A deep learning based cross-wavelength spectrum prediction method

Abstract

The application discloses a kind of cross wavelength spectrum prediction methods based on deep learning, applied to spectral prediction technical field, constructs training dataset, utilizes modeling hypersurface unit structure in simulation software to obtain optical response result;Generation-elimination network architecture is constructed;Training dataset is input to generation network and elimination network for training, obtains trained generation network and trained elimination network, forms respective latent space, and can produce multiple solutions.After training, at least one group of solutions in the first frequency band candidate solution generated by generation network is consistent with actual real result, and the elimination network also successfully filters out this group of solutions, which is consistent with the true value.The generation-elimination network proposed in the application successfully solves the cross-wavelength optical response prediction and overcomes the difficulty of many-to-many mapping.In addition, the application reduces the dimension of high-dimensional data, and successfully realizes the visualization of the mapping relationship between low-frequency and high-frequency data in the latent space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spectrum prediction, and more particularly to a cross-wavelength spectrum prediction method based on deep learning. BACKGROUND

[0002] Metamaterial is a kind of artificial material with special properties proposed in the 21st century, which can realize optical properties such as negative refractive index that traditional materials do not have. Through specific design, it can control light and electromagnetic waves, break through the limitation of some apparent natural laws, and ultimately realize functions such as electromagnetic invisibility, optical computing, and wireless communication. Electromagnetic metasurface is a two-dimensional metamaterial structure composed of many sub-wavelength electromagnetic supercell structures arranged in a specific way, as shown in FIG. 1. Each unit structure is composed of an upper resonant structure, a middle non-metal layer, and a lower metal layer. According to specific requirements, by designing the upper metasurface unit structure pattern and parameterizing each layer (such as thickness, dielectric constant, etc.), the phase and amplitude of electromagnetic waves in the required waveband are controlled to obtain the corresponding optical response. Figure 1

[0003] By combining deep learning and applying artificial intelligence models to electromagnetic metasurface design, the low efficiency and high limitations of traditional design strategies can be overcome. In the prior art, intelligent electromagnetic metasurface design mainly includes two types: 1. Forward design, which quickly and batch predicts and infers the corresponding optical response according to a specific metasurface unit structure / arrangement through a trained model, rather than calculating the optical response by solving Maxwell's equations through a large number of numerical iterations; 2. Reverse design, which is the opposite of forward design, and designs the metasurface structure / arrangement according to the required optical response through a network model.

[0004] However, the application and development of these two design concepts have basically reached saturation, and nothing more than applying the same or similar models (such as transfer learning models, generative adversarial models, hybrid density models, self-encoder or variational autoencoder models) to different nano or photonic structure design physical scenarios (such as photonic crystals, metamaterials, plasmonic nanostructures), both of which are designed to obtain one item from the other through design parameters and optical responses. No one has proposed to achieve a direct conversion from spectrum to spectrum without design parameters.

[0005] ​Unlike the one-to-one mapping problem in forward design (in which a structure design corresponds to a single optical response, i.e. a single solution) and the one-to-many mapping problem in inverse design (in which a desired optical response can be realized by multiple designs, i.e. multiple solutions), the spectrum-to-spectrum conversion involves a tricky many-to-many mapping relationship. If a traditional fully connected layer network is used to handle the many-to-many mapping problem, the network training result will not converge due to the complex mapping relationship, and the training loss value will remain at a high level, which cannot achieve the cross-wavelength optical response prediction.

[0006] Therefore, how to realize the many-to-many mapping relationship in cross-wavelength spectrum prediction without design parameters is a problem that those skilled in the art need to solve. SUMMARY

[0007] Therefore, the present application provides a cross-wavelength spectrum prediction method based on deep learning, which constructs a generate-eliminate network architecture to realize the cross-wavelength spectrum prediction method, solves the many-to-many mapping problem in cross-wavelength spectrum prediction, and uses a neural network to reduce the dimensionality of high-dimensional data, successfully realizes the visualization of the mapping relationship between the first frequency band data and the second frequency band data in the latent space.

