiconductor technology, thereby improving the integration between the nonlinear excitation layer and the diffraction layer.
[0091] Figure 7 FIG. 2 shows a schematic diagram of the architecture of a diffractive neural network in an embodiment of the present application. Figure 7In one embodiment, the optical signal modulation module in the diffractive neural network is provided with at least one metasurface 1. On the light emitting surface of one of the metasurfaces 1, a nonlinear optical waveguide is integrated by using semiconductor technology, thereby obtaining a nonlinear excitation layer 2 integrated on the light emitting surface of the metasurface 1.
[0092] In addition to using nonlinear optical waveguides to construct nonlinear excitation layers, phase change materials, two-dimensional materials, lithium niobate, etc. can also be used to construct nonlinear excitation layers.
[0093] In one embodiment, the optical signal input module includes: at least two signal input ports; at least two first optical waveguides. The first optical waveguides are matched with the signal input ports one by one, and the first optical waveguides are used to couple the optical signals received by the corresponding signal input ports to the optical signal modulation module.
[0094] Figure 8 FIG. 2 shows a schematic diagram of the structure of an optical signal input module in an embodiment of the present application. Figure 8 In one embodiment, the end of the first optical waveguide close to the optical signal modulation module is tapered, and the tip of the tapered waveguide faces the same direction as the propagation direction of the optical signal. In this embodiment, by configuring the first optical waveguide in this way, the spatial coverage of each optical signal input can be effectively controlled, so that different optical signal inputs can avoid crossing in space, thereby avoiding mutual interference between different optical signal inputs.
[0095] In one embodiment, the optical signal output module includes: at least two signal output ports; at least two second optical waveguides. The second optical waveguides are matched with the signal output ports one by one, and the second optical waveguides are used to couple the optical signal from the optical signal modulation module to the corresponding signal output port.
[0096] Fig. 9 FIG. 2 shows a schematic diagram of the structure of an optical signal output module in an embodiment of the present application. Fig. 9 In one embodiment, the second optical waveguide is tapered near the head end of the optical signal modulation module, and the tip of the tapered waveguide faces the opposite direction of the propagation direction of the optical signal. In this embodiment, by configuring the second optical waveguide in this way, each optical signal output can be accurately sampled.
[0097] The present application also provides a three-dimensional diffraction neural network. Fig.10 The schematic diagram of the architecture of the three-dimensional diffraction neural network provided by the present application is shown. Fig.10, the three-dimensional diffraction neural network provided in the present application includes at least two diffraction neural networks provided in any of the above embodiments. In the present application, a diffraction neural network can be regarded as a two-dimensional diffraction neural network; it should be noted that the two-dimensional diffraction neural network referred to here is mainly a two-dimensional abstract expression used to describe the connection method between the optical signal input module, the optical signal modulation module and the optical signal output module, and does not mean that the optical signal in the two-dimensional diffraction neural network is propagated in two-dimensional space; obviously, the optical signal in the two-dimensional diffraction neural network is propagated in three-dimensional space.
[0098] The three-dimensional diffraction neural network provided in this application is to further expand the two-dimensional diffraction neural network from a two-dimensional abstract expression to a three-dimensional abstract expression. Specifically, perpendicular to the propagation direction of the light signal, each two-dimensional diffraction neural network is stacked and arranged to obtain a three-dimensional diffraction neural network. A two-dimensional diffraction neural network can usually only perform a single task, while a three-dimensional diffraction neural network can perform multiple tasks in parallel, thereby improving the task execution capability of the diffraction neural network.
