A design method for a multifunctional vortex light generator

Abstract

The present invention discloses a design method for a multifunctional vortex light generator. Based on optical requirements, the nanostructure of a metasurface is reversely designed through a deep neural network, and multifunctional vortex light is obtained in the reflected light of the metasurface. The method is flexible in design and can quickly complete the design of the metasurface according to the required wavelength, polarization state and topological charge to generate the required multifunctional vortex light. Compared with vortex light generators with a single function, the method of the present invention uses deep learning to achieve multifunctional integration of a single device, which is conducive to the development of the potential infinite-dimensional Hilbert space of vortex light, greatly promoting the application of vortex light in the field of information optics, and promoting the application of vortex light in cutting-edge fields such as photonic chips and quantum optics. Compared with traditional reverse design methods, the method of the present invention can generate a large amount of multifunctional vortex light in a short time, greatly reducing the consumption of computing resources and computing time, and opening up the possibility of developing the advantageous characteristics of vortex light to the maximum extent.

Description

Technical Field

[0001] The present invention relates to a method for generating multifunctional vortex light by designing using a deep neural network, and specifically to a method for designing a multifunctional vortex light generator by designing a nanostructure that meets specific multifunctional requirements based on the optical requirements for generating vortex light by using an inverse neural network. Background Art

[0002] Vortex light is a beam of light that spirals along its propagation axis, characterized by a helical phase wavefront and a dark spot at the center of the beam. Vortex light can be generated using methods such as whispering gallery cavities, spiral phase disks, forked gratings, and artificial metasurfaces. Considerable progress has been made in the application of optical vortices in applications such as optical tweezers, super-resolution imaging, optical multiplexing, and optical communications. Optical vortices can manipulate large amounts of optical information, offering great potential for applications in high-sensitivity sensing and quantum communications. These promising advantages stem from the unique physical properties of optical vortices. A key property is orbital angular momentum (OAM), which provides infinite degrees of freedom for manipulating light. Topological charge, a physical quantity that describes the helicity of vortex light around the phase singularity, has an infinite number of possible values, forming an infinite-dimensional Hilbert space. Based on this, vortex light provides an unprecedented platform for sensitive information manipulation. To achieve excellent modulation, high capacity and wide modulation range are essential for vortex light carrying OAM. Therefore, multifunctional vortex light with independent topological charge is highly desirable in these applications.

[0003] Through appropriately designed nanostructures, metasurfaces can achieve a variety of light manipulation functions due to their powerful light modulation capabilities. Metasurfaces are composed of a large number of different nanostructures, each designed to achieve a specific function. The enormous complexity of these structures makes traditional reverse engineering, which involves scanning geometric parameters and performing extensive simulations to obtain the appropriate structure, extremely difficult. To overcome this reverse engineering difficulty, deep learning has been introduced into nanophotonics, offering an opportunity to break through bottlenecks in spectral prediction, structural design, and device optimization. Deep neural networks, a representative artificial intelligence algorithm, can rapidly solve complex optical systems without solving the complex Maxwell equations. They can autonomously explore useful information and establish connections between optical responses and geometric structures. Therefore, based on the input target optical response, deep neural networks can efficiently generate a suitable nanostructure. They can handle a large number of nanostructures under complex optical requirements, require minimal computational resources, and have negligible computation time. Therefore, deep neural networks are an appropriate approach to meet the enormous design needs of multifunctional vortex light generation. Summary of the Invention

[0004] The present invention aims to provide a design method for a multifunctional vortex light generator. This method, based on optical requirements, reverse-engineers the nanostructure of a metasurface using a deep neural network to generate multifunctional vortex light from the reflected light of the metasurface. This method offers flexibility and can rapidly design a metasurface to generate the desired multifunctional vortex light, tailored to the desired wavelength, polarization state, and topological charge.

[0005] The method of the present invention realizes the intelligent design of a multifunctional vortex light generator through a deep neural network. In some embodiments, the interaction between the nanostructure and different polarization states is independently modulated through a metal V-shaped antenna to obtain a variety of optical responses. According to the topological charge requirements, the deep neural network quickly generates each nanostructure to achieve a specific phase at a certain position of the light wave. A large number of different nanostructures are generated in an instant to form a functional metasurface. Therefore, the method of the present invention can realize arbitrary multifunctional vortex light. The method of the present invention realizes the efficient generation of multifunctional vortex light, opening up the possibility of developing the advantageous characteristics of vortex light to the maximum extent. The method of the present invention provides a universal solution for the practical application of vortex light, such as optical chips and quantum communications.

