A design method of a super-structured grating

Abstract

The application relates to the technical field of optical engineering design, in particular to a design method of a super-structured grating, the super-structured grating comprising a grating layer, the grating layer comprising a periodic cell, the design method comprising the following steps: constructing a neural network of deep reinforcement learning, taking the current state of the cell unit as the input of the input layer of the neural network, and taking the output of the output layer of the neural network as the action of filling the cell unit; obtaining a new state of the cell unit after the action is performed; calculating the diffraction efficiency of each diffraction order of the super-structured grating in the new state; obtaining the reward return of the action according to the diffraction efficiency; updating the input of the neural network based on the reward return and propagating reversely until the terminal state of the current deep reinforcement learning is reached; and repeatedly performing the above deep reinforcement learning steps until the requirement of the optimized design is met, so that the rapid optimized design of the super-structured grating comprising a complex cell structure can be realized.

Description

Technical Field

[0001] This application belongs to the field of optical engineering design technology, specifically a design method for meta-gratings. Background Technology

[0002] Metagratings are optical devices with subwavelength structures, exhibiting excellent performance in controlling the phase, amplitude, and polarization of light. They have been used in various applications to replace traditional lenses and diffractive optical elements. A one-dimensional metagrating consists of a substrate layer and a grating layer, with the grating layer period on the order of wavelength. A periodic cell is composed of several structural units. For simple cellular structures, such as a periodic cell consisting of two units, there are four parameters: h (unit height), w1 (width of the left unit), w2 (distance between the two units), and w3 (width of the right unit). Optimizing the design of a one-dimensional metagrating structure requires scanning the entire continuous parameter space to find parameter values ​​that satisfy the light field control capability. Since the entire parameter space is too large, optimization methods such as genetic algorithms and particle swarm optimization are commonly used. These optimization methods demonstrate good optimization capabilities for simple cellular structures.

[0003] Complex cellular structures are typically quite complex. As the number of cellular units increases, the dimension of the parameter space also increases accordingly. The aforementioned optimization methods suffer from poor convergence, long computation time, and slow optimization speed. Furthermore, with the increasing demand for optical field manipulation capabilities, the development trend of metagratings inevitably leans towards more complex cellular structures. Therefore, how to design metagratings incorporating complex cellular structures has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of one or more embodiments of this specification is to provide a design method for metagratings that enables rapid optimization design of metagratings including complex cellular structures.

[0005] To solve the above-mentioned technical problems, one or more embodiments of this specification are implemented as follows:

[0006] In a first aspect, a method for designing a metagrating is provided. The metagrating includes a grating layer comprising periodic cells. The design method includes the following steps: constructing a deep reinforcement learning neural network; using the current state of the cell as the input of the input layer of the neural network; and identifying the output of the output layer of the neural network as an action to fill the cell; obtaining a new state of the cell after executing the action; calculating the diffraction efficiency of each diffraction order of the metagrating in the new state; obtaining a reward for the action based on the diffraction efficiency; updating the input of the neural network based on the reward through backpropagation until the final state of the current deep reinforcement learning is reached; and repeating the above deep reinforcement learning steps until the requirements of the optimized design are met.

[0007] As can be seen from the technical solutions provided by one or more embodiments of this specification above, the metagrating design method provided by the embodiments of the present invention is applicable to cases where the metagrating includes a grating layer, and the grating layer includes periodic cells. This design method first constructs a deep reinforcement learning neural network, uses the current state of the cell as the input of the input layer of the neural network, and the output of the output layer of the neural network is identified as the action of filling the cell. Here, the output can be a symbol representing the action. After executing the action based on the output, a new state of the cell is obtained. The diffraction efficiency of each diffraction order of the metagrating in the new state is calculated. The reward for the action is obtained based on the diffraction efficiency. The input of the neural network is updated based on the reward, until the final state of this deep reinforcement learning is reached. The above deep reinforcement learning steps are repeated until the requirements of the optimization design are met, which can realize the rapid optimization design of metagratings including complex cell structures. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in one or more embodiments or prior art of this specification, the accompanying drawings used in the description of one or more embodiments or prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart of a metagrating design method provided according to an embodiment of the present invention;

[0010] Figure 2 This is a schematic diagram of the structure of a metagrating designed using the metagrating design method provided in an embodiment of the present invention;