[0008] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0009] A cross-wavelength spectrum prediction method based on deep learning, the specific steps include the following:

[0010] A training data set is constructed, and the optical response result is obtained by modeling the metasurface unit structure in the simulation software. The training data set includes first frequency band training data and second frequency band training data;

[0011] A generate-eliminate network architecture is constructed. The generate-eliminate network architecture includes two sub-networks: a generate network and an eliminate network;

[0012] The training data set is input into the generate network for training to obtain a trained generate network, and the training data set is input into the eliminate network for training to obtain a trained eliminate network. The generate network and the eliminate network form respective latent spaces and can produce multiple solutions;

[0013] The mapping relationship between the original first frequency band data and the set of second frequency band candidate solutions is obtained according to the trained generate network. The mapping relationship between the set of second frequency band candidate solutions and the set of first frequency band candidate solutions is obtained according to the trained eliminate network;

[0014] An optimal first frequency band candidate solution is obtained by calculating the difference between the set of first frequency band candidate solutions and the original first frequency band data, and the second candidate solution to which the optimal first frequency band candidate solution belongs is the final result predicted from the original first frequency band data.

[0015] If the original first frequency band data is original low frequency data, the second frequency band candidate solution is a high frequency candidate solution, and the first frequency band candidate solution is a low frequency candidate solution.

[0016] Similarly, if the original data is original high frequency data, the first frequency band candidate solution is a low frequency candidate solution, and the second frequency band candidate solution is a high frequency candidate solution.

[0017] Preferably, the constructing a training data set comprises the following steps:

[0018] Step 1: Design different sizes and rotation angles of elliptical shapes as the upper metal pattern of the metasurface unit structure.

[0019] Step 2: Model the metasurface unit structure in simulation software to obtain the optical response results.

[0020] Step 3: Split each group of optical responses into first frequency band training data and second frequency band training data.

[0021] According to the above steps, the training data set capable of training the generation network and the elimination network is obtained, i.e. the first frequency band training data and the second frequency band training data split from each group of optical responses, which are independently trained for the generation network and the elimination network, respectively.

[0022] Preferably, the generation network is composed of a first encoder, a first latent space and a first decoder.

[0023] During the training process, the first frequency band training data and the second frequency band training data are first passed through the first encoder and encoded into the first latent space, and then probabilistically sampled from the first latent space as first sampling points. The first frequency band training data and the first sampling points are taken as the input of the first decoder, and the second frequency band training data is taken as the output of the first decoder.

[0024] That is, the first frequency band training data and the second frequency band training data obtain the probabilistic mapping relationship between the first frequency band training data and the second frequency band training data through the training and encoding results of the generation network, and form the first latent space according to the probabilistic mapping relationship between the first frequency band training data and the second frequency band training data. The training of the generation network is completed, and the mapping relationship from the first frequency band training data to the second frequency band training data is realized.

[0025] Preferably, the elimination network is composed of a second encoder, a second latent space and a second decoder.

[0026] In the training process, the first frequency band training data and the second frequency band training data are first encoded into the second latent space by the second encoder, and then are probabilistically sampled from the second latent space as second sampling points. The second frequency band training data and the second sampling points are taken as inputs of the second decoder, and the first frequency band training data is taken as an output of the second decoder.

[0027] That is, the first frequency band training data and the second frequency band training data are trained by the elimination network to obtain a probabilistic mapping relationship between the second frequency band training data and the first frequency band training data, and the second latent space is obtained according to the probabilistic mapping relationship between the second frequency band training data and the first frequency band training data. The training of the elimination network is completed, and the mapping relationship from the second frequency band training data to the first frequency band training data is realized.

[0028] Preferably, the first latent space and the second latent space are both two-dimensional latent spaces, and the second frequency band training data and the first frequency band training data are both high-dimensional data. The high-dimensional data is reduced to one-dimensional values, and the points in the two-dimensional latent space are colored according to the one-dimensional values.

[0029] The high-dimensional data is reduced to form continuous one-dimensional values, and the points in the two-dimensional latent space are colored according to these values to visualize the distribution of the related data in the two-dimensional latent space.