[0006] The technical solutions of the present invention are as follows:

[0007] A method for designing a multifunctional vortex light generator, wherein the multifunctional vortex light generator is a metasurface composed of a large number of different nanostructures, and the design of the metasurface comprises the following steps:

[0008] 1) Based on the desired operating wavelength of the vortex light generated and the device fabrication process, the overall dimensions of the vortex light generator, the unit period of its constituent units, and the minimum size of the nanostructure are determined. Based on these conditions, the geometric constraints for generating the nanostructure are determined to complete the parameterization of the nanostructure.

[0009] 2) Determine the polarization state and topological charge of the vortex light in different polarization channels based on the function of the multifunctional vortex light generator, further complete the phase profile design corresponding to the multifunctional vortex light generator, and then determine the optical properties that the nanostructures at different locations on the metasurface need to meet based on the phase profiles corresponding to the vortex light in different polarization channels;

[0010] 3) Constructing an artificial intelligence algorithm platform through a deep neural network, and then using this algorithm platform to find the nanostructures with specific optical properties required in step 2) to complete the metasurface design;

[0011] 4) Utilize the algorithm platform constructed in step 3) to obtain optical response data of the metasurface and evaluate device performance.

[0012] In the above step 1), since the nanostructure can interact with light at the subwavelength scale and thus modulate the phase and amplitude of the light, the unit period of the constituent unit should be set at the subwavelength scale. The minimum size of the nanostructure is usually limited by the device preparation process, and the processing line width can usually be set to about 30nm. The parameterization of the nanostructure refers to the use of a set of structural parameters to describe the geometric shape of the structure. For example, for a V-shaped nanostructure, its geometric structure is parameterized into four quantities (w1, w2, a, b), where: w1 and w2 are the lengths of the two nanorods that make up the V shape, a is the azimuth angle of the angle bisector of the angle between the two nanorods, and b is the angle between the two nanorods.

[0013] The phase profile design corresponding to the multifunctional vortex light generator in step 2) above involves calculating the optical properties of the nanostructures at different locations on the metasurface based on the polarization state, topological charge, generator size, and unit structure period of the vortex light in different polarization channels. Because the multifunctional vortex light generator can generate a variety of vortex lights with different topological charges in multiple polarization channels, the optical properties of the nanostructures must simultaneously meet the phase of the reflected light corresponding to different polarization channels.

[0014] The artificial intelligence algorithm platform described in step 3) above is composed of a forward neural network and a reverse neural network, wherein the forward neural network can predict the optical properties of the nanostructure based on its geometric parameters, while the reverse neural network can reversely generate structural geometric parameters that meet the conditions based on the optical properties required by the nanostructure.

[0015] In step 4), the algorithmic platform is used to evaluate device performance. For nanostructures designed using the inverse neural network, the forward neural network can rapidly predict their phase and amplitude. The predicted phase can be compared with the target phase to verify the desired phase profile and determine whether it meets design expectations. The amplitude, on the other hand, can be used to assess the device's operating efficiency, specifically the intensity of the generated vortex light.

[0016] Furthermore, the different polarization channels described in the above step 2) refer to the fact that under the incidence of incident light with different polarizations, the designed metasurface can change the polarization state and phase of the reflected light through polarization-dependent interactions. In short, the multifunctional vortex light generator can switch the polarization of the generated vortex light and information such as topological charge according to the polarization state of the incident light. In order to reduce the noise crosstalk of the incident light, the incident polarization state is generally orthogonal to the output polarization state. In addition, the incident polarizations under different functions should be orthogonal to each other or evenly distributed as much as possible to avoid crosstalk. For example, the dual-channel can select left-handed (LCP) and right-handed (RCP), and the three-channel can select three linear polarizations of 0°, 120° and 240°.

[0017] Furthermore, the workflow of the artificial intelligence algorithm platform constructed in the above step 3) mainly includes four parts: data set construction, forward neural network and reverse neural network training, metasurface design and device performance evaluation.

