Threshold screening based multi-wavelength channel diffraction neural network system and processing method

Abstract

The application discloses a kind of multi-wavelength channel diffraction neural network systems and processing method based on threshold screening, belong to artificial intelligence and optical computing technical field. Including: encoding module, for the task information corresponding to different to-be-processed tasks is encoded to the light beam of multiple different operating wavelengths of light source module output respectively, forms multiple task channel light field;Optical diffraction module includes multiple modulation layers;Optical diffraction module is used to receive task channel light field, and the diffraction modulation and threshold screening of each task channel light field are realized by changing the super surface structure on different modulation layers;Each task channel light field shares the same set of diffraction modulation parameters, i.e. phase, mask generation and threshold parameter etc.;Threshold screening refers to comparing mask generation parameter with threshold parameter, according to the comparison result, the binary selection of phase parameter is carried out, and the target phase parameter is screened out;Output detection module detects output light intensity.The application expands task channel and improves system integration.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and optical computing technology, specifically relating to a multi-wavelength channel diffraction neural network system and processing method based on threshold screening. Background Technology

[0002] Optical diffraction neural networks (ODNs), as a specific computational model, essentially utilize optical elements to simulate and implement neural network inference functions. From a physical mechanism perspective, this computational model relies on the physical processes such as interference and diffraction that occur when light waves propagate in a medium, thereby achieving parallel matrix operations. Based on this computational mode, ODNs exhibit advantages such as high speed, low energy consumption, and inherent parallelism. Compared to traditional artificial neural networks based on electronic hardware, ODNs manipulate the light field (e.g., phase modulation) through pixel units on different modulation layers. Specifically, they utilize the complex amplitude weight distribution formed during interlayer diffraction propagation to construct an optical physical computational network equivalent to a fully connected neural network; this architecture overcomes the bottlenecks of traditional electronic computing in energy efficiency optimization and parallel processing capabilities.

[0003] Currently, optical diffraction neural networks have demonstrated good performance in single-task processing scenarios such as image classification and target recognition. For example, patent application CN117040623A discloses an angular momentum classification and recognition system based on a diffraction neural network, which constructs an optical diffraction neural network. Meanwhile, metasurfaces, with their subwavelength structural characteristics, can flexibly control multidimensional parameters such as amplitude, phase, and polarization of incident light waves by designing the geometry and spatial arrangement of these nanostructure units. This allows for complex optical field manipulation functions in a compact planar form, thus providing crucial support for the miniaturization and high integration of optical diffraction neural networks.

[0004] While optical diffraction neural networks have made progress in single-task processing, real-world applications are far more complex, often requiring the ability to handle multiple tasks simultaneously. Most current computing systems are constrained by fixed physical structures, lacking flexible dynamic reconfiguration or multi-path parallel computing capabilities when faced with multi-task demands. To address these issues, current research explores three main directions:

[0005] Firstly, some computing systems adopt a mechanical reconfigurable scheme, which achieves task switching by physically replacing optical components. However, this scheme mainly relies on mechanical reconfigurable operations, which require high alignment accuracy and have a complex operation process. Reconfiguration and replacement can also make the system structure more complex.

[0006] Secondly, a polarization multiplexing scheme is adopted, which encodes different tasks into different polarization channels to achieve parallel computing in the same optical system; however, due to the polarization degrees of freedom of light, the number of reusable channels is extremely limited, and there is also crosstalk between polarization channels, which further limits the scalability of the computing system under multi-task channels.

[0007] Thirdly, some studies address the multi-task learning problem of optical diffraction neural networks at the algorithmic level. Specifically, they introduce a flexible weight retention mechanism to avoid catastrophic forgetting, enabling a single optical diffraction neural network to execute multiple tasks sequentially. However, such a computing system can only process a single task at any given time, and each new task requires retraining and updating. The training cost increases linearly with the number of tasks, and the increase in the number of tasks also leads to an exponential increase in training complexity, while the channel expansion capability remains limited.

[0008] Therefore, the bottlenecks encountered by existing optical multitasking systems in terms of channel expansion, system integration, and mechanical reconfiguration have become a major technical challenge in this field. Summary of the Invention

[0009] To address the problems in related technologies, this invention proposes a multi-wavelength channel diffraction neural network system and processing method based on threshold screening, thereby overcoming the aforementioned technical issues in existing related technologies. This invention employs a threshold screening mechanism during the training phase, comparing mask generation parameters with threshold parameters and performing binarization selection of phase parameters based on the mask results. This dynamically filters out phase parameters that play a crucial role in multi-task processing. Furthermore, by utilizing a combination of a wavelength-dimensional multi-task coding structure and a shared multilayer metasurface diffraction network, multi-task optical computation is achieved while suppressing redundant noise information. This expands the number of task channels while maintaining a compact optical path structure, improving system integration, and maintaining high recognition accuracy and result stability.

[0010] The technical solution of the present invention is implemented as follows: a multi-wavelength channel diffraction neural network system based on threshold screening, the system comprising a light source module, an encoding module, an optical diffraction module and an output detection module connected in sequence;

[0011] The light source module is used to generate N beams of light with different operating wavelengths;

[0012] The encoding module is used to encode the task information corresponding to M different tasks to be processed onto the N beams with different working wavelengths, forming a multi-channel input optical field with a one-to-one correspondence between the working wavelength and the task to be processed; the multi-channel input optical field includes N task channel optical fields;

[0013] The optical diffraction module includes a first modulation layer, ..., an Lth modulation layer arranged sequentially along the optical axis, with a metasurface structure on the surface of each modulation layer. The optical diffraction module receives the light fields of the N task channels and achieves diffraction modulation and threshold filtering of the light fields of each task channel by changing the metasurface structure on different modulation layers. And N, M, and L are all positive integers;

[0014] On each modulation layer, the optical field of each task channel shares the same set of diffraction modulation parameters; the diffraction modulation parameters include phase parameters, mask generation parameters, and threshold parameters.