[0030] Preferably, the two-dimensional latent space includes a two-dimensional Gaussian distribution obtained by mapping any one data input point, and each point in the two-dimensional Gaussian distribution can be decoded into a first frequency band candidate solution or a second frequency band candidate solution.

[0031] That is, only the latent space with two dimensions can effectively encode the input and output spectra into a compact space with a large amount of information in the form of conditional probability.

[0032] Preferably, the trained generation network takes the original first frequency band data as an input, randomly samples an arbitrary value from the two-dimensional Gaussian distribution, and inputs the two values into the decoder of the trained generation network to obtain a second frequency band candidate solution. One sampling point corresponds to one second frequency band candidate solution.

[0033] Through the trained generation network, multiple second frequency band candidate solutions are obtained from the original first frequency band data, realizing a one-to-many mapping relationship, and ensuring that each second frequency band candidate solution meets the requirements as much as possible. However, since there is a deviation, the trained elimination network is needed to filter the deviation data.

[0034] Preferably, the trained elimination network takes the second frequency band candidate solution as input, and randomly samples an arbitrary value from the two-dimensional Gaussian distribution, and inputs both into the decoder of the trained elimination network to obtain a first frequency band candidate solution, one sampling point corresponding to one first frequency band candidate solution.

[0035] Through the trained elimination network, multiple second frequency band candidate solutions obtain multiple first frequency band candidate solutions, realizing a many-to-many mapping relationship. By calculating the difference between all first frequency band candidate solutions and the original first frequency band data, the optimal first frequency band candidate solution is obtained, and the second frequency band candidate solution to which the optimal first frequency band candidate solution belongs is the final result predicted for the original first frequency band data, completing the cross-wavelength optical response prediction.

[0036] Through the above technical solution, compared with the prior art, the present application provides a cross-wavelength spectrum prediction method based on deep learning, constructs a training data set, uses the modeling metasurface unit structure in the simulation software to obtain the optical response result; constructs a generate-eliminate network architecture: a generation network and an elimination network; inputs the training data set into the generation network and the elimination network for training to obtain a trained generation network and a trained elimination network, and forms respective latent spaces, which can produce multiple solutions; obtains the mapping relationship between the original first frequency band data and the set of second frequency band candidate solutions according to the trained generation network; obtains the mapping relationship between the set of second frequency band candidate solutions and the set of first frequency band candidate solutions according to the trained elimination network; obtains the optimal first frequency band candidate solution by calculating the difference between the set of first frequency band candidate solutions and the original first frequency band data, and the second candidate solution to which the optimal first frequency band candidate solution belongs is the final result predicted for the original first frequency band data. After training, at least one set of solutions in the second frequency band candidate solutions generated by the generation network is consistent with the actual true result, and the elimination network also successfully filters out this set of solutions, which is consistent with the true value. That is, the present application realizes the many-to-many mapping relationship in the cross-wavelength spectrum prediction and completes the cross-wavelength optical response prediction under the condition that the metasurface unit structure is unknown. In addition, the present application uses a neural network to realize the method of reducing the dimensionality of high-dimensional data, and also successfully realizes the visualization of the mapping relationship between the first frequency band data and the second frequency band data in the latent space, which has the advantage of nonlinear dimensionality reduction and is more general and effective compared with the traditional PCA (principal component analysis) and other dimensionality reduction methods.

[0037] The cross-wavelength spectrum prediction method based on deep learning provided by the application opens up a new research direction in the field of combination of optics and deep learning, and is also applicable to the field of high-frequency information recovery after Fourier transform, Raman spectroscopy, and high-frequency noise removal in protein spectrum analysis, has very great application prospect, and the thought for solving the many-to-many mapping problem can not only complete the cross-wavelength spectrum conversion, but also provides a general and effective method for the fields also facing the two-way non-unique problem. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, below the drawings needed to be used in the embodiment or prior art description will be briefly introduced, obviously, the drawings in the following description are only embodiments of the application, and for those skilled in the art, without creative labor, other drawings can also be obtained according to the provided drawings.

