Spectrum polarization encoding metasurface device and reverse design optimization method thereof

Abstract

The application discloses a kind of spectral polarization encoding metasurface devices and its reverse design optimization method, it is related to spectral polarization imaging field, alleviate the existing spectral polarization encoding metasurface design technique and the problem of low design efficiency with high computing cost.One kind of spectral polarization encoding metasurface device, structural unit A and structural unit B are used, and the two structures are respectively used double split ring and L type + split ring structure.A kind of reverse design optimization method of spectral polarization encoding metasurface device, forward network and reverse network are combined, and the reverse network is used to optimize the initial structure parameter and related coefficient in coordination, and the obtained four groups of structure parameters are sent to forward network module;Forward network is used to predict structure parameters, and the correlation coefficient of initial structure parameter and four optimized output structure parameters and spectral response result are obtained as reverse design optimization result.The method disclosed in the application is suitable for the field of spectral polarization encoding metasurface and design method thereof.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spectral polarization imaging, in particular to the field of metasurfaces. BACKGROUND

[0002] Spectral polarization imaging plays an important role in target detection. It combines spectral information and polarization information of a target, significantly expands the information perception dimension of the target, and exhibits unique advantages in improving target recognition and detection capability. However, current spectral polarization imaging systems rely heavily on the combination of spectral elements and polarization elements such as gratings, interferometers, filters and polarizers, resulting in a large system volume, a complex structure, a limited imaging efficiency, and difficulty in ensuring measurement accuracy.

[0003] The emergence of metasurface technology provides new possibilities for innovative design of spectral polarization imaging systems. By carefully designing and reasonably arranging nanostructures, spectral and polarization information can be directly encoded in the optical field, and complete spectral and polarization information can be extracted through decoding. This strategy significantly reduces the system volume, simplifies the optical path design, and improves the imaging efficiency and real-time measurement capability. However, traditional metasurface design methods highly depend on researchers' prior knowledge and are optimized through iterative trial and error of full-wave simulation. Especially when encoding spectral and polarization multidimensional information, nanostructure parameters need to be optimized for different wavelengths and polarizations, resulting in a sharp increase in computational cost and low optimization efficiency.

[0004] In recent years, deep learning (DL) has shown great potential in metasurface design, providing an efficient design paradigm. Among them, various generative networks (including generative adversarial networks, variational autoencoders, diffusion models) and Transformer architecture are used to realize high-dimensional mapping of metasurfaces. However, in practical applications, these networks are difficult to achieve effective balance between different networks, and introduce many hyperparameters, making the computational complexity rise, the network difficult to converge, and ultimately affecting the actual availability of the model.

[0005] In summary, the existing spectral polarization encoding metasurface design technology has the problems of high computational cost and low design efficiency. SUMMARY

[0006] The present application alleviates the problem of high computational cost and low design efficiency of the existing spectral polarization encoding metasurface design technology. The present application provides the following solutions:

[0007] Scheme one, a spectral polarization encoding metasurface device, the metasurface device is arranged in an array structure by a plurality of metasurface structure units, the metasurface structure units have two types, which are structure unit A or structure unit B respectively; the structure unit A and the structure unit B are both double-atom nanocolumn structures with the same period; the structure unit A and the structure unit B sequentially include a dielectric layer and a substrate layer from top to bottom;

[0008] The structure unit A includes two split ring cross-section atom nanocolumns A1 and A2; the atom nanocolumn A1 is in the shape of a split ring, and the geometric size parameters include an opening angle θ1, a rotation angle θ2, an outer radius r1 and an inner radius r2; the atom nanocolumn A2 is in the shape of a split ring, and the geometric size parameters include an opening angle θ3, a rotation angle θ4, an outer radius r3 and an inner radius r4;

[0009] The structure unit B includes an L-shaped cross-section atom nanocolumn B1 and a split ring cross-section atom nanocolumn B2;

[0010] The atom nanocolumn B1 is in the shape of a right angle, and the geometric size parameters include horizontal lengths L1 and L3, and vertical lengths L2 and L4; the atom nanocolumn B2 is in the shape of a split ring, and the geometric size parameters include an opening angle θ5, a rotation angle θ6, an outer radius r5 and an inner radius r6.