[0011] Figure 3 This is a schematic diagram of the neural network in the design method of metagrating provided according to an embodiment of the present invention;

[0012] Figure 4 This is a schematic diagram of the initial structure of the metagrating to be designed according to the metagrating design method provided in the embodiment of the present invention;

[0013] Figure 5 It corresponds Figure 4 The neural network constructed using the metagrating shown;

[0014] Figure 6 This is a schematic diagram illustrating the optimization of the neural network based on backpropagation of the reward in the design method of metagrating provided according to an embodiment of the present invention;

[0015] Figure 7 The design method of the meta-grating provided in this invention is derived from... Figure 4 The diagram shows the optimal design of the meta-grating. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described one or more embodiments are merely some embodiments of this specification, and not all embodiments. All other embodiments obtained by those skilled in the art based on one or more embodiments of this specification without creative effort should fall within the protection scope of this document.

[0017] The metagrating design method provided in this invention employs a deep reinforcement learning neural network to rapidly optimize the complex cellular structure of periodic cells, reducing optimization time and improving design efficiency. The design method and its steps provided in this specification will be described in detail below.

[0018] Example 1

[0019] Reference Figure 1 The diagram illustrates a meta-grating design method provided in this embodiment of the invention. This method is an optimization design approach based on deep reinforcement learning, where a neural network represents the deep reinforcement learning strategy. Based on this strategy, the current state of a cell is mapped to a corresponding action. Executing this action yields a new state for the cell. The new state of the cell provides a reward to update the neural network. This training loop continues until the input to the neural network, i.e., the new state of the cell, meets the optimization design target.

[0020] This invention is applicable to the design of metagratings, which include grating layers and periodic cells. The design method includes the following steps:

[0021] Step S10: Construct a deep reinforcement learning neural network. The current state of each cell is used as the input to the input layer of the neural network, and the output of the output layer represents the action that fills the cell. Deep reinforcement learning requires constructing three elements: the state space of the cell, the action space, and the neural network itself. The state space represents the set of all possible environment states, and the action space represents the set of all executable actions. The current state of each cell is set as the environment state for deep reinforcement learning, which becomes the input to the input layer of the neural network. The output of the output layer represents the action that fills the cell; this action can be filling the cell with a grating medium or filling it with air.

[0022] Step S20: After performing the action, the new state of the cell is obtained; after performing the action, the cell that is currently in the state of air can be filled with grating medium, or the cell that is currently in the state of air can remain air after performing the action.

[0023] Step S30: Calculate the diffraction efficiency of each diffraction order of the metagrating in the new state; calculate the diffraction efficiency of each diffraction order of the metagrating based on the cell unit of the new state. The calculation method is not limited in this embodiment of the invention.

[0024] Step S40: Obtain the reward for the action based on the diffraction efficiency; the purpose of obtaining the reward is to update the neural network. More precisely: calculate the reward function value based on the diffraction efficiency; this function value represents the reward for taking the action in the new state.

[0025] Step S50: Update the input of the neural network based on backpropagation of reward rewards until the final state of this deep reinforcement learning session is reached. Update the input of the application network based on backpropagation of reward rewards, i.e., update the new state of the cells, until the new state of the cells reaches the final state of this deep reinforcement learning session. The final state (done) is a boundary condition used to determine whether the current deep reinforcement learning training should end. In this example, when all array elements are 0, it means that all cells in the raster have been filled with air, and no action can change the cell state, reaching a final state. Alternatively, if the array elements no longer change, it indicates that the training has converged or fallen into a local optimum, which is also a final state.