[0015] The threshold screening refers to comparing the mask generation parameters with the threshold parameters, performing binarization selection on the phase parameters based on the comparison results, and finally screening out the target phase parameters.

[0016] The output detection module is used to detect the output light intensity of each task channel after the optical field is modulated by the optical diffraction module, and to obtain the processing results corresponding to each task information.

[0017] Furthermore, the optical diffraction module also includes a free propagation space located between two adjacent modulation layers.

[0018] Furthermore, each of the modulation layers has a modulation plane on the light incident side;

[0019] In the optical diffraction module, at least one of the modulation planes of the modulation layer is provided with a first metasurface structure, and the first metasurface structure includes a plurality of phase modulation units located on the modulation plane;

[0020] On each of the modulation layers other than the Lth modulation layer, continuous and independent spatial phase modulation is applied to the incident light by rotating each of the phase modulation units on the modulation plane;

[0021] Furthermore, on each of the modulation layers other than the Lth modulation layer, continuous and independent spatial phase modulation is applied to the incident light by rotating each of the phase modulation units clockwise or counterclockwise on the modulation plane;

[0022] Furthermore, multiple phase modulation units are uniformly arranged on the corresponding modulation plane.

[0023] Furthermore, in the optical diffraction module, the first metasurface structure is provided on each of the modulation layers except for the Lth modulation layer. The (L-1) modulation layers perform optical field modulation by using continuous phase modulation and do not perform threshold screening operation.

[0024] The modulation plane of the Lth modulation layer is provided with a second metasurface structure, the second metasurface structure including a plurality of threshold filtering units located on the modulation plane; the plurality of threshold filtering units are used to filter out the phase modulation units corresponding to the target phase parameter;

[0025] Furthermore, the phase modulation unit corresponding to the target phase parameter refers to the phase modulation unit that is retained after threshold filtering and that plays a role in the multi-task classification result;

[0026] Furthermore, the phase modulation unit corresponding to the target phase parameter specifically refers to the phase modulation unit that is retained after threshold screening and plays a key role in the multi-task classification result.

[0027] Furthermore, each of the threshold filtering units is a hole structure disposed on the surface of the Lth modulation layer;

[0028] Furthermore, the shape of the hole structure includes, but is not limited to, rectangle, circle, and ellipse; in this embodiment, the hole structure is preferably a nanometer-scale rectangular hole structure.

[0029] Each of the phase modulation units is a bump structure disposed on the surface of the corresponding modulation layer;

[0030] Furthermore, the three-dimensional shape of the bump structure includes, but is not limited to, cuboid, cube, and cylinder; in this embodiment, the bump structure is preferably a cuboid at the nanometer level.

[0031] Furthermore, the light source module includes N independently configured monochromatic light sources for generating N beams of discrete working wavelengths that are different from each other;

[0032] The encoding module includes M wavelength channel encoding units, used to encode the first wavelength channel. The task information for each pending task is loaded into the working wavelength. On the beam of light, to form the carrying of the first Encoding beams of task information to achieve working wavelength With the Each pending task corresponds to one other task; among them... It is a positive integer;

[0033] The output detection module includes multiple detection units.

[0034] An optical multitasking method is applied to a threshold-based multi-wavelength channel diffraction neural network system. The method includes the following steps:

[0035] Step S1: Preconstruct a trainable model containing L modulation layers, each modulation layer having a geometric phase metasurface; each metasurface layer in the trainable model is regarded as a trainable phase layer, and the phase distribution of each phase layer is the parameter to be optimized; set N independent monochromatic light sources in the light source module and generate N beams with different working wavelengths, and then transmit these N beams to the encoding module.

[0036] Step S2: The encoding module uses M wavelength channel encoding units to independently encode the task information of multiple tasks to be processed onto the beams of the corresponding working wavelengths, forming multiple independent encoded beams; then the encoded beams carrying the corresponding task information are transmitted to the optical diffraction module.

[0037] Step S3: The optical diffraction module receives and processes multiple encoded beams from the encoding module. Multiple encoded beams of different wavelengths propagate optically simultaneously in free propagation space and pass through L modulation layers in sequence for diffraction modulation. During the optical propagation process, multiple tasks are calculated. Among them, the first modulation layer to the (L-1)th modulation layer applies continuous and independent phase control to the incident encoded beams by rotating phase modulation units without performing threshold filtering. A threshold filtering unit is set on the Lth modulation layer to dynamically filter out the phase modulation units corresponding to the target phase parameters. Only the effective phase modulation units are retained to participate in subsequent training, thereby determining the final phase distribution of each modulation layer.

[0038] Step S4: After multiple coded beams propagate sequentially through L modulation layers and the free propagation space between the layers, they reach the detection plane of the output detection module through the last free propagation space; the output light intensity distribution of each working wavelength on the output detection plane is detected by multiple detection units to obtain the processing results of each wavelength channel.

[0039] Step S5: By calculating the mean square error between the output light intensity distribution of each working wavelength and the corresponding target light intensity distribution, the task loss of each working wavelength is obtained, and the loss of all task channels is weighted and summed to obtain the composite total loss function;

[0040] Step S6: During the training process, the error backpropagation algorithm is used to calculate the gradient of the composite total loss function with respect to the phase parameters, mask generation parameters and threshold parameters of each modulation layer under all working wavelengths, and the gradient descent algorithm is used to iteratively update these parameters.

[0041] Step S7: By repeating steps S1 to S6, the network learns to transform the light field of each task channel into the desired output light intensity distribution until the network optimization training reaches convergence or reaches the preset number of iterations.

[0042] Step S8: After training, for each target working wavelength, the geometric parameters of the metasurface are scanned and optimized using electromagnetic simulation software to obtain a set of geometric parameters at each working wavelength; based on the simulation results, the broadband half-wave plate is scanned, and the structures of multiple phase modulation units and / or multiple threshold screening units are rotated according to the required phase distribution to determine the overall first metasurface structure and / or second metasurface structure.