[0039] Figure 1 It is a schematic diagram of electromagnetic super surface unit structure.

[0040] Figure 2 It is a randomly selected four electromagnetic super surface unit upper layer metal pattern structure.

[0041] Figure 3 It is the optical response result after modeling.

[0042] Figure 4 It is a schematic diagram of the generation network and elimination network structure of the application.

[0043] Figure 5 It is a generation-elimination network architecture implementation process diagram of the application.

[0044] Figure 6 It is a schematic diagram of the low-frequency data to high-frequency data prediction process of the embodiment of the application.

[0045] Figure 7 It is a schematic diagram of the high-frequency data to low-frequency data prediction process of the embodiment of the application.

[0046] Figure 8 It is a tree structure diagram of the many-to-many mapping relationship of the application.

[0047] Figure 9 It is a two-dimensional latent space after coloring with high-frequency data as features.

[0048] Among them, in the figure, a-upper layer, b-middle layer, c-lower layer. DETAILED DESCRIPTION

[0049] With reference to the drawings and the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0050] The embodiment of the present application discloses a cross-wavelength spectrum prediction method based on deep learning, which is suitable for cross-wavelength spectrum prediction problems in all fields and even all many-to-many mapping problems. Taking intelligent metasurface design as an example, each electromagnetic metasurface structure has a single optical response corresponding thereto, but in the case of unknown structure, each low-frequency optical response has multiple high-frequency optical responses, and at the same time, each high-frequency optical response also has multiple low-frequency optical responses. In order to solve the many-to-many mapping problem in cross-wavelength spectrum prediction, the present application proposes a general and effective generation-elimination network architecture.

[0051] Before training the generation network and the elimination network, the training data set needs to be collected first, and the specific steps are as follows:

[0052] Step 1: design different sizes and rotation angles of elliptical shapes as the upper metal pattern of the metasurface unit structure; as shown in Figure 2 Figure 2 As shown in the figure, four randomly selected elliptical patterns are shown, each of which is composed of 64*64 black and white pixels, wherein the white part represents the metal medium, and the black part represents the air.

[0053] Step 2: model the metasurface unit structure in the simulation software, as shown in Figure 1 Figure 3 As shown in the figure, the optical response result is the amplitude of the reflected wave, which is composed of three reflection coefficient curves R xx , R xy and R yy , and the waveband range is from 40THz to 100THz, and each curve is discretized into 1001 points.

[0054] Step 3: split each group of optical responses into first frequency band training data and second frequency band training data; in this embodiment, each group of optical responses is split into low-frequency training data and high-frequency training data; 40-60THz is taken as the input of the generation network, which is a 999-dimensional vector, and 60-100THz is taken as the output of the generation network, which is a 2004-dimensional vector.

[0055] ​​According to the first frequency band training data and the second frequency band training data, the generation network and the elimination network are independently trained, respectively, as shown in Figure 4 The generation network is composed of a first encoder, a first latent space and a first decoder. In the training process, the first frequency band training data and the second frequency band training data are first encoded by the first encoder to the first latent space, and then are probabilistically sampled from the first latent space as first sampling points. The first frequency band training data and the first sampling points are taken as the input of the first decoder, the second frequency band training data is taken as the output of the first decoder, and the trained generation network is obtained.

[0056] The elimination network is composed of a second encoder, a second latent space and a second decoder. In the training process, the first frequency band training data and the second frequency band training data are first encoded by the second encoder to the second latent space, and then are probabilistically sampled from the second latent space as second sampling points. The second frequency band training data and the second sampling points are taken as the input of the second decoder, the first frequency band training data is taken as the output of the second decoder, and the trained elimination network is obtained.

[0057] In order to combine the two respectively trained generation network and elimination network, as shown in Figure 5 The trained generation network takes the original first frequency band data as the input, and then randomly samples an arbitrary value from a two-dimensional Gaussian distribution. The two are input into the decoder of the trained generation network to obtain a second frequency band candidate solution. One sampling point corresponds to one second frequency band candidate solution, so theoretically, an infinite number of second frequency band candidate solutions can be generated, realizing a one-to-many mapping relationship.

