Short-circuit link based metasurface optical neural network and parameter design, configuration method

Abstract

The application discloses a kind of short-circuit link-based metasurface light neural network and parameter design, configuration method, the parameter design method includes: receiving detection image and corresponding light intensity distribution target configuration;Corresponding to the light field simulation model of metasurface is constructed, according to the light field simulation model, the light intensity distribution data of the detection image and the light intensity distribution target configuration, determine loss function;Iterate and adjust the anti-mask film configuration parameter and target microstructure parameter of the metasurface, until the loss function converges, determine target anti-mask film parameter and target microstructure parameter.This method uses end-to-end design, by adjusting the element geometry parameter and anti-mask film area distribution until the loss function converges, to determine the optimal mask film area distribution and element geometry parameter.Enhances the expression ability and robustness to noise of metasurface light neural network, improves its optical computing precision and accuracy, while reducing process cost.

Description

Technical Field

[0001] This invention relates to the field of optical design technology, and in particular to a metasurface optical neural network based on short-circuit links and a method for parameter design and configuration. Background Technology

[0002] Metasurfaces have a wide range of applications. They utilize optical principles and material properties to modulate, reflect, transmit, and split light waves. Common applications include metasurface lenses. In the use of metasurface lenses, by designing the shape, size, and arrangement of metasurface elements, it is possible to more effectively encode the intensity of incident light and control the light field, leading to entirely new physical phenomena and application scenarios, and making integrated hardware and software programming possible.

[0003] However, as the layer structure of a metasurface lens gradually increases, its optical computational power tends to saturate, resulting in layer saturation and / or unit saturation. In other words, increasing the number of metasurface layers and / or increasing the number of microstructural units on a particular metasurface layer does not effectively improve the optical expressive power of the metasurface lens. Summary of the Invention

[0004] One of the objectives of this invention is to provide a method for designing parameters of a metasurface optical neural network based on short-circuit connections, so as to solve the technical problem that metasurfaces in the prior art cannot simultaneously take into account the number of optical layers and optical computing power.

[0005] One of the objectives of this invention is to provide a method for configuring metasurface optical neural networks based on short-circuit links.

[0006] One of the objectives of this invention is to provide a metasurface optical neural network based on short-circuit connections.

[0007] To achieve one of the aforementioned objectives, this invention provides a method for designing parameters of a metasurface optical neural network based on short-circuit links, comprising: receiving a detection image and a corresponding target configuration of light intensity distribution; constructing a light field simulation model corresponding to the metasurface; determining a loss function based on the light intensity distribution data of the detection image and the target configuration of light intensity distribution according to the light field simulation model; iterating and adjusting the anti-mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determining the target anti-mask parameters and target microstructure parameters; wherein, the anti-mask configuration parameters characterize the configuration of the anti-mask region on the metasurface; the target microstructure parameters represent the geometric parameters constituting the microstructure unit on the metasurface; the anti-mask region allows incident light to directly pass through the metasurface to achieve short-circuit links on the metasurface.

[0008] As a further improvement of one embodiment of the present invention, the metasurface includes a plurality of metasurface layers; the "constructing an optical field simulation model corresponding to the metasurface" specifically includes: constructing an optical field simulation model corresponding to the plurality of metasurface layers; the "iteratio and adjustment of the anti-mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determining the target anti-mask parameters and target microstructure parameters" specifically includes: iterating and adjusting the anti-mask configuration parameters and target microstructure configuration parameters of the first metasurface layer until the loss function corresponding to the first metasurface layer converges, and determining the target anti-mask parameters and target microstructure parameters corresponding to the first metasurface layer; iterating and adjusting the anti-mask configuration parameters and microstructure configuration parameters of other metasurface layers one by one, and determining a plurality of target anti-mask parameters and a plurality of target microstructure parameters respectively corresponding to other metasurface layers.

[0009] As a further improvement of one embodiment of the present invention, the target mask parameters include position configuration parameters, quantity configuration parameters, and area configuration parameters; the target microstructure parameters include position configuration parameters, quantity configuration parameters, and size configuration parameters.

[0010] As a further improvement of one embodiment of the present invention, the number of anti-masking regions configured on each metasurface layer of the metasurface does not exceed one.

[0011] As a further improvement of one embodiment of the present invention, the "construction of an optical field simulation model corresponding to the metasurface" specifically includes: determining the initial configuration parameters of the microstructure units on the metasurface using a random initialization method or a Gaussian distribution method; and constructing an optical field simulation model corresponding to the metasurface based on the initial configuration parameters.