[0011] Further, in an embodiment of the present application, the material of the dielectric layer is titanium dioxide material;

[0012] The material of the substrate layer is silicon oxide material.

[0013] Scheme two, a reverse design optimization method of a spectral polarization encoding metasurface device, the spectral polarization encoding metasurface device is the spectral polarization encoding metasurface device in scheme one, inputting the input structure parameters and the target correlation coefficient of the spectral polarization encoding metasurface device into a reverse design optimization model to obtain the correlation coefficient of the input structure parameters and four optimized output structure parameters and the spectral response result as the reverse design optimization result;

[0014] The structure parameters include the type of the metasurface structure unit and the corresponding geometric size parameters.

[0015] Further, in an embodiment of the present application, the training set of the reverse design optimization model includes a structure unit A training set and a structure unit B training set:

[0016] The structure unit A and the structure unit B are distinguished by introducing 4-dimensional Gaussian parameters with different mean values but the same standard deviation in front of the geometric size parameters, and the 12-dimensional geometric size parameters after introducing the 4-dimensional Gaussian parameters are taken as network training set data.

[0017] Further, in an embodiment of the present application, the inverse design optimization model comprises:

[0018] an inverse network module, configured to obtain four output structural parameters satisfying the target correlation coefficient based on the input structural parameters, and send the four output structural parameters to a forward network module;

[0019] the forward network module, configured to respectively predict the four output structural parameters to obtain corresponding correlation coefficients and spectral response results;

[0020] The spectral response results comprise transmittance curves of 0° polarization state, 45° polarization state, 90° polarization state, 135° polarization state, left circular polarization LCP and right circular polarization RCP.

[0021] Further, in an embodiment of the present application, the inverse network module adopts a multi-task learning strategy.

[0022] Further, in an embodiment of the present application, the forward network module adopts a multi-task learning strategy and a self-encoder network.

[0023] Further, in an embodiment of the present application, the inverse design optimization model is obtained by the following training method:

[0024] Step S01, training the forward network by using a forward network loss function to obtain a trained forward network;

[0025] The forward network loss function is:

[0026]

[0027] wherein, is the number of forward network tasks; the subscript represents the th task; is a forward network weight factor; is a predicted spectral response result; is a forward network function; is a structural parameter; is a weight; is a bias;

[0028] Step S02, connecting the obtained forward network and inverse network in series to obtain a series network, training the series network by using an inverse design optimization loss function to obtain a trained series network as an inverse design optimization model;

[0029] The inverse design optimization loss function is:

[0030]

[0031] wherein, is a latent vector weight factor; is a correlation coefficient weight factor; is a series network task number; is a reverse network weight factor; is a series network weight factor; is a latent vector of the predicted 4-group output structural parameters; represents a reverse network function; is a predicted correlation coefficient; is a correlation coefficient, is a series network function.

[0032] The spectral polarization coding metasurface device and the reverse design optimization method thereof disclosed by the application alleviate the problems of high calculation cost and low design efficiency existing in the prior art spectral polarization coding metasurface design technology. The specific beneficial effects include:

[0033] 1. The spectral polarization imaging system in the prior art adopts a structure composed of a prism or a grating combined with a polarization device, resulting in a large system volume, a complex structure, and limited imaging efficiency. The spectral polarization coding metasurface device proposed by the application does not need to be regulated by a dispersion device and a polarization device, has a smaller volume, and has higher imaging efficiency. The spectral polarization coding metasurface device proposes two structures, i.e., structure unit A and structure unit B. The two structures adopt a double split ring and an L type + split ring structure, respectively. By adjusting the opening angle and ring width of the split ring and the length and width size of the door gap of the L type, the electromagnetic coupling strength and resonance frequency of the resonant unit can be effectively adjusted, the anisotropy and chirality response required can be realized, and finally the metasurface device with different spectral and polarization responses can be realized.

