A large-aperture meta-lens based on a deep learning network model and its design method

Abstract

The present invention provides a large-aperture metalens based on a deep learning network model and a design method thereof, including: S1, using a traditional design method to design a small-aperture metalens M1, which is an initial model; S2, using a traditional optimization method to optimize the geometric dimensions of the unit structure of the initial model; S3, training a deep learning network model: using the initial model as input and the optimized metalens as output; S4, using a traditional design method to design a large-aperture metalens M2, and optimizing the metalens M2 through the trained deep learning network model. The device efficiency loss caused by non-periodic local electromagnetic coupling is reduced by correcting the size of the metasurface unit structure. On this basis, the deep learning network model is used to solve the problem of limited design efficiency of large-aperture metalenses. This method is independent of the device function and provides a very effective solution for improving the design efficiency of devices with complex unit structure arrangements.

Description

Technical Field

[0001] The present invention relates to the technical field of metasurfaces, and in particular to a large-aperture metalens based on a deep learning network model and a design method thereof. Background Art

[0002] At present, the optimization design of metasurfaces mainly adopts the traditional design paradigm of parameter scanning and manual optimization, that is, firstly, the electromagnetic response of the unit structure is obtained according to the periodic boundary conditions, and then the unit structure is arranged according to a certain rule, and finally the design of the metasurface device is completed. This design method faces problems such as long time consumption and difficulty in optimizing performance. In addition, the initial structure generated by the traditional "trial and error method" or reverse intelligent design method ignores the electromagnetic coupling effect between micro-nano structures and the influence of boundary conditions. The metasurface devices constructed in this way often cannot obtain high device performance, that is, the efficiency is usually low. The boundary optimization method based on adjoint optimization is an optimization method based on the full-mode structure. Through two forward and adjoint calculations, the gradient information of the objective function in the entire parameter optimization space can be obtained. This method not only saves computing resources when designing small-aperture metalenses, but also effectively incorporates the physical information of the interaction between micro-nano structures into the algorithm, which can effectively improve the focusing efficiency of high numerical aperture metalenses. Although the above optimization method based on full-mode structure simulation alleviates the influence of electromagnetic coupling effects and boundary conditions between micro-nano structures. However, for metasurface devices with larger apertures, the computational resources they occupy are huge, resulting in low computational efficiency. In addition, due to the limitation of computer computing power, the current optimization method of full-mode structural simulation is only applicable to metasurface devices with small apertures. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a meta-lens design method that combines traditional optimization methods with deep learning network models, thereby quickly and efficiently designing large-aperture, high-numerical-aperture meta-lenses. Deep learning is a data-driven algorithm. The input data undergoes a series of nonlinear changes between neural networks. With appropriate network structures, optimizers, loss functions and other hyperparameters, an accurate mathematical mapping relationship with the output data can be established. For meta-surface devices, a mathematical mapping model between the initial structure and the structure after boundary optimization can be effectively constructed using a deep learning neural network, and the complex physical functions involved in the traditional optimization process can be fitted. Subsequently, the large-aperture, high-numerical-aperture meta-lenses designed in the forward direction are sampled and used as the input of the trained deep learning network model, and the output is the size of the optimized unit structure. This method can effectively solve the problem of limited design methods for large-aperture, high-numerical-aperture meta-lenses.

[0004] The technical solution proposed by the present invention is as follows:

[0005] A design method for a large-aperture meta-lens based on a deep learning network model, comprising:

[0006] S1, design a small-aperture metalens M1, which is the initial model;

[0007] S2, optimizing the geometric dimensions of the unit structure of the initial model by using an optimization method;

[0008] S3, training a deep learning network model, wherein the deep learning network model takes the initial model as input and outputs the optimized metalens;

[0009] S4, designing a large-aperture meta-lens M2, and optimizing the meta-lens M2 through the trained deep learning network model.

[0010] Furthermore, in S1, designing the metalens M1 includes: calculating the focal length f according to the numerical aperture NA and the metalens diameter D:

[0011]

[0012] The phase distribution of each discrete coordinate (x, y) is obtained according to f:

[0013]

[0014] In the formula, is the phase under discrete coordinates (x, y), λ is the incident wavelength, x is the horizontal coordinate, y is the vertical coordinate, x0 is the horizontal coordinate of the device center, and y0 is the vertical coordinate of the device center;

[0015] By matching the I-order unit structure with the discretized phase obtained above, where I≥2, a meta-lens M1 that meets the target numerical aperture NA, aperture D, and focal length f is obtained.

