Coding module, input port and three-dimensional input port for optical neural network

Abstract

The utility model discloses a coding module for an optical neural network, an input port and a three-dimensional input port. The coding module comprises a plane wave generator; an information encoder; the plane wave generator is arranged at the upstream of the information encoder along the light path propagation direction; the plane wave generator is used for converting the received light beams into plane waves, and the information encoder is used for carrying out information encoding on the plane waves; wherein the information encoder is a reconfigurable metasurface, and the reconfigurable metasurface comprises at least two reconfigurable micro-nano units. According to the method, the light beam is converted into the plane wave before information coding, so that the precision of information coding is improved, and the accuracy of processing a light calculation task is further improved; moreover, the adjustable metasurface serves as an information encoder, so that the number of information which can be borne in a unit area is larger, the input information density of the optical neural network is improved, a data structure with higher dimensionality or more complexity can be processed, and the optical neural network can process a large-scale optical calculation task.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical chips, in particular to an encoding module, an input port and a three-dimensional input port for an optical neural network. BACKGROUND

[0002] In recent years, with the in-depth development of optical technology and metamaterial research, people have carried out various researches on simulating neural networks using optical structures, that is, optical neural networks. Optical neural networks use photons as information carriers and realize signal transmission, modulation and processing through optical elements.

[0003] However, due to the current limitations of hardware and technology, such as insufficient integration of photonic chips and insufficient modulation capacity of optical elements on light beams, the input information density of optical neural networks is relatively low, which makes the optical neural networks process less data in unit time and is difficult to meet the demand of large-scale optical computing tasks. CONTENT OF THE UTILITY MODEL

[0004] In view of the above technical problems, the present application provides an encoding module, an input port and a three-dimensional input port for an optical neural network. The present application converts the light beam into a plane wave before encoding the information, which avoids the influence of the original wavefront curvature of the light beam on the accuracy of information encoding, resulting in a decrease in the accuracy of the final optical computing task. Moreover, by using an adjustable metasurface as an information encoder, more information can be carried in a unit area, which improves the input information density of the optical neural network, enables the optical neural network to process more data streams, and thus can process higher-dimensional or more complex data structures, so that the optical neural network can process large-scale optical computing tasks.

[0005] According to an aspect of an embodiment of the present application, an encoding module for an optical neural network is disclosed, the encoding module comprising: a plane wave generator; an information encoder;

[0006] The plane wave generator is arranged upstream of the information encoder along the direction of light propagation;

[0007] The plane wave generator is configured to convert a received light beam into a plane wave, and the information encoder is configured to encode information on the plane wave;

[0008] The information encoder is a reconfigurable metasurface, and the reconfigurable metasurface comprises at least two reconfigurable micro-nano units.

[0009] In an exemplary embodiment of the present application, one reconfigurable micro-nano unit on the reconfigurable metasurface constitutes one data point.

[0010] In an example embodiment of the present application, at least two reconfigurable micro-nano units on the reconfigurable metasurface constitute a data point.

[0011] In an example embodiment of the present application, the plane wave generator is constituted by at least one metasurface and / or at least one refractive lens.

[0012] In an example embodiment of the present application, the equivalent refractive index of the reconfigurable micro-nano unit changes when subjected to external excitation.

[0013] In an example embodiment of the present application, the external excitation is one or more of optical excitation, electrical excitation, thermal excitation, and mechanical excitation.

[0014] According to an aspect of an embodiment of the present application, an input port for an optical neural network is disclosed, characterized in that the input port comprises: an encoding module according to any one of the above; an incident coupler; an exit coupler;

[0015] The incident coupler and the exit coupler are connected through an optical waveguide.

[0016] The incident coupler is used to receive a light beam emitted by a light source, and the exit coupler is used to emit the light beam to the encoding module.

[0017] According to an aspect of an embodiment of the present application, a three-dimensional input port for a three-dimensional optical neural network is disclosed, characterized in that the three-dimensional input port comprises: at least two input ports described above.

[0018] In an example embodiment of the present application, the three-dimensional input port further comprises: a diffraction element.

[0019] The diffraction element is used to converge the light beams emitted by the light source to the incident couplers of the input ports, respectively.

