Diffraction neural network and three-dimensional diffraction neural network

Abstract

The utility model provides a diffraction neural network and a three-dimensional diffraction neural network. The diffraction neural network provided by the utility model comprises an input port, a diffraction system and an output port, wherein the input port is used for receiving input light and guiding the input light to the diffraction system; the diffraction system comprises a diffraction layer and a focusing light control system; the diffraction layer comprises a metasurface, and the metasurface is used for modulating the input light; the metasurface comprises at least two reconfigurable micro-nano units based on the phase change material; the focusing light control system is used for performing phase change regulation and control on the phase change material in the diffraction layer by controlling laser; and the output port is used for receiving the output light from the diffraction layer subjected to phase change regulation and control by the focusing light control system and transmitting the output light to the detector array for receiving the output light. The diffraction neural network provided by the utility model is high in training efficiency, and the speed of processing the light calculation task is accelerated.

Description

Technical Field

[0001] This application relates to the field of optical chips, and particularly to a diffractive neural network and a three-dimensional diffractive neural network. Background Art

[0002] With the development of optical technology and metamaterial research, various studies on using optical structures to simulate neural networks, that is, optical neural networks, have been carried out. Among them, the diffractive optical neural network (DONN), as a representative of optical neural networks, uses light as the information medium and utilizes the diffraction of light to perform arithmetic processing on information, so as to execute the tasks matched during training.

[0003] In the prior art, using phase change materials as the diffractive layer of a reconfigurable diffractive neural network has achieved certain results. The phase change material can switch between two states, amorphous and crystalline, under the stimulation of external conditions, so as to realize the regulation of the equivalent refractive index of the phase change material. When training a reconfigurable diffractive neural network or when it is necessary to use a reconfigurable diffractive neural network to process multiple different optical computing tasks, it is necessary to regulate multiple phase change materials in the diffractive layer of the reconfigurable diffractive neural network to meet different requirements. However, in the existing method of regulating phase change materials through a high-magnification microscope objective lens, if it is desired to regulate phase change materials at multiple different positions, it is necessary to control the high-magnification microscope objective lens to translate and rotate, so that the high-magnification microscope objective lens can sequentially regulate each phase change material to complete the regulation of all phase change materials. This regulation method takes a long time and has low regulation efficiency, resulting in a long training time for the reconfigurable diffractive neural network and a long time for processing large-scale optical computing tasks containing multiple different tasks, seriously affecting the efficiency of optical computing. Summary of the Utility Model

[0004] An object of this application is to propose a diffractive neural network and a three-dimensional diffractive neural network, which can achieve efficient parallel regulation of the diffractive layer composed of phase change materials in the diffractive neural network and the three-dimensional diffractive neural network, improve the training efficiency of the diffractive neural network and the three-dimensional diffractive neural network, and improve the speed of the diffractive neural network and the three-dimensional diffractive neural network for processing optical computing tasks.

[0005] According to one aspect of the embodiments of this application, a diffractive neural network is disclosed, and the diffractive neural network includes: an input port, a diffractive system, and an output port;

[0006] Wherein, the input port is used for receiving input light and guiding the input light to the diffractive system;

[0007] The diffraction system includes a diffraction layer and a focusing light control system; the diffraction layer includes a metasurface, and the metasurface is used for modulating the input light; the metasurface includes at least two reconfigurable micro-nano units based on phase change materials;

[0008] The focusing light control system is used for phase change regulation of the phase change material in the diffraction layer by controlling a laser;

[0009] The output port is used for receiving the output light from the diffraction layer that has been phase change regulated by the focusing light control system, and emitting the output light to a detector array for receiving the output light.

[0010] In an exemplary embodiment of the present application, the diffraction layer includes at least two layers of metasurfaces, and the metasurfaces are arranged at intervals in the light path propagation direction in sequence.

[0011] In an exemplary embodiment of the present application, the reconfigurable micro-nano unit includes a substrate, a micro-nano structure, and a spacer structure;

[0012] Wherein, the material of the micro-nano structure is a phase change material.

[0013] In an exemplary embodiment of the present application, the reconfigurable micro-nano unit includes a substrate, a micro-nano structure, and a spacer structure;

[0014] Wherein, the material of the spacer structure is a phase change material.

[0015] In an exemplary embodiment of the present application, the diffraction layer is disposed at the first focal length of the focusing light control system.

[0016] In an exemplary embodiment of the present application, the focusing light control system is a multi-focus focusing light control system or a single-focus focusing light control system.

[0017] In an exemplary embodiment of the present application, the multi-focus focusing light control system includes a phase modulation module and a lens module;

[0018] The phase modulation module is used for modulating the wavefront of the control laser, and transmitting the control laser with the modulated wavefront to the lens module;

[0019] The lens module is used for focusing the control laser with the modulated wavefront to the phase change material of the diffraction layer.

[0020] In an exemplary embodiment of the present application, the phase modulation module is a spatial light modulator or a digital micromirror device;

[0021] The phase modulation module is disposed at the entrance pupil of the lens module.

[0022] According to another aspect of the embodiments of the present application, a three-dimensional diffraction neural network is disclosed. The three-dimensional diffraction neural network includes a three-dimensional input port, a three-dimensional diffraction system, and a three-dimensional output port;

[0023] Wherein, the three-dimensional input port includes at least two input ports as described in any one of the above, and the input port is used to receive input light and guide the input light to the three-dimensional diffraction system;

[0024] The three-dimensional diffraction system includes at least two diffraction layers as described in any one of the above. The diffraction layer includes a metasurface, and the metasurface is used to modulate the input light; the metasurface includes at least two reconfigurable micro-nano units based on phase change materials;

[0025] The three-dimensional diffraction system further includes at least two focusing light control systems as described in any one of the above. The focusing light control system is used to perform phase change regulation on the phase change material in the diffraction layer by controlling a laser;

[0026] The three-dimensional output port includes at least two output ports as described in any one of the above. The output port is used to receive the output light from the diffraction layer that has been phase change regulated by the focusing light control system and emit the output light to a detector array for receiving the output light.