[0016] Furthermore, in S2, the optimization method is any one of a boundary optimization algorithm based on adjoint simulation, a genetic algorithm or a particle swarm optimization algorithm; and the optimization goal is to maximize the focusing efficiency of the metalens.

[0017] Furthermore, S3 specifically includes: sampling the unit structure of the initial model, and using the sampled super unit B1 as the input sample for training the deep learning network model; sampling the unit structure of the initial model optimized in S2, and using the sampled super unit B2 as the output sample for training the deep learning network model.

[0018] Further, the supercell B1 is a group of unit structures after sampling the unit structures in a (2n + 1)×(2n + 1) manner, where n is a positive integer, and the step length between two adjacent supercells B1 is 1 unit structure; the supercell B2 is the unit structure optimized at the (n + 1, n + 1) position in each supercell B1.

[0019] Further, the supercells B1 are classified. According to the variation relationship of the geometric dimensions of the supercells B1, they are divided into smooth region samples and non-smooth region samples. Among them, for the supercell B1, when the width or length of the unit structure increases, decreases, or the dimensions of adjacent unit structures remain unchanged during the change process, it is a smooth region sample; otherwise, it is a non-smooth region sample.

[0020] Further, the deep learning network model in S3 is any one of a fully connected network, an autoencoder neural network, a convolutional network, or a trained neural network model.

[0021] Further, S4 specifically includes sampling the metasurface lens M2, taking the geometric dimensions of the unit structures in the sampled supercell B3 as the input of the deep learning network model, optimizing the metasurface lens M2 through the trained deep learning network model, and splicing the optimized unit structures in the sampling order to obtain the optimized metasurface lens M2.

[0022] Further, the supercell B3 is a group of unit structures after sampling the unit structures in a (2n + 1)×(2n + 1) manner, where n is a positive integer, and the step length between two adjacent supercells B3 is 1 unit structure; the optimized unit structure is the unit structure optimized at the (n + 1, n + 1) position in each supercell B3.

[0023] The present invention also proposes another technical solution:

[0024] A large-aperture metasurface lens is designed by using the design method described above.

[0025] Further, the aperture D1 of the metasurface lens M1 ≤ 50μm; the unit structure period of the metasurface lens M1 is square, and the nanocolumns of the unit structure are rectangular, square, triangular, circular, or elliptical.

[0026] Further, the unit structure period of the metasurface lens M1 is p1 and 0.1λ1 < p1 < 0.5λ1, where λ1 is the central working wavelength; the height of the nanocolumn of the unit structure of the metasurface lens M1 is h1 and 0.1λ1 < h1 < λ1, the width is w1 and 0.01p1 < w1 < p1, the length is l1 and 0.01p1 < l1 < p1; the number of unit structures of the metasurface lens M1 is (D1 / p1)×(D1 / p1).

[0027] Furthermore, the material of the metasurface lens M1 is a pure dielectric material, which is selected from any one of Si, SiN, SiO2, TiO2, or GaAs.

[0028] Furthermore, the aperture of the large-aperture metasurface lens M2 is D2 and D2 > D1; the unit structure period is p2 = p1; the height of the unit structure nanocolumn is h2 = h1; the width of the unit structure nanocolumn is w2 and 0.01p2 < w2 < p2; the length of the unit structure nanocolumn is l2 and 0.01p2 < l2 < p2; the number of unit structures is (D2 / p2) × (D2 / p2).

[0029] The beneficial effects of the present invention are as follows:

[0030] By cross-combining the deep learning network design method and the metasurface, two cutting-edge fields, and using two effective intelligent inverse design methods, namely traditional optimization methods such as adjoint optimization and deep learning, the design of large-aperture metasurface devices is realized. By modifying the size of the unit structure of the small-aperture metasurface, the device efficiency loss caused by non-periodic local electromagnetic coupling is reduced, and the geometric size of the unit structure of the large-aperture and high numerical aperture metasurface lens is inversely calculated through the deep learning network model learning the change relationship of the structure size before and after optimization, so as to solve the problem that due to the limitation of computer computing power, the optimization method of full-mode structure simulation is only applicable to the design of small-aperture metasurface devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematically shows a flowchart of a design method for a large-aperture metasurface lens based on a deep learning network model according to an embodiment of the present invention;