[0020] In an example embodiment of the present application, the diffraction element is a metasurface or a diffraction grating.

[0021] The application discloses an encoding module, an input port and a three-dimensional input port for an optical neural network, and the encoding module comprises: a plane wave generator; an information encoder; the plane wave generator is arranged upstream of the information encoder in the direction of the optical path propagation; the plane wave generator is used for converting a received light beam into a plane wave, and the information encoder is used for information encoding on the plane wave; wherein the information encoder is a reconfigurable metasurface, and the reconfigurable metasurface comprises at least two reconfigurable micro-nano units. The application provides an encoding module, an input port and a three-dimensional input port for an optical neural network. The application converts a light beam into a plane wave before information encoding, avoids the influence of the original wave front curvature of the light beam on the accuracy of information encoding, and reduces the accuracy of the final optical computing task, that is, the accuracy of information encoding is improved, and the accuracy of processing the optical computing task is improved; and the reconfigurable metasurface is used as the information encoder, so that more information can be carried in a unit area, the input information density of the optical neural network is improved, the optical neural network can process more data streams, and thus a higher-dimensional or more complex data structure can be processed, so that the optical neural network can process large-scale optical computing tasks.

[0022] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0023] It should be understood that the foregoing general description and the following detailed description are only exemplary and are not restrictive of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0024] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0025] Figure 1 A schematic diagram of an encoding module provided by an embodiment of the present application is shown.

[0026] Figure 2 A working schematic diagram of an encoding module provided by an embodiment of the present application is shown.

[0027] Figure 3 A schematic diagram of a reconfigurable metasurface provided by an embodiment of the present application is shown.

[0028] Figure 4 A schematic diagram of an input port provided by an embodiment of the present application is shown.

[0029] Figure 5 A schematic diagram of an input port provided by an embodiment of the present application is shown.

[0030] Figure 6A schematic diagram of a three-dimensional input port is shown.

[0031] Figure 7 A schematic diagram of a three-dimensional input port is shown.

[0032] Reference signs:

[0033] 100 - encoding module; 101 - plane wave generator; 102 - information encoder; 200 - plane wave; 300 - input port; 301 - incoupler; 302 - outcoupler; 400 - three-dimensional input port; 500 - optical neural network; 600 - three-dimensional optical neural network; 700 - diffractive element. DETAILED DESCRIPTION

[0034] Example implementations are now described with reference to the drawings; however, these descriptions are not intended to limit the scope of the application, but are intended to provide example examples of implementing the application. Examples of the application can be implemented in any number of ways, and are not limited to the examples described herein. It should be noted that the examples described herein can be used in a variety of applications, including but not limited to the examples described herein.

[0035] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more example implementations. In the following description, numerous specific details are provided to give a thorough understanding of example implementations. One skilled in the relevant art will recognize, however, that the techniques described herein can be practiced without one or more of the specific details, or with other methods, components, materials, etc. In other instances, well-known structures, materials, or operations are not shown or described in detail in order to avoid obscuring aspects of the application.

[0036] Optical neural networks utilize photons as information carriers, and achieve the transmission, modulation and processing of optical signals through optical elements such as modulators, optical waveguides, and photodetectors. Optical neural networks combine optical elements and machine learning algorithms, aiming to achieve more efficient computing through the physical properties of light such as high speed, parallelism, and low energy consumption. Optical neural networks have shown potential in various computing tasks, especially in large-scale data processing tasks.

[0037] However, due to current hardware and technological limitations, such as insufficient integration of photonic chips leading to a limited number of parallel channels, and the difficulty in balancing modulation efficiency and bandwidth of existing modulators (such as Mach-Zehnder modulators), which easily introduces losses and errors during modulation, and the lack of optimization of existing deep learning algorithms for the multi-dimensional characteristics of light, resulting in wasted hardware resources, the input information density of current optical neural networks is relatively low. This limits the amount of data that optical neural networks can process per unit time, thus restricting the practical application scope of optical neural networks. As a result, optical neural networks are only suitable for simple optical computing tasks and cannot meet the needs of large-scale optical computing tasks.