[0043] Step S9: Samples of various metasurface structures are prepared using micro-nano fabrication technology, and a multi-wavelength laser source with switchable wavelengths and a high-precision optical field detection system are built. The mission efficiency, wavelength crosstalk and mission fidelity of the samples at different wavelengths are experimentally characterized. The mission fidelity is evaluated by the deviation between the measured focal position and the designed position.

[0044] Further, in step S3, each phase modulation unit generates parameters in the Lth modulation layer using its corresponding mask, and compares them with the corresponding threshold parameters after mapping by a normalization function to obtain a mask with a value of 0 or 1.

[0045] When the mask value is 1, the corresponding phase parameters are retained and participate in phase modulation during training;

[0046] When the mask value is 0, the corresponding phase parameters are suppressed during training, thereby dynamically selecting the phase parameters that play a key role in multi-task processing.

[0047] Furthermore, the normalization function is a Sigmoid function, which is used to map the mask generation parameters to a continuous range of values ​​between 0 and 1, and compares the parameters mapped by the normalization function with the corresponding threshold parameters to determine the mask value;

[0048] When the result after normalization function mapping is greater than or equal to the threshold parameter, the mask at the corresponding position is set to 1; when the result after normalization function mapping is less than the threshold parameter, the mask at the corresponding position is set to 0, thereby achieving binarization selection.

[0049] Further, in step S4, the detection plane of the output detection module is divided into A physical sub-regions, each physical sub-region corresponding to a preset target category; and has B detection units for detecting the output light intensity of N different wavelength channels respectively; for each wavelength channel, the maximum value of its output light intensity in the A physical sub-regions is determined, and the target category corresponding to the physical sub-region where the maximum value is located is output as the prediction result of the wavelength channel; where A and B are both positive integers.

[0050] Further, in step S5, the composite total loss function The task channel loss corresponding to each working wavelength According to weighting coefficients We obtain the result by weighted summation, using the following formula:

[0051] ;

[0052] in, This represents the total number of wavelength channels. Indicates the first The loss of each working wavelength corresponds to the task channel. For the task channel loss The corresponding weighting coefficients; when the importance of each task is equal, Values .

[0053] The beneficial effects of this invention are:

[0054] (1) This invention adopts a multi-wavelength channel diffraction neural network system and processing method based on threshold screening. During the training process, the phase modulation units of the metasurface are dynamically screened, and only the phase modulation units that play a key role in multi-task processing are retained. At the algorithm level, efficient selection of task feature information is achieved, and the final phase distribution of each metasurface structure is determined, thus ensuring the accuracy of multi-task processing.

[0055] (2) This invention employs multiple working wavelengths and encoding modules to load the input information of different tasks onto independent wavelength channels, so that different tasks to be processed correspond one-to-one with different working wavelengths. Furthermore, it utilizes a shared diffraction neural network composed of multi-layer metasurface structures to modulate and propagate all wavelength channels, thereby expanding the number of task channels processed by the system. At the same time, it simplifies the hardware structure and improves the system's integration and reliability.

[0056] (3) This invention optimizes the phase response of a single metasurface to all working wavelengths during the training phase by constructing a multi-task joint optimization framework. In each iteration update, the phase gradient of each metasurface unit under all wavelengths is calculated using the composite total loss function, and these gradient information are integrated to update the unit structure parameters, enabling it to adaptively learn and be compatible with the diffraction transformations of multiple independent tasks. In a multi-task classification scenario based on multiple standard datasets such as MNIST, Fashion-MNIST, and EMNIST, after adopting the threshold filtering method, the classification accuracy of each task channel on the corresponding datasets is 92.11%, 90.28%, 88.48%, 74.07%, 70.81%, 78.59%, 76.88%, 73.84%, 76.31%, and 75.24%, respectively. Compared with the case without the introduction of the threshold filtering mechanism, the recognition performance of most task channels is improved to varying degrees. The processing method in this invention can effectively improve the overall performance and task adaptability of the multi-task optical computing system under dynamic filtering.

[0057] (4) In typical multi-task classification tests, this invention, while ensuring a compact hardware structure, also expands the number of task channels processed by the system, achieving a relatively stable improvement in recognition accuracy, and verifying the effectiveness of the multi-wavelength architecture and threshold screening mechanism in multi-task optical diffraction neural networks; this invention is applicable to low-power, highly integrated multi-task optical intelligent computing and other application scenarios. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the overall structure of the multi-wavelength channel diffraction neural network system based on threshold screening according to the present invention;

[0059] Figure 2 This is a schematic diagram illustrating the principle of multi-task processing in the multi-wavelength channel diffraction neural network system of the present invention.

[0060] Figure 3 This is a schematic diagram of the process flow of an optical multitasking method according to the present invention;

[0061] Figure 4 This is a schematic diagram of the phase modulation unit of the present invention;

[0062] Figure 5 This is a top view of the phase modulation unit of the present invention;

[0063] Figure 6 This is a top view of the modulation plane of the present invention, in which multiple phase modulation units are distributed.

[0064] Figure 7 This is a schematic diagram of the threshold filtering unit of the present invention;

[0065] Figure 8 This is a schematic diagram illustrating the principle of the threshold filtering mechanism of the present invention;

[0066] Figure 9 This is a schematic diagram of the task pattern corresponding to the multiple input wavelengths of the output detection module of the present invention;

[0067] Figure 10 This is a schematic diagram showing the layout of multiple detection units on the detection plane of the output detection module of the present invention;

[0068] Figure 11 This is a task pattern light intensity distribution diagram of the multi-wavelength output of the output detection module of the present invention;

[0069] Figure 12 This is a schematic diagram of the multi-task classification confusion matrix of the present invention under different wavelengths without the use of a threshold screening mechanism;

[0070] Figure 13 This is a schematic diagram of the multi-task classification confusion matrix under the threshold screening mechanism of the present invention at different wavelengths;

[0071] Figure 14 This is a comparison chart showing the multi-task classification accuracy of the present invention at different wavelengths with and without the threshold screening mechanism.