[0027] In an exemplary embodiment of the present application, the focusing light control system and the diffraction layer are in one-to-one correspondence;

[0028] The diffraction layer is disposed at the focal length of the focusing light control system.

[0029] In an exemplary embodiment of the present application, when the focal lengths of all the focusing light control systems are the same, all the focusing light control systems are not in the same plane.

[0030] In an exemplary embodiment of the present application, when the focal lengths of all the focusing light control systems are different, all the focusing light control systems are in the same plane.

[0031] The diffractive neural network provided by this application includes: an input port, a diffractive system, and an output port; wherein, the input port is used to receive input light and guide the input light to the diffractive system; the diffractive system includes a diffractive layer and a focusing light control system; the diffractive layer includes a metasurface, and the metasurface is used to modulate the input light; the metasurface includes at least two reconfigurable micro-nano units based on phase change materials; the focusing light control system is used to perform phase change regulation on the phase change materials in the diffractive layer by controlling a laser; the output port is used to receive the output light from the diffractive layer that has been phase change regulated by the focusing light control system and emit the output light to a detector array for receiving the output light. The diffractive neural network and the three-dimensional diffractive neural network provided by this application can perform independent and efficient parallel phase change regulation on the reconfigurable micro-nano units based on phase change materials, improve the training efficiency of the diffractive neural network, and accelerate the speed of processing optical computing tasks.

[0032] Other features and advantages of this application will become apparent from the following detailed description, or will be learned in part through the practice of this application.

[0033] It should be understood that the above general description and the following detailed description are only exemplary and do not limit this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] By referring to the accompanying drawings and describing its exemplary embodiments in detail, the above and other objectives, features, and advantages of this application will become more apparent.

[0035] Figure 1 FIG. 1 shows a first schematic diagram of a diffractive neural network in an embodiment of this application.

[0036] Figure 2 FIG. 2 shows a second schematic diagram of a diffractive neural network in an embodiment of this application.

[0037] Figure 3 FIG. 3 shows a schematic structural diagram of a reconfigurable micro-nano unit in an embodiment of this application.

[0038] Figure 4 FIG. 4 shows a third schematic diagram of a diffractive neural network in an embodiment of this application.

[0039] Figure 5 FIG. 5 shows a fourth schematic diagram of a diffractive neural network in an embodiment of this application.

[0040] Figure 6 FIG. 6 shows a first schematic diagram of a three-dimensional diffractive neural network in an embodiment of this application.

[0041] Figure 7 FIG. 7 shows a second schematic diagram of a three-dimensional diffractive neural network in an embodiment of this application.

[0042] Figure 8 Shows a third schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application.

[0043] Figure 9 Shows a fourth schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application.

[0044] Figure 10 Shows a fifth schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application.

[0045] Figure 11 Shows a sixth schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application.

[0046] Reference numerals:

[0047] 1 - Focusing optical control system; 2 - Metasurface; 3 - Control laser; 4 - Reconfigurable micro-nano unit;

[0048] 41 - Micro-nano structure; 42 - Spacing structure; 43 - Substrate; 5 - Base layer;

[0049] 6 - First diffraction layer; 7 - Second diffraction layer; 8 - Third diffraction layer. Detailed implementation manners

[0050] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this application will be more thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art. The drawings are merely schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus their repeated description will be omitted.

[0051] In addition, the described features, structures, or characteristics can be combined in any suitable manner in one or more example embodiments. In the following description, numerous specific details are provided to give a thorough understanding of the example embodiments of the present application. However, those skilled in the art will realize that the technical solutions of the present application can be practiced without one or more of the specific details, or other methods, components, steps, etc. can be adopted. In other cases, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring the various aspects of the present application.

[0052] Some of the block diagrams shown in the drawings are functional entities and do not necessarily have to correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processing unit devices and / or microcontroller devices.

[0053] In order to make the diffractive neural network reconfigurable and thus have the ability of dynamic training, the diffractive layer in the diffractive neural network that provides phase to the input light needs to have the ability of being reconfigured. In the related art, it is proposed to use phase change materials to form the diffractive layer of the reconfigurable diffractive neural network, so as to realize the regulation of the state of the phase change materials through external stimuli, so as to realize the training of the reconfigurable diffractive neural network, or realize the switching of the types of processable tasks of the reconfigurable diffractive neural network. In the related art, the regulation of the phase change materials is usually to focus the control laser on the phase change materials of the diffractive layer of the reconfigurable neural network by using a high-magnification microscope objective lens to cause the phase change materials to undergo phase change. However, as the optical computing tasks to be processed become more and more complex, the number of diffractive layers based on phase change materials in the reconfigurable diffractive neural network is increasing, that is, the number of phase change materials that need to be phase-regulated in the reconfigurable diffractive neural network is increasing. The time required for the method of sequentially regulating each phase change material by using a high-magnification microscope objective lens is very long, resulting in slow processing speed and low efficiency of the reconfigurable diffractive neural network for optical computing tasks.

[0054] In consideration of overcoming the above-mentioned defects existing in the related art, the present application provides a diffractive neural network, which can realize the efficient parallel regulation of the diffractive layer composed of phase change materials, thereby greatly improving the processing speed of the diffractive neural network for optical computing tasks.

[0055] The diffractive neural network provided by the present application includes: an input port, a diffractive system, and an output port.

[0056] Specifically, the input port is used to receive the input light and guide the input light to the diffractive system. The working wavelength band of the input light can be one or more of the visible light band, the near-infrared band, and the terahertz band, or can also be an optical signal of other working wavelength bands, which is not limited in the present application.

[0057] The diffractive system includes a diffractive layer and a focusing light control system. The diffractive layer is used to perform corresponding optical computing by using the input light; the diffractive layer includes a metasurface, and the metasurface is used to modulate the received input light; that is, in the embodiments of the present application, the metasurface is the physical carrier of the neurons in the diffractive neural network. Specifically, the metasurface includes at least two reconfigurable micro-nano units based on phase change materials; the reconfigurable micro-nano units are used as the functional units of the diffractive layer and further constitute reconfigurable neurons.