[0058] Among them, since the generation network and the elimination network are both based on the principle of conditional variational autoencoder, it is ensured that the obtained candidate solutions meet the requirements as much as possible. However, whether it is the generation network or the elimination network, after the training is completed, in the actual generation process, the actual sampling region during the training is replaced by the standard Gaussian distribution. Even if the loss function is defined to make the actual sampling region similar to the standard Gaussian distribution, it cannot be completely the same, so there will be a deviation, so the trained elimination network is needed to screen the deviation data.

[0059] The second frequency band candidate solution is screened by the trained elimination network, as shown in Figure 5The data, indicated by the bold arrow 1, is fed into the trained elimination network. The trained elimination network takes the second-band candidate solution as input and then randomly samples an arbitrary value from a two-dimensional Gaussian distribution. Both are then input into the decoder of the trained elimination network to obtain the first-band candidate solution. One sampling point corresponds to one first-band candidate solution, so theoretically, an infinite number of first-band candidate solutions can be generated. Through the trained elimination network, multiple first-band candidate solutions are derived from multiple second-band candidate solutions, achieving a many-to-many mapping relationship. By calculating the difference between all first-band candidate solutions and the original first-band data, the optimal first-band candidate solution is obtained. The second-band candidate solution to which the optimal first-band candidate solution belongs is the final result predicted from the original first-band data, completing the cross-wavelength optical response prediction.

[0060] Furthermore, in one embodiment of the present invention, to achieve optical response prediction from low frequency to high frequency, the generator network consists of a first encoder, a first latent space, and a first decoder; low-frequency data and high-frequency data are used to obtain a probability mapping relationship between low-frequency data and high-frequency data through the training and encoding results of the generator network; the first latent space is obtained based on the probability mapping relationship between low-frequency data and high-frequency data; and the first latent space includes a two-dimensional Gaussian distribution mapped from any low-frequency data input point. Each point in the two-dimensional Gaussian distribution can be decoded into a high-frequency candidate solution, so theoretically, an infinite number of high-frequency candidate solutions can be generated; the generator network is trained using low-frequency data and high-frequency data to obtain a trained generator network, thereby realizing the mapping relationship from low frequency to high frequency.

[0061] The elimination network consists of a second encoder, a second latent space, and a second decoder. Low-frequency and high-frequency data are trained and encoded through the elimination network to obtain a probability mapping relationship between them. Based on this probability mapping, a second latent space is obtained, comprising a two-dimensional Gaussian distribution mapped from any high-frequency data input point. Each point in this two-dimensional Gaussian distribution can be decoded into a low-frequency candidate solution, so theoretically, an infinite number of low-frequency candidate solutions can be generated from an infinite number of high-frequency candidate solutions. The elimination network is trained using both low-frequency and high-frequency data to obtain a trained elimination network, thus realizing the mapping relationship from high to low frequencies.

[0062] In order to combine two separately trained subnetworks, such as Figure 6 As shown, for any original low-frequency data, firstly, the trained generator network takes the original low-frequency data as input, and then randomly samples an arbitrary value from a two-dimensional Gaussian distribution. Both are then input into the decoder of the trained generator network to obtain high-frequency candidate solutions. One sampling point corresponds to one high-frequency candidate solution, so theoretically, an infinite number of high-frequency candidate solutions can be generated. These high-frequency candidate solutions are then processed... Figure 6the bold arrow 2 in the figure is sent to the trained elimination network.

[0063] Secondly, the trained elimination network takes the high-frequency candidate solution as input, and randomly samples an arbitrary value from the two-dimensional Gaussian distribution, and inputs both into the decoder of the trained elimination network to obtain a low-frequency candidate solution. One sampling point corresponds to one low-frequency candidate solution, so theoretically, an infinite number of low-frequency candidate solutions can be generated. A many-to-many mapping from high frequency to low frequency is achieved.

[0064] Finally, by calculating the difference between all low-frequency candidate solutions generated by the trained elimination network and the original low-frequency data, the optimal low-frequency candidate solution is obtained. The high-frequency candidate solution to which the optimal low-frequency candidate solution belongs is the best high-frequency candidate solution predicted from the original low-frequency data.