[0072] Marker explanation:

[0073] 100, Light source module; 200, Encoding module; 300, Optical diffraction module; 301, First modulation layer; 302, Second modulation layer; 303, Third modulation layer; 400, Output detection module; 500, Bump structure; 600, Hole structure. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0076] like Figure 1-2 As shown, this embodiment provides a multi-wavelength channel diffraction neural network system based on threshold screening. The system includes a light source module 100, an encoding module 200, an optical diffraction module 300, and an output detection module 400 connected in sequence.

[0077] The light source module 100 includes N independently configured monochromatic light sources for generating N beams of discrete working wavelengths that are different from each other.

[0078] The encoding module 200 is used to encode the task information corresponding to M different tasks to be processed onto the N beams of different working wavelengths, forming a multi-channel input optical field with a one-to-one correspondence between the working wavelength and the task to be processed; the multi-channel input optical field includes N task channel optical fields.

[0079] More specifically, the encoding module 200 includes M wavelength channel encoding units, used to encode the first wavelength channel. The task information for each pending task is loaded into the working wavelength. On the beam of light, to form the carrying of the first Encoding beams of task information to achieve working wavelength With the Each pending task corresponds to one other task; among them... It is a positive integer;

[0080] It should be noted that in this embodiment, the encoding module 200 can use DMD (Digital Micromirror Device) or SLM (Spatial Light Modulator) to encode the task information onto the corresponding working wavelength.

[0081] In this embodiment, M=N=10, L=3;

[0082] The optical diffraction module 300 includes a first modulation layer 301, ..., a third modulation layer 303 arranged sequentially along the optical axis, and each modulation layer has a metasurface structure on its surface; the optical diffraction module 300 is used to receive the optical fields of 10 task channels, and to achieve diffraction modulation and threshold screening of the optical fields of each task channel by changing the metasurface structure on different modulation layers.

[0083] On each modulation layer, the optical field of each task channel shares the same set of diffraction modulation parameters; the diffraction modulation parameters include phase parameters, mask generation parameters, and threshold parameters.

[0084] The threshold screening refers to comparing the mask generation parameters with the threshold parameters, performing binarization selection on the phase parameters based on the comparison results, and finally screening out the target phase parameters.

[0085] Specifically, the optical diffraction module 300 also includes a free propagation space located between two adjacent modulation layers.

[0086] Specifically, each of the modulation layers has a modulation plane on the light incident side;

[0087] In the optical diffraction module 300, at least one of the modulation planes of the modulation layer is provided with a first metasurface structure, and the first metasurface structure includes a plurality of phase modulation units located on the modulation plane;

[0088] On each of the modulation layers other than the third modulation layer 303, continuous and independent spatial phase modulation is applied to the incident light by rotating each of the phase modulation units on the modulation plane;

[0089] Specifically, on each of the modulation layers other than the third modulation layer 303, the incident light is subjected to continuous and independent spatial phase modulation by rotating each of the phase modulation units clockwise or counterclockwise on the modulation plane;

[0090] More specifically, multiple phase modulation units are uniformly arranged on the corresponding modulation plane.

[0091] Specifically, in the optical diffraction module 300, the first metasurface structure is provided on each of the modulation layers except the third modulation layer 303. The first two modulation layers (i.e. the first modulation layer 301 and the second modulation layer 302) perform optical field modulation by using continuous phase modulation without performing threshold screening.

[0092] The modulation plane of the third modulation layer 303 is provided with a second metasurface structure, the second metasurface structure including a plurality of threshold filtering units located on the modulation plane; the plurality of threshold filtering units are used to filter out the phase modulation units corresponding to the target phase parameter;

[0093] Specifically, the phase modulation unit corresponding to the target phase parameter refers to the phase modulation unit that is retained after threshold screening and that plays a role in the multi-task classification result;

[0094] More specifically, the phase modulation unit corresponding to the target phase parameter specifically refers to the phase modulation unit that is retained after threshold screening and plays a key role in the multi-task classification result.

[0095] like Figure 7 As shown, each threshold filtering unit is a hole structure 600 disposed on the modulation plane of the third modulation layer 303;

[0096] More specifically, the shape of the hole structure 600 includes, but is not limited to, rectangle, circle, and ellipse; in this embodiment, the hole structure 600 is preferably a nanometer-scale rectangular hole structure.

[0097] Specifically, each of the phase modulation units is a bump structure 500 disposed on the corresponding modulation plane;

[0098] More specifically, the three-dimensional shape of the bump structure 500 includes, but is not limited to, cuboid, cube, and cylinder; in this embodiment, the bump structure 500 is preferably a cuboid at the nanometer level.

[0099] In this embodiment, as Figure 5-6 As shown, the projections of each phase modulation unit or threshold filtering unit onto the corresponding modulation plane are all rectangles;

[0100] The length extension direction of the rectangle is decomposed into mutually orthogonal x-axis and y-axis, and a triangular coordinate system is established with the optical axis as the z-axis; wherein the modulation plane is parallel to the xoy plane;

[0101] On each modulation layer, the angle between the length extension direction of the rectangle and the x-axis is the rotation angle. The rotation angle is adjusted by rotating the angle of each of the phase modulation units. This allows for continuous and independent spatial phase modulation of the incident light. and satisfy That is, to achieve This allows for phase coverage across a wide range, enabling the targeting of multiple beams at different operating wavelengths without significantly depending on the material's refractive index dispersion. Continuous and independent phase modulation within the range;

[0102] It should be noted that the rotation angle of each phase modulation unit on each modulation plane... They may be the same or different. Furthermore, this embodiment obtains the parameters (such as rotation angle) of each phase modulation unit in each modulation layer through training, allowing the light field to naturally form energy concentrations in certain regions after propagation; the phase of each phase modulation unit is obtained through training and then corresponds to different rotation angles in the modulation layer. That is, the formula ;

[0103] like Figure 4-6 As shown, on the first modulation layer 301, the rotation angle is changed by rotating each of the phase modulation units. By different rotation angles To achieve Phase modulation;

[0104] like Figure 7 As shown, on the third modulation layer 303, the rotation angle can also be changed by rotating the threshold filtering unit on the modulation plane of the third modulation layer. ;

[0105] It should be emphasized that the phase modulation unit can be retained if the beam passes through the threshold filtering unit, and the phase modulation unit should be suppressed if the beam does not pass through.