[0034] 2. Based on the spectral polarization coding metasurface device of the application, the reverse design method composed of a forward network and a reverse network is highlighted. The reverse network is used for collaborative optimization of initial structural parameters and correlation coefficients. The obtained optimized four groups of structural parameters are sent to the forward network module.

[0035] The forward network is used for predicting the structural parameters. The initial structural parameters and the 0°, 45°, 90°, 135°, left circularly polarized light LCP and right circularly polarized light RCP transmittance data of the four output structural parameters of the reverse network are simultaneously output in the form of multi-task learning, and the correlation coefficients of the initial structural parameters and the four output structural parameters are further calculated.

[0036] The inverse design method can be applied to optimization of various super surface structures, and directly predicts the structure parameters meeting the correlation coefficient requirements. The inverse design optimization method can directly obtain the structure parameters meeting the given correlation coefficient with the input structure parameters, without relying on the iteration and trial of full-wave simulation for optimization.

[0037] The method described in the application is suitable for the field of spectral polarization coding super surface and its design method. BRIEF DESCRIPTION OF DRAWINGS

[0038] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the following drawings, in which:

[0039] Figure 1 is a planar schematic diagram of the super surface structure unit described in embodiment one, wherein (a) is a schematic diagram of structure unit A, and (b) is a schematic diagram of structure unit B.

[0040] Figure 2 is a three-dimensional schematic diagram of structure unit A described in embodiment one.

[0041] Figure 3 is a three-dimensional schematic diagram of structure unit B described in embodiment one.

[0042] Figure 4 is a three-dimensional schematic diagram of the spectral polarization coding super surface device including structure unit A and structure unit B described in embodiment one.

[0043] Figure 5 is a flowchart of the inverse design optimization model described in embodiment five.

[0044] Figure 6 is the spectral response result of structure unit A described in embodiment nine under different structure parameters.

[0045] Figure 7 is the spectral response result of structure unit B described in embodiment nine under different structure parameters.

[0046] Figure 8 is the training result diagram of the autoencoder network of the forward network module described in embodiment ten.

[0047] Figure 9 is the training loss result of the forward network described in embodiment ten.

[0048] Figure 10 is the training result of the forward network described in embodiment ten.

[0049] Figure 11 is the training loss result of the series network described in embodiment ten.

[0050] Figure 12 This is a comparison chart of the input structural parameters (A)para1:[95.00, 60.00, 75.00, 30.00, 180.00, 110.00, 50.00, -110.00] of the reverse network module described in Implementation Method 10, and the parameter results with a target correlation coefficient C of 0.1278.

[0051] Figure 13 This is a comparison chart of the input structural parameters (A)para1: [85.00, 50.00, 80.00, 60.00, 90.00, 0.00, 0.00, 0.00] predicted by the reverse network module described in Implementation Method 10, and the parameter results with a target correlation coefficient C of 0.2992.

[0052] Figure 14 This is a comparison chart of the spectral response results of the input structural parameter (A)para1:[95.00, 60.00, 75.00, 30.00, 180.00, 110.00, 50.00, -110.00] and the four output structural parameters described in Implementation Method 10, with the simulation results.

[0053] Figure 15 This is a comparison chart of the spectral response results of the input structural parameter (A)para1: [85.00, 50.00, 80.00, 60.00, 90.00, 0.00, 0.00, 0.00] and the four output structural parameters described in Implementation Method 10, and the simulation results.

[0054] Figure label:

[0055] Dielectric layer 1; Substrate layer 2. Detailed Implementation

[0056] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0057] Implementation Method 1: The spectral polarization-encoded metasurface device described in this implementation method, such as... Figures 1 to 4 The metasurface device consists of an array of multiple metasurface structural units. There are two types of metasurface structural units: structural unit A or structural unit B. Both structural unit A and structural unit B are two-atom nanopillar structures with the same period. From top to bottom, structural unit A and structural unit B include a dielectric layer 1 and a substrate layer 2.