[0026] Step S60: Repeat the above deep reinforcement learning steps until the optimization design requirements are met. Based on the new state of the cell, re-execute the above deep reinforcement learning steps S10-S60 until the optimization design requirements of the periodic cells of the metagrating are met. Figure 6As shown, after a certain number of iterations, deep reinforcement learning can obtain a stable and desired target value. For example, if the diffraction efficiency of the +1 order grating is maximized, as indicated in the figure, then the requirements for optimization design are considered met. Optionally, the meta-grating design method provided in this application embodiment includes step S10: constructing a deep reinforcement learning neural network, using the current state of the cell unit as the input to the input layer of the neural network, specifically including: constructing the input layer, output layer, and a fully connected layer between the input and output layers of the neural network;

[0027] like Figure 2 As shown, a grating layer is set on the base layer 4. The structure within one grating period is called a periodic cell. The periodic structure of the grating is repeated. The grating layer includes multiple periodic cells. The periodic cell (shown by the dashed box in the figure) is set to include equally spaced discretized cell units (as shown in Figures 5 and 6), that is, one periodic cell is divided into N cell units. The environmental state of the cell unit being filled with grating medium (Figure 5) or air (Figure 6) is used as the current state of the cell unit. A one-dimensional array is used to represent the current state of the cell units within the periodic cell and serves as the input to the input layer. For a one-dimensional metagrating, the structure of N cell units within the periodic cell can be represented by a one-dimensional array containing N elements. Each element of this array takes the value 0 or 1, so there are a total of 2 N The width of the state space in deep reinforcement learning is 2. N In this array, an element of 1 indicates that the cell at the corresponding position is filled with grating medium, and an element of 0 indicates that the cell at the corresponding position is filled with air.

[0028] Furthermore, the metagrating design method provided in this application, which constructs a fully connected layer of a neural network positioned between the input and output layers, specifically includes: adjusting the number of fully connected layers and the number of neurons in each layer according to the number of cell units in the metagrating. For example... Figure 3 As shown, the neural network includes an input layer, an output layer, and at least one fully connected layer. The input layer takes in a one-dimensional array representing the current state of each cell. When the number of cells is N, the number of neurons in the input layer is adjusted to N, and the number of neurons in each fully connected layer is adjusted to m×N, where m is an integer. In our example, N = 10, m = 3, and there is only one fully connected layer. The adjustment process is as follows: if increasing the number of cells N causes the optimization process to no longer converge, then the number of fully connected layers and the value of m can be increased to improve the neural network's perceptual ability. This adjustment can be made based on the optimization effect.

[0029] Optionally, in the metagrating design method provided in this application embodiment, step S10: the output identifier of the neural network output layer is the action of filling cell units, specifically including:

[0030] Filling a cell with air is set as an action in deep reinforcement learning, with an action space of width N. The Q-value represents the expected reward of a cell performing an action in a given state. A neural network is used to fit the Q-function to calculate the Q-value for each action, and the output of the neural network's output layer is set as the Q-value of the action. An ε-greedy algorithm is used, where the action corresponding to the largest Q-value is executed with a probability ε, and any action in the corresponding action space is randomly selected with a probability of 1-ε to change the current state of the cell. The specific operation is as follows: Set the ε-value, where 0 ≤ ε ≤ 1. Generate a random number x between [0, 1]. If x < ε, execute the action corresponding to the largest Q-value; if x ≥ ε, generate a random integer y between [1, N] and execute the corresponding action. In this example, the y-th array element is set to 0. Initialize the one-dimensional array, setting all elements to 1, meaning all cells are filled with the raster medium. After entering the deep reinforcement learning update loop, the current one-dimensional array is set as the environment state of deep reinforcement learning. The current state of the cell is input into the neural network. Based on the output Q value, the ε greedy algorithm is used to select an action, and the array element at the corresponding position is changed to 0, that is, the grating medium of the corresponding cell is removed and filled with air.

[0031] Optionally, in the metagrating design method provided in this application embodiment, step S30: calculating the diffraction efficiency of each diffraction order of the metagrating in the new state, specifically includes: calculating the diffraction efficiency of each diffraction order of the metagrating in the new state using a rigorous coupled-wave analysis method. Further, in the metagrating design method provided in this application embodiment, step S40: obtaining the reward for the action based on the diffraction efficiency, specifically includes: substituting the diffraction efficiency into the optimization objective function to obtain the reward. After obtaining the new state of the cell unit, the diffraction efficiency of each diffraction order of the one-dimensional metagrating in the new state is calculated using a rigorous coupled-wave analysis method, and substituted into the optimization objective function; the obtained calculation result is then set as the reward.

[0032] In this example, the objective function F is to maximize the diffraction efficiency of the +1 diffraction order, i.e., F = max[η]. +1 [(s)], where s is the cell state, η +1 This represents the diffraction efficiency of the +1 diffraction order.