[0032] Figure 2 Schematically shows a small-aperture metasurface lens M1 and a unit structure schematic diagram according to an embodiment of the present invention;

[0033] Figure 3 Schematically shows a flowchart of an optimization process of a boundary optimization algorithm of a traditional optimization method according to an embodiment of the present invention;

[0034] Figure 4 Schematically shows a schematic diagram of the structural change of the unit structure of the small-aperture metasurface lens M1 before and after optimization according to an embodiment of the present invention;

[0035] Figure 5 Schematically shows a diagram of the change in geometric size of the unit structure of the small-aperture metasurface lens M1 before and after optimization at the xz plane and the y = 0 cross-section according to an embodiment of the present invention;

[0036] Figure 6A schematic diagram showing the comparison of the normalized electric field curve distribution on the xz plane before and after the optimization of the small-aperture meta-lens M1 according to an embodiment of the present invention;

[0037] Figure 7 The structural distribution of a small-aperture meta-lens M1 according to an embodiment of the present invention is schematically shown, wherein a white dotted circle represents a rough area of ​​the meta-lens structure;

[0038] Figure 8 The method for establishing a data set using a small-aperture meta-lens M1 as a sample set and the corresponding relationship between super-unit B1 and super-unit B2 according to an embodiment of the present invention are schematically shown. In this embodiment, n=2;

[0039] Fig. 9 A schematic diagram of an autoencoding neural network framework according to an embodiment of the present invention is shown;

[0040] Fig.10 Schematically showing the electric field distribution diagrams of the xz plane and xy plane of the large-aperture metalens M2 before and after optimization according to an embodiment of the present invention;

[0041] Fig.11 A schematic diagram shows a comparison of the normalized electric field curve distribution in the xz plane before and after optimization of a large-aperture metalens according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited to the following embodiments, but should include all the contents in the claims. Moreover, those skilled in the art can implement all the contents in the claims from the following embodiment.

[0043] Figure 1 The schematic diagram of the design method of a large-aperture meta-lens based on a deep learning network model according to an embodiment of the present invention is shown, including the following steps:

[0044] S1, using traditional design methods to design a small-aperture meta-lens M1, which is the initial model;

[0045] S2, using traditional optimization methods to optimize the geometric dimensions of the initial model unit structure;

[0046] S3, training based on deep learning network model: taking the initial model as input and the optimized meta-lens as output;

[0047] S4, uses traditional design methods to design a large-aperture meta-lens M2, and optimizes the meta-lens M2 through a trained deep learning network model.

[0048] Among them, in S1, the traditional design method for designing a small-aperture meta-lens in this embodiment is to calculate the focal length f according to the relationship between the numerical aperture NA, the focal length f and the aperture D of the meta-lens.

[0049]

[0050] Then, the phase distribution of each discrete coordinate (x, y) is obtained according to f:

[0051]

[0052] In the formula, is the phase under discrete coordinates (x, y), λ is the incident wavelength, x is the horizontal coordinate, y is the vertical coordinate, x0 is the horizontal coordinate of the device center, and y0 is the vertical coordinate of the device center.

[0053] In this embodiment, by matching the 8th-order unit structure with the discretized phase obtained above, a forward-designed meta-lens that meets the target numerical aperture NA, aperture D, and focal length f can be obtained. This meta-lens is the initial model for subsequent performance optimization using traditional optimization methods.

[0054] Figure 2 A small-aperture metalens M1 and a unit structure schematic diagram according to an embodiment of the present invention are schematically shown; the period of the unit structure is p, the height of the nanocolumns on the unit structure is h, the width of the nanocolumns is w, and the length of the nanocolumns is l.

[0055] Figure 3 The schematic diagram of the optimization process of the boundary optimization algorithm of the traditional optimization method according to the embodiment of the present invention is schematically shown; the geometric dimensions of the unit structure of the small-aperture meta-lens M1 are optimized by the traditional optimization method.

[0056] The conventional optimization method may be any one of a boundary optimization algorithm based on adjoint simulation, a genetic algorithm or a particle swarm optimization algorithm; in this embodiment, a boundary optimization algorithm based on adjoint simulation is used. The optimization goal is to maximize the focusing efficiency of the meta-lens.