[0038] To overcome the aforementioned deficiencies in related technologies, this application provides an encoding module, an input port, and a three-dimensional input port for an optical neural network. Before information encoding, this application converts the light beam into a plane wave, avoiding the impact of the original wavefront curvature of the beam on the accuracy of information encoding, which would otherwise lead to a decrease in the accuracy of the final optical computing task. Furthermore, by using a tunable metasurface as the information encoder, a greater amount of information can be carried per unit area, increasing the input information density of the optical neural network. This allows the optical neural network to process more data streams, thereby enabling it to handle higher-dimensional or more complex data structures, ultimately allowing it to handle large-scale optical computing tasks.

[0039] This application provides an encoding module 100 for optical neural networks, such as... Figure 1 As shown, Figure 1 The diagram illustrates an encoding module 100 according to an embodiment of this application. The encoding module 100 includes a plane wave generator 101 and an information encoder 102, wherein the plane wave generator 101 is disposed upstream of the information encoder 102 along the optical path propagation direction. Figure 1 The arrow in the diagram indicates the direction of light propagation. A plane wave generator is used to generate a received light beam (such as...) Figure 1 The middle arrow pointing to the plane wave generator 101 converts the plane wave into a plane wave 200 and transmits the plane wave 200 to the information encoder 102. After receiving the plane wave 200, the information encoder 102 encodes the plane wave 200. It should be noted that a plane wave is an electromagnetic wave whose equiphase surface is a plane during propagation.

[0040] And in the embodiments of the present application, the information encoder 102 is a reconfigurable metasurface, which can also be referred to as a tunable metasurface. The reconfigurable metasurface refers to a metasurface based on a tunable material, which can change the equivalent refractive index when subjected to external excitation, thereby changing the phase at each position on the reconfigurable metasurface, and realizing dynamic regulation of physical parameters such as the phase and amplitude of light. The information encoder 102 can realize phase regulation of N data points, so as to be able to encode a 1xN size information vector onto the light field of the plane wave 200 and serve as the input of the subsequent optical neural network 500.

[0041] As an example, as shown in Figure 2 , Figure 2 The working schematic diagram of the encoding module 100 provided by an embodiment of the present application is shown, the plane wave generator 101 converts the received light beam into a plane wave 200, the plane wave 200 propagates a certain distance and reaches the information encoder 102, the reconfigurable metasurface corresponding to the information encoder 102 has been regulated according to the externally input information vector of the tunable material at each position thereon, and then the information encoder 102 modulates the plane wave 200 after receiving the plane wave 200 (as shown by φ in Figure 2 , which corresponds to the externally input information vector), loads the information corresponding to the information vector onto the light field of the plane wave 200, and inputs the modulated plane wave 200 into the calculation layer of the optical neural network 500.

[0042] In an embodiment, the calculation layer of the optical neural network 500 is realized in a diffractive manner, that is, the calculation layer of the optical neural network 500 is one or more diffraction layers, and the optical neural network 500 is a diffractive neural network, but it should be noted that this is only used for subsequent description in the present application, and does not mean that the optical neural network 500 of the present application is limited to a diffractive neural network. It should be noted that the size of the effective area of the reconfigurable metasurface is determined by the size of the input information required by the calculation layer of the optical neural network 500, and the effective area refers to the area of the reconfigurable metasurface that can modulate the light beam.

[0043] In an embodiment, the reconfigurable micro-nano unit is composed of a nanostructure of a tunable material and a corresponding substrate, and when the tunable material is subjected to external excitation, the equivalent refractive index of the tunable material will change. As shown in Figure 3 , Figure 3 The schematic diagram of the reconfigurable metasurface provided by an embodiment of the present application is shown, Figure 3The dashed box in the figure represents a reconfigurable micro-nano unit, the dark block in the dashed box represents a nano structure, and the substrate is the adjustable material. In this case, the reconfigurable metasurface is used as the information encoder 102, which can adjust the equivalent refractive index of the adjustable material at different positions on the reconfigurable metasurface in real time through external excitation. The encoding mode can be dynamically changed when processing different optical computing tasks, which improves the flexibility and adaptability of the optical neural network and brings better performance and a wider range of applications to the optical neural network.