[0033] Optionally, the metagrating design method provided in this application embodiment includes step S50: updating the input of the neural network based on backpropagation of reward until the final state of this deep reinforcement learning is reached. Specifically, this includes: updating the Q value based on reward; executing actions based on the Q value until the new state of the cell is completely filled with air or the new state of the cell no longer changes. The diffraction efficiency of each diffraction order of the one-dimensional metagrating in the new state is calculated using rigorous coupled-wave analysis, and substituted into the optimization objective function. The calculated result is set as the reward, the Q value is updated, and the neural network is updated via backpropagation. When all array elements are 0 or the array elements no longer change, the training ends. The next round of deep reinforcement learning training is started based on the cell in the new state until the effect of deep reinforcement learning meets the requirements of the optimization design.

[0034] The update process here follows a standard procedure in deep reinforcement learning. Specifically, it requires introducing a loss function. Assuming the current state is s, the current action is a, and after performing the action, the new state becomes s', then the loss function is:

[0035] L=(R+γmaxQ(s',a')-Q(s,a)) 2 .

[0036] Where R is the reward for performing the current action, γ is a set discount factor, maxQ(s',a') is the maximum Q-value output after inputting the new state s' into the neural network, and Q(s,a) is the Q-value corresponding to the current action a in the output after inputting the current state s into the neural network. Then, the backpropagation mechanism of the neural network is used, combined with the calculation of the loss function, to update the neural network parameters. When the neural network is updated, and the input s is the same as the current state, the output Q-value is also updated.

[0037] A neural network can be simply viewed as a Q-function, where the output is the Q-value obtained from each action. After each action is executed, the current state of a cell may change. Inputting the current state s into the neural network yields a new Q-value, and the ε-greedy algorithm is used to select the next action to execute. Therefore, the loop occurs between "executing an action – changing to a new state and calculating the reward for that action – executing the new action." Within this loop, the loss function L is continuously calculated based on the reward, and the parameters of the neural network are updated using backpropagation.

[0038] The above process involves updating the parameters of the neural network in a loop during a deep reinforcement learning training session; this is the inner loop. If a final state condition is reached, the training session is terminated. Then, the current state of the cells is initialized, but the current parameters of the neural network are retained, and the next deep reinforcement learning training session begins; this is the outer loop.

[0039] Optionally, before step S10: constructing the deep reinforcement learning neural network, the metagrating design method provided in this application embodiment further includes setting the periodic cell as comprising equally spaced discretized cell units, specifically: assuming the periodic length of the periodic cell along the length direction is d, setting the number of cell units to N, and the width of the cell unit along the length direction is... And ensure that the value of N is such that Greater than the actual machining accuracy.

[0040] For example, a periodic cell consists of two units: h is the unit height, w1 is the width of the left unit, w2 is the distance between the two units, and w3 is the width of the right unit, totaling four parameters. When optimizing the cellular structure of a one-dimensional metagrating, these four parameters need to be scanned throughout the entire continuous parameter space to find parameter values ​​that satisfy the light field manipulation capability. However, for metagratings with more cellular units (i.e., more complex cellular structures), the design of the cellular structure requires searching within a larger continuous parameter space, consuming significant time and computational resources. Furthermore, due to limitations in existing micro / nano fabrication technologies, the actual fabricated cellular structure often deviates significantly from the design result, failing to meet design specifications. This invention, combining practical fabrication capabilities, discretizes the periodic cells of a one-dimensional metagrating, determining the width of the smallest discretized cellular unit along its length. Based on deep reinforcement learning, optimization is performed in this discrete space, satisfying the design requirements for complex periodic cell structures and ensuring consistency between the fabricated and designed results.

[0041] Furthermore, the metagrating design method provided in this application sets the periodic cell as including equally spaced discretized cell units, specifically including: assuming that all cell units have the same height and are h, and setting the size of the parameter space to S. h ×2 N S h Let h be the range of values ​​for height h; based on the longitudinal tolerance Δh of the machining accuracy, the height h of the cell element is further discretized into n finite values, ensuring that h0+nΔh∈[h0, h0+S]. h ], where h0 is the minimum height of the cell; the size of the parameter space is 2. N There are n finite values. For example, if h∈[100,300]nm and the processing accuracy is 50nm, then h can only take the five finite values ​​{100, 150, 200, 250, 300}, and n=5.