[0058] The focused light control system is used to perform phase change regulation on the phase change material in the diffraction layer by controlling the laser. Specifically, the focused light control system actually controls the phase change material in the reconfigurable micro-nano unit to switch between the crystalline state and the amorphous state, so as to realize the phase change regulation of the reconfigurable micro-nano unit based on the phase change material. The equivalent refractive index of the reconfigurable micro-nano unit will change with the phase change of the phase change material, that is, the phase provided by the reconfigurable micro-nano unit will also change accordingly. Correspondingly, the phase provided by the metasurface composed of the reconfigurable micro-nano units will also change, and the modulation effect of the metasurface on the input light will also change. Thus, the diffraction layer of the entire diffraction neural network can be reconstructed according to the actual task requirements to achieve the task requirements.

[0059] Specifically, the focused light control system can focus the incident light with a high energy density (such as a laser) onto a light focus smaller than or equal to the size of the reconfigurable micro-nano unit in the metasurface, so as to heat the phase change material at this light focus and cause it to undergo a phase change, thereby realizing the phase change regulation of the phase change material of the reconfigurable micro-nano unit and changing the equivalent refractive index of the micro-nano structure.

[0060] The output port is used to receive the output light from the diffraction layer that has undergone phase change regulation by the focused light control system, and send the output light to a pre-set detector array, and finally analyze the content detected by the detector array to obtain the final optical calculation result.

[0061] In one embodiment, the diffraction layer includes at least two metasurfaces, and the metasurfaces are arranged at intervals in sequence along the optical path propagation direction. Moreover, all the metasurfaces are arranged on the same side of the base layer of the diffraction neural network. In fact, in the diffraction neural network, the diffraction layer uses the input light to simulate the forward propagation process in the fully connected neural network; according to the Fresnel diffraction theorem, when the input light passes through the diffraction layer, it can be used as a secondary radiation source and propagate to the next diffraction layer again. That is to say, all the network matrix operations are completed during the diffraction process of the input light, and it is modulated into the corresponding output light. It should be noted that the arrangement intervals of the metasurfaces in the diffraction layer can be equal or unequal, and are specifically set according to actual needs, which are not limited herein.

[0062] As an example, Figure 1 shows the first schematic diagram of the diffraction neural network in an embodiment of the present application. As Figure 1As shown, in the diffraction neural network provided by an embodiment of the present application, the diffraction layer in the diffraction neural network includes five metasurfaces 2. The arrow direction is the optical path propagation direction. The five metasurfaces 2 are arranged at intervals in sequence along the optical path propagation direction, and the five metasurfaces 2 are all disposed on the base layer 5. During the process of optical calculation, the input light is first guided by the input port to the first metasurface along the optical path direction, and the first metasurface modulates the input light; the input light modulated by the first metasurface continues to propagate along the optical path propagation direction to the second metasurface, and the second metasurface modulates the input light again; the input light modulated by the second metasurface continues to propagate along the optical path propagation direction to the third metasurface,..., and the fifth metasurface modulates the input light modulated by the fourth metasurface to obtain the output light; the output light propagates along the optical path propagation direction to the output port, and the output port then emits the output light to the detector array for receiving the output light. Finally, the content detected by the detector array is analyzed to obtain the final optical calculation result.

[0063] Figure 2 FIG. shows a second schematic diagram of the diffraction neural network in an embodiment of the present application. As Figure 1 and Figure 2 shown, in the diffraction neural network, the diffraction layer includes at least two metasurfaces 2, and the two metasurfaces 2 are both arranged on the same side of the base layer 5 of the diffraction neural network; each metasurface 2 includes at least two reconfigurable micro-nano units based on phase change materials. Therefore, the focusing optical control system 1 will control the laser 3 to perform phase change regulation on each reconfigurable micro-nano unit 4 respectively.

[0064] In one embodiment, the diffraction layer may also only include a single diffraction layer, that is, the diffraction neural network provided by this embodiment is a single-layer neural network.

[0065] In one embodiment, the reconfigurable micro-nano unit 4 includes a substrate 43, a micro-nano structure 41, and a spacer structure 42. Figure 3 FIG. shows a structural schematic diagram of the reconfigurable micro-nano unit in an embodiment of the present application. As Figure 3 shown, in the reconfigurable micro-nano unit, the micro-nano structure 41 is disposed on one side of the substrate 43, and the spacer structure 42 is disposed around the micro-nano structure 41.

[0066] In one embodiment, the material of the micro-nano structure of the reconfigurable micro-nano unit is a phase change material. Correspondingly, the spacer structure of the reconfigurable micro-nano unit can be air or other materials that are transparent in the wavelength band of the input light.

[0067] In one embodiment, the material of the spacer structure in the reconfigurable micro-nano unit is a phase change material. Correspondingly, the micro-nano structure of the reconfigurable micro-nano unit can be air or other materials that are transparent in the wavelength band of the input light.

[0068] When the focusing light control system performs phase change control on the reconfigurable micro-nano unit, the focusing light control system will focus the control laser on the reconfigurable micro-nano unit. That is, the focusing light control system generates a light focus that matches each reconfigurable micro-nano unit in the diffraction layer one by one, ensuring that the focusing light control system can focus the control laser on the corresponding reconfigurable micro-nano unit, thereby realizing phase change control of all reconfigurable micro-nano units in the diffraction layer.

[0069] In one embodiment, since the diffractive neural network in the present application realizes the regulation of the reconfigurable micro-nano units based on phase change materials in the diffractive layer through the focusing light control system, the diffractive layer is set at one focal length of the focusing light control system. Figure 1 As shown, the focal length of the focusing light control system is f, and the diffraction layer is arranged at one times the focal length of the focusing light control system, that is, each metasurface in the diffraction layer is arranged at one times the focal length of the focusing light control system.

[0070] It should be noted that the one-focal-length position in the embodiment of the present application does not refer to the one-focal-length position in a strict sense, but is within a reasonable tolerance range near the one-focal-length. It can be understood that if the focal length of the focusing light control system is f, the one-focal-length position can be expressed as: f±Δx, where Δx represents a reasonable tolerance. That is to say, in the present application, the one-focal-length position is equivalent to the one-focal-length plus / minus a reasonable tolerance.