[0058] The structural unit A includes two atomic nanorods A1 and A2 with split-ring cross sections; the atomic nanorod A1 has a split-ring shape, and geometric size parameters include an opening angle θ1, a rotation angle θ2, an outer radius r1, and an inner radius r2; the atomic nanorod A2 has a split-ring shape, and geometric size parameters include an opening angle θ3, a rotation angle θ4, an outer radius r3, and an inner radius r4.

[0059] The structural unit B includes one atomic nanorod B1 with an L-shaped cross section and one atomic nanorod B2 with a split-ring cross section.

[0060] The atomic nanorod B1 has a right-angle shape, and geometric size parameters include horizontal lengths L1 and L3, and vertical lengths L2 and L4; the atomic nanorod B2 has a split-ring shape, and geometric size parameters include an opening angle θ5, a rotation angle θ6, an outer radius r5, and an inner radius r6.

[0061] This embodiment is a further limitation of the spectral polarization encoding metasurface device, and the structural unit A and the structural unit B are exemplified. Both the structural unit A and the structural unit B regulate their optical responses through the breaking of in-plane symmetry, and the different spectral and polarization responses of the shape mode are affected by the complex scattering and coupling of the Bloch modes propagating in the metasurface layer, so that the units exhibit different spectral behaviors under different polarization states, thereby widening the functional design space of the structure.

[0062] Specifically, the atomic nanorods A1 and A2 in the structural unit A; and the atomic nanorods B1 and B2 in the structural unit B. Both the two structural units adopt a double-split ring and split gate ring structure, and are both double-atom structures, and different spectral and polarization information responses are obtained by breaking the face symmetry distribution.

[0063] The spectral polarization encoding metasurface device includes the structural unit A and the structural unit B, selects structural parameters with strong spectral and polarization responses, and ensures low correlation between each other, so as to combine to form a metasurface device with strong spectral polarization information encoding capability. Compared with only using a single structural unit, using two unit structures increases the structural parameters with different spectral and polarization responses, strengthens the spectral and polarization response capability, and increases the probability of spectral polarization information encoding capability. At the same time, the metasurface device can be extended to use three or more unit structures, but greatly increases the design cost, and the processability risk probability is further increased.

[0064] By adjusting the geometric size parameters of the above-mentioned atomic nanorods, the spectral response and polarization characteristics can be flexibly regulated, thereby constructing a metasurface unit with unique optical behavior, and realizing sensitive response to different polarized light.

[0065] Implementation Method Two: This implementation method further defines the spectral polarization-encoded metasurface device described in Implementation Method One. In this implementation method, as... Figure 2 and Figure 3 As shown, the material of the dielectric layer 1 is titanium dioxide.

[0066] The substrate layer 2 is made of silicon oxide.

[0067] This embodiment further defines the spectral polarization-coded metasurface device, providing an example of dielectric layer 1 and substrate layer 2. Dielectric layer 1 is made of TiO2 material with high refractive index and low loss, due to its excellent electromagnetic response performance in the visible and near-infrared bands and good process compatibility in photolithography and deposition processes. Furthermore, dielectric layer 1 is placed on a SiO2 substrate, ensuring not only the optical stability of the entire structure but also facilitating subsequent fabrication and integration.

[0068] Implementation Method 3: This implementation method further defines the spectral polarization-coded metasurface device described in Implementation Method 1. In this implementation method, the spectral polarization-coded metasurface device is any one of the spectral polarization-coded metasurface devices described in Implementation Method 1 to 2, such as... Figure 4 As shown, the input structural parameters and target correlation coefficients of the spectral polarization encoded metasurface device are input into the reverse design optimization model, and the correlation coefficients and spectral response results of the input structural parameters and four optimized output structural parameters are obtained as the reverse design optimization results.

[0069] The structural parameters include the type of the metasurface structural unit and its corresponding geometric dimensions.

[0070] In this embodiment, the design principle of the target correlation coefficient is as follows:

[0071] Mathematical models are used to elucidate the spectral and polarization encoding principles of metasurfaces, thus providing theoretical guidance for determining the target correlation coefficient. Most polarization measurements utilize Stokes parameters. To describe the total polarization information. The Stokes parameter is expressed as...