[0044] As an example, the adjustable material of the reconfigurable micro-nano unit can be a phase change material. The phase change material changes the crystal lattice inside the material when subjected to external excitation (such as heat, laser, or applied voltage), which can greatly change the dielectric constant of the material. Commonly used phase change materials include GST (germanium antimony tellurium) and V02 (vanadium dioxide).

[0045] Further, in an embodiment, one reconfigurable micro-nano unit on the reconfigurable metasurface constitutes a data point. The data point refers to a minimum information unit in the data set of the information vector that can be processed individually. Each data point performs phase modulation on the light beam received by the corresponding position according to the input information vector, and then loads the information corresponding to the information vector onto the light field of the plane wave. That is, when there are M reconfigurable micro-nano units on the reconfigurable metasurface, the information encoder 102 can realize phase modulation of M data points. In this case, since the size of the reconfigurable micro-nano unit on the reconfigurable metasurface is small, the size of a single reconfigurable micro-nano unit can be controlled within 100-500 nm, and a single reconfigurable micro-nano unit can carry a piece of information as a data point. Therefore, more information can be carried in a limited space, so that the information encoder 102 can maximize the input information density of the optical neural network, thereby maximizing the information processing capacity of the optical neural network, and enabling the optical neural network to meet the demand of large-scale optical computing tasks.

[0046] In another embodiment, at least two reconfigurable micro-nano units on the reconfigurable metasurface constitute a data point, that is, a plurality of reconfigurable micro-nano units on the reconfigurable metasurface constitute a data point as a whole, and then perform phase modulation on the light beam received by the data point. In this case, the information carrying capacity of the reconfigurable metasurface can be ensured while considering the difficulty of dynamic adjustment of the adjustable metasurface, and the input information density of the optical neural network is also greatly improved.

[0047] In an embodiment, the external excitation can be one or more of optical excitation, electrical excitation, thermal excitation, and mechanical excitation.

[0048] Continuing with the above examples, if GST (Gaseous Transformer Stem Cell) is used as the tunable material, the external excitation methods can be thermal or optical. It should be noted that different external excitations have their own advantages. For example, optical excitation eliminates the need for physical contact with the tunable material, avoiding mechanical interference, and the light spot can be focused to the nanometer scale, satisfying the control requirements of small-sized tunable materials. Electrical excitation allows for easy digital adjustment of voltage and current, facilitating relevant switching, and reduces the difficulty of electronic device integration, enabling direct integration with reconfigurable metasurfaces, thereby improving the integration level of optical neural networks. Thermal excitation addresses the fact that most phase change materials are temperature-sensitive, exhibiting strong adaptability, and carries a low risk of irreversible damage to the tunable material, making it suitable for long-term stable operation. Mechanical excitation, triggering phase transitions through mechanical deformation, is unaffected by temperature, laser, or current, making it suitable for complex working environments.

[0049] In one embodiment, considering both the integration level and computational efficiency of the optical neural network, this embodiment preferably uses optical excitation or electrical excitation to control the tunable material. In this case, the tunable material on the information encoder 102 provided in this application can be controlled more quickly, thereby enabling the optical neural network to receive input information and perform subsequent computational tasks more quickly, thus improving the computational efficiency of optical computing tasks.

[0050] In one embodiment, the plane wave generator 102 is composed of at least one metasurface and / or at least one refractive lens. That is, the plane wave generator 102 may be composed of at least one metasurface, at least one refractive lens, or a combination of at least one metasurface and at least one refractive lens.

[0051] like Figure 1 The plane wave generator 102 shown is a plane wave generator 102 composed of a metasurface. In this case, the phase on this metasurface is a collimated phase, thereby enabling collimation of the beam received by this metasurface. An example of a plane wave generator 102 being a refractive lens is not shown in the accompanying drawings of this application.

[0052] It should be noted that, since the size of the encoding module 100 needs to be considered in practical applications, the plane wave generator 102 is preferably composed of a metasurface. A single metasurface can convert the light beam into a plane wave. In this case, the size of the encoding module can be minimized, thereby improving the integration of the optical chip and making it more conducive to the large-scale deployment of the optical chip.

