A spatiotemporal encoding matrix generation method based on a physical driving vector quantization autoencoder

Abstract

The application discloses a kind of space-time coding matrix generation methods based on physical drive vector quantization automatic encoder, with target harmonic scattering mode as input, physical drive vector quantization automatic encoder can quickly output optimal discrete space-time coding matrix.The physical drive vector quantization automatic encoder includes an encoder module, a vector quantization layer and a physically driven decoder module, the output of the encoder module is passed to the vector quantization layer, which is converted into a discrete vector to obtain a discrete space-time coding matrix, while the physical operating mechanism between the space-time coding matrix and the harmonic scattering mode is introduced into the decoder module of the automatic encoder for reconstructing the harmonic scattering mode, so the physical drive vector quantization automatic encoder is trained in an unsupervised manner, without preparing a large amount of artificially labeled data.The application can help space-time coding digital metasurface to realize flexible real-time multi-harmonic beamforming.

Description

Technical Field

[0001] This invention belongs to the field of artificial electromagnetic materials and communication technology, and particularly relates to a method for generating a spatiotemporal coding matrix based on a physical-driven vector quantization autoencoder. Background Technology

[0002] Novel artificial electromagnetic materials are metamaterials composed of artificially designed subwavelength structures. Due to their ability to finely modulate electromagnetic waves, they have been extensively studied in both scientific and engineering communities. As a two-dimensional version of metamaterials, metasurfaces possess advantages such as ultrathin thickness, high integration, and lower insertion loss than metamaterials, thus enabling more flexible manipulation of electromagnetic waves. Digitally coded and programmable metasurfaces were proposed in 2014. They can digitally manipulate electromagnetic waves using only a limited number of unit types (e.g., only two in the case of 1 bit). Digitally coded metasurfaces can be integrated with active devices, where each unit can be independently controlled by a control module, thus creating programmable metasurfaces. Digitally programmable metasurfaces have become a powerful and versatile platform for realizing various functions. Introducing a time-based control dimension allows for digital manipulation of electromagnetic waves in the spatial and frequency domains, resulting in spatiotemporally coded metasurfaces, further expanding the capabilities of digitally coded metasurfaces to manipulate electromagnetic waves.

[0003] The spatiotemporal coding matrix describes the dynamic coding state of each unit of a spatiotemporally coded metasurface in both the spatial and temporal domains. By rationally designing the spatiotemporal coding matrix, electromagnetic waves can be precisely modulated in both the spatial and frequency domains. The traditional method for obtaining the spatiotemporal coding matrix is ​​iterative optimization (e.g., particle swarm optimization), but this requires significant computational resources and several hours of computation time, making the efficiency of obtaining the spatiotemporal coding matrix very low, especially as the matrix dimension increases, thus hindering the practical application of spatiotemporally coded metasurfaces to some extent. Besides iterative optimization algorithms, another coding strategy involves constructing temporally interleaved coding subsequences within the spatiotemporal coding matrix, which can independently and simultaneously synthesize the target scattering modes of the spatiotemporally coded metasurface at a given harmonic frequency. However, this method has specific limitations on the length of the temporal coding sequence and the number of target harmonics, and suffers from low spectral efficiency due to the inability to suppress unwanted higher-order harmonics. Summary of the Invention

[0004] The purpose of this invention is to provide a spatiotemporal coding matrix generation method based on a physical-driven vector quantization autoencoder, so as to solve the technical problem of low efficiency in traditional methods for obtaining spatiotemporal coding matrices.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0006] A spatiotemporal coding matrix generation method based on a physical-driven vector quantization autoencoder includes the following steps:

[0007] Step 1: Construct a dataset consisting of multiple sets of target harmonic scattering modes;

[0008] Step 2: Based on the required spatiotemporal coding matrix size, construct an artificial neural network model. This model is a physical-driven vector quantization autoencoder, which includes an encoder module, a vector quantization layer, and a physical-driven decoder module.

[0009] Step 3: Train the physical-driven vector quantization autoencoder in an unsupervised manner using the dataset constructed in Step 1;

[0010] Step 4: After the physical-driven vector quantization autoencoder has been trained, it quickly outputs the corresponding optimal discrete-time coding matrix using the actual target harmonic scattering mode as input.

[0011] Furthermore, in step 1, each group of target harmonic scattering modes contains a scattering pattern with 2K+1 harmonic frequencies, wherein the power of each harmonic frequency is preserved or suppressed, and the beam direction of each harmonic pattern is randomly set.

[0012] Furthermore, in step 2, the physical-driven vector quantization autoencoder takes the target harmonic scattering mode as input, first outputs a continuous spatiotemporal coding matrix through the encoder module, then obtains a discrete spatiotemporal coding matrix through the vector quantization layer, and then inputs it into the physical-driven decoder module to output the reconstructed harmonic scattering mode.

[0013] Furthermore, in step 2, the encoder module of the physical drive vector quantization autoencoder is an artificial neural network model, the activation function of the output layer is tanh, and the output value is continuous, in the range [-1,+1], corresponding to the reflection phase [-π,+π].