[0072] (1)

[0073] in, Represents linearly polarized light at 0°, 90°, 45°, and 135°; This indicates the intensity of right-handed and left-handed circularly polarized light.

[0074] Therefore in At wavelength, the polarization of the entire Stokes spectrum can be determined by the Mueller matrix. Contact

[0075] (2)

[0076] where, and denote the incident and emergent Stokes parameters at different wavelengths, respectively.

[0077] Since the detector can only receive intensity information, we use the first row of the Mueller matrix to represent the output spectral polarization information. The light intensity after the th super surface structure unit is represented as

[0078] (3)

[0079] where, is the number of spectral channels; ; denotes the noise of each super surface structure unit corresponding to the detector pixel. Then The spectral polarization transmission equation composed of

[0080] (4)

[0081] where,

[0082] (5)

[0083] From the above formula, it can be seen that the spectral polarization information can be obtained directly by inversely solving the spectral polarization encoding matrix .

[0084] However, in order to save space and reduce operating costs, the number of super surface structure units (n) is usually much smaller than the number of spectral channels (m). Therefore, in such a framework, in order to enable the super surface array to capture sufficient spectral and polarization information and achieve accurate reconstruction, it is necessary to design highly anisotropic super surface structure units with different spectral and polarization characteristics, which means each row must exhibit as much mutual difference as possible to ensure sufficient information decoupling and reconstruction accuracy. The specific expression is as follows

[0085] (6)

[0086] where, denote the transmittance of 0°\90°\45° and left-handed circularly polarized light, respectively.

[0087] In order to ensure the accuracy and stability of the reconstruction process, it is necessary to design a group of highly anisotropic super surface units with significant differences in spectral and polarization response, that is, The correlation coefficient between rows should be as small as possible. Defined as follows

[0088] (7)

[0089] in, and for The row and number The rows are the i-th metasurface structural unit and the j-th metasurface structural unit; and These represent covariance and standard deviation, respectively. A larger value indicates a greater correlation, smaller differences between the two sides, and lower accuracy in spectral polarization reconstruction; conversely, a smaller value indicates a smaller correlation, greater differences between the two sides, and higher accuracy in spectral polarization reconstruction.

[0090] Implementation Method Four: This implementation method further defines the reverse design optimization method described in Implementation Method Three. In this implementation method, the training set of the reverse design optimization model includes a training set of structural unit A and a training set of structural unit B.

[0091] The structural unit A and the structural unit B are distinguished by introducing 4-dimensional Gaussian parameters with different means but the same standard deviation before the geometric dimension parameters, and the 12-dimensional geometric dimension parameters after introducing the 4-dimensional Gaussian parameters are used as the network training set data.

[0092] This embodiment further defines the reverse design optimization model and provides an example of the training set for the reverse design optimization model. This method improves the ability of the deep learning model to distinguish different metasurface structure data by generating Gaussian distributions with different parameters, enabling it to accurately predict the corresponding spectral data for input parameters of different structures.

[0093] Implementation Method Five: This implementation method further defines the reverse design optimization method described in Implementation Method Three. In this implementation method, as... Figure 5 As shown, the reverse design optimization model includes:

[0094] The reverse network module is used to obtain four output structure parameters that satisfy the target correlation coefficient based on the input structure parameters, and send the four output structure parameters to the forward network module;

[0095] The feedforward network module is used to predict the four output structural parameters respectively to obtain the corresponding correlation coefficients and spectral response results;

[0096] The spectral response result includes transmittance curves of 0° polarization state, 45° polarization state, 90° polarization state, 135° polarization state, left circular polarization LCP and right circular polarization RCP.

[0097] Embodiment six, this embodiment is a further limitation of the reverse design optimization method of embodiment five, in this embodiment, the reverse network module adopts a multi-task learning strategy.