[0106] like Figure 1-2 As shown, the output detection module 400 uses multiple detection units to detect the output light intensity of each task channel light field after being modulated by the optical diffraction module 300, and obtains the processing results corresponding to each task information.

[0107] like Figure 1-3 As shown, this embodiment also provides an optical multitasking processing method applied to a multi-wavelength channel diffraction neural network system based on threshold screening. The method includes the following steps:

[0108] Step S1: Preconstruct a trainable model containing 3 modulation layers, each modulation layer having a geometric phase metasurface; each metasurface layer in the trainable model is regarded as a trainable phase layer, and the phase distribution of each phase layer is the parameter to be optimized; set 10 independent monochromatic light sources in the light source module 100 and generate 10 beams with different working wavelengths, and then transmit these 10 beams to the encoding module 200;

[0109] Step S2: The encoding module 200 uses 10 wavelength channel encoding units to independently encode the task information of multiple tasks to be processed onto the beams of the corresponding working wavelengths, forming multiple independent encoded beams; then the encoded beams carrying the corresponding task information are transmitted to the optical diffraction module 300.

[0110] Step S3: The optical diffraction module 300 receives and processes multiple coded beams from the encoding module 200. These multiple coded beams of different wavelengths propagate optically simultaneously in free propagation space and sequentially pass through three modulation layers for diffraction modulation. Multiple calculations are performed during the optical propagation process. The first modulation layer 301 to the second modulation layer 302 adjusts the rotation angle using a rotating phase modulation unit. This allows for continuous and independent phase modulation of the incident coded beam without threshold filtering. A threshold filtering unit is set on the third modulation layer 303 to dynamically filter out the phase modulation units corresponding to the target phase parameters, retaining only the effective phase modulation units for subsequent training, thereby determining the final phase distribution of each modulation layer.

[0111] Step S4: After multiple coded beams propagate sequentially through three modulation layers and the free propagation space between the layers, they reach the detection plane of the output detection module 400 through the last free propagation space; the output light intensity distribution of each working wavelength on the output detection plane is detected by multiple detection units to obtain the processing results of each wavelength channel.

[0112] Step S5: By calculating the mean square error between the output light intensity distribution of each working wavelength and the corresponding target light intensity distribution, the task loss of each working wavelength is obtained, and the loss of all task channels is weighted and summed to obtain the composite total loss function;

[0113] Step S6: During the training process, the backpropagation algorithm is used to calculate the gradient of the composite total loss function with respect to the phase parameters, mask generation parameters and threshold parameters of each modulation layer under all working wavelengths, and the gradient descent algorithm is used to iteratively update these parameters; in this embodiment, the initial learning rate of the optimizer is set to 0.001 and the batch size is 64.

[0114] Step S7: By repeating steps S1 to S6, the network learns to transform the light field of each task channel into the desired output light intensity distribution until the network optimization training reaches convergence or reaches the preset number of iterations (epochs).

[0115] Step S8: After training, for each target working wavelength, the geometric parameters of the metasurface are scanned and optimized using electromagnetic simulation software to obtain a set of geometric parameters at each working wavelength; based on the simulation results, the broadband half-wave plate is scanned, and the structures of multiple phase modulation units and / or multiple threshold screening units are rotated according to the required phase distribution to determine the overall first metasurface structure and / or second metasurface structure.

[0116] Step S9: Samples of various metasurface structures are prepared using micro-nano fabrication technology, and a multi-wavelength laser source with switchable wavelengths and a high-precision optical field detection system are built. The mission efficiency, wavelength crosstalk and mission fidelity of the samples at different wavelengths are experimentally characterized. The mission fidelity is evaluated by the deviation between the measured focal position and the designed position.

[0117] In this embodiment, the 10 different working wavelengths in step S1 correspond to 10 task channels; in the electromagnetic spectrum, 10 monochromatic light sources located in the visible light band are preferentially selected, and the working wavelengths of these 10 monochromatic light sources are 450nm, 470nm, 490nm, 510nm, 530nm, 550nm, 570nm, 590nm, 610nm and 630nm respectively.

[0118] In this embodiment, step S2 further includes steps S2-1 and S2-2, as follows:

[0119] Step S2-1: The system is configured to assign different computing tasks or subsets of data to different specific operating wavelengths for processing. In this embodiment, multiple wavelengths are used ( to This example handles multiple image classification tasks. The following task assignments are merely examples and not intended to limit the scope of this embodiment.

[0120] definition Indicates the wavelength assigned The dataset and its corresponding category subsets;

[0121] wavelength and Different subsets of digits are used to process handwritten digit recognition tasks (MNIST, short for Modified National Institute of Standards and Technology database);

[0122] ,

[0123] ;

[0124] wavelength and Different category subsets are used to process clothing image recognition tasks (Fashion-MNIST);

[0125] ,

[0126] ;

[0127] wavelength to Different subsets of letters are used to process character recognition tasks (EMNIST, short for Extended MNIST);

[0128] ,

[0129] ,

[0130] ,

[0131] ,

[0132] ,

[0133] ;

[0134] Step S2-2: To adapt to the input resolution of the system, the original 28×28 pixel input image is first expanded to 70×70 pixels by padding, and then scaled to 200×200 pixels by bilinear interpolation to match the modulation resolution of the modulation layer in the subsequent optical diffraction module 300.