[0014] Furthermore, in step 2, the vector quantization layer of the physical drive vector quantization autoencoder contains a discrete latent space with D discrete variables. The encoder output is passed to the vector quantization layer, which uses nearest neighbor lookup to transform the output into discrete vectors using the discrete latent space, thereby obtaining a discrete spatiotemporal coding matrix.

[0015] Furthermore, in step 2, the decoder module of the physical drive vector quantization autoencoder introduces a physical mechanism between the spatiotemporal coding matrix and the harmonic scattering mode, and outputs a reconstructed harmonic scattering mode.

[0016] Furthermore, in step 3, the physical-driven vector quantization autoencoder is trained in an unsupervised manner. Its overall loss function consists of two terms: the first term is the reconstruction loss, which represents the difference between the reconstructed harmonic scattering pattern and the target scattering pattern; the second term is the encoding quantization loss, which represents the quantization difference between the continuous spatiotemporal encoding matrix and the discrete spatiotemporal encoding matrix.

[0017] Furthermore, in step 4, using the actual target harmonic scattering mode as input, the physical-driven vector quantization autoencoder quickly outputs the corresponding optimal discrete-time coding matrix. The trained physical-driven vector quantization autoencoder has good generalization ability, so the actual target harmonic scattering mode does not need to be included in the dataset constructed in step 1. The target harmonic scattering mode can be directly constructed according to actual needs.

[0018] The spatiotemporal coding matrix generation method based on a physical-driven vector quantization autoencoder of the present invention has the following advantages: This method takes the target harmonic scattering mode as input, and the physical-driven vector quantization autoencoder can quickly output the optimal discrete spatiotemporal coding matrix. This autoencoder requires no large amount of computational resources, has high output efficiency, and is trained in an unsupervised manner, thus eliminating the need for a large amount of manually labeled data. It also exhibits good generalization ability and is flexible in its requirements regarding the length of the time-coded sequence and the number of target harmonics. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a spatiotemporally encoded digital metasurface of the present invention, which uses a physics-driven vector quantization autoencoder to quickly acquire a spatiotemporally encoded matrix and modulate electromagnetic waves.

[0020] Figure 2 This is a schematic diagram of the structure of the physical drive vector quantization autoencoder of the present invention;

[0021] Figure 3 (a) is the target harmonic scattering pattern of the present invention as the input diagram of the encoder module;

[0022] Figure 3 (b) is a diagram of the continuous spatiotemporal coding matrix generated by the encoder module of the present invention;

[0023] Figure 3 (c) is a diagram of the discrete-time coding matrix of the present invention after the vector quantization layer;

[0024] Figure 3 (d) is the output diagram of the decoder module, which is the reconstructed harmonic scattering mode corresponding to the generated discrete spatiotemporal coding matrix of this invention. Detailed Implementation

[0025] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, provides a more detailed account of a spatiotemporal coding matrix generation method based on a physically driven vector quantization autoencoder.

[0026] This invention includes the following steps:

[0027] Step 1: Construct a dataset consisting of multiple sets of target harmonic scattering modes;

[0028] In each target harmonic scattering mode, there is a scattering pattern containing (2K+1) harmonic frequencies, where the power of each harmonic frequency can be preserved or suppressed, and the beam direction of each harmonic pattern is randomly set. We uniformly sample P points in space, and the scattering pattern of each harmonic frequency can be represented as a vector of dimension P. Then, the target scattering mode containing (2K+1) harmonic frequencies is represented as a vector of dimension [(2K+1)×P].

[0029] Step 2: Based on the required spatiotemporal coding matrix size, construct an artificial neural network model. This model is a physical-driven vector quantization autoencoder, which includes an encoder module, a vector quantization layer, and a physical-driven decoder module.

[0030] This autoencoder comprises an encoder module, a vector quantization layer, and a physically driven decoder module. The physically driven vector quantization autoencoder takes the target harmonic scattering mode x as input, where x has a dimension of (2K+1)×P. First, x passes through the encoder module to output a continuous spatiotemporal coding matrix e(x), where e(x) has a dimension of (N×L), where N is the number of units corresponding to the spatiotemporal coding matrix, and L is the length of the time series corresponding to the spatiotemporal coding matrix. e(x) then passes through the vector quantization layer to obtain a discrete spatiotemporal coding matrix q(x), which is then used by the physically driven decoder to reconstruct the harmonic scattering mode y.

[0031] The encoder module of the physical-driven vector quantization autoencoder is an artificial neural network model. The activation function of the output layer is tanh, so the output value is continuous and in the range [-1,+1], corresponding to the reflection phase [-π,+π].

[0032] The physical-driven vector quantization autoencoder includes a vector quantization layer, which makes the quantization process non-differentiable. In order for gradient descent to proceed normally during model training, the gradient of the decoder is directly propagated to the encoder.

[0033] The decoder of a physics-driven vector quantization autoencoder introduces a physical mechanism between the spatiotemporal coding matrix and harmonic scattering modes, outputting a reconstructed harmonic scattering mode. Taking a 2-bit reflective programmable metasurface as an example, each element has four operating states with the same reflection amplitude but different reflection phases, corresponding to -π / 2, 0, π / 2, and π, respectively. The spatiotemporal coding matrix consists of N columns of time-coded sequences, with each element in the nth column sharing the same time-coded sequence, which can be represented as a periodic function of time t.