[0012] As a further improvement to one embodiment of the present invention, the step of "constructing a light field simulation model corresponding to the metasurface, and determining a loss function based on the light intensity distribution data of the detected image and the light intensity distribution target configuration according to the light field simulation model" specifically includes: constructing a light field simulation model corresponding to the metasurface; performing traversal calculations based on the detected image to obtain the light intensity distribution data; calculating light intensity distribution target data based on the light intensity distribution data under the constraint of the light intensity distribution target configuration; and determining the loss function based on the difference between the light intensity distribution target data and the light intensity distribution data.

[0013] As a further improvement to one embodiment of the present invention, the step of "constructing a light field simulation model corresponding to the metasurface and performing traversal calculations based on the detected image to obtain the light intensity distribution data" specifically includes: constructing a light field simulation model equipped with the time-domain finite difference method and the Rayleigh-Sommerfeld formula; obtaining a light field parameter matrix by parsing the detected image; wherein each element in the light field parameter matrix points to the light field intensity value of each pixel in the detected image; using the light field parameter matrix as the input of the light field simulation model, first substituting its corresponding light intensity data into the Rayleigh-Sommerfeld formula to obtain the distribution of the light intensity data on the metasurface, and then performing the time-domain finite difference method to substitute the light intensity data near the metasurface into the Rayleigh-Sommerfeld formula, and traversally calculating to obtain the light intensity distribution data; wherein the near light intensity data characterizes the light field intensity before photoelectric conversion, and the light intensity distribution data characterizes the light field intensity after photoelectric conversion.

[0014] As a further improvement of one embodiment of the present invention, the step of "calculating light intensity distribution target data based on the light intensity distribution data under the constraint of the light intensity distribution target configuration" specifically includes: parsing the light intensity distribution target configuration to obtain at least a target high light intensity region; configuring at least the part of the light intensity distribution data corresponding to the target high light intensity region as high light intensity, and updating the light intensity distribution data to obtain the light intensity distribution target data.

[0015] As a further improvement of one embodiment of the present invention, the step of "analyzing the light intensity distribution target configuration to obtain at least a target high light intensity region" specifically includes: parsing the light intensity distribution target configuration to obtain a target high light intensity region and a target low light intensity region; the step of "configuring at least the portion of the light intensity distribution data corresponding to the target high light intensity region as high light intensity, and updating the light intensity distribution data to obtain the light intensity distribution target data" specifically includes: configuring the portion of the light intensity distribution data corresponding to the target high light intensity region as a first light field intensity, configuring the portion of the light intensity distribution data corresponding to the target low light intensity region as a second light field intensity; and updating the light intensity distribution data to obtain the light intensity distribution target data.

[0016] As a further improvement of one embodiment of the present invention, the light intensity distribution data is formed into a light intensity distribution matrix; the light intensity distribution target data is formed into a light intensity distribution target matrix; the step of "determining the loss function based on the light intensity distribution data of the detected image and the light intensity distribution target data corresponding to the detected image according to the neural network model" specifically includes: subtracting the elements at corresponding positions in the light intensity distribution matrix and the light intensity distribution target matrix to obtain an error matrix; calculating the square of each element in the error matrix in turn to obtain a squared error matrix; and calculating and determining the loss function based on the average value of all elements in the squared error matrix.

[0017] As a further improvement to one embodiment of the present invention, the step of "iteratio and adjustment of the anti-mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determination of the target anti-mask parameters and target microstructure parameters" specifically includes: calculating the gradient vectors of the anti-mask configuration parameters and the microstructure configuration parameters in the light field simulation model using a backpropagation algorithm based on the loss function; wherein the gradient vectors characterize the changes of the light field simulation model parameters relative to the loss function; inputting the gradient vectors into an optimization algorithm to calculate the updated values ​​of the light field simulation model parameters, and updating the anti-mask configuration parameters and target microstructure parameters; wherein the optimization algorithm includes at least one of stochastic gradient descent and the Adam algorithm; iteratively executing the backpropagation algorithm and the optimization algorithm until the loss function converges, and determining the target anti-mask parameters and the target microstructure parameters.

[0018] To achieve one of the above-mentioned objectives, an embodiment of the present invention provides a method for configuring a metasurface optical neural network based on short-circuit links, comprising: performing the parameter design method for a metasurface optical neural network based on short-circuit links as described in any one of the above-mentioned methods to obtain a first target design parameter and a second target design parameter; wherein the first target design parameter includes a target anti-mask parameter, and the second target design parameter includes a target microstructure parameter; and configuring a first metasurface according to the first target design parameter and the second target design parameter.

[0019] To achieve one of the above-mentioned objectives, an embodiment of the present invention provides a metasurface optical neural network based on short-circuit links, comprising: performing the parameter design method for a metasurface optical neural network based on short-circuit links as described in any one of the above-mentioned methods to obtain a third target design parameter and a fourth target design parameter, and configuring a second metasurface according to the third target design parameter and the fourth target design parameter; the second metasurface includes a substrate and microstructure units, the microstructure units including a plurality of primitives.