[0065] Further, in another embodiment of the present application, to realize the prediction of optical response from high frequency to low frequency, the generation network is composed of a first encoder, a first latent space and a first decoder; the low-frequency data and the high-frequency data are used to obtain the probability mapping relationship between the low-frequency data and the high-frequency data through the training and encoding results of the generation network, and the first latent space is obtained according to the probability mapping relationship between the low-frequency data and the high-frequency data, and the first latent space includes a two-dimensional Gaussian distribution obtained by mapping any one high-frequency data input point. Each point in the two-dimensional Gaussian distribution can be decoded into a low-frequency candidate solution, so theoretically, an infinite number of low-frequency candidate solutions can be generated.

[0066] The elimination network is composed of a second encoder, a second latent space and a second decoder; the low-frequency data and the high-frequency data are used to obtain the probability mapping relationship between the low-frequency data and the high-frequency data through the training and encoding results of the elimination network, and the second latent space is obtained according to the probability mapping relationship between the low-frequency data and the high-frequency data. The second latent space includes a two-dimensional Gaussian distribution obtained by mapping any one low-frequency data input point. Each point in the two-dimensional Gaussian distribution can be decoded into a high-frequency candidate solution, so theoretically, an infinite number of high-frequency candidate solutions can be generated from an infinite number of low-frequency candidate solutions.

[0067] In order to combine the two respectively trained sub-networks, as shown in the figure, for any one original high-frequency data, firstly, the trained generation network takes the original high-frequency data as input, and then randomly samples an arbitrary value from the two-dimensional Gaussian distribution, and inputs both into the decoder of the trained generation network to obtain a low-frequency candidate solution. One sampling point corresponds to one low-frequency candidate solution, so theoretically, an infinite number of low-frequency candidate solutions can be generated. Figure 7 These low-frequency candidate solutions are sent to the trained elimination network through the bold arrow 3 in the figure. Figure 7

[0068] ​Secondly, the trained elimination network takes the above low-frequency candidate solution as input, and randomly samples an arbitrary value from a two-dimensional Gaussian distribution, and inputs the two into the decoder of the trained elimination network to obtain a high-frequency candidate solution. One sampling point corresponds to one high-frequency candidate solution, so theoretically, an infinite number of high-frequency candidate solutions can be generated.

[0069] Finally, by calculating the difference between all high-frequency candidate solutions generated by the trained elimination network and the original high-frequency data, the optimal high-frequency candidate solution is obtained. The low-frequency candidate solution to which the optimal high-frequency candidate solution belongs is the best low-frequency candidate solution predicted from the original high-frequency data.

[0070] As shown in Figure 8 is a tree structure diagram of the many-to-many mapping relationship of the application. In combination with Figure 5 , the generation network first generates multiple sets of second frequency band candidate solutions according to the original first frequency band data. These second frequency band candidate solutions are sent into the elimination network through the bold arrow 1 in Figure 5 . Each second frequency band candidate solution will generate multiple sets of first frequency band candidate solutions, and finally form a tree structure as shown in Figure 8 . By calculating the difference between all first frequency band candidate solutions (leaf nodes) and the original first frequency band data input (root node), the optimal first frequency band candidate solution (leaf node) is obtained. The second frequency band candidate solution (intermediate node) that produces the optimal first frequency band candidate solution is the final prediction result. That is, in the generation-elimination network architecture, the generation network plays the role of a producer to generate different candidate solutions, and the elimination network plays the role of a screener to find the optimal candidate solution, and finally realizes the prediction of cross-wavelength optical response in the case of unknown super surface unit structure.

[0071] Further, in an embodiment of the application, the original data is original low-frequency data. Correspondingly, in combination with Figure 6 , the generation network first generates multiple sets of high-frequency candidate solutions according to the original low-frequency data. These high-frequency candidate solutions are sent into the elimination network through the bold arrow 2 in Figure 6 . Each high-frequency candidate solution will generate multiple sets of low-frequency candidate solutions. By calculating the difference between all low-frequency candidate solutions (leaf nodes) and the original low-frequency input (root node), the optimal low-frequency candidate solution (leaf node) is obtained. The high-frequency candidate solution (intermediate node) that produces the optimal low-frequency candidate solution is the best high-frequency candidate solution.