[0135] In this embodiment, step S3 further includes steps S3-1 and S3-11, as follows:

[0136] Step S3-1: As Figure 4-6 As shown, by adjusting the rotation angle of the metasurface This allows for continuous phase modulation of the incident light field. and satisfy ,accomplish Phase coverage of the range; such as Figure 7 As shown, the threshold filtering unit is a nanometer-scale rectangular aperture structure. The threshold filtering unit filters the phase modulation unit by the presence or absence of the rectangular aperture structure (i.e., light beams that pass through the rectangular aperture structure can retain the phase modulation unit, while light beams that do not pass through must suppress the phase modulation unit). Each modulation layer consists of 200×200 diffractive optical units, with a unit period (pixel size) of 1.2μm, corresponding to a single-layer physical size of 0.24mm×0.24mm. The axial spacing between adjacent modulation layers is optimized based on scalar diffraction theory to ensure the coherence of light field propagation and modulation efficiency.

[0137] Step S3-2: Initialize the following three types of trainable parameters for each modulation layer:

[0138] Phase parameters It is used to perform phase modulation on the input features received by the modulation layer. L represents the number of optical modulation layers, and in this embodiment, L is preferably 3.

[0139] Mask generation parameters Used to generate dynamic masks. ;

[0140] Threshold parameter This is used to set a trainable discrimination threshold for the modulation layer. Based on this threshold, the network filters and determines whether to pass the corresponding features to subsequent layers. ;

[0141] For the first (L-1) layers, each layer uses continuous phase modulation to control the light field for all incident wavelengths, without threshold filtering, and the mask generation parameters are... Fixed as an all-one matrix;

[0142] Step S3-3: For the l-th layer in the first (L-1) layers, apply a shared spatial phase modulation to the light field of all incident wavelengths, with the phase modulation function being... The calculation formula is: ;

[0143] Step S3-4: In the first modulation layer, all task channels share the same physical modulation layer, and for the... Input light field of each task channel With phase modulation function respectively Phase modulation operations are performed independently, thus obtaining the phase-modulated optical field of the task channel. The calculation formula is:

[0144] ;

[0145] in, ; This represents the total number of task channels.

[0146] Step S3-5: Since the first (L-1) layer only performs phase modulation, the mask generation parameters are... The output light field of the first modulation layer 301 is always 1. This refers to the phase-modulated optical field, namely:

[0147] ;

[0148] The third modulation layer 303 in the optical diffraction module 300 determines the retained phase modulation units through a threshold screening method, such as... Figure 8 As shown;

[0149] Step S3-6: In the final third modulation layer 303, phase modulation is first applied to the input light field to obtain the modulated light field. The calculation formula is:

[0150] ;

[0151] in, ;

[0152] Step S3-7: Generate mask parameters for the third modulation layer 303 Normalization is performed using the Sigmoid function, compressing its value to the (0,1) interval, denoted as . ;

[0153] Step S3-8: Map the normalized result With threshold parameter By comparing data, a binarization selection operation is performed to obtain the mask. The specific comparison is as follows: ;

[0154] Step S3-9: Apply the obtained mask By applying phase-modulated light fields, dynamic selection of the effective modulation region of the third modulation layer 303 is achieved. After phase modulation and region selection are completed, the third modulation layer 303 is obtained. Output light field of each task channel The calculation formula is:

[0155] ;

[0156] By using the above method, while maintaining the linearity of physical propagation, a threshold constraint is introduced during network training, enabling the network to automatically select regions that contribute significantly to multi-task classification.

[0157] Step S3-10: Free-space propagation between modulation layers is calculated using the angular spectrum method, and its transfer function is... The calculation formula is:

[0158] ;

[0159] In the formula, For the distance of propagation; The operating wavelength; and Spatial frequency;

[0160] Step S3-11: For the first The first task channel, its light field from the first... The modulation layer output surface propagates to the first The process of the modulation layer input surface, It can be represented as:

[0161] ;

[0162] In the formula, For the first Optical field distribution at the output surface of the modulation layer; For the first Optical field distribution at the input surface of the modulation layer; Fourier transform; This is the inverse Fourier transform; For the first Layer and First Axial spacing between layers.

[0163] In this embodiment, step S4 further includes steps S4-1 and S4-3, as follows:

[0164] Step S4-1: The detection plane of the output detection module 400 is divided into A physical sub-regions, each corresponding to a preset target category; and has B detection units for detecting the output light intensity of 10 different wavelength channels respectively; where A and B are both positive integers; the entire output detection module 400 contains a total of In this embodiment, there are one detection unit. The specific distribution of the detection units is as follows: Figure 10 As shown; in this embodiment, the output detection module 400 can use a CMOS camera for detection;

[0165] Step S4-2: The system first calculates Light intensity distribution of each task channel in each physical sub-region on the output detection plane The calculation formula is:

[0166] ;

[0167] in, For the first The wavelength corresponding to each task channel; ; To output the index of the physical sub-region of the detection plane, ; For the first Output light field distribution of each task channel; For the first Light intensity of each task channel;

[0168] Step S4-3: For each task channel, the system independently collects the light intensity response values ​​measured by the corresponding detection unit in each of the A physical sub-regions. By comparing these A light intensity response values, the preset target category corresponding to the physical sub-region with the largest light intensity response value is output as the final prediction result for this task channel; for example, for the first... Each task channel has a corresponding operating wavelength of If measured If the maximum value is found, and the third physical sub-region corresponds to "Category 2", then the system outputs the task channel. The identification result is "Category 2".