[0020] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0021] This invention employs a metasurface optical neural network parameter design method based on short-circuit links. It constructs an end-to-end optical field simulation model corresponding to the metasurface and defines a corresponding loss function. The model parameters are iteratively adjusted until the loss function converges to determine the optimal target mask parameters and target microstructure parameters (i.e., the geometric dimensions of the metasurface primitives). This method eliminates the need for manual trial and error, automatically obtaining the optimal design parameters. It can generate mask configuration parameters and target primitive sizes that meet the design objectives, reducing design costs, saving time, and exhibiting good versatility.

[0022] Furthermore, this method introduces an adaptively configured anti-masking region distribution on several metasurface layers, allowing incident light to directly pass through the metasurface with the anti-masking regions, isomorphic to the residual connections of an artificial neural network. That is, by configuring the metasurface according to the target anti-masking parameters, some of the original optical data can be transmitted through the metasurface. This not only simplifies the calculation process but also enhances the expressive power and robustness to noise of the metasurface-based optical field simulation model by preserving the original optical data, thereby improving its optical calculation accuracy and precision. This approach is beneficial for metasurfaces to balance the number of optical layers and optical calculation capabilities. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the steps of a metasurface optical neural network parameter design method based on short-circuit links in one embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the steps of the metasurface optical neural network parameter design method based on short-circuit links in the first embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the steps of the metasurface optical neural network parameter design method based on short-circuit links in the second embodiment of the present invention.

[0026] Figure 4 This is a detailed schematic diagram of step S211 of the metasurface optical neural network parameter design method based on short-circuit links in one embodiment of the present invention.

[0027] Figure 5 This is a detailed schematic diagram of step S22 of the metasurface optical neural network parameter design method based on short-circuit links in one embodiment of the present invention.

[0028] Figure 6 This is a detailed schematic diagram of step S221 of the metasurface optical neural network parameter design method based on short-circuit links in one embodiment of the present invention.

[0029] Figure 7This is a schematic diagram of some steps in the parameter design method for metasurface optical neural networks based on short-circuit links in the third embodiment of the present invention.

[0030] Figure 8 This is a detailed schematic diagram of step S3 in the method for designing parameters of a metasurface optical neural network based on short-circuit links according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0032] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Furthermore, the terms "first," "second," "third," "fourth," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0033] Traditional optical medium lenses are often designed in a fixed and constrained manner, making it difficult to meet the needs of different application scenarios. However, by constructing an optical field simulation model, at least by adjusting or configuring the masking parameters of the metasurface, the target imaging region can achieve the target optical field distribution. This high degree of adjustability and reconfigurability provides great freedom in the manufacturing process, thereby enabling the application of the nonlinear characteristics of the metasurface. Therefore, such metasurfaces are more competitive in terms of improving computational accuracy and metasurface hierarchical structure.

[0034] Based on this, the present invention provides a method for designing parameters of a metasurface optical neural network based on short-circuit links, such as... Figure 1 As shown, the specific steps include the following:

[0035] Step S1: Receive the detected image and the corresponding light intensity distribution target configuration;

[0036] Step S2: Construct a light field simulation model corresponding to the metasurface, and determine the loss function based on the light intensity distribution data of the detected image and the light intensity distribution target configuration according to the light field simulation model;

[0037] Step S3: Iterate and adjust the anti-mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determine the target anti-mask parameters and target microstructure parameters.

[0038] Thus, by constructing an optical field simulation model and adjusting the masking configuration parameters and target microstructure parameters of the metasurface, the optimal target masking parameters and target microstructure parameters can be determined. This method can efficiently generate masking configuration parameters and metasurface primitive geometry dimensions that meet the design objectives, reducing design costs and demonstrating good universality. Furthermore, configuring the metasurface according to the target masking parameters allows some of the original optical data to pass through the metasurface, which not only simplifies the calculation process but also enhances the metasurface's expressive power and robustness to noise by preserving the original optical data, thereby improving computational accuracy and precision.

[0039] The anti-masking configuration parameters characterize the configuration of the anti-masking region on the metasurface. The anti-masking region allows incident light to pass directly through the metasurface, thus achieving a short-circuit connection. In other words, on the metasurface layer covered with the anti-masking region, incident light can directly and freely pass through the metasurface layer or metasurface lens with the anti-masking region configured therein. There are no micro / nano units within the anti-masking region, meaning that the incident light in this region is not modulated by the units on the metasurface.