[0057] Figure 3 The shape and size of the unit structure are only examples to explain the optimization process, and are not limitations on the shape and size of the unit structure. In order to maximize the focusing efficiency of the metalens, the maximization of the focal spot electric field intensity can be used as the optimization goal, and the width and length of the unit structure can be used as the optimization parameters. In this embodiment, the evaluation function F of the boundary optimization can be defined as:

[0058] F=|E(x)| 2 =E(x)E * (x) (3)

[0059] Where F represents the electric field intensity of the focal spot, E represents the electric field in the focal plane, and * represents the conjugate of the complex number. Here x is the horizontal coordinate, where E(x) is obtained when calculating the forward field, and E*(x) is obtained when calculating the adjoint field. According to the boundary conditions of the Maxwell equations, the gradient of the evaluation function relative to the width of the nanocolumn can be expressed as:

[0060]

[0061] Where W represents the width of the nanocolumn, the subscript old represents the structure parameters before iteration, and the subscript new represents the structure parameters after iterative update. Similarly, the gradient of the valence function with respect to the length of the nanocolumn can be expressed as:

[0062]

[0063] Figure 4 The schematic diagram of the structural change of the small-aperture metalens M1 unit structure before and after optimization according to an embodiment of the present invention is shown; after a limited number of iterative updates of the nanocolumn width and length, the geometric dimensions of the optimized unit structure are obtained, and the specific number of iterative updates is selected by those skilled in the art based on experience. Figure 4 It can be seen that the aforementioned boundary optimization algorithm directly optimizes the geometric dimensions of the unit structure. During the optimization process, the height and period of the initial structure are fixed, and the optimized unit structure only changes in length and width compared to the initial structure.

[0064] On the basis of the above embodiment, the unit structure period of the metalens M1 is p1, which is 500nm, and λ1, which is 1550nm; the unit structure nanocolumn height of the metalens M1 is h1, which is 1000nm; the unit structure nanocolumn width of the metalens M1 is w1, which is 100-400nm; the unit structure nanocolumn length of the metalens M1 is l1, which is 100-400nm. The material of the metalens M1 is Si.

[0065] On the basis of the above embodiment, the unit structure of the meta-lens with an aperture of 50 μm is optimized. According to the actual simulation process, it takes about 20 days to optimize all unit structures using the above boundary optimization algorithm. Figure 5 The diagram schematically shows the geometrical dimension changes of the small-aperture meta-lens M1 before and after the unit structure optimization at the xz plane and the y=0 section according to an embodiment of the present invention; Figure 5 It can be seen that the length and width of the unit structure have changed before and after optimization. Compared with the unit structure size of the metalens M1, the unit structure width size change range Δw after optimization using the traditional optimization method is -400nm~400nm, and the unit structure length size change range Δl after optimization is -400nm~400nm, and the unit structure nanocolumn height and period remain unchanged.

[0066] Figure 6 A schematic diagram shows a comparison of the normalized electric field curve distribution in the xz plane before and after the optimization of the small-aperture metalens M1 according to an embodiment of the present invention; it can be clearly seen from the figure that after the optimization, the normalized maximum field intensity at the focus of the metalens M1 is 10% higher than before the optimization.

[0067] The main function of the deep learning neural network is to learn the relationship between the changes in the unit structure before and after optimization. Using the trained deep learning neural network, the optimized unit structure is obtained based on the unit structure of the large-aperture meta-lens designed using the traditional design method, and then spliced ​​to form a meta-lens. The structure optimized by the boundary changes in both width and length, so two types of neural networks need to be constructed specifically. At the same time, the inventors found that the optimization of the non-smooth area samples in the meta-lens plane changes the size significantly, such as Figure 5 and Figure 7 Therefore, we separate the data of smooth area samples and non-smooth area samples and train them separately.

[0068] On the basis of the above embodiments, a self-encoding neural network model is specifically selected as the deep learning network model in step S3, and the design method of the neural network will be described in detail below. In this embodiment, the period and height of all unit structures are fixed at 500nm and 1000nm. In the training of the neural network model, the meta-lens obtained by the traditional design method is used as input, and the meta-lens after boundary optimization is used as output. Specifically, the unit structure of the initial model is sampled, and the sampled super unit B1 is used as the input sample for training the deep learning network model; the super unit B1 is a unit structure group after the unit structure is sampled by (2n+1)×(2n+1), where n is a positive integer, and n is 2 in this embodiment, that is, the width and length of 5×5 unit structures are used as one input. The step length of two adjacent super units B1 is 1 unit structure. The unit structure of the meta-lens after S2 optimization is sampled, and the sampled super unit B2 is used as the output sample for training the deep learning network model. Super unit B2 is a unit structure group obtained by optimizing the position (n+1,n+1) in each super unit B1. In this embodiment, the width and length of the unit structure at the middle position of the optimized 5×5 output sample are used as output, that is, the size of the unit structure at the position (3,3) is super unit B2.