[0053] This application also provides an input port 300 for an optical neural network, such as... Figure 4 As shown, Figure 4An input port 300 provided by an embodiment of the present application is shown in the schematic diagram, which comprises the encoding module 100 as described in any of the above embodiments, an incident coupler 301, and an exit coupler 302.

[0054] Specifically, the incident coupler 301, the exit coupler 302, the plane wave generator 101 and the information encoder 102 of the encoding module 100 are sequentially arranged along the light path propagation direction, the incident coupler 301 is connected with the exit coupler 302 through an optical waveguide, the incident coupler 301 is used to receive the light beam emitted by the light source (as shown by the arrow pointing to the incident coupler 301), the light beam propagates to the exit coupler 302 through the optical waveguide, and the exit coupler 302 emits the light beam to the plane wave generator 102 of the encoding module 100. Figure 4

[0055] It should be noted that the light beam emitted by the light source received by the incident coupler 301 is a converging light beam, and the convergence point of the light beam is located on the incident coupler 301, and the light beam output by the output coupler 302 is a diverging light beam, as shown by the arrow pointing to the exit coupler 302. Figure 5 Figure 5 An input port 300 provided by an embodiment of the present application is shown in the schematic diagram, the light beam emitted by the exit coupler 302 is a diverging light beam, the plane wave generator 101 receives the diverging light beam and modulates it to generate the plane wave 200.

[0056] Further, the incident coupler 301, the exit coupler 302 and the optical waveguide connecting the two in the input port 300 are all materials with high transmittance in the working waveband. For example, when the working center wavelength is 1550nm in the communication waveband, the incident coupler 301, the exit coupler 302 and the optical waveguide connecting the two are polysilicon, and the substrate bearing the three can be quartz glass. It should be noted that this is only an example and does not mean that the working waveband of the input port 300 provided by the present application is limited to this. In fact, it can be one or more of the visible light waveband, the near-infrared waveband and the terahertz waveband, or other working wavebands, which are not limited in the present application.

[0057] The optical neural network 500 in the present application can be regarded as a two-dimensional optical neural network, and a plurality of optical neural networks 500 can constitute a three-dimensional optical neural network 600; the three-dimensional optical neural network can perform a plurality of tasks in parallel. It should be noted that the two-dimensional optical neural network referred to here is mainly a two-dimensional abstract representation of the connection mode between the input port 100 and the calculation layer of the optical neural network 500, and does not mean that the optical signal in the two-dimensional optical neural network propagates in a two-dimensional space; obviously, the optical signal in the two-dimensional optical neural network propagates in a three-dimensional space.

[0058] ​​To this end, the application further provides a three-dimensional input port 400 for a three-dimensional optical neural network, as shown in Figure 6 Figure 6 A schematic diagram of the three-dimensional input port 400 provided by an embodiment of the application is shown, which includes at least two input ports 300 as described in the above embodiments.

[0059] Since the three-dimensional input port 400 is composed of multiple input ports 300, each input port 300 also corresponds to an optical neural network 500 in the three-dimensional optical neural network 600, that is, in the process of actually performing optical computing tasks by the three-dimensional optical neural network 600, each input port 300 in the three-dimensional input port 400 can perform the same / different information encoding processing on the respective corresponding plane wave 200, thereby being able to perform multiple optical computing tasks in parallel.

[0060] As an example, as shown in Figure 6 Figure 6 the respective input ports 300 in the three-dimensional input port 400 are denoted as a first port, a second port and a third port from top to bottom, and the optical neural networks 500 in the three-dimensional optical neural network 600 in Figure 6 are denoted as a first optical neural network, a second optical neural network and a third optical neural network from top to bottom, and the first port, the second port and the third port are matched with the first optical neural network, the second optical neural network and the third optical neural network one by one respectively. The first port is used to perform a first task, the second port is used to perform a second task, and the third port is used to perform a third task, so that in the actual optical computing process, the information of the first port can be encoded by the information encoder in the first port to load the information corresponding to the first task onto the plane wave light field, and the second port and the third port are the same, in this case, the information encoding processing can be performed according to the actual optical computing task requirement for different plane waves, thereby being able to more fully utilize the parallelism characteristic of light, realize parallel processing of the three-dimensional optical neural network, and improve the flexibility and computing efficiency of the three-dimensional optical neural network.