[0098] This embodiment is a further limitation of the reverse design optimization method, and the reverse network module is exemplified. The reverse network adopts a multi-task learning strategy, and outputs a structure parameter corresponding to the correlation coefficient by inputting the correlation coefficient and the structure parameter. If only one structure parameter is output, the model expression ability is limited and there is a high probability that it does not meet the processing requirements. Therefore, the multi-task learning strategy is adopted, and four structure parameters respectively satisfying the correlation coefficient condition of the input structure parameter are output at the same time, which greatly increases the model robustness and parameter availability.

[0099] Embodiment seven, this embodiment is a further limitation of the reverse design optimization method of embodiment five, in this embodiment, the forward network module adopts a multi-task learning strategy and an autoencoder network.

[0100] In this embodiment, the multi-task learning strategy is preferably used to output 200-dimensional transmittance curves of 0°, 45°, 90°, 135°, left circular polarization LCP and right circular polarization RCP, and then the transmittance is upgraded to 501 by the autoencoder network.

[0101] This embodiment is a further limitation of the reverse design optimization method, and the forward network module is exemplified. Since the forward network needs to convert the 1x12-dimensional structure parameter into 501x6 transmittance data, 6 dimensions represent the transmittance of 6 different polarization directions, and there is no correlation between different polarization transmittances. If a single-task network form is adopted, it will be difficult for the network to learn the characteristics of different transmittances. Therefore, a multi-task learning strategy is adopted to predict different polarization states separately. Multiple tasks can adapt to their own transmittance characteristics during the training process, effectively avoiding the mutual interference between polarization states and improving the training accuracy. At the same time, the network is combined with the autoencoder to convert the large dimension into a distributed data processing method, effectively avoiding the instability of the forward network training.

[0102] Embodiment eight, this embodiment is a further limitation of the reverse design optimization method of embodiment three, in this embodiment, the reverse design optimization model is trained by the following training method:

[0103] Step S01, training the forward network by using the forward network loss function to obtain the trained forward network;

[0104] The forward network loss function is:

[0105]

[0106] wherein, is the number of forward network tasks; the lower index represents the th task; is the forward network weight factor; is the predicted spectral response result; is the forward network function; is the structure parameter; is the weight; is the bias;

[0107] Step S02, the obtained forward network and the reverse network are connected in series to obtain a series network, the series network is trained by using a reverse design optimization loss function, and the trained series network is obtained as a reverse design optimization model;

[0108] The reverse design optimization loss function is:

[0109]

[0110] wherein, is the latent vector weight factor; is the correlation coefficient weight factor; =4 is the number of series network tasks; is the reverse network weight factor; is the series network weight factor; is the predicted latent vector of the four groups of output structure parameters; represents the reverse network function; is the predicted correlation coefficient; is the correlation coefficient, is the series network function.

[0111] In the present embodiment, the preferred key parameters for training the forward network in step S01 are: the number of training rounds Epoch is 600, the initial value of the learning rate is 0.001, the optimizer is Adma, the batch size is 128, the number of tasks is 6, and the learning rate changes according to

[0112]

[0113] changes with the number of rounds epoch .

[0114] In this embodiment, the preferred key parameters for training the series network in step S02 are: 500 Epochs, 0.001 initial value of learning rate, Adma optimizer, 256 batch size, 4 tasks, and the learning rate is changed according to

[0115]

[0116] with the number of rounds epoch .

[0117] In this embodiment, the forward network weight factor is preferably 1, which ensures that the model can learn the characteristics of different polarization directions in a balanced manner.

[0118] In this embodiment, the latent vector weight factor is preferably 0.3, and the correlation coefficient weight factor is preferably 0.7.

[0119] In this embodiment, the reverse network weight factor and the series network weight factor are preferably 1, which ensures that the model can learn the characteristics of different groups of data in a balanced manner.

[0120] In this embodiment, the spectral polarization encoding metasurface device is based on the spectral polarization encoding metasurface device of embodiment one, combined with the structural parameters of structural unit A and structural unit B optimized in embodiment two, the hierarchical structure of structural unit A optimized in embodiment three, and the hierarchical structure of structural unit B optimized in embodiment four.