[0069] Here combined Figure 5 and Figure 7The smooth area samples and non-smooth area samples are described. In this embodiment, the super unit B1 is a unit structure group composed of 5×5 unit structures. The super unit B1 has a smooth area sample when the unit structure width or length increases, decreases or changes, and the adjacent unit structure size remains unchanged. Otherwise, it is a non-smooth area sample. When only the width change is considered, such as Figure 5 As shown in the first figure, with the increase of radius, 5 unit structures are grouped together, and the smooth area samples and non-smooth area samples are classified according to their size changes. The width values ​​of the 5 unit structures in the first group remain unchanged, so the super unit B1 is a smooth area sample. For the 5 unit structures in the second group, except for the first two width values ​​which decrease, the other width values ​​remain unchanged, so it is a smooth area sample. For the 5 unit structures in the third group, the first three width values ​​remain unchanged, the fourth decreases compared to the third, and the fifth and fourth width values ​​remain unchanged, so they are smooth area samples. For the 5 unit structures in the fourth group, the first width value is the smallest, the second width value increases, and from the second value onwards, it begins to show a trend of decreasing from large to small, that is, as the radius increases in the 5 unit structures, the broadband value increases and decreases at the same time, so it is a non-smooth area sample. According to this principle, the super unit B1 can be classified as a smooth area sample and a non-smooth area sample according to all width and length changes. From Figure 7 It can be seen that the length and width of the nanopillars in the non-smooth area samples will change suddenly.

[0070] The collected data is put into the built autoencoder neural network model as the input and output of the network, and then trained. The neural network model is as follows Figure 8 In this embodiment, the specific autoencoder neural network model is trained after 30,000 iterations, and the error between the structure obtained by the neural network model and the structure of the test set is within 1.5 nm.

[0071] The trained neural network model is used to optimize the larger-aperture meta-lens. In this embodiment, the meta-lens M2 with an aperture of 100 μm and a focal length of 100 μm is optimized and designed, and the obtained meta-lens is simulated and verified. The unit structure period is p2=p1=500nm; the unit structure nano-column height is h2=h1=1000nm; the unit structure nano-column width is w2 and 100-400nm; the unit structure nano-column length is l2 and 100-400nm; the number of unit structures is 200×200.

[0072] The meta-lens M2 is sampled, and the geometric dimensions of the unit structure in the super-unit B3 are used as the input of the deep learning network model; the super-unit B3 is a unit structure group after the unit structure is sampled at 5×5, and the step length of two adjacent super-units B3 is 1 unit structure. The unit structure width variation range Δw2 of the large-aperture meta-lens after optimization is -400nm~400nm, and the unit structure length variation Δl2 range after optimization is -400nm~400nm, and the unit structure nano-column height and period remain unchanged. The optimized unit structures are spliced ​​in the sampling order to obtain the optimized meta-lens.

[0073] The electric field distribution of the metalens at the focal plane and the electric field intensity distribution in the yz direction after optimization by the neural network model are shown in the figure below. Fig. 9 shown. Fig.10 The figure shows the normalized electric field distribution at the focal plane. Compared with the simulation results of the meta-lens with the same aperture and focal length designed by the traditional method, the focusing efficiency is improved by 6%. After the above-mentioned large-aperture meta-lens has been trained with a good neural network model, the focusing efficiency can be effectively improved by changing the length and width of the unit structure. Once the neural network model is completed, the optimization time is in the order of seconds. Compared with the accompanying optimization algorithm, the optimization time is in the order of days, and the optimization efficiency is greatly improved.

[0074] Therefore, the embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above specific embodiments, which are merely illustrative and not restrictive. The present invention does not elaborate on some of the common technologies that belong to those skilled in the art.