[0061] It should be noted that in this embodiment, each input port in the three-dimensional input port has a one-to-one matched light source, and each light source is used to emit a light beam to the corresponding input port incident coupler; and as with the two-dimensional optical neural network, the light beams emitted by each light source are convergent, and the convergent points are respectively on the corresponding incident couplers 301.

[0062] In an embodiment, the three-dimensional input port further includes a diffraction element 700, as shown in Figure 7 Figure 7 ​​​A schematic diagram of the three-dimensional input port 400 provided by an embodiment of the present application is shown. The diffractive element 700 is used to converge the light beams emitted by the light source onto the incoupling elements 301 of each input port of the three-dimensional input port 400, respectively. In this case, the entire three-dimensional optical neural network 600 can be driven by a single light source, avoiding the complex optical path alignment problem and the synchronous / asynchronous control problem caused by the use of an array of light sources, reducing the complexity of the system, improving the stability and consistency of the system, and also reducing the manufacturing cost of the three-dimensional input port.

[0063] It should be noted that, in an embodiment, the light beams emitted by the light source to the diffractive element 700 are collimated laser beams.

[0064] In an embodiment, the diffractive element 700 can be a metasurface or a diffraction grating. It should be noted that, when a diffraction grating is used as the diffractive element 700, due to the mature process of the diffraction grating, a diffraction grating that meets the beam splitting requirement can be designed and manufactured more conveniently when the beam splitting requirement is low, reducing the manufacturing difficulty of the three-dimensional optical neural network; when a metasurface is used as the diffractive element 700, due to the characteristics of low manufacturing cost, nanoscale volume, and high design freedom of the metasurface, the cost of mass-producing the optical neural network can be reduced, and the volume of the optical neural network can also be reduced while controlling the cost, and the periodicity limitation of the grating can be broken, supporting complex beam splitting modes (such as asymmetric beam splitting, custom energy distribution), thereby improving the flexibility of the arrangement of each optical neural network in the three-dimensional optical neural network, and the integration of the three-dimensional optical neural network can also be improved, further reducing the volume of the three-dimensional optical neural network.

[0065] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

Claims

1. An encoding module for optical neural networks, characterized in that, The encoding module comprises: a plane wave generator; an information encoder; The plane wave generator is arranged upstream of the information encoder along the direction of propagation of the light path; The plane wave generator is configured to convert the received light beam into a plane wave, and the information encoder is configured to encode information on the plane wave; The information encoder is a reconfigurable metasurface, and the reconfigurable metasurface comprises at least two reconfigurable micro-nano units.

2. The coding module of claim 1, wherein, One reconfigurable micro-nano unit on the reconfigurable metasurface constitutes one data point.

3. The coding module of claim 1, wherein, At least two reconfigurable micro-nano units on the reconfigurable metasurface constitute one data point.

4. The coding module of claim 1, wherein, The plane wave generator is composed of at least one metasurface and / or at least one refractive lens.

5. The coding module of claim 1, wherein, The equivalent refractive index of the reconfigurable micro-nano unit changes when subjected to external excitation.

6. The coding module of claim 5, wherein, The external excitation is one or more of optical excitation, electrical excitation, thermal excitation, and mechanical excitation.

7. An input port for an optical neural network, the input port comprising: The input port comprises: an encoding module as claimed in any one of claims 1-6; an incident coupler; an exit coupler; The incident coupler and the exit coupler are connected by an optical waveguide; The incident coupler is configured to receive a light beam emitted by a light source, and the exit coupler is configured to emit the light beam to the encoding module.

8. A three-dimensional input port for a three-dimensional optical neural network, characterized by The three-dimensional input port comprises: at least two input ports as claimed in claim 7.

9. The three-dimensional input port of claim 8, wherein, The three-dimensional input port further comprises: a diffraction element; The diffraction element is configured to converge the light beam emitted by the light source onto the incident coupler of each input port.

10. The three-dimensional input port of claim 9, wherein, The diffraction element is a metasurface or a diffraction grating.

Images