[0042] The metagrating design method provided in this invention uses deep reinforcement learning to discretize the height of the cell units of the periodic cells. This discretizes the one-dimensional metagrating with complex cell structures in a discrete space after both the length and height of the cell units are discretized, which can reduce the difficulty of processing and improve the product yield.

[0043] The parameter space refers to the range of values ​​for the parameter variables of a metagrating. The purpose of defining the parameter space is to determine the state space in deep reinforcement learning. The parameters of the metagrating are its height *h* and the medium distribution of the grating layers, i.e., what material fills each cell. A periodic cell contains *N* cells, each of which can be filled with either air or the grating medium, resulting in two possibilities. Therefore, a cell has 2n... N Possible distribution of various media. S h The value range of the cell height is h, but deep reinforcement learning is only applicable to discrete parameter spaces. Therefore, we need to further discretize h into n values, which can be achieved by using the tolerance method.

[0044] Optionally, the metagrating design method provided in this application embodiment includes step S10: constructing a deep reinforcement learning neural network, specifically comprising: constructing a deep reinforcement learning neural network for each of the n finite values ​​of the discretized cell unit. When the height h of the cell unit is a discrete value, which can be the finite value mentioned above, parallel computation can be used, that is, deep reinforcement learning is executed independently and in parallel for each height h value to ensure optimization speed.

[0045] Because actions in the action space of deep reinforcement learning only change the filling material of individual cells without any operation on the height parameter, a trained neural network can only optimize cells with a specific height h. Parallel computing refers to the ability to perform deep reinforcement learning independently on n height h values ​​simultaneously.

[0046] Example:

[0047] like Figure 4 As shown, the substrate layer 10 of the one-dimensional metagrating has a refractive index of 1.5, the grating dielectric layer 11 has a refractive index of 3.42, the length of the periodic cell of the metagrating is 500 nm, and the periodic cell consists of 10 equidistant cell units, each with a width of 50 nm. The height of the cell unit ranges from 200 to 350 nm with a tolerance of 50 nm. A 532 nm, TE mode plane wave is perpendicularly incident on the one-dimensional metagrating, and the optimization design requirement is to maximize the diffraction efficiency of the metagrating at the +1 diffraction order.

[0048] Figure 5The neural network constructed to meet the above design requirements is demonstrated, comprising an input layer 13, an output layer 15, and a fully connected layer 14. The input layer has 10 neurons to handle the input environment state; the output layer has 10 neurons to output the Q-value of the action; and the fully connected layer has 30 neurons. Using parallel computation, four identical neural networks were built to train deep reinforcement learning on a one-dimensional metagrating with cell heights of 200nm, 250nm, 300nm, and 350nm, respectively. Figure 4 The one-dimensional meta-grating shown is optimized, and the optimization process using deep reinforcement learning is as follows: Figure 6 As shown, Figure 6 The x-axis represents the number of deep reinforcement learning training steps; one training session includes... Figure 1 The training process involves several iterations (exiting the loop when the final state condition is met indicates that one training cycle is complete). Figure 6 The ordinate refers to the maximum diffraction efficiency of the +1 order achieved during one training cycle. From Figure 6 As can be seen, with the increase of the number of training steps, the maximum diffraction efficiency of the +1 order that can be achieved during training also increases and tends to converge. After the neural network training is completed, the optimal cell structure is finally obtained with a cell height of 300nm, and the array corresponding to the periodic cells is [1001001011]. The meta-grating structure is as follows: Figure 7 As shown, the diffraction efficiency of the +1 diffraction order is 0.79555, which meets the optimization design requirements.

[0049] The configuration of the discretized cell structure of the metagrating periodic cell obtained by the design method of the metagrating provided in the embodiments of the present invention, the width of the cell unit, the dielectric material, the number of cell units in the periodic cell, and the height of the cell unit are all within the scope of protection.