[0018] During the dataset construction process, the geometric parameters of the nanostructure can be randomly generated, and numerical simulations can be performed using the finite-difference time-domain (FDTD) method to obtain the corresponding optical response data of the nanostructure. The optical response includes the amplitude and phase data of the reflected light corresponding to the multiplexed polarization state at the operating wavelength. The optical response data and the geometric parameter data of the nanostructure together constitute the dataset. The size of the dataset can affect the training effect of the neural network. Too small a dataset may lead to underfitting and reduce network performance, while too large a dataset consumes more computing resources. Typically, a dataset size of thousands to tens of thousands can meet the training requirements of the neural network. The dataset is then used to train a forward neural network to construct a forward prediction model from the nanostructure geometric parameters to the reflected light amplitude and phase data. To address the problem of difficulty in reducing the loss function during neural network training due to phase periodicity, the phase data values ​​actually used in the network are the sine and cosine values ​​of the phase. The trained forward neural network is then attached to the reverse neural network, and the reverse neural network is trained with the dataset to construct a reverse generation model from the optical response data to the nanostructure geometric parameters. During the training of the reverse neural network, a forward neural network with fixed network parameters is introduced to assist in training. This is because different nanostructures with significant differences can have similar or identical optical responses, making it difficult to directly train an inverse neural network to map the optical response to the geometric structure. During neural network training, gradient descent is used to adjust the network node weights so that the neural network output approaches the true value of the dataset.

[0019] The trained inverse neural network can generate the geometric parameters corresponding to the nanostructures at different spatial positions of the metasurface one by one according to the phase profile of the multifunctional OAM, thereby completing the design of the metasurface. Due to the powerful parallel computing capability of the neural network, the method of the present invention can design a metasurface composed of millions of nanostructures within a time of the order of milliseconds. After completing the structural design of the target multifunctional vortex light generator, the forward neural network can be used to evaluate the performance of the device. That is, for the nanostructure designed by the inverse neural network, the forward neural network can quickly predict its phase and amplitude. The predicted phase can be compared with the target phase. The closer the phase value is, the higher the mode purity of the generated vortex light. The amplitude can evaluate the working efficiency of the device. The larger the amplitude, the higher the intensity of the generated vortex light under the same incident light intensity.

[0020] The design method of the multifunctional vortex light generator proposed in the present invention is the first method to use deep neural networks to reversely design metasurfaces and obtain polarization-multiplexed vortex light. The method of the present invention quickly generates a large number of nanostructures that meet the requirements through an inverse neural network based on the phase requirements of polarization-multiplexed multifunctional vortex light. The method of the present invention quickly evaluates the phase accuracy and amplitude efficiency of a given nanostructure through a forward neural network. The method proposed in the present invention is flexible in design and can extract and calculate the corresponding data set for training from the initial database according to the required working wavelength and multiplexed polarization, thereby completing the training of the neural network. The trained inverse neural network quickly completes the reverse design of all nanostructures, greatly improving the design efficiency and reducing the required computing resources and computing time. The multiple information capacity and high design efficiency will boost the generation of a large number of multifunctional vortex lights, which can be applied to high-capacity information communications and quantum optics with high quantum numbers, and has a very broad market application prospect.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] Compared to single-function vortex light generators, the proposed method leverages deep learning to achieve multifunctional integration within a single device. This facilitates the exploitation of the potential infinite-dimensional Hilbert space of vortex light, significantly promoting its application in information optics and advancing its applications in cutting-edge fields such as photonic chips and quantum optics. Furthermore, by constructing the initial dataset using a multi-wavelength Jones matrix, it ensures transferable applications at other operating wavelengths and with other multiplexed polarizations, eliminating the need to collect datasets from scratch for new nanostructures through numerical simulation.

[0023] Compared to traditional reverse design methods, the method of the present invention does not require physical intuition or known physical knowledge, and the optical properties of the nanostructure are developed to the extreme through a neural network algorithm. As long as it is within the scope allowed by physical principles, all possible optical responses can be reverse-engineered through neural networks. The trained neural network can generate a large number of multifunctional vortex lights in a short period of time, without the need to reverse-engineer the parameters of the nanostructure one by one as in traditional methods, thus greatly reducing the consumption of computing resources and computing time.

[0024] Compared to directly training the network using phase, this method uses the sine and cosine values ​​of the phase as input and output, effectively avoiding the challenges caused by phase periodicity. Essentially, this is because there are some phases with different values ​​but the same physical meaning, such as -π and π. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1Schematic diagram and design principle diagram of generating vortex light with a multifunctional metasurface in a specific embodiment of the present invention, wherein: (a) is a functional schematic diagram of the device; (b) shows the phase accumulation of vortex light with different topological charges after one revolution. Since the phase has a periodicity of 2π, the phase of the dual-function vortex light can be folded into the 2π×2π phase space; (c) shows the phase space corresponding to the vortex light of different multiplexed channels; (d) is an algorithm framework diagram of the deep neural network.