[0099] It should be noted that Fig.10 The three-dimensional diffraction neural network shown, the internal composition of each two-dimensional diffraction neural network seems to be completely consistent (the number of signal input ports of each two-dimensional diffraction neural network is consistent, the number of diffraction layers and the reconfigurable micro-nano structural units contained are also consistent, and the number of signal output ports is also consistent), and each two-dimensional diffraction neural network is vertically opposite to each other (the signal input ports of each two-dimensional diffraction neural network are vertically opposite to each other, the diffraction layers of the same level are also vertically opposite to each other, and the signal output ports are also vertically opposite to each other). However, this does not mean that the three-dimensional diffraction neural network provided by this application is limited to Fig.10 Specifically, the internal compositions of different two-dimensional diffractive neural networks of the three-dimensional diffractive neural network provided by the present application may be different from each other, and each two-dimensional diffractive neural network may be interlaced with each other.
[0100] In one embodiment, the three-dimensional diffractive neural network is packaged using a stacking packaging process, thereby making the three-dimensional diffractive neural network have a higher degree of integration.
[0101] Example 1
[0102] In Example 1, the central wavelength of the wavelength band of the optical signal is 1550nm. In the diffraction neural network provided in this embodiment, the optical signal input module includes 784 signal input ports. The optical signal modulation module includes 5 layers of metasurfaces, that is, 5 diffraction layers; the spacing between adjacent layers of metasurfaces is 30μm; each layer of metasurface includes 2000 reconfigurable micro-nano structure units; the arrangement period of the reconfigurable micro-nano structure units is 1000nm, and the length in the direction perpendicular to the metasurface is 2000nm; the substrate in the reconfigurable micro-nano structure unit is quartz glass, the material of the micro-nano structure is GST, and the material of the spacing structure around the micro-nano structure is silicon dioxide, and the micro-nano structure is controlled by the direction of laser heating to achieve phase change between crystalline and amorphous states. The optical signal output module includes 10×1 signal output ports.
[0103] In this embodiment, the MNIST data set is used as a training data set to train the diffractive neural network to perform a handwritten digit recognition task. Fig.11 The confusion matrix of the trained diffractive neural network in this embodiment is shown. Fig.11 The ordinate of the confusion matrix represents the true label of the training data set, that is, the label that the diffractive neural network should correctly identify; the abscissa represents the predicted label output by the diffractive neural network, that is, the label actually output by the diffractive neural network; the diagonal line from the upper left to the lower right represents the situation where the predicted label is correctly matched to the true label.
[0104] It can be seen from the confusion matrix that after the training is completed, the average accuracy of the diffractive neural network provided in this embodiment for the handwritten digit recognition task can reach 92.53%, and it has excellent handwritten digit recognition task capability.
[0105] Example 2
[0106] The diffractive neural network provided in Example 2 is obtained by further adding a nonlinear excitation layer to the diffractive neural network provided in Example 1. Specifically, a nonlinear excitation layer is added after the last diffractive layer of the diffractive neural network provided in Example 1. The nonlinear excitation element in the nonlinear excitation layer is a germanium-silicon coupled waveguide; wherein the length, width and height of the silicon waveguide are 2000nm, 400nm and 2000nm respectively, and the length, width and height of the germanium waveguide are 500nm, 300nm and 2000nm respectively.
[0107] In this embodiment, the MNIST data set is also used as a training data set to train the diffractive neural network to perform a handwritten digit recognition task. Fig.12 The confusion matrix of the trained diffractive neural network in this embodiment is shown. Fig.12, it can be seen from the confusion matrix that after the training of the diffractive neural network provided in this embodiment, the average accuracy of the handwritten digit recognition task is improved to 97.19%. It can be seen that the addition of the nonlinear excitation layer further improves the task execution capability of the diffractive neural network.
[0108] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the utility model disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the appended claims.
Claims
1. A diffractive neural network based on a metasurface, characterized in that: The diffractive neural network comprises: an optical signal input module; an optical signal modulation module; an optical signal output module; Wherein, the optical signal input module is used to guide the optical signal to the optical signal modulation module; The optical signal modulation module includes a metasurface; the metasurface is used to modulate the received optical signal; the metasurface includes at least two reconfigurable micro-nano structure units; the reconfigurable micro-nano structure unit includes a micro-nano structure and a spacing structure arranged around the micro-nano structure; The optical signal output module is used to receive the optical signal from the optical signal modulation module and output the optical signal to a detector used for detecting signals.