Claims

1. A design method for a large-aperture meta-lens based on a deep learning network model, characterized in that: include: S1, design a small-aperture metalens M1, which is the initial model; S2, optimizing the geometric dimensions of the unit structure of the initial model by using an optimization method; S3, training a deep learning network model, wherein the deep learning network model takes the initial model as input and outputs the optimized metalens, including: The unit structure of the initial model is sampled, and the sampled super unit B1 is used as an input sample for training the deep learning network model; the unit structure of the initial model optimized in S2 is sampled, and the sampled super unit B2 is used as an output sample for training the deep learning network model; Super unit B1 is a unit structure group obtained by sampling the unit structure at (2n+1)×(2n+1), where n is a positive integer and the step length between two adjacent super units B1 is 1 unit structure; super unit B2 is a unit structure obtained by optimizing the position (n+1, n+1) in each super unit B1; The super cell B1 is classified into smooth area samples and non-smooth area samples according to the relationship between the geometrical size changes of the super cell B1, wherein the super cell B1 is a smooth area sample when the cell structure width or length increases, decreases or changes while the adjacent cell structure size remains unchanged, otherwise, it is a non-smooth area sample; S4, designing a large-aperture meta-lens M2, and optimizing the meta-lens M2 through the trained deep learning network model, including: Sampling the metalens M2, taking the geometric dimensions of the unit structure in the sampled super unit B3 as the input of the deep learning network model, optimizing the metalens M2 through the trained deep learning network model, and splicing the optimized unit structure in the sampling order to obtain the optimized metalens; The super unit B3 is a unit structure group after the unit structure is sampled as (2n+1)×(2n+1), where n is a positive integer, and the step length between two adjacent super units B3 is 1 unit structure; the optimized unit structure is the unit structure optimized at the (n+1,n+1) position in each super unit B3.

2. The method for designing a large-aperture meta-lens based on a deep learning network model according to claim 1, characterized in that: In S1, designing the meta-lens M1 includes: calculating the focal length f according to the numerical aperture NA and the meta-lens diameter D: (1) The phase distribution of each discrete coordinate (x, y) is obtained according to f: (2) Wherein, φ is the phase under discrete coordinates (x, y), λ is the incident wavelength, x is the horizontal coordinate, y is the vertical coordinate, x0 is the horizontal coordinate of the device center, and y0 is the vertical coordinate of the device center; By matching the I-order unit structure with the discretized phase obtained above, where I≥2, a meta-lens M1 that meets the target numerical aperture NA, aperture D, and focal length f is obtained.

3. The design method of a large-aperture meta-lens based on a deep learning network model according to claim 1, characterized in that: In S2, the optimization method is any one of a boundary optimization algorithm based on adjoint simulation, a genetic algorithm or a particle swarm optimization algorithm; and the optimization goal is to maximize the focusing efficiency of the meta-lens.

4. The method for designing a large-aperture meta-lens based on a deep learning network model according to claim 1, characterized in that: The deep learning network model in S3 is any one of a fully connected network, an autoencoding neural network, a convolutional network or a trained neural network model.

5. A large-aperture meta-lens, characterized in that: The design is performed using the design method described in any one of claims 1 to 4.

6. The large-aperture metalens according to claim 5, characterized in that: The aperture D1 of the meta-lens M1 is ≤50 μm; the unit structure period of the meta-lens M1 is a square, and the nano-columns of the unit structure are rectangular, square, triangular, circular or elliptical.

7. The large-aperture metalens according to claim 5, characterized in that: The unit structure period of the metasurface lens M1 is p1, and 0.1λ1 < p1 < 0.5λ 1， λ1 is the central working wavelength; the height of the nanocolumn of the unit structure of the metasurface lens M1 is h1, and 0.1λ1 < h1 < λ1, the width is w1, and 0.01p1 < w1 < p1, the length is l1, and 0.01p1 < l1 < p1; the number of unit structures of the metasurface lens M1 is (D1 / p1)×(D1 / p1).

8. The large-aperture metalens according to claim 5, characterized in that: The material of the metalens M1 is a pure dielectric material, selected from any one of Si, SiN, SiO2, TiO2 or GaAs.

9. The large-aperture metalens according to claim 5, characterized in that: The aperture of the large-aperture metasurface lens M2 is D2; the unit structure period is p2 and 0.1λ1 < p2 < 0.5λ 1， λ1 is the central working wavelength; the height of the unit structure nanocolumn is h2 and 0.1λ1 < h2 < λ1; the width of the unit structure nanocolumn is w2 and 0.01p2 < w2 < p2; the length of the unit structure nanocolumn is l2 and 0.01p2 < l2 < p2; the number of unit structures is (D2 / p2) × (D2 / p2).

Images