[0121] The spectral response results of structural unit A under different structural parameters are shown in Figure 6 , the spectral response results of structural unit B under different structural parameters are shown in Figure 7 , Figure 6 and Figure 7 show that the transmittance between the two structural units under 0°, 45°, 90°, and 135° polarization states has a large difference, although the difference in transmittance between LCP and RCP is relatively small, but still has a certain difference. It shows that the two proposed metasurface structures have different spectral responses under different polarization conditions.

[0122] Implementation Method 10: The reverse design optimization method for the spectral polarization-encoded metasurface device used in this implementation method is based on the spectral polarization-encoded metasurface device described in Implementation Method 6, combined with the reverse design optimization model optimized in Implementation Method 7, the training set of the reverse design optimization model optimized in Implementation Method 8, the forward network module optimized in Implementation Method 9, and the training method optimized in Implementation Method 10.

[0123] The training result diagram of the autoencoder network of the feedforward network module in this embodiment is shown in the figure below. Figure 8 As shown in the figure, (ad) represents four sets of prediction results. It can be seen from the figure that the autoencoder can accurately predict high-dimensional data from low-dimensional data. While ensuring prediction accuracy, it also avoids the training instability caused by directly transforming the feedforward network from 1×12 dimensions to 501×6 dimensions.

[0124] The training loss result of the feedforward network in this embodiment is as follows: Figure 9 As shown in the figure, the training and validation losses drop sharply before 50 rounds, then the rate of decline slows down after 250 rounds, and finally the loss decreases to 0.001 as the learning rate decreases.

[0125] The training results of the feedforward network in this embodiment are as follows: Figure 10 As shown, the results compare the true and predicted values ​​of different metasurface structures under various structural parameters, with incident light at 0°, 45°, 90°, 135°, and with left-handed circularly polarized (LCP) and right-handed circularly polarized (RCP) light. The results demonstrate that the feedforward network can accurately predict the spectral response under different polarization conditions for various metasurface structures, proving its superiority in data discrimination, prediction accuracy, and robustness.

[0126] The training loss result of the concatenated network in this embodiment is as follows: Figure 11 As shown in the figure, the inverse loss decreases rapidly before 50 rounds, then slowly levels off, reaching its lowest point of 0.0025 around 500 rounds. Although the loss curve fluctuates somewhat within 500 rounds during training, it shows an overall downward trend, indicating that the model is continuously learning and converging. Furthermore, the difference between the training loss and the validation loss is small, indicating that the model has good generalization ability between the training and validation sets and does not exhibit significant overfitting.

[0127] The inverse network module of this embodiment has a set of input structural parameters (A)para1: [95.00, 60.00, 75.00, 30.00, 180.00, 110.00, 50.00, -110.00] and four output structural parameters with a target correlation coefficient C of 0.1278, as shown below. Figure 12As shown, another set of input structural parameters (A) para1 : [85.00, 50.00, 80.00, 60.00, 90.00, 0.00, 0.00, 0.00] and the target correlation coefficient C is 0.2992. The four output structural parameters are as shown in Figure 13 The output structural parameters are compared with the simulation results, which shows that the four sets of predicted geometric parameters are highly consistent with the true values in numerical value, indicating that the constructed inverse network has good generalization ability and stability.

[0128] To verify the physical validity of the predicted parameters, the input structural parameters of the inverse network module of the embodiment Figure 12 The comparison chart of the spectral response results of the input structural parameters and the four output structural parameters with the simulation results of the transmittance distribution of the FDTD algorithm in different polarization directions is as shown in Figure 14 Figure 13 The comparison chart of the spectral response results of the input structural parameters and the four output structural parameters with the simulation results of the transmittance distribution of the FDTD algorithm in different polarization directions is as shown in Figure 15 It can be seen that the transmittance curve calculated based on the predicted parameters is highly consistent with the simulation value, that is, even under different correlation coefficient conditions, the network can still accurately distinguish and predict the corresponding structure type and parameter combination, verifying its robustness to the difference between structures and indicating that the inverse prediction result has physical realizability.​