[0026] Figure 2 This is an algorithm flow chart for a specific embodiment of the present invention, wherein: (a)-(b) are data set displays, including a schematic diagram of the device unit structure (a) and its optical response (b); (c) is the main flow chart for the design of a multifunctional vortex light generator, including the target phase profile, the architecture of the reverse and forward neural networks, and a schematic diagram of the final device structure.

[0027] Figure 3 Schematic diagram of the dual-function vortex light with an operating wavelength of 1064nm in a specific embodiment of the present invention, wherein: (a)-(d) are the nanostructures generated by the reverse neural network in the phase space corresponding to the dual-function vortex light generator, and the amplitude efficiency (a)-(b) and phase (c)-(d) prediction results given by the forward neural network; (e) shows the phase prediction values ​​corresponding to the eight dual-function vortex light generators designed by the algorithm platform.

[0028] Figure 4 Figures 1 and 2 show images of experimental samples and interferometry results for a specific embodiment of the present invention operating at a wavelength of 1064 nm. (a) is a scanning electron microscope image of the fabricated device, with the scale bar representing 1.5 μm; (b) is a zoomed-in image of the boxed area in (a), with the scale bar representing 0.4 μm; (c)-(e) show the interference patterns between the generated vortex beam and the spherical reference beam, with the experimental results shown on the left and the theoretical calculations derived from the forward network predictions on the right. The scale bar represents 10 μm.

[0029] Figure 5 Schematic diagram of the three-function vortex light in a specific embodiment of the present invention, wherein: (a) is a schematic diagram of three working linear polarization states; (b) is a schematic diagram of the three-function phase space; (c) is an algorithm framework diagram of the three-function vortex light generator; and (d) is the interference pattern of the vortex light generated by four three-channel vortex light generators.

[0030] Figure 6Schematic diagram of the dual-function vortex light with an operating wavelength of 1550nm in a specific embodiment of the present invention, wherein: (a)-(d) are the nanostructures generated by the reverse neural network in the phase space corresponding to the dual-function vortex light generator, and the amplitude efficiency (a)-(b) and phase (c)-(d) prediction results given by the forward neural network; (e) shows the phase prediction values ​​corresponding to the three dual-function vortex light generators designed by the algorithm platform.

[0031] Figure 7 Schematic diagram of the dual-function vortex light with an operating wavelength of 700nm in a specific embodiment of the present invention, wherein: (a)-(d) are the nanostructures generated by the reverse neural network in the phase space corresponding to the dual-function vortex light generator, and the amplitude efficiency (a)-(b) and phase (c)-(d) prediction results given by the forward neural network; (e) shows the phase prediction values ​​corresponding to the three dual-function vortex light generators designed by the algorithm platform. DETAILED DESCRIPTION

[0032] The present invention will be described in further detail below with reference to the accompanying drawings through several design cases of a multifunctional vortex light generator in the near-infrared band, so that those skilled in the art can more clearly understand the present invention.

[0033] Functional schematic diagram of the multifunctional vortex light generating device Figure 1 As shown in (a), dual-function vortex light is realized by two circularly polarized lights. In the scheme of the present invention, any two orthogonal polarization states can be used to realize dual-function vortex light with polarization multiplexing. Here, we use left-handed light and right-handed light to demonstrate the design of dual-function vortex light. When the incident light is left-handed light, the right-handed component of its reflected light is a beam of vortex light; when the incident light is right-handed light, the left-handed component of its reflected light is another independent beam of vortex light. Taking the function of incident right-handed light and emitted left-handed light as an example, the vortex light carries a spiral phase factor where l LR is the topological charge, defined as:

[0034]

[0035] where θ is the angular coordinate of the beam cross section, Under the illumination of right-circularly polarized (RCP) incident light, the interaction between the nanostructure and the incident light causes the optical phase mutation of left-circularly polarized (LCP) reflected light. This loop integral can follow any path surrounding the phase singularity and satisfy:

[0036]

[0037]

[0038] like Figure 1 As shown in (b), a two-dimensional phase space is established with the dual-functional phase of each nanostructure. To realize any dual-functional vortex light, the required nanostructure forms a line segment starting from the origin in the phase space. Due to the periodicity of the phase, the eight samples in the figure are folded into a square with a side length of 2π. Figure 1 As shown in (c), for single-function vortex light, the phase of the required nanostructure is on a line segment with a length of 2π; for dual-function vortex light, the phase of the required nanostructure is within a square with a side length of 2π; for triple-function vortex light, the phase of the required nanostructure is within a cube with a side length of 2π. Figure 1 As shown in (d), the trained neural network consists of a reverse neural network and a forward neural network. The reverse neural network generates structural parameters based on the desired optical response, and the forward neural network then verifies the optical response.