[0071] Continuation of the above Figure 1 For example, after obtaining the final optical calculation result, the diffractive neural network can be further trained based on the final optical calculation result, or different optical calculation tasks can be performed based on the diffractive neural network with reconfigurable capabilities. Therefore, by performing phase change control on the diffractive layer in the diffractive neural network according to the focusing optical control system, the following can be achieved, including but not limited to:

[0072] (1) Efficient training of reconfigurable diffractive neural networks;

[0073] (2) Efficient switching of reconfigurable diffractive neural networks for different subtasks in large-scale optical computing tasks.

[0074] During the training process of the reconfigurable diffraction neural network, the input light irradiates on the input port and is conducted to the diffraction system through the input port. After being modulated by the diffraction layer in the diffraction system, the output light is obtained. The output light is emitted through the output port to the detector array set up in advance, and by analyzing the content detected by the detector array, the result of this optical calculation can be obtained. Subsequently, based on the result of this optical calculation, a regulation strategy for the diffraction layer is formulated, and according to the regulation strategy, the focusing light control system is used to regulate each metasurface in the diffraction neural network respectively, so that the phase of each metasurface in the diffraction layer changes correspondingly. After the regulation is completed, the input light is emitted to the input port again, the result of this optical calculation is recorded again, and according to the result of this optical calculation, the focusing light control system is used to regulate each metasurface in the diffraction neural network respectively. Repeat the above process until the computing ability of the diffraction neural network meets the preset conditions, thereby completing the training of the diffraction neural network. It should be noted that when the computing ability of the diffraction neural network meets the preset conditions, it is not necessarily to stop the training of the diffraction neural network, and the diffraction neural network can still be continuously optimized. During this training process, through the focusing light control system of the present application, each metasurface in the diffraction layer can be quickly regulated, which actually greatly shortens the training time of the diffraction neural network, improves the training efficiency of the diffraction neural network, and also makes the subsequent adjustment and optimization of the diffraction neural network more efficient.

[0075] In large-scale optical computing tasks, there are often various different types of optical computing subtasks. The reconfigurable diffraction neural network has the reconfigurable ability and can realize the change of the phase of the diffraction layer, so as to adapt to different optical computing subtasks. When the reconfigurable diffraction neural network has been trained for multiple different optical computing subtasks respectively, the regulation states of each metasurface in the diffraction layer when the diffraction neural network is applicable to different optical computing subtasks can be recorded. For example, if the large-scale optical computing task includes the first subtask, the second subtask and the third subtask, and the three subtasks are different optical computing tasks respectively, and the initial state of the diffraction neural network has been set to the state that can complete the first subtask, then after the diffraction neural network completes the first subtask, the diffraction layer needs to be regulated through the focusing light control system to change the phase of the diffraction layer, so that the diffraction neural network can complete the second subtask, and the same is true for completing the third subtask subsequently. In the application, through the focusing light control system, all the metasurfaces in the diffraction layer can be quickly regulated according to the recorded regulation states of the metasurfaces for different optical computing subtasks, so that the diffraction neural network can quickly change the phase of the diffraction layer to adapt to different optical computing tasks, thereby greatly shortening the time to complete the large-scale optical computing task, improving the computing efficiency of the diffraction neural network, and also increasing the flexibility of the diffraction neural network.

[0076] In one embodiment, the focused light control system is a multi - focus focused light control system. During the process of realizing phase change regulation, the multi - focus focused light control system uses a multi - focus parallel regulation method; that is, phase change regulation is performed on multiple reconfigurable micro - nano units simultaneously. Specifically, Figure 1 and Figure 2 The focused light control systems shown are multi - focus focused light control systems. The multi - focus focused light control system will simultaneously focus the control laser to the corresponding reconfigurable micro - nano unit according to the light focus corresponding to each reconfigurable micro - nano unit based on phase - change material on the metasurface, so as to realize rapid phase - change regulation of the entire metasurface.

[0077] In one embodiment, the multi - focus focused light control system includes a phase modulation module and a lens module; among them, the phase modulation module is used to modulate the wavefront of the control laser and send the control laser with the modulated wavefront to the lens module; the lens module is used to focus the control laser with the modulated wavefront to the phase - change material of the diffraction layer, that is, to focus the control laser input into the multi - focus focused light control system, so that the control laser can be focused to each reconfigurable micro - nano unit respectively.

[0078] In one embodiment, the phase modulation module is arranged at the entrance pupil of the lens module.

[0079] That is to say, the control laser first passes through the phase modulation module. The phase modulation module modulates the wavefront of the control laser and transmits the control laser with the modulated wavefront to the lens module. The lens module focuses the control laser with the modulated wavefront, thereby forming multiple light foci. All the light foci are in the same plane, and each light focus corresponds to each reconfigurable micro - nano unit in the diffraction layer respectively, so as to realize independent regulation of each reconfigurable micro - nano unit, and can rapidly perform phase - change regulation on each reconfigurable micro - nano unit in the diffraction layer, improving the efficiency of diffraction neural network reconstruction.

[0080] In one embodiment, the control laser can be parallel light, and the phase modulation module can modulate the wavefront of the parallel light into a converging wavefront and transmit the control laser with the modulated wavefront to the lens module.

[0081] In one embodiment, the phase modulation module can be a Spatial Light Modulator (SLM). The spatial light modulator includes a liquid - crystal spatial light modulator and also includes a spatial light modulator composed of a tunable metasurface.

[0082] In one embodiment, the phase modulation module can also be a Digital Micromirror Devices (DMD). DMD is a device based on semiconductor manufacturing technology and composed of a high - speed digital light reflection switch array.

[0083] In one embodiment, the lens module includes at least one metasurface lens or at least one microlens. That is, in the embodiments of the present application, the lens module may be composed of a metasurface lens, a metasurface lens group, a metasurface lens array, or may be composed of a microlens, a microlens group, a microlens array.