[0050] The above analysis shows that the metagrating design method provided in this invention is applicable to metagratings comprising grating layers, where the grating layers comprise periodic cells. This design method first constructs a deep reinforcement learning neural network, using the current state of the cell as the input to the neural network's input layer. The output of the neural network's output layer is identified as the action of filling the cell; the output can be a symbol representing the action. Based on this output, the action is executed to obtain a new state of the cell. The diffraction efficiency of each diffraction order of the metagrating in the new state is calculated. The reward for the action is obtained based on the diffraction efficiency. The input to the neural network is updated based on the reward, until the final state of this deep reinforcement learning is reached. The above deep reinforcement learning steps are repeated until the requirements of the optimized design are met, enabling rapid optimized design of metagratings including complex cellular structures.

[0051] In summary, the above description is merely a preferred embodiment of this specification and is not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of protection of this specification.

[0052] The systems, apparatuses, modules, or units described in one or more of the above embodiments may be implemented by a computer chip or entity, or by a product having a certain function. A typical implementation device is a computer. Specifically, a computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0053] It should also be noted that the terms "comprising," "including," or any other variations thereof are 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0055] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A method for designing a meta-grating, characterized in that, Metagratings include grating layers, the grating layers comprising periodic cells, and the design method includes the following steps: A deep reinforcement learning neural network is constructed, in which the current state of the cell is used as the input of the input layer of the neural network, and the output of the output layer of the neural network is identified as the action of filling the cell. After performing the action, a new state of the cell is obtained; Calculate the diffraction efficiency of each diffraction order of the metagrating in the new state; The reward for the action is determined based on the diffraction efficiency. The input of the neural network is updated by backpropagation based on the reward, until the final state of this deep reinforcement learning is reached; Repeat the above deep reinforcement learning steps until the requirements of the optimized design are met; Constructing a deep reinforcement learning neural network, using the current state of the cell as the input to the input layer of the neural network, specifically includes: Construct the neural network's input layer, output layer, and a fully connected layer positioned between the input layer and the output layer; The periodic cell is configured to include equally spaced discretized cell units, and the environmental state of the cell unit being filled with grating medium or air is taken as the current state of the cell unit. A one-dimensional array is used to represent the current state of the cell unit within the periodic cell and serves as the input to the input layer; The fully connected layer that constructs the neural network, positioned between the input and output layers, specifically includes: The number of fully connected layers and the number of neurons in each layer are adjusted according to the number of cellular units of the metagrating. The output identifier of the output layer of the neural network is the action of filling the cell unit, specifically including: The output of the output layer of the neural network is set as the Q value of the action, and the action corresponding to the maximum Q value is executed with a certain probability ε. The remaining probabilities 1-ε will randomly select any action in the action space to be executed.

2. The design method according to claim 1, characterized in that, Calculating the diffraction efficiency of each diffraction order of the metagrating in the new state specifically includes: The diffraction efficiency of each diffraction order of the metagrating in the new state was calculated using the rigorous coupled-wave analysis method.

3. The design method according to claim 1, characterized in that, The reward for the action is obtained based on the diffraction efficiency, specifically including: The diffraction efficiency is then substituted into the objective function for optimization to obtain a reward.

4. The design method according to claim 3, characterized in that, The input to the neural network is updated based on the reward, through backpropagation, until the final state of this deep reinforcement learning is reached. Specifically, this includes: The Q value is updated based on the aforementioned reward. The action is performed based on the Q value until the new state of the cell is either completely filled with air or the new state of the cell no longer changes.

5. The design method according to any one of claims 1 to 4, characterized in that, Before constructing the deep reinforcement learning neural network, the method further includes setting the periodic cell to include equally spaced discretized cell units, specifically: Assuming the periodic length of the periodic cell along the length direction is d, and the number of cell units is N, and the width of the cell unit along the length direction is... And ensure that the value of N is such that Greater than the actual machining accuracy.

6. The design method according to claim 5, characterized in that, The periodic cell is configured to include equally spaced discretized cell units, specifically including: Assuming the height of the cell is h, and the size of the parameter space is set to be... ,in The range of values ​​for height h; Based on the longitudinal tolerance of machining accuracy Further discretizing the height h of the cell unit into n finite values ​​ensures ,in Set the minimum height for the cell unit; The size of the parameter space is obtained as follows A finite value.

7. The design method according to claim 6, characterized in that, Constructing a deep reinforcement learning neural network specifically includes: A deep reinforcement learning neural network is constructed using n finite values ​​of the discretized cell units.

Images