[0039] In this example, the initial data set consists of 10,000 random nanostructures and their optical responses. The V-shaped nanostructure is a unit structure of a metal-dielectric-metal (MIM) metasurface with a period of 400 nm. Figure 2 As shown in (a), its geometric structure is parameterized into four quantities (w1, w2, a, b), where w1 and w2 are the lengths of the two nanorods that form the V shape, a is the azimuth angle of the angle bisector between the two nanorods, and b is the angle between the two nanorods. Figure 2 As shown in (b), the Jones matrix of the nanostructure at 530nm to 1550nm is obtained by FDTD, and the phase mutation of the dual channel is further calculated. And amplitude efficiency η. The data set is divided into 90% and 10% ratios to train the network, respectively as the training set for training the network and the test set for verifying the network performance. For the forward neural network, the input is a one-dimensional tensor (w1, w2, a, b) with a length of 4, and the output structure phase and amplitude data The network is trained by minimizing the root mean square error (MSE) between the output and the optical properties in the dataset. For the training of the reverse neural network, the trained forward neural network is connected to it and the forward neural network parameters are fixed. The reverse neural network is input, and the four structural parameters output by it are then used by the forward neural network to predict its optical response. The reverse neural network training is completed by minimizing the difference between the target response and the predicted result. The computing platform for completing the training has the following workflow: Figure 2As shown in (c), the target phase of each nanostructure is obtained based on the phase requirements of the dual-channel vortex light. The inverse neural network generates structural parameters, and the forward neural network predicts the optical response. In this example, both the forward neural network predictions and the FDTD simulation results meet the target, demonstrating the reliability of this method.

[0040] Taking the 1064nm operating wavelength as an example, the entire phase space is reversely designed and predicted by the forward neural network. Figure 3 As shown in (a) and (b), each graph shows uniform and efficient characteristics, indicating the excellent efficiency of each channel; at the same time, the similarity between the two indicates the uniformity of efficiency between different channels. Figure 3 As shown in (c) and (d), this method is accurate for the reverse design of any phase point, indicating that this method can be used to design any dual-channel vortex light. Figure 3 (e) shows Figure 1 The predicted results of the reverse design of 8 samples shown in (b) show the reliability of this method. 3 samples were processed by electron beam etching. Figure 4 The results of the scanning electron microscope (SEM) images (a) and (b) show that the processing results are highly consistent with the shape definition of the V-shaped antenna. The spherical reference light and the sample reflected light are interfered by the Michelson interferometer. The results are as follows Figure 4 The experimental results on the left in panels (c)-(e) are shown, while the calculated results on the right are the predictions of the forward neural network and the results of spherical wave interferometry. The number of cantilevers is consistent with the absolute value of the topological charge, and the cantilever orientation is consistent with the sign of the topological charge. The clear target vortex characteristics demonstrate the reliability of this method.

[0041] like Figure 5 As shown in (a) and (b), three channels of vortex light are realized by three linearly polarized incident light with an angle of 120°. The phases of all structures are folded into a cube with a side length of 2π. It is not necessary to re-simulate the data set. The initial data set of the two channels is used to obtain the neural network for training the three channels using the Jones matrix. Figure 5 As shown in (c), the input of the reverse neural network is adjusted to 6 components, corresponding to the sine and cosine of the three-channel phase, plus the amplitude of the three channels, which is the output of the forward neural network. The training of the three-channel network is completed with a similar training method as the two-channel network. The four samples are shown as follows Figure 5 As shown in (d). The number of cantilevers corresponds to the absolute value of the topological charge, and the direction corresponds to the sign of the topological charge, demonstrating the reliability of this method. Figure 3 akin, Figure 6 The results for the 1550nm operating wavelength are shown. Figure 7Results for a 700nm operating wavelength are presented. Both demonstrate the transferability of this method to other operating wavelengths. The above design examples demonstrate the flexibility of our method, which can be transferred to multifunctional vortex light defined by other polarization states and other operating wavelengths based on a single initial dataset.