[0084] When the lens module is a metasurface lens array, there is a one-to-one correspondence between each metasurface lens in the metasurface lens array and each reconfigurable micro-nano unit in the diffraction layer. That is, each metasurface lens in the metasurface lens array is respectively used to perform phase change regulation on the corresponding reconfigurable micro-nano unit, and each metasurface lens respectively focuses the control laser onto the corresponding reconfigurable micro-nano unit.

[0085] When the lens module is a microlens array, similarly to the metasurface lens array, there is a one-to-one correspondence between each microlens in the microlens array and each reconfigurable micro-nano unit in the diffraction layer. That is, each microlens in the microlens array is respectively used to perform phase change regulation on the corresponding reconfigurable micro-nano unit, and each microlens respectively focuses the control laser onto the corresponding reconfigurable unit.

[0086] In one embodiment, the lens module includes at least one metasurface lens and at least one microlens. That is, in another embodiment of the present application, the lens module may also be an optical system combined with microlenses and metasurface lenses.

[0087] It should be noted that the lens module is an optical system with a large numerical aperture to be able to generate light foci with a spacing of hundreds of nanometers. The large numerical aperture ensures that the light foci are small and the energy is concentrated, which is more conducive to pixel-level precise regulation.

[0088] It should be noted that if the lens module includes a metasurface lens, the focusing phase of the metasurface lens needs to satisfy:

[0089]

[0090] where φ(x,y) is the focusing phase of the metasurface lens, (x,y) is the position on the metasurface lens, λ represents the working wavelength band of the control laser, and f is the focal length of the metasurface lens. In the present application, the focusing phases of all metasurface lenses satisfy this formula.

[0091] In one embodiment of the present application, for the light source used to emit the control laser to the multi-focus focusing optical control system, it may be a single light source; that is, the corresponding phase change regulation can be achieved by controlling the on / off of the light source. When the light source is turned on, the multi-focus focusing optical control system performs phase change regulation on the reconfigurable micro-nano units in the diffraction layer of the diffraction neural network. And the on-time of the light source can be controlled to control the regulation time of the reconfigurable micro-nano units, so as to enable the phase change material in the reconfigurable micro-nano units to reach the actual required crystallization ratio and complete the precise regulation of the equivalent refractive index of the reconfigurable micro-nano units.

[0092] In another embodiment of the present application, when a metasurface lens array or a microlens array is used as the lens module, the light source for emitting the control laser can be a light source array composed of multiple sub-light sources; it should be noted that there is a one-to-one correspondence between each sub-light source in the light source array and each metasurface lens in the metasurface lens array, or there is a one-to-one correspondence between each sub-light source in the light source array and each microlens in the microlens array. In this embodiment, the on / off of each sub-light source can be controlled individually, so as to realize independent phase change regulation of a certain reconfigurable micro-nano unit in the diffraction layer. That is to say, by controlling whether each sub-light source is turned on or off, it is possible to control whether the corresponding reconfigurable micro-nano unit is regulated, and independent regulation of each reconfigurable micro-nano unit in the diffraction layer of the diffraction neural network can be achieved. Moreover, by setting different turn-on times of each sub-light source, the regulation time of each reconfigurable micro-nano unit can be realized, so that the phase change materials in each reconfigurable micro-nano unit can reach different crystallization ratios respectively, that is, each regulated reconfigurable micro-nano unit can have different equivalent refractive indices, which can better meet the actual reconfigurable requirements, thereby realizing high-precision reconstruction of the diffraction neural network and greatly enhancing the flexibility of the reconstruction of the diffraction neural network.

[0093] In one embodiment, the focusing optical control system can also be a single-focus focusing optical control system. By way of example, the single-focus focusing optical control system includes a laser scanning module and a lens module; the laser scanning module is used to deflect the control laser; the lens module is used to focus the control laser. That is to say, along the optical path direction of the control laser, the setting order of the laser scanning module and the lens module is not unique. The laser scanning module can be arranged upstream of the lens module, or the lens module can be arranged upstream of the laser module.

[0094] Along the optical path direction of the control laser, when the laser scanning module is arranged upstream of the lens module, the laser scanning module is used to deflect the control laser and transmit the deflected control laser to the lens module, and the lens module is used to focus the deflected control laser onto the phase change material of the diffraction layer; when the lens module is arranged upstream of the laser scanning module, the lens module first converges the control laser, and then the laser scanning module deflects the control laser converged by the lens module.

[0095] Specifically, Figure 4 shows a third schematic diagram of the diffraction neural network in an embodiment of the present application, Figure 5 shows a fourth schematic diagram of the diffraction neural network in an embodiment of the present application, as Figure 4 and Figure 5 shown, Figure 4 and Figure 5 the focusing optical control system 1 in is a single-focus focusing optical control system.

[0096] Different from the multi-focus focusing optical control system, in the process of realizing phase change regulation, the single-focus focusing optical control system uses the method of rapid single-focus scanning of light; that is, phase change regulation is performed on each reconfigurable micro-nano unit in turn. Specifically, Figure 4 The multiple control lasers in are emitted by the single-focus focusing optical control system in a certain order. The single-focus focusing optical control system will deflect the control laser in a certain order according to the light focus corresponding to each reconfigurable micro-nano unit through the laser scanning module, and the deflected control laser will be focused on the corresponding reconfigurable unit by the lens module, and phase change regulation is performed on each reconfigurable micro-nano unit in turn, so as to realize the rapid regulation of the entire metasurface. It should be noted that Figure 4 The dotted line part in represents the control laser emitted by the single-focus focusing optical control system for the reconfigurable micro-nano units of different metasurfaces.

[0097] Figure 5 Only the diffraction layer composed of a single-layer metasurface is shown in, and the diffraction layer is arranged on the base layer of the diffraction neural network. Figure 5 The input port and output port arranged on the base layer at the same time are not shown in. Figure 5 In the single-focus focusing optical control system in, the control laser is focused on the reconfigurable micro-nano unit through the single-focus focusing optical control system, so as to realize the phase change regulation of the reconfigurable micro-nano unit. After the phase change regulation of this reconfigurable micro-nano unit is completed, the single-focus focusing optical control system will deflect the control laser through the laser scanning module, and through the cooperation with the lens module, the control laser is focused on another reconfigurable micro-nano unit that needs to perform phase change regulation. Repeat the above process until the phase change regulation of all reconfigurable micro-nano units is completed.