[0040] Furthermore, the light intensity distribution target configuration mentioned in step S1 can be understood as a target requirement or a set configuration for the light intensity distribution of the detected image. Specifically, in one embodiment, the light intensity distribution target configuration may include the light intensity distribution of the target imaging area; in another embodiment, the light intensity distribution target configuration may include several constraints that the light intensity distribution of the imaging area needs to meet; in yet another embodiment, the light intensity distribution target configuration may include the relationship between the light intensity distribution of the target imaging area and other imaging areas. In a preferred embodiment, any two or three of the above embodiments may be used in combination, without specific limitations. Of course, the present invention does not exclude other target requirements or configurations, and can be freely adjusted according to the actual application scenario.

[0041] The metasurface can be a metasurface lens, or other metasurface optical devices. For example... Figure 2 As shown, in the first embodiment, the metasurface includes several metasurface layers, and the metasurface optical neural network parameter design method based on short-circuit links may specifically include the following steps:

[0042] Step S1: Receive the detected image and the corresponding light intensity distribution target configuration;

[0043] Step S2': Construct a light field simulation model corresponding to several metasurface layers, and determine the loss function based on the light intensity distribution data of the detected image and the light intensity distribution target configuration according to the light field simulation model;

[0044] Step S311: Iterate and adjust the anti-mask configuration parameters and microstructure configuration parameters of the first metasurface layer until the loss function corresponding to the first metasurface layer converges, and determine the target anti-mask parameters and target microstructure parameters corresponding to the first metasurface layer.

[0045] Step S312: Iterate through and adjust the anti-mask configuration parameters and microstructure configuration parameters of other metasurface layers one by one, and determine a number of target anti-mask parameters and a number of target microstructure parameters corresponding to other metasurface layers respectively.

[0046] Thus, by performing model calculations on the anti-masking configuration parameters of several metasurface layers, several target anti-masking parameters and several target microstructure parameters are obtained. This method has stronger generalization ability and higher computational accuracy, which is beneficial to improving the fault tolerance of metasurfaces.

[0047] It should be noted that the first metasurface layer does not necessarily refer to the first metasurface layer, but can refer to any one of the plurality of metasurface layers. When the metasurface includes a plurality of metasurface layers, the configuration parameters of the masking layer and the target microstructure parameters on the plurality of metasurface layers can be adjusted simultaneously or sequentially, and the present invention does not impose specific limitations on this.

[0048] The target anti-mask parameters include location configuration parameters, quantity configuration parameters, and area configuration parameters. The target microstructure parameters include location configuration parameters, quantity configuration parameters, and size configuration parameters. The quantity configuration parameter can be understood as whether or not the anti-mask region is configured on several metasurface layers. In this invention, the number of anti-mask regions configured on each metasurface layer does not exceed one. In other words, each metasurface layer may or may not have an anti-mask region configured.

[0049] In this way, by determining the parameters of various types of target masking through simulation calculations, the characteristics of the masking region can be described more comprehensively, and the parameters can form a constraint relationship. Cross-validation can also improve the accuracy of parameter simulation.

[0050] Understandably, the initial parameters of the metasurface can be obtained in various ways before constructing the optical field simulation model corresponding to the metasurface. Therefore, the "constructing the optical field simulation model corresponding to the metasurface" part in step S2 can specifically include:

[0051] Step P2: Determine the initial configuration parameters of the microstructure units on the metasurface using a random initialization method or a Gaussian distribution method;

[0052] Step S2': Based on the initial configuration parameters, construct a light field simulation model corresponding to the metasurface.

[0053] Thus, determining the initial configuration parameters of the light field simulation model through random initialization does not require manual specification and is easy to implement; using the Gaussian distribution method to determine the initial configuration parameters of the light field simulation model allows the initial parameters to be distributed around a certain expected target value, which can accelerate the optimization process.

[0054] The random initialization method mentioned here refers to using a random method to generate initial values ​​for model parameters in machine learning and optimization algorithms, rather than specifying initial values ​​for the parameters. The Gaussian distribution method can be understood as making the initial parameters follow a Gaussian distribution (also called a normal distribution). The specific choice between the two initialization methods can be made freely according to the actual application scenario, without specific limitations. Of course, this invention does not exclude other methods for initializing configuration parameters.

[0055] Furthermore, the initial configuration parameters of the microstructure units represent the geometric parameters of the microstructure units constituting the metasurface, specifically including at least one of the following: the radius of the unit on the metasurface layer, the height of the unit, the geometric shape of the unit (e.g., cylinder or cuboid), the positional distribution of the unit, and material parameters. Additionally, step S2' can be understood as a derivative step of step S2.