[0042] The present invention combines bidirectional neural networks for the first time to design polarization-multiplexed multifunctional vortex light. With the help of deep neural networks, the design potential of the structural parameter space is fully explored, and the phase of the same nanostructure under different polarization incidences is independently modulated. In addition, the method of the present invention can be easily migrated to multifunctional vortex light defined by various polarizations and arbitrary wavelengths in the near-infrared band. Compared with traditional methods, the method of the present invention saves a lot of computing time and resources, promotes the large-scale application of multifunctional vortex light, and fully explores the application potential of the infinite-dimensional Hilbert space of vortex light in information optics. Therefore, the method of the present invention has promoted the development of vortex light in cutting-edge fields such as optical communications, photonic chips and quantum optics.

[0043] Finally, it should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. Those skilled in the art should understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments; the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

Claims

1. A method for designing a multifunctional vortex light generator, wherein the multifunctional vortex light generator is a metasurface composed of a large number of different nanostructures, wherein the design of the metasurface comprises the following steps: 1) Based on the desired operating wavelength of the vortex light generated and the device fabrication process, the overall size of the vortex light generator, the unit period of the constituent units, and the minimum size of the nanostructure are determined. Based on the above conditions, the geometric constraints for generating the nanostructure are determined to complete the parameterization of the nanostructure. The unit period of the constituent units is set at a subwavelength scale, and the nanostructure is a V-shaped nanostructure. Its geometric structure is parameterized into four quantities (w1, w2, a, b), where w1 and w2 are the lengths of the two nanorods forming the V-shape, a is the azimuth of the angle bisector of the angle between the two nanorods, and b is the angle between the two nanorods. 2) Determine the polarization state and topological charge of the vortex light in different polarization channels based on the function of the multifunctional vortex light generator, further complete the design of the phase profile corresponding to the multifunctional vortex light generator, and then determine the optical properties that the nanostructures at different locations on the metasurface need to meet based on the phase profiles corresponding to the vortex light in different polarization channels; 3) Constructing an artificial intelligence algorithm platform through a deep neural network, and then using this algorithm platform to find the nanostructure with specific optical properties required in step 2) to complete the metasurface design; the artificial intelligence algorithm platform is composed of a forward neural network and a reverse neural network, wherein the forward neural network predicts the optical properties of the nanostructure based on its geometric parameters, and the reverse neural network reversely generates structural geometric parameters that meet the conditions based on the optical properties required by the nanostructure; this step includes: Constructing a data set: first randomly generate the geometric structural parameters of the nanostructure, and perform numerical simulation through the time-domain finite difference method to obtain the optical response data corresponding to the nanostructure, the optical response includes the amplitude and phase data of the reflected light corresponding to the multiplexed polarization state at the working wavelength, and the optical response data and the geometric parameter data of the nanostructure together constitute a data set; Training the forward neural network through the data set to construct a forward prediction model from the geometric parameters of the nanostructure to the amplitude and phase data of the reflected light; Attaching the trained forward neural network to the reverse neural network, and training the reverse neural network with the data set to construct a reverse generation model from the optical response data to the geometric parameters of the nanostructure; The trained inverse neural network generates the geometric parameters corresponding to the nanostructures at different spatial locations on the metasurface based on the phase profile of the target multifunctional vortex light, thereby completing the metasurface design. The sine and cosine values ​​of the phase are used as input and output phase data when training the neural network. 4) Use the algorithm platform constructed in step 3) to obtain the optical response data of the metasurface and evaluate the device performance.

2. The design method according to claim 1, wherein: The multifunctional vortex light generator can generate a variety of vortex lights containing different topological charges under multiple different polarization channels. The optical properties of the nanostructures constituting the metasurface in step 2) need to simultaneously satisfy the phases of the reflected light corresponding to different polarization channels.

3. The design method according to claim 1, wherein: Step 4) For the nanostructure geometric parameters generated by the inverse neural network, the forward neural network predicts their phase and amplitude. The predicted phase is compared with the target phase to verify the required phase profile and determine whether it meets the design expectations. The device's operating efficiency is evaluated by the amplitude, that is, the intensity information of the generated vortex light.

4. The design method according to claim 1, wherein: In the phase profile design in step 2), the incident polarizations under different functions are as orthogonal as possible or evenly distributed.

5. The design method according to claim 4, wherein: In step 2), for the dual-function vortex light generator, the dual polarization channels select left-hand polarization and right-hand polarization; for the triple-function vortex light generator, the triple polarization channels select three linear polarizations of 0°, 120°, and 240°.

6. The design method according to claim 1, wherein: In step 3), the initial data set is constructed using a multi-wavelength Jones matrix.

Images