[0098] It should be noted that in this application, the focusing optical control system can realize phase change regulation through the photothermal effect of the control laser, or can realize phase change regulation by irradiating the control laser in the form of laser pulses, which is not limited in this application.

[0099] It should be noted that in the regulation process of the above single - focus focusing optical control system, the method of quickly scanning the single optical focus to perform phase - change regulation on each reconfigurable micro - nano unit quickly refers to being able to meet the requirements of actual optical computing tasks and complete the phase - change regulation that meets the requirements within the task - specified time. In other words, the regulation speed of the single - focus focusing optical control system in this application can be dynamically adjusted according to the time required to complete the phase - change regulation and the number of reconfigurable micro - nano units to be regulated in the optical computing task, so as to meet the target requirements of the optical computing task. As an example, during the training process of a certain diffraction neural network, if the total number of reconfigurable micro - nano units in the diffraction layer is 500 and the total time given for the training process is 1 second, then the single - focus focusing optical control system needs to complete the phase - change regulation of 500 reconfigurable micro - nano units within 1 second. That is, for each reconfigurable micro - nano unit, the single - focus focusing optical control system needs to complete the deflection of the control laser within 1 / 500 second.

[0100] The diffraction neural network in this application can be regarded as a two - dimensional diffraction neural network, and multiple diffraction neural networks can form a three - dimensional diffraction neural network; a three - dimensional diffraction neural network can execute multiple tasks in parallel. For a reconfigurable three - dimensional diffraction neural network, the problem brought by executing multiple tasks in parallel is that when the types of tasks executed in parallel by the three - dimensional diffraction neural network change and all diffraction layers of the three - dimensional diffraction neural network need to be reconfigured, since there are multiple diffraction layers in the three - dimensional diffraction neural network and there are multiple layers of metasurfaces in each diffraction layer, it is difficult for the prior art to perform independent, flexible, and efficient phase - change regulation on the multiple diffraction layers in the three - dimensional diffraction neural network.

[0101] Therefore, this application proposes a three - dimensional diffraction neural network, which includes a three - dimensional input port, a three - dimensional diffraction system, and a three - dimensional output port.

[0102] Specifically, the three - dimensional diffraction neural network can be vertically stacked and encapsulated by multiple diffraction neural networks. Therefore, the three - dimensional input port includes at least two of the above - mentioned input ports. The three - dimensional diffraction system includes at least two of the above - mentioned diffraction layers; the three - dimensional diffraction system also includes at least two of the above - mentioned focusing optical control systems; the three - dimensional output port includes at least two of the above - mentioned output ports.

[0103] Since the three - dimensional diffraction neural network is vertically stacked and encapsulated by multiple diffraction neural networks, at least two of the above - mentioned diffraction layers are vertically stacked and encapsulated in the diffraction layer part of the three - dimensional diffraction neural network.

[0104] In an embodiment, the focusing optical control system is in one - to - one correspondence with the diffraction layer; that is, in the three - dimensional diffraction neural network, each diffraction layer has a corresponding focusing optical control system, and each focusing optical control system can perform phase - change regulation on the corresponding diffraction layer.

[0105] Moreover, each diffraction layer is disposed at the focal length of the corresponding focusing optical control system. Similar to the two-dimensional diffraction neural network, the focal length here is not strictly equal to the focal length, but is equal to the focal length plus / minus a reasonable tolerance.

[0106] It should be noted that in the three-dimensional diffraction neural network, the input ports in the three-dimensional input port, the diffraction layers, and the output ports in the three-dimensional output port are in one-to-one correspondence. That is to say, in the process of the three-dimensional diffraction neural network executing multiple tasks in parallel, each input port in the three-dimensional input port respectively guides the input light to the corresponding diffraction layer, and the corresponding output port then guides the output light modulated by the corresponding diffraction layer to the detector array, and analyzes the content detected by the detector, so as to obtain the calculation results of multiple tasks.

[0107] In the three-dimensional diffraction neural network provided in the present application, similar to the two-dimensional diffraction neural network, the focusing optical control system can be a multi-focus focusing optical control system or a single-focus focusing optical control system.

[0108] In one embodiment, when the focal lengths of all the focusing optical control systems are the same, all the focusing optical control systems are not in the same plane.

[0109] Specifically, Figure 6 FIG. 1 shows a first schematic diagram of a three-dimensional diffraction neural network according to an embodiment of the present application, as Figure 6 shown, Figure 6 only the three-dimensional diffraction system part of the three-dimensional diffraction neural network in this embodiment is shown. In this embodiment, the three-dimensional diffraction system includes a first diffraction layer 6, a second diffraction layer 7, and a third diffraction layer 8, and the focusing optical control system 1 is a multi-focus focusing optical control system. The first diffraction layer, the second diffraction layer, and the third diffraction layer respectively have corresponding multi-focus focusing optical control systems, and the first diffraction layer, the second diffraction layer, and the third diffraction layer are respectively disposed at the focal length of the corresponding multi-focus focusing optical control system, f 1 、f 2 and f 3 are respectively the focal lengths of the multi-focus focusing optical control systems corresponding to the first diffraction layer, the second diffraction layer, and the third diffraction layer. The first diffraction layer, the second diffraction layer, and the third diffraction layer are respectively phase-change regulated by the corresponding multi-focus focusing optical control system according to actual needs, so as to realize independent and efficient phase-change regulation of multiple diffraction layers in the three-dimensional diffraction neural network. In this embodiment, the focal lengths of the multi-focus focusing optical control systems corresponding to each diffraction layer are equal, that is, f 1 =f 2 =f 3 , and since the first diffraction layer, the second diffraction layer, and the third diffraction layer are in different planes, the multi-focus focusing optical control systems corresponding to each diffraction layer are also in different planes. In Figure 6In [the figure], the dashed part in the focusing optical control system 1 is used to indicate that the focusing optical control system corresponding to the diffraction layer is an integral whole. The focusing optical control systems at both ends of the dashed part are the same system, and the dashed part can be air or a material that is transparent in the wavelength band of the control laser, as long as it can ensure that the multi-focus focusing optical control system corresponding to the diffraction layer is located in the same plane.