[0169] In step S5, after obtaining the output light intensity of each task channel through model forward propagation, the loss is calculated in the training environment according to the corresponding task label, as follows:

[0170] Step S5-1: Calculate the loss for each task channel using the mean squared error loss function. :

[0171] ;

[0172] In the formula, It is the number of samples in the dataset; For the first Predicted light intensity distribution for each task channel; For training time The true target light intensity distribution of each task channel;

[0173] Step S5-2: Sum the loss functions of all task channels according to their weights to obtain the composite total loss function. The calculation formula is:

[0174] ;

[0175] in, This represents the total number of wavelength channels. Indicates the first The loss of each working wavelength corresponds to the task channel. For the task channel loss The corresponding weighting coefficients; when the importance of each task is equal, Values ;

[0176] In this embodiment, to achieve the processing of 10 tasks ( The system calculates the independent loss values ​​for ten task channels, which are denoted as follows: If the weight coefficient of each task is set to 0.1, then the composite total loss function is... The calculation result is:

[0177] ;

[0178] Complete the calculation of the composite total loss function.

[0179] After completing multi-wavelength training and obtaining convergence results, the task output results of each wavelength channel are analyzed; such as... Figure 9-11 As shown, this embodiment loads multiple different operating wavelengths onto a single metasurface structure; Figure 9 The diagram illustrates the task pattern light intensity distribution formed on the output plane when several representative wavelengths (such as 470nm, 490nm, 530nm, and 610nm) are incident, which is used to explain that different wavelength channels correspond to different tasks.

[0180] Figure 10 The specific distribution of the detection units on the detection plane of the output detection module 400 is shown. The detection plane is divided into multiple physical sub-regions, which correspond to different task categories and different wavelength channels by row or column. Each white rectangular area represents an independent detection unit. Multi-wavelength and multi-task readout is achieved through spatial division.

[0181] from Figure 11 It can be seen that each wavelength channel obtains a higher normalized light intensity at its corresponding correct category detection position than at other positions, while the light intensity at non-target category positions is lower; this shows that the multi-wavelength output detection module 400 of this embodiment can spatially separate and read out the results of multiple tasks after a single optical propagation, and still maintain low inter-channel crosstalk and stable category discrimination ability under multi-wavelength conditions.

[0182] like Figure 12-13 As shown, this embodiment compares the confusion matrix results of multi-task classification under two conditions: no threshold filtering and threshold filtering. Figure 12 This is the confusion matrix for each working wavelength corresponding to the task without threshold filtering. Figure 13 The image shows the confusion matrix after threshold filtering under the same structure and training configuration. Compared with the absence of threshold filtering, the diagonal elements of the confusion matrix for each wavelength channel are enhanced and the off-diagonal error elements are reduced after using the threshold filtering method. In particular, the number of misjudgments is reduced in channels with higher task difficulty. This shows that by using the threshold filtering method during the training phase to dynamically filter the phase modulation units of the metasurface, this embodiment can effectively focus on the key phase region.

[0183] like Figure 14As shown, this embodiment presents curves comparing the multi-task classification accuracy under 10 working wavelengths with and without threshold filtering. The dashed square lines correspond to the results after threshold filtering, while the solid circular lines correspond to the results without threshold filtering. At multiple working wavelengths from 450nm to 630nm, the classification accuracy after threshold filtering is higher than that without threshold filtering in all task channels, with classification accuracies of 92.11%, 90.28%, 88.48%, 74.07%, 70.81%, 78.59%, 76.88%, 73.84%, 76.31%, and 75.24% respectively on the corresponding datasets.

[0184] Depend on Figure 13-14 As can be seen, the threshold screening mechanism proposed in this embodiment sets threshold constraints during the network training stage, enabling the metasurface structure to adaptively select phase modulation units that are more beneficial to the task, thereby improving the classification accuracy and result stability of the multi-wavelength channel diffraction neural network system as a whole.

[0185] In this embodiment, addressing the bottlenecks faced by existing optical multitasking systems in channel expansion, system integration, and mechanical reconfiguration, metasurfaces provide a feasible path for hardware integration. This enables optical neural networks to achieve efficient optical field modulation with a compact planar structure, and also possesses the ability to precisely control multi-dimensional optical field parameters such as phase and amplitude, thus supporting the complex optical field modulation required for multitasking computation. Simultaneously, a multitasking optical diffraction neural network utilizing wavelength dimensions as task carriers can process multiple tasks within the same optical system through a shared network architecture, achieving information sharing and collaborative optimization between tasks. Based on the integration of these two technologies, this embodiment also introduces a threshold filtering mechanism to construct a multi-wavelength channel diffraction neural network system based on threshold filtering, and designs an optical multitasking processing method. By dynamically filtering the phase modulation units of the metasurface, only the phase parameters that contribute critically to multitasking computation are retained, thereby effectively expanding the number of task channels while maintaining a compact optical path structure, improving system integration and computational efficiency, and ultimately achieving highly integrated, multi-channel, and low-energy-consumption multitasking optical processing on the same computer system.

[0186] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A multi-wavelength channel diffraction neural network system based on threshold screening, characterized in that, It includes a light source module, an encoding module, an optical diffraction module, and an output detection module connected in sequence; The light source module is used to generate N beams of light with different operating wavelengths; The encoding module is used to encode the task information corresponding to M different tasks to be processed onto the N beams with different working wavelengths, forming a multi-channel input optical field with a one-to-one correspondence between the working wavelength and the task to be processed; the multi-channel input optical field includes N task channel optical fields; The optical diffraction module includes a first modulation layer, ..., an Lth modulation layer arranged sequentially along the optical axis, with a metasurface structure on the surface of each modulation layer. The optical diffraction module receives the light fields of the N task channels and achieves diffraction modulation and threshold filtering of the light fields of each task channel by changing the metasurface structure on different modulation layers. And N, M, and L are all positive integers; On each modulation layer, the optical field of each task channel shares the same set of diffraction modulation parameters; the diffraction modulation parameters include phase parameters, mask generation parameters, and threshold parameters. The threshold screening refers to comparing the mask generation parameters with the threshold parameters, performing binarization selection on the phase parameters based on the comparison results, and finally screening out the target phase parameters. The output detection module is used to detect the output light intensity of each task channel after the optical diffraction module modulates the light field, and to obtain the processing results corresponding to each task information. The optical diffraction module also includes a free propagation space located between two adjacent modulation layers; Each of the modulation layers has a modulation plane on the light incident side; In the optical diffraction module, a first metasurface structure is provided on each modulation layer except for the Lth modulation layer. The first metasurface structure includes a plurality of phase modulation units located on the modulation plane. On each modulation layer except for the Lth modulation layer, the incident light is subjected to continuous and independent spatial phase modulation by rotating each of the phase modulation units on the modulation plane. In this case, the L-1 modulation layers perform optical field modulation by using continuous phase modulation without threshold filtering. The Lth modulation layer is provided with a second metasurface structure, which includes a plurality of threshold filtering units; the plurality of threshold filtering units are used to filter out the phase modulation units corresponding to the target phase parameter.