[0056] Combination Figure 1 and Figure 3 As shown, in the second embodiment, the method for designing parameters of a metasurface optical neural network based on short-circuit links may specifically include the following steps:

[0057] Step S211: Construct a light field simulation model corresponding to the metasurface, and perform traversal calculations based on the detected image to obtain the light intensity distribution inference data;

[0058] Step S212: Under the constraint of the light intensity distribution target configuration, calculate the light intensity distribution target data based on the light intensity distribution data;

[0059] Step S213: Determine the loss function based on the difference between the target light intensity distribution data and the light intensity distribution data.

[0060] Thus, by constructing a light field simulation model and calculating the light intensity distribution data and the target light intensity distribution data, the loss function is determined. This facilitates subsequent iterative updates to the metasurface parameters, allowing the light intensity distribution data to gradually approach the target light intensity distribution result, achieving automatic parameter tuning, high automation, fast calculation speed, and low computational cost.

[0061] Based on the above description of the target microstructure parameters and target mask parameters, the configuration of the light intensity distribution target can also consider the configuration of metasurfaces or microstructures on metasurfaces, such as metasurface materials, the radius of the basic units in the microstructure, and the height of the basic units. Under these constraints, the light intensity distribution target data is calculated based on the light intensity distribution data.

[0062] Furthermore, combined Figure 3 and Figure 4 As shown, for step S211, the present invention provides a refined step, which may specifically include the following steps:

[0063] Step S2111: Construct a light field simulation model equipped with the time-domain finite difference method and the Rayleigh-Sommerfeld formula;

[0064] Step S2112: Based on the detected image, the light field source parameter matrix is ​​obtained by analysis;

[0065] Step S2113: Using the light source field parameter matrix as the input of the light field simulation model, firstly, substitute the corresponding light intensity data into the Rayleigh-Sommerfeld formula to obtain the distribution of the light intensity data on the metasurface. Then, execute the finite-difference time-domain method to substitute the light intensity data near the metasurface into the Rayleigh-Sommerfeld formula and iterate through the calculations to obtain the light intensity distribution data.

[0066] Thus, by constructing a light field simulation model that incorporates both the time-domain finite difference method and the Rayleigh-Sommerfeld formula, the advantages of both algorithms can be combined to improve the speed and accuracy of light field analysis and calculation, providing reliable light intensity distribution data support for subsequent work.

[0067] In the calculation process of the light field simulation model, simulated data of the light source is required. In this invention, a light field parameter matrix can be obtained by parsing the detected image. Each element in the light field parameter matrix points to the light field intensity value of each pixel in the detected image. The nearby light intensity data represents the light field intensity before photoelectric conversion, and the light intensity distribution data represents the light field intensity after photoelectric conversion.

[0068] In other words, the input light intensity data is substituted into the Rayleigh-Sommerfeld formula to calculate the light field intensity before the light intensity data is processed by the metasurface; after modulation by the metasurface, the light intensity data near the metasurface is obtained (i.e., the light field intensity before photoelectric conversion); finally, the finite-difference time-domain method is executed, and the light intensity data near the metasurface is substituted into the Rayleigh-Sommerfeld formula to calculate the light field intensity on CMOS (Complementary Metal Oxide Semiconductor), thus obtaining the light intensity distribution data.

[0069] Specifically, the aforementioned Finite-Difference Time-Domain Method (FDTD) is a numerical analytical method. It primarily discretizes the spatial region, dividing a continuous space into small grid cells, each called a grid point; then, by discretizing the time region, it calculates the electromagnetic field value at each grid point within each time interval, iteratively updating the electromagnetic field at each grid point to simulate the propagation of incident light throughout space. In this invention, this method can simulate the light propagation and distribution on metasurfaces or their attachments, and it can quickly determine the near-field light intensity.

[0070] The Rayleigh-Sommerfeld formula can be understood as a formula describing the propagation of light, which gives the relationship between the light received at an observation point and the incident light source. In this invention, there may be a certain distance difference between the metasurface or metasurface layer and the CMOS. By using the Rayleigh-Sommerfeld formula, the energy attenuation and phase change caused by the distance can be taken into account. This facilitates accurate reasoning or prediction of the light field distribution of the detected image on the CMOS imaging plane, and it can quickly and accurately determine the light field intensity in the far field.

[0071] Furthermore, the terms "before photoelectric conversion" and "after photoelectric conversion" can be understood as the process where a light signal is generated by a primitive on a simulated metasurface after being exposed to light, and then the photoelectric conversion is completed by CMOS. There is no logical sequence between steps S2111 and S2112, and changing their execution order will not affect the technical effect achieved by the solution.

[0072] Furthermore, such as Figure 5 As shown, for step S212, the present invention provides a refined step, which may specifically include the following steps:

[0073] Step S2121: Analyze the light intensity distribution target configuration to obtain at least the target high light intensity area;

[0074] Step S2122: Configure at least the portion of the light intensity distribution data corresponding to the target high light intensity region as high light intensity, and update the light intensity distribution data to obtain the light intensity distribution target data.