Claims

1. A spectrally polarizing encoding metasurface device, characterized in that, The super surface device is arranged in an array structure by a plurality of super surface structure units, the super surface structure units include structure unit A and structure unit B; the structure unit A and the structure unit B are both two-atom nanometer column structures; the structure unit A and the structure unit B sequentially include a dielectric layer and a substrate layer from top to bottom; The structure unit A includes two atom nanometer columns A1 and A2 with split ring cross sections; the atom nanometer column A1 has a split ring shape, and geometric size parameters include an opening angle θ1, a rotation angle θ2, an outer radius r1, and an inner radius r2; the atom nanometer column A2 has a split ring shape, and geometric size parameters include an opening angle θ3, a rotation angle θ4, an outer radius r3, and an inner radius r4; The structure unit B includes an atom nanometer column B1 with an L-shaped cross section and an atom nanometer column B2 with a split ring cross section; The atom nanometer column B1 has a right angle shape, and geometric size parameters include horizontal lengths L1 and L3, and vertical lengths L2 and L4; the atom nanometer column B2 has a split ring shape, and geometric size parameters include an opening angle θ5, a rotation angle θ6, an outer radius r5, and an inner radius r6; The structure unit A and the structure unit B both control optical responses through breaking of in-plane symmetry.

2. The spectrally polarisation-encoded metasurface device of claim 1, wherein, The material of the dielectric layer is titanium dioxide material; The material of the substrate layer is silicon oxide material.

3. A method for inverse design optimization of a spectrally polarizing encoding metasurface device, the spectrally polarizing encoding metasurface device being the spectrally polarizing encoding metasurface device of any one of claims 1 or 2, characterized in that, Input the input structure parameters and the target correlation coefficient of the spectrum polarization coding super surface device into the inverse design optimization model to obtain the correlation coefficient of the input structure parameters and the four optimized output structure parameters and the spectrum response result as the inverse design optimization result; The structure parameters include the type of the super surface structure unit and the corresponding geometric size parameters.

4. The inverse design optimization method of claim 3, wherein, The training set of the inverse design optimization model includes a structure unit A training set and a structure unit B training set: The structure unit A and the structure unit B are distinguished by introducing 4-dimensional Gaussian parameters with different mean values but the same standard deviation in front of the geometric size parameters, and the 12-dimensional geometric size parameters after introducing the 4-dimensional Gaussian parameters are used as network training set data.

5. The inverse design optimization method of claim 3, wherein, The inverse design optimization model includes: An inverse network module, configured to obtain four output structure parameters satisfying the target correlation coefficient based on the input structure parameters, and send the four output structure parameters to a forward network module; A forward network module, configured to predict the four output structure parameters respectively to obtain corresponding correlation coefficients and spectrum response results; The spectrum response result includes transmittance curves of 0° polarization state, 45° polarization state, 90° polarization state, 135° polarization state, left circular polarization LCP, and right circular polarization RCP.

6. The inverse design optimization method of claim 5, wherein, The inverse network module adopts a multi-task learning strategy.

7. The inverse design optimization method of claim 5, wherein, The forward network module adopts a multi-task learning strategy and a self-encoder network.

8. The inverse design optimization method of claim 3, wherein, The inverse design optimization model is trained by the following training method: Step S01, training the forward network by using a forward network loss function to obtain a trained forward network; The forward network loss function is: wherein, is the number of forward network tasks; the lower index denotes the task; is the forward network weight factor; is the predicted spectral response result; is the forward network function; is the structure parameter; is the weight; is the bias; In step S02, the obtained forward network and reverse network are connected in series to obtain a series network, the series network is trained by using a reverse design optimization loss function, and a trained series network is obtained as a reverse design optimization model. The reverse design optimization loss function is: wherein, is a latent vector weight factor; is a correlation coefficient weight factor; is a series network task number; is an inverse network weight factor; is a series network weight factor; is a latent vector of the predicted 4-group output structure parameters; denotes an inverse network function; is a predicted correlation coefficient; is a correlation coefficient, is a series network function.

Images