2. The diffractive neural network according to claim 1, characterized in that The arrangement period of the reconfigurable micro-nano structure units is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 2 times the central wavelength.
3. The diffractive neural network according to claim 1, characterized in that The length of the reconfigurable micro-nano structure unit in a direction perpendicular to the metasurface is greater than or equal to 0.3 times the central wavelength of the optical signal and less than or equal to 5 times the central wavelength.
4. The diffractive neural network according to claim 1, characterized in that The optical signal modulation module includes at least two layers of metasurfaces; in the direction from the optical signal input module to the optical signal output module, the metasurfaces are hierarchically arranged in sequence; the optical signal modulated by the metasurface propagates along the hierarchical arrangement direction of the metasurface.
5. The diffractive neural network according to claim 4, characterized in that The spacing between adjacent levels of metasurfaces is greater than or equal to 0.5 times the central wavelength of the optical signal and less than or equal to 1000 times the central wavelength.
6. The diffractive neural network according to claim 1, characterized in that: In each layer of the super surface, one of the reconfigurable micro-nano structure units constitutes a neuron, or at least two adjacent reconfigurable micro-nano structure units constitute a neuron.
7. The diffractive neural network according to claim 1, characterized in that: In each layer of the hypersurface, the number of neurons J is less than 2M, where M is the number of signal input ports in the optical signal input module.
8. The diffractive neural network according to claim 1, characterized in that: The material of the micro-nano structure is a phase change material; the reconfigurable micro-nano structure unit also includes a phase change excitation element for applying excitation to the micro-nano structure.
9. The diffractive neural network according to claim 1, characterized in that: The reconfigurable micro-nano structure unit further includes a refractive index matching structure, and the refractive index matching structure is used to match the refractive index of the micro-nano structure and the refractive index of the spacing structure to control the transmittance of the reconfigurable micro-nano structure unit.
10. The diffractive neural network according to claim 1, characterized in that: The optical signal modulation module further includes: a nonlinear excitation layer, and the nonlinear excitation layer is used to exert a nonlinear response on the light intensity of the optical signal.
11. The diffractive neural network according to claim 10, characterized in that: The nonlinear excitation layer includes a nonlinear optical waveguide.
12. The diffractive neural network according to claim 11, characterized in that: The nonlinear optical waveguide is integrated on the light emitting surface of at least one layer of the metasurface by using semiconductor technology.
13. The diffractive neural network according to claim 1, characterized in that: The optical signal input module comprises: at least two signal input ports; at least two first optical waveguides; The first optical waveguides are matched one-to-one with the signal input ports, and the first optical waveguides are used to couple the optical signal received by the corresponding signal input port to the optical signal modulation module.
14. The diffractive neural network according to claim 13, characterized in that: The end of the first optical waveguide close to the optical signal modulation module is tapered, and the direction of the tapered tip is the same as the propagation direction of the optical signal.
15. The diffractive neural network according to claim 1, characterized in that: The optical signal output module comprises: at least two signal output ports; at least two second optical waveguides; The second optical waveguides are matched one-to-one with the signal output ports, and the second optical waveguides are used to couple the optical signal from the optical signal modulation module to the corresponding signal output port.
16. The diffractive neural network according to claim 15, characterized in that: The head end of the second optical waveguide close to the optical signal modulation module is tapered, and the tip of the tapered waveguide faces in the opposite direction to the propagation direction of the optical signal.
17. A three-dimensional diffraction neural network, characterized in that: The three-dimensional diffractive neural network comprises at least two diffractive neural networks according to any one of claims 1 to 16; The diffractive neural networks are stacked and arranged perpendicular to the propagation direction of the optical signal.
18. The three-dimensional diffractive neural network according to claim 17, characterized in that: The three-dimensional diffraction neural network is packaged using a stacking packaging process.