[0075] Thus, based on the target configuration of the light intensity distribution, the location of the target high light intensity area in the light intensity distribution data is determined, so as to make targeted adjustments to the light field intensity in the future, avoid processing the entire light intensity distribution data, and improve processing efficiency.

[0076] Furthermore, such as Figure 6 As shown, in one embodiment, step S2121 may specifically include:

[0077] Step S2121': Analyze the target configuration of the light intensity distribution to obtain the target high light intensity area and the target low light intensity area.

[0078] Based on this, step S2122 can specifically include the following steps:

[0079] Step S21221: Configure the portion of the light intensity distribution data corresponding to the high light intensity region of the target as the first light field intensity, and configure the portion of the light intensity distribution data corresponding to the low light intensity region of the target as the second light field intensity;

[0080] Step S21222: Update the light intensity distribution data to obtain the light intensity distribution target data.

[0081] Thus, by analyzing the target configuration of the light intensity distribution, the different light intensity distributions required for different target areas can be further clarified, and optimization can be achieved by configuring different light field intensities. This method realizes customized and automatic optimization of the light field intensity distribution, with strong adaptability and a high degree of intelligence.

[0082] In this invention, step S2121' can be understood as a derivative step of step S2121, and steps S21221 to S21222 can be understood as derivative steps of step S2122. The target high-intensity region and the target low-intensity region can include several instances, and the first and second light field intensities can also include several instances. In this invention, the objective of the metasurface optical neural network parameter design method based on short-circuit links includes configuring the light field intensity of the target high-intensity region to be greater than the light field intensity of other target low-intensity regions, that is, maximizing the light field intensity of the target high-intensity region.

[0083] Combination Figure 1 and Figure 7 As shown, the present invention provides a third embodiment, wherein the light intensity distribution data is formed into a light intensity distribution matrix; the light intensity distribution target data is formed into a light intensity distribution target matrix; based on this, the present invention provides a refined step for step S2, which may specifically include:

[0084] Step S221: Subtract the elements at corresponding positions in the light intensity distribution matrix and the light intensity distribution target matrix to obtain the error matrix;

[0085] Step S222: Calculate the square of each element in the error matrix in sequence to obtain the squared error matrix;

[0086] Step S223: Calculate and determine the loss function based on the average value of all elements in the squared error matrix.

[0087] Thus, by employing the mean square error, which is simple and easy to understand, the update direction of the light field simulation model parameters can be clearly guided, with the aim of minimizing the mean square error. This optimizes the model parameters, making the calculation results as close as possible to the true light field intensity distribution, resulting in high reliability.

[0088] The mean squared error (MSE) is a metric used by users to evaluate the prediction quality of a model. It reflects the degree of deviation between the predicted and actual values. A smaller MSE indicates a closer approximation between the predicted and actual values, resulting in better model prediction and higher accuracy. Furthermore, squaring the error between the predicted and actual values ​​allows samples with larger errors to receive higher weights, facilitating the identification of parameters with significant errors and improving the sensitivity of the model's simulation process. Of course, this invention does not preclude other calculation methods for the loss function.

[0089] It should be emphasized that calculating the loss function once or a finite number of times does not necessarily make the loss function reach its minimum or converge. Therefore, the loss function obtained in each round of calculation can be used as the input parameter for the next round of calculation to gradually reduce the value of the loss function.

[0090] Based on this, combined Figure 1 and Figure 8 As shown, in the fourth embodiment, step S3 may specifically include the following steps:

[0091] Step S321: Based on the loss function, the gradient vectors of the anti-mask configuration parameters and the microstructure configuration parameters in the light field simulation model are calculated using the backpropagation algorithm.

[0092] Step S322: Input the gradient vector into the optimization algorithm to calculate the updated values ​​of the light field simulation model parameters, and update the anti-mask configuration parameters and the microstructure configuration parameters;

[0093] Step S323: Iteratively execute the backpropagation algorithm and the optimization algorithm until the loss function converges, and determine the target anti-mask parameters and the target microstructure parameters.

[0094] Thus, by employing backpropagation and optimization algorithms, the convergence speed of the loss function can be accelerated, allowing the loss function to continuously decrease, reducing the error generated by the light field simulation model, and improving the accuracy of the fitting parameters.

[0095] The gradient vector represents the change of the light field simulation model parameters relative to the loss function; the optimization algorithm includes at least one of stochastic gradient descent and the Adam algorithm. The backpropagation algorithm is a learning algorithm that can be used in multilayer neural networks. This algorithm calculates the gradient of the loss function for each known output value obtained from the input value, and this gradient is fed back to the optimization algorithm to update the weights to minimize the loss function.