2. The multi-wavelength channel diffraction neural network system based on threshold screening according to claim 1, characterized in that, Each of the threshold filtering units is a hole structure disposed on the surface of the Lth modulation layer; Each of the phase modulation units is a bump structure disposed on the surface of the corresponding modulation layer.

3. The multi-wavelength channel diffraction neural network system based on threshold screening according to claim 1, characterized in that, The light source module includes N independently configured monochromatic light sources, which are used to generate N beams of light with different discrete working wavelengths. The encoding module includes M wavelength channel encoding units, used to encode the first wavelength channel. The task information for each pending task is loaded into the working wavelength. On the beam of light, to form carrying the first Encoding beams of task information to achieve working wavelength With the Each pending task corresponds to one other task; among them... It is a positive integer; the output detection module includes multiple detection units.

4. An optical multitasking method, characterized in that, The method, applied to a threshold-based multi-wavelength channel diffraction neural network system as described in any one of claims 1 to 3, comprises the following steps: Step S1: Preconstruct a trainable model containing L modulation layers, each modulation layer having a geometric phase metasurface; each metasurface layer in the trainable model is regarded as a trainable phase layer, and the phase distribution of each phase layer is the parameter to be optimized; set N independent monochromatic light sources in the light source module and generate N beams with different working wavelengths, and then transmit these N beams to the encoding module. Step S2: The encoding module uses M wavelength channel encoding units to independently encode the task information of multiple tasks to be processed onto the beams of the corresponding working wavelengths, forming multiple independent encoded beams; then the encoded beams carrying the corresponding task information are transmitted to the optical diffraction module. Step S3: The optical diffraction module receives and processes multiple encoded beams from the encoding module. Multiple encoded beams of different wavelengths propagate optically simultaneously in free propagation space and pass through L modulation layers in sequence for diffraction modulation. During the optical propagation process, multiple tasks are calculated. Among them, the first modulation layer to the (L-1)th modulation layer applies continuous and independent phase control to the incident encoded beams through rotating phase modulation units without threshold screening. A threshold screening unit is set on the Lth modulation layer to dynamically screen out the phase modulation units corresponding to the target phase parameters. Only the effective phase modulation units are retained to participate in subsequent training, thereby determining the final phase distribution of each modulation layer. Step S4: After multiple coded beams propagate sequentially through L modulation layers and the free propagation space between the layers, they reach the detection plane of the output detection module through the last free propagation space; then, multiple detection units are used to detect the output light intensity distribution of each working wavelength on the output detection plane to obtain the processing results of each wavelength channel. Step S5: By calculating the mean square error between the output light intensity distribution of each working wavelength and the corresponding target light intensity distribution, the task loss of each wavelength is obtained, and the loss of all task channels is weighted and summed to obtain the composite total loss function; Step S6: During the training process, the error backpropagation algorithm is used to calculate the gradient of the composite total loss function with respect to the phase parameters, mask generation parameters and threshold parameters of each modulation layer under all working wavelengths, and the gradient descent algorithm is used to iteratively update these parameters. Step S7: By repeating steps S1 to S6, the network learns to transform the light field of each task channel into the desired output light intensity distribution until the network optimization training reaches convergence or reaches the preset number of iterations, and the training is completed.

5. The optical multitasking method according to claim 4, characterized in that, In step S3, each phase modulation unit in the Lth modulation layer uses the mask generation parameters corresponding to the Lth modulation layer, maps them through a normalization function, and compares them with the corresponding threshold parameters to obtain a mask with a value of 0 or 1. When the mask value is 1, the corresponding phase parameters are retained and participate in phase modulation during training; When the mask value is 0, the corresponding phase parameters are suppressed during training.

6. The optical multitasking method according to claim 5, characterized in that, The normalization function is the Sigmoid function, which is used to map the mask generation parameters to a continuous range of values ​​between 0 and 1, and compares the parameters mapped by the normalization function with the corresponding threshold parameters to determine the mask values. When the result after normalization function mapping is greater than or equal to the threshold parameter, the mask at the corresponding position is set to 1; when the result after normalization function mapping is less than the threshold parameter, the mask at the corresponding position is set to 0, thereby achieving binarization selection.

7. The optical multitasking method according to claim 4, characterized in that, In step S4, the detection plane of the output detection module is divided into A physical sub-regions, each physical sub-region corresponding to a preset target category; and has B detection units for detecting the output light intensity of N different wavelength channels respectively. For each wavelength channel, the maximum value of its output light intensity in A physical sub-regions is determined, and the target category corresponding to the physical sub-region where the maximum value is located is used as the prediction result of that wavelength channel and output; where A and B are both positive integers.

8. The optical multitasking method according to claim 4, characterized in that, In step S5, the composite total loss function The task channel loss corresponding to each working wavelength According to weighting coefficients We obtain the result by weighted summation, using the following formula: ； in, This represents the total number of wavelength channels. Indicates the first The loss of each working wavelength corresponds to the task channel. For the task channel loss The corresponding weighting coefficients.

Images