[0072] Further, in another embodiment of the application, the original data is original high-frequency data. Correspondingly, in combination with Figure 7 , the generation network first generates multiple sets of low-frequency candidate solutions according to the original high-frequency data. These low-frequency candidate solutions are sent into the elimination network through the bold arrow 2 in Figure 7The bold arrow 3 in the figure is sent into the elimination network, each low-frequency candidate solution generates multiple groups of high-frequency candidate solutions, the optimal high-frequency candidate solution (leaf node) is obtained by calculating the difference between all high-frequency candidate solutions (leaf nodes) and the original high-frequency input (root node), and the low-frequency candidate solution (intermediate node) generating the optimal high-frequency candidate solution is the best low-frequency candidate solution.

[0073] The first latent space and the second latent space are both two-dimensional latent spaces, and the first frequency band training data and the second frequency band training data are both high-dimensional data, and the first frequency band data and the second frequency band data are also high-dimensional data, the high-dimensional data is reduced to one-dimensional values, and the points in the two-dimensional latent space are colored according to the one-dimensional values, as shown in the figure. Figure 9 As shown in the figure, the high-dimensional data is processed by dimension reduction to form continuous one-dimensional values, and the points in the two-dimensional latent space are colored according to the values, so that the distribution of related data in the two-dimensional latent space is visualized, the two-dimensional latent space includes a two-dimensional Gaussian distribution obtained by mapping any one data input point, and each point in the two-dimensional Gaussian distribution can be decoded into a second frequency band candidate solution or a first frequency band candidate solution.

[0074] The continuous and high-dimensional first frequency band data or second frequency band data is processed by dimension reduction through a self-encoder model, the self-encoder model is composed of an encoder, an intermediate layer and a decoder, the intermediate layer is set to one dimension as one-dimensional values after dimension reduction, the first frequency band data or the second frequency band data is reduced to one-dimensional values, and the points in the two-dimensional latent space are colored according to the one-dimensional values, so that the data points of different values in the two-dimensional latent space after the high-dimensional data is reduced to one-dimensional values can be directly observed; not only the effectiveness of the self-encoder for data dimension reduction is verified, but also it is proved that the latent space with only two dimensions can effectively encode the input and output spectra into a compact and large information space in the form of conditional probability, which can be realized without three-dimensional or higher-dimensional latent space, further proving the ability of the generation network and the elimination network to process complex mapping relationship.

[0075] Further, in an embodiment of the present application, the original data is original low-frequency data, the continuous and high-dimensional high-frequency data is processed by dimension reduction through a self-encoder model, the self-encoder model is composed of an encoder, an intermediate layer and a decoder, the intermediate layer is set to one dimension as one-dimensional values after dimension reduction, the high-frequency data is reduced to one-dimensional values, and the points in the first latent space are colored according to the one-dimensional values. The colored result is as shown in the figure. Figure 9As shown, it can be intuitively seen that the data points of different values of the high-frequency data after being reduced to one-dimensional values are represented in the first latent space, and the points in the first latent space after being colored according to the high-frequency data show a gradient trend, effectively separating the data points; similarly, the continuous and high-dimensional low-frequency data can also be reduced by the autoencoder model, the intermediate layer is set to one dimension as the one-dimensional value after dimension reduction, the low-frequency data is reduced to one-dimensional values, and the points in the second latent space are colored according to the one-dimensional values.

[0076] Further, in another embodiment of the present application, the original data is original high-frequency data, the continuous and high-dimensional low-frequency data is reduced by the autoencoder model, the autoencoder model is composed of an encoder, an intermediate layer and a decoder, the intermediate layer is set to one dimension as the one-dimensional value after dimension reduction, the low-frequency data is reduced to one-dimensional values, and the points in the first latent space are colored according to the one-dimensional values. Similarly, the continuous and high-dimensional high-frequency data can also be reduced by the autoencoder model, the intermediate layer is set to one dimension as the one-dimensional value after dimension reduction, the high-frequency data is reduced to one-dimensional values, and the points in the second latent space are colored according to the one-dimensional values.