[0110] Figure 7 The second schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application is shown. As Figure 7 shown, Figure 7 the focusing optical control system 1 in [the figure] is a single-focus focusing optical control system, and f 1 , f 2 and f 3 are the focal lengths of the single-focus focusing optical control systems corresponding to the first diffraction layer 6, the second diffraction layer 7, and the third diffraction layer 8 respectively, and the focal lengths of the single-focus focusing optical control systems corresponding to each diffraction layer are equal, that is, f 1 = f 2 = f 3 . Since the first diffraction layer, the second diffraction layer, and the third diffraction layer are located in different planes, the single-focus focusing optical control systems corresponding to each diffraction layer are also located in different planes. It should be noted that, Figure 7 the dashed part in [the figure] represents the control laser emitted by the single-focus focusing optical control system for the reconfigurable micro-nano units of different metasurfaces.

[0111] In an embodiment, when all the focusing optical control systems are located in the same plane, the focal lengths of all the focusing optical control systems are different.

[0112] Specifically, Figure 8 the third schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application is shown, Figure 9 and the fourth schematic diagram of the three-dimensional diffraction neural network according to an embodiment of the present application is shown. As Figure 8 and Figure 9 shown, in this embodiment, the three-dimensional diffraction system includes a first diffraction layer 6, a second diffraction layer 7, and a third diffraction layer 8. All the focusing optical control systems are located in the same plane, and the focusing optical control systems are multi-focus focusing optical control systems. A is the multi-focus focusing optical control system corresponding to the first diffraction layer, B is the multi-focus focusing optical control system corresponding to the second diffraction layer, and C is the multi-focus focusing optical control system corresponding to the third diffraction layer. A, B, and C are all in the same plane. It should be understood that, since the distances from the planes where the first diffraction layer, the second diffraction layer, and the third diffraction layer are located to the plane where the multi-focus focusing optical control system is located are different, and each diffraction layer is disposed at the focal length of the corresponding multi-focus focusing optical control system, the focal lengths of the three multi-focus focusing optical controls A, B, and C are different, that is, f 1 ≠ f 2 ≠ f 3, so that A, B, and C can respectively generate optical foci in different planes, realizing the phase change regulation of all reconfigurable micro-nano units in the three-dimensional diffraction neural network.

[0113] Figure 10 FIG. 5 shows a fifth schematic diagram of a three-dimensional diffraction neural network according to an embodiment of the present application. As Figure 10 shown, in this embodiment, the three-dimensional diffraction system includes a first diffraction layer 6, a second diffraction layer 7, and a third diffraction layer 8. All the focusing light control systems 1 are located in the same plane, and the focusing light control system is a single-focus focusing light control system. Similar to Figure 8 and Figure 9 the embodiment shown, in the Figure 10 corresponding embodiment, f 1 ≠f 2 ≠f 3 , so that each single-focus focusing light control system can perform phase change regulation on the corresponding diffraction layer. It should be noted that Figure 10 the dashed part in

[0114] represents the control laser emitted by the single-focus focusing light control system for the reconfigurable micro-nano units of different metasurfaces.

[0115] Furthermore, when each focusing light control system is located in a different plane, the focal lengths of each focusing light control system can also be different.

[0115] It should be emphasized that since the diffraction layer part in the three-dimensional diffraction system is vertically stacked by multiple diffraction layers, and in the Figures 6 to 10 embodiment shown above, each diffraction layer needs to be subjected to phase change regulation. Therefore, if you want to perform phase change regulation on the diffraction layer through the focusing light control system corresponding to each diffraction layer, it is necessary to ensure that the control laser output by each focusing light control system will not be blocked, so as to be effectively focused on the target reconfigurable micro-nano unit that needs to be subjected to phase change regulation.

[0116] In one embodiment, in the direction of the vertical stacking of the diffraction layers, by performing corresponding misalignment settings on the metasurface to be regulated, and the base layer of the diffraction neural network should be a material that is transparent in the wavelength band of the control laser, so that the base layer can at least transmit the control light, so that each focusing light control system can focus the control laser on the reconfigurable micro-nano unit of the corresponding diffraction layer, realizing the phase change regulation of the entire three-dimensional diffraction neural network. Specifically, when the focusing light control system is a multi-focus focusing light control system, as Figure 6 and Figure 9 shown, Figure 6 and Figure 9 all the metasurfaces in

[0117] are misaligned. Figure 7 and Figure 10 shown, sinceFigure 7 and Figure 10 the positional relationship of the single-focus focusing optical control system in Figure 7 and Figure 10 only the left metasurface in may be blocked, so misalignment settings are required; while Figure 7 and Figure 10 the right metasurface in does not have a blocking situation, so misalignment settings are not required.

[0118] Similarly, in a three-dimensional diffractive neural network, there are also cases where not all diffractive layers need to be phase-transition regulated. In this case, the metasurfaces of the diffractive layers that do not need to be regulated can also not be misaligned. Figure 11 shows the sixth schematic diagram of a three-dimensional diffractive neural network according to an embodiment of the present application. As Figure 11 shown, in this embodiment, the first diffractive layer 6 and the second diffractive layer 7 need to be phase-transition regulated, the third diffractive layer 8 does not need to be phase-transition regulated, and the focusing optical control system in this embodiment is a multi-focus focusing optical control system. Therefore, only the first diffractive layer and the second diffractive layer have corresponding multi-focus focusing optical control systems A and B, and the metasurfaces in the first diffractive layer and the second diffractive layer are both misaligned. The metasurfaces in the third diffractive layer do not need to be misaligned with the metasurfaces in the first diffractive layer and the second diffractive layer.

[0119] That is to say, in the three-dimensional diffractive neural network of the present application, regardless of whether the focusing optical control system is a multi-focus focusing optical control or a single-focus focusing optical control system, and regardless of how many diffractive layers need to be phase-transition regulated, the positional relationship between the metasurfaces only needs to satisfy: each metasurface that needs to be phase-transition regulated can be phase-transition regulated by the corresponding focusing optical control system.