[0096] The stochastic gradient descent algorithm is an optimization algorithm that updates weight parameters by calculating the average of the error gradient. Specifically, in each iteration, the stochastic gradient descent algorithm randomly selects a set of sample parameters, calculates the gradient of the loss function for that sample, and then updates the weight parameters along the gradient direction. The Adam algorithm is an adaptive learning rate optimization algorithm. Specifically, in each iteration, it calculates the average and variance of the gradient, and then uses these values ​​as a basis to update the learning rate of each weight parameter.

[0097] In practical applications, the target design parameters obtained by the above-mentioned metasurface optical neural network parameter design method based on short-circuit links can be used to complete the configuration of the metasurface. Based on this, the present invention also provides a metasurface optical neural network configuration method based on short-circuit links, comprising: executing a metasurface optical neural network parameter design method based on short-circuit links to obtain first target design parameters and second target design parameters; configuring a first metasurface according to the first target design parameters and the second target design parameters. Wherein, the first target design parameters include target masking parameters, and the second target design parameters include target microstructure parameters.

[0098] In this way, configuring metasurfaces according to target design parameters can avoid repetitive work, improve production efficiency, reduce unnecessary trial and error costs, save time, and greatly improve work efficiency.

[0099] The present invention also provides a metasurface optical neural network based on short-circuit links, comprising: executing a metasurface optical neural network parameter design method based on short-circuit links to obtain a third target design parameter and a fourth target design parameter, and configuring a second metasurface according to the third target design parameter and the fourth target design parameter; the second metasurface includes a substrate and microstructure units, the microstructure units including a plurality of primitives.

[0100] Thus, the metasurface can be configured according to the parameter design method described above. The production process is simple and convenient, and it has the characteristics of low energy consumption, fast calculation speed and high classification accuracy, which is conducive to improving the product quality and work efficiency of metasurfaces.

[0101] In summary, the metasurface optical neural network parameter design method based on short-circuit links provided by this invention adopts an end-to-end design. It constructs an optical field simulation model corresponding to the metasurface and simulates and calculates light intensity distribution data based on this model. Under the constraint of the target light intensity distribution configuration, target light intensity distribution data is calculated based on the light intensity distribution data. The loss function is determined based on the difference between the target light intensity distribution data and the actual light intensity distribution data. The method iteratively adjusts, at least the anti-mask configuration parameters of the metasurface, until the loss function converges, thus determining the target anti-mask parameters and target microstructure parameters. This method, through defining a loss function and iterative parameter tuning, can automatically obtain the optimal design parameters without manual trial and error, reducing design costs, saving time, exhibiting good universality, and improving computational accuracy and design efficiency.

[0102] Furthermore, by introducing an adaptive mask distribution on several hierarchical structures of the metasurface, specifically by setting anti-mask regions, incident light can directly pass through the metasurface where these anti-mask regions are located. That is, by configuring the metasurface according to the target anti-mask parameters, some of the original optical data can be transmitted through the metasurface. This not only simplifies the calculation process but also enhances the metasurface's expressive power and robustness to noise by preserving the original optical data, thereby improving its optical calculation accuracy and precision. This approach allows the metasurface to balance the number of optical layers and optical calculation capabilities while simultaneously reducing manufacturing costs.

[0103] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0104] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for designing parameters of a metasurface optical neural network based on short-circuit links, characterized in that, include: Receive and detect images and corresponding light intensity distribution target configurations; A light field simulation model corresponding to the metasurface is constructed, and a loss function is determined based on the light intensity distribution data of the detected image and the light intensity distribution target configuration according to the light field simulation model. Iterate and adjust the mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determine the target mask parameters and target microstructure parameters; Wherein, the anti-mask configuration parameters characterize the configuration of the anti-mask regions on the metasurface; the target microstructure parameters represent the geometric parameters of the microstructure units constituting the metasurface; the anti-mask regions allow incident light to pass directly through the metasurface to achieve short-circuit connections on the metasurface; and the number of anti-mask regions configured on each metasurface layer of the metasurface does not exceed one. The construction of the light field simulation model corresponding to the metasurface, and the determination of the loss function based on the light intensity distribution data of the detected image and the target configuration of the light intensity distribution according to the light field simulation model, include: A light field simulation model incorporating the time-domain finite difference method and the Rayleigh-Sommerfeld formula was constructed. Based on the detected image, a light field parameter matrix is ​​obtained by parsing; wherein, each element in the light field parameter matrix points to the light field intensity value of each pixel in the detected image; Using the light field parameter matrix as the input of the light field simulation model, firstly, the corresponding light intensity data is substituted into the Rayleigh-Sommerfeld formula to obtain the distribution of the light intensity data on the metasurface. Then, the finite-difference time-domain method is executed, and the light intensity data near the metasurface is substituted into the Rayleigh-Sommerfeld formula to iterate and calculate the light intensity distribution data. The nearby light intensity data characterizes the light field intensity before photoelectric conversion, and the light intensity distribution data characterizes the light field intensity after photoelectric conversion. Analyze the target configuration of the light intensity distribution to obtain the high light intensity area and the low light intensity area of ​​the target; The region in the light intensity distribution data corresponding to the high light intensity region of the target is configured as the first light field intensity, and the region in the light intensity distribution data corresponding to the low light intensity region of the target is configured as the second light field intensity. The light intensity distribution data is updated to obtain the light intensity distribution target data.

2. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The metasurface comprises several metasurface layers; the "construction of an optical field simulation model corresponding to the metasurface" specifically includes: Construct optical field simulation models corresponding to several metasurface layers; The phrase "iterate and adjust the mask configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determine the target mask parameters and target microstructure parameters" specifically includes: Iterate and adjust the anti-mask configuration parameters and microstructure configuration parameters of the first metasurface layer until the loss function corresponding to the first metasurface layer converges, and determine the target anti-mask parameters and target microstructure parameters corresponding to the first metasurface layer. The masking configuration parameters and microstructure configuration parameters of other metasurface layers are traversed and adjusted one by one to determine several target masking parameters and several target microstructure parameters corresponding to other metasurface layers.

3. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The target mask parameters include position configuration parameters, quantity configuration parameters, and area configuration parameters; the target microstructure parameters include position configuration parameters, quantity configuration parameters, and size configuration parameters.

4. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The "construction of an optical field simulation model corresponding to a metasurface" specifically includes: The initial configuration parameters of the microstructure units on the metasurface are determined by using a random initialization method or a Gaussian distribution method. Based on the initial configuration parameters, a light field simulation model corresponding to the metasurface is constructed.

5. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The phrase "determining the loss function based on the light intensity distribution data of the detected image and the light intensity distribution target configuration according to the light field simulation model" specifically includes: The light intensity distribution data is obtained by performing a traversal calculation based on the detected image; Under the constraints of the light intensity distribution target configuration, the light intensity distribution target data is calculated based on the light intensity distribution data; The loss function is determined based on the difference between the target light intensity distribution data and the light intensity distribution data.

6. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The light intensity distribution data is formed into a light intensity distribution matrix; the light intensity distribution target data is formed into a light intensity distribution target matrix; the step of "determining the loss function based on the light intensity distribution data of the detected image and the light intensity target data corresponding to the detected image using the neural network model" specifically includes: Subtracting the corresponding elements in the light intensity distribution matrix and the light intensity distribution target matrix yields the error matrix; The square of each element in the error matrix is ​​calculated sequentially to obtain the squared error matrix; The loss function is calculated and determined based on the average of all elements in the squared error matrix.

7. The method for designing parameters of a metasurface optical neural network based on short-circuit links according to claim 1, characterized in that, The phrase "iterate and adjust the masking configuration parameters and microstructure configuration parameters of the metasurface until the loss function converges, and determine the target masking parameters and target microstructure parameters" specifically includes: Based on the loss function, the gradient vectors of the anti-mask configuration parameters and the microstructure configuration parameters in the light field simulation model are calculated using the backpropagation algorithm; wherein, the gradient vectors characterize the changes of the light field simulation model parameters relative to the loss function; The gradient vector is input into the optimization algorithm to calculate the updated values ​​of the parameters of the light field simulation model, and to update the anti-mask configuration parameters and the target microstructure parameters; wherein, the optimization algorithm includes at least one of stochastic gradient descent and the Adam algorithm; The backpropagation algorithm and the optimization algorithm are executed iteratively until the loss function converges, thereby determining the target anti-mask parameters and the target microstructure parameters.

8. A metasurface configuration method based on short-circuit connections, characterized in that, include: The method for designing parameters of a metasurface optical neural network based on short-circuit links according to any one of claims 1 to 7 is used to obtain a first target design parameter and a second target design parameter; wherein the first target design parameter includes a target anti-mask parameter and the second target design parameter includes a target microstructure parameter; Configure the first metasurface according to the first target design parameters and the second target design parameters.

9. A metasurface optical neural network based on short-circuit links, characterized in that, include: The method for designing metasurface optical neural network parameters based on short-circuit links according to any one of claims 1 to 7 is used to obtain third target design parameters and fourth target design parameters, and a second metasurface is configured according to the third target design parameters and the fourth target design parameters; the second metasurface includes a substrate and microstructure units, and the microstructure units include a plurality of primitives.

Images