[0077] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0078] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based cross-wavelength spectrum prediction method, characterized in that, Includes the following steps: A training dataset is constructed, and optical response results are obtained by modeling metasurface unit structures in simulation software. The training dataset includes training data for the first frequency band and training data for the second frequency band. Construct a generation-elimination network architecture; The generation-elimination network architecture includes two sub-networks: a generation network and an elimination network; Input the training dataset into the generator network for training to obtain the trained generator network; The training dataset is input into the elimination network for training to obtain the trained elimination network. The generator network and the elimination network form their respective latent spaces and can generate multiple solutions. The mapping relationship between the original first frequency band data and the set of candidate solutions for the second frequency band is obtained based on the trained generator network; the mapping relationship between the set of candidate solutions for the second frequency band and the set of candidate solutions for the first frequency band is obtained based on the trained elimination network. The optimal first frequency band candidate solution is obtained by calculating the difference between the set of candidate solutions for the first frequency band and the original first frequency band data. The second candidate solution to which the optimal first frequency band candidate solution belongs is the final result predicted from the original first frequency band data.

2. The deep learning-based cross-wavelength spectrum prediction method according to claim 1, characterized in that, The construction of the training dataset includes the following steps: Step 1: Design elliptical shapes of different sizes and rotation angles as the upper metal pattern of the metasurface unit structure; Step 2: Model the metasurface unit structure in simulation software and obtain the optical response results; Step 3: Split each set of optical responses into training data for the first frequency band and training data for the second frequency band.

3. The deep learning-based cross-wavelength spectrum prediction method according to claim 1, characterized in that, The generative network consists of a first encoder, a first latent space, and a first decoder; During training, the first frequency band training data and the second frequency band training data are first encoded into the first latent space by the first encoder, and then probabilistically sampled from the first latent space as the first sampling point. The first frequency band training data and the first sampling point are used together as the input of the first decoder, and the second frequency band training data is used as the output of the first decoder.

4. The deep learning-based cross-wavelength spectrum prediction method according to claim 3, characterized in that, The elimination network consists of a second encoder, a second latent space, and a second decoder; During training, the first frequency band training data and the second frequency band training data are first encoded into the second latent space by the second encoder, and then probabilistically sampled from the second latent space as the second sampling points. The second frequency band training data and the second sampling points are used together as the input of the second decoder, and the first frequency band training data is used as the output of the second decoder.

5. The deep learning-based cross-wavelength spectrum prediction method according to claim 4, characterized in that, Both the first latent space and the second latent space are two-dimensional latent spaces. Both the second frequency band training data and the first frequency band training data are high-dimensional data. The high-dimensional data is reduced to one-dimensional values, and the points in the two-dimensional latent space are colored according to the one-dimensional values.

6. The deep learning-based cross-wavelength spectrum prediction method according to claim 5, characterized in that, The two-dimensional potential space includes a two-dimensional Gaussian distribution mapped from any data input point, and each point in the two-dimensional Gaussian distribution can be decoded into a first frequency band candidate solution or a second frequency band candidate solution.

7. The deep learning-based cross-wavelength spectrum prediction method according to claim 6, characterized in that, The trained generator network takes the original first frequency band data as input, and then randomly samples an arbitrary value from the two-dimensional Gaussian distribution. Both are then input into the decoder of the trained generator network to obtain a second frequency band candidate solution. One sampling point corresponds to one second frequency band candidate solution.

8. The deep learning-based cross-wavelength spectrum prediction method according to claim 7, characterized in that, The trained elimination network takes the second frequency band candidate solution as input, and then randomly samples an arbitrary value from the two-dimensional Gaussian distribution. Both are then input into the decoder of the trained elimination network to obtain the first frequency band candidate solution. One sampling point corresponds to one first frequency band candidate solution.

Images