[0120] In another embodiment, control lasers with different wavelengths can be set for the focusing optical control systems corresponding to each diffractive layer to avoid the problem that the control laser focused on the target reconfigurable micro-nano unit is blocked during the phase-transition regulation process, resulting in the inability of the target reconfigurable micro-nano unit to be phase-transition regulated.

[0121] Figures 6 to 11 The three-dimensional diffractive neural network shown is only an example and does not represent the actual setting method of the diffractive layer part and the focusing optical control system in the three-dimensional diffractive neural network of the present application. That is, the number of diffractive layers and the number of metasurfaces in the three-dimensional diffractive neural network of the present application are not limited.

[0122] As an example, such as Figure 9In the three-dimensional diffraction neural network shown, the diffraction layer part includes a first diffraction layer, a second diffraction layer, and a third diffraction layer. Assuming that the tasks performed by the three diffraction layers are all different, the first diffraction layer needs to perform a first subtask, the second diffraction layer needs to perform a second subtask, and the third diffraction layer needs to perform a third subtask. The following is an example:

[0123] (1) Before the three diffraction layers perform three corresponding tasks respectively, each of the three diffraction layers needs to be trained. Since each diffraction layer has a corresponding multi-focus focusing optical control system, each diffraction layer can be trained efficiently simultaneously, greatly improving the training efficiency of the three-dimensional diffraction neural network and shortening the overall time for the three-dimensional diffraction neural network to complete the optical computing task.

[0124] (2) After the training of the three diffraction layers is completed, if there is a certain diffraction layer that still needs to be adjusted and optimized, the corresponding multi-focus focusing optical control system can be used to adjust and optimize this diffraction layer again, making the regulation of the diffraction layer part of the three-dimensional diffraction neural network more targeted and improving the flexibility of the reconstruction of the three-dimensional diffraction neural network.

[0125] (3) If, after the first diffraction layer has completed the first subtask, it is necessary to change the optical computing task to be performed, from the original first subtask to the second subtask, then the phase change regulation of the first diffraction layer can be carried out through the multi-focus focusing optical control system corresponding to the first diffraction layer, so that the phase of the first diffraction layer can be quickly changed, thereby processing the second subtask in a timely manner and improving the overall efficiency of the optical computing task.

[0126] In summary, for the three-dimensional diffraction neural network provided by this application, phase change regulation is carried out on different diffraction layers through multiple focusing optical control systems respectively, improving the training efficiency of the three-dimensional diffraction neural network, increasing the flexibility of the reconstruction of the three-dimensional diffraction neural network, and improving the overall efficiency of the optical computing task.

[0127] After considering the specification and practicing the utility model disclosed herein, those skilled in the art will readily conceive of other embodiments of this application. This application is intended to cover any variations, uses, or adaptations of this application, which follow the general principles of this application and include the common general knowledge or conventional technical means in this technical field that are not disclosed in this application. The specification and examples are only regarded as exemplary, and the true scope and spirit of this application are pointed out by the appended claims.

Claims

1. A diffractive neural network, characterized in that: The diffractive neural network comprises: an input port, a diffractive system and an output port; Wherein, the input port is used to receive input light and guide the input light to the diffraction system; The diffraction system includes a diffraction layer and a focusing light control system; the diffraction layer includes a metasurface, and the metasurface is used to modulate the input light; the metasurface includes at least two reconfigurable micro-nano units based on phase change materials; The focusing light control system is used to control the phase change of the phase change material in the diffraction layer by controlling the laser; The output port is used to receive the output light from the diffraction layer that is phase-changed and regulated by the focusing light control system, and emit the output light to a detector array for receiving the output light.

2. The diffractive neural network according to claim 1, characterized in that The diffraction layer includes at least two layers of metasurfaces, and the metasurfaces are sequentially arranged at intervals along the propagation direction of the light path.

3. The diffractive neural network according to claim 1, characterized in that The reconfigurable micro-nano unit comprises a substrate, a micro-nano structure and a spacer structure; Wherein, the material of the micro-nano structure is a phase change material.

4. The diffractive neural network according to claim 1, characterized in that The reconfigurable micro-nano unit comprises a substrate, a micro-nano structure and a spacer structure; Wherein, the material of the spacing structure is phase change material.

5. The diffractive neural network according to claim 1, characterized in that: The diffraction layer is arranged at one focal length of the focusing light control system.

6. The diffractive neural network according to claim 1, characterized in that: The focusing light control system is a multi-focus focusing light control system or a single-focus focusing light control system.

7. The diffractive neural network according to claim 6, characterized in that: The multi-focal point focusing light control system comprises a phase modulation module and a lens module; The phase modulation module is used to modulate the wavefront of the control laser and transmit the control laser after wavefront modulation to the lens module; The lens module is used to focus the control laser after wavefront modulation onto the phase change material of the diffraction layer.

8. The diffractive neural network according to claim 7, characterized in that: The phase modulation module is a spatial light modulator or a digital micromirror device; The phase modulation module is arranged at the entrance pupil of the lens module.

9. A three-dimensional diffraction neural network, characterized in that: The three-dimensional diffraction neural network includes a three-dimensional input port, a three-dimensional diffraction system and a three-dimensional output port; Wherein, the three-dimensional input port includes at least two input ports, and the input port is used to receive input light and guide the input light to the three-dimensional diffraction system; The three-dimensional diffraction system includes at least two diffraction layers, the diffraction layer includes a metasurface, and the metasurface is used to modulate the input light; the metasurface includes at least two reconfigurable micro-nano units based on phase change materials; The three-dimensional diffraction system further comprises at least two focusing light control systems, and the focusing light control systems are used to control the phase change of the phase change material in the diffraction layer by controlling the laser; The three-dimensional output port includes at least two output ports, and the output ports are used to receive the output light from the diffraction layer controlled by the phase change of the focusing light control system, and emit the output light to a detector array for receiving the output light.

10. The three-dimensional diffractive neural network according to claim 9, characterized in that: The focusing light control system and the diffraction layer are matched one by one; The diffraction layer is arranged at one focal length of the focusing light control system.

Images