[0034]

[0035] Where L is the length of the time-coded sequence corresponding to the spatiotemporal coding matrix, each time-coded sequence contains L time-coded intervals, and the length of each time interval is τ. U is the reflection coefficient for the l-th time interval. l (t) is a periodic pulse function with period T0 corresponding to the l-th time interval, therefore the modulation frequency is f0 = 1 / T0:

[0036]

[0037] For a 2-bit reflective programmable metasurface, there is a reflection coefficient. These correspond to reflection phases of -π / 2, 0, π / 2, and π, respectively. The time-domain harmonic scattering modes of the metasurface can be approximated as:

[0038]

[0039] Where E(θ) is the far-field scattering mode of the coding unit, approximately cosθ, and the frequency of the incident wave is f. c , λ c Let U be the wavelength of the incident wave in vacuum. The periodic pulse function U... l (t) can be extended to a Fourier series, therefore, the nth column element at the vth harmonic frequency f c The scattering mode at +vf0 can be written as

[0040]

[0041] We define two vectors z and h(θ,v) as follows:

[0042]

[0043] in The dimensions of z and h(θ,v) are (N×L). Therefore, Equation 4 can be transformed into matrix form, i.e.

[0044] G(v,θ)=h(θ,v) T ·z (6)

[0045] We uniformly select P points (θ1, θ2, θ3, ..., θ) in space. P Sampling is performed, and the harmonic frequency f c The scattering mode under +vf0 can be represented as a vector y of dimension P. (v)

[0046]

[0047] Then, (2K+1) harmonic frequency points f c The scattering mode on +vf0 (v=0,±1,±2...±K) can be represented as a vector y with dimension [(2K+1)×P], i.e.

[0048]

[0049] Where H∈C [(2K+1)×P]×(N×L) It can be represented as

[0050]

[0051] Equation 8 gives the physical operating mechanism between the spatiotemporal coding matrix and the harmonic scattering mode. This mechanism is introduced into the decoder module of the physical drive vector quantization autoencoder for reconstructing the harmonic scattering mode.

[0052] Step 3: Train the physical-driven vector quantization autoencoder in an unsupervised manner using the dataset constructed in Step 1;

[0053] The physics-driven vector quantization autoencoder is trained in an unsupervised manner. Its overall loss function consists of two terms: the first term is the reconstruction loss, which represents the difference between the reconstructed harmonic scattering pattern and the target scattering pattern; the second term is the encoding quantization loss, which represents the quantization difference between the continuous spatiotemporal encoding matrix and the discrete spatiotemporal encoding matrix.

[0054] Step 4: After the physical-driven vector quantization autoencoder is trained, it takes the actual target harmonic scattering mode as input and quickly outputs the corresponding optimal discrete-time coding matrix.

[0055] Using the actual target harmonic scattering mode as input, the physical-driven vector quantization autoencoder quickly outputs the corresponding optimal discrete-time coding matrix. The actual target harmonic scattering mode may not be included in the dataset constructed in step 1. The physical-driven vector quantization autoencoder has good generalization ability and can directly construct the target harmonic scattering mode according to actual needs.

[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0057] This invention provides a spatiotemporal coding matrix generation method based on a physical-driven vector quantization autoencoder. The method takes the target harmonic scattering mode as input, and the physical-driven vector quantization autoencoder can quickly output the optimal discrete spatiotemporal coding matrix. This physical-driven vector quantization autoencoder is trained in an unsupervised manner, thus eliminating the need for large amounts of manually labeled data. It exhibits good generalization ability and is flexible in its requirements regarding the length of the time-coded sequence and the number of target harmonics. The specific design steps are as follows:

[0058] Step 1: Construct a dataset consisting of multiple sets of target harmonic scattering modes;

[0059] Each target harmonic scattering mode contains a scattering pattern with (2K+1) harmonic frequencies, where the power of each harmonic frequency can be preserved or suppressed, and the beam direction of each harmonic pattern is randomly set.

[0060] Step 2: Based on the required spatiotemporal coding matrix size, construct an artificial neural network model. This model is a physical-driven vector quantization autoencoder, which includes an encoder module, a vector quantization layer, and a physical-driven decoder module.

[0061] This physical-driven vector quantization autoencoder takes the target harmonic scattering mode as input. First, it outputs a continuous spatiotemporal coding matrix through the encoder module, then passes through the vector quantization layer to obtain a discrete spatiotemporal coding matrix, and then inputs it into the physical-driven decoder module to output the reconstructed harmonic scattering mode.

[0062] The encoder module of the physical-driven vector quantization autoencoder is an artificial neural network model. The activation function of the output layer is tanh, and the output value is continuous, ranging from [-1, +1], corresponding to the reflection phase [-π, +π].

[0063] The vector quantizer layer of a physics-driven vector quantization autoencoder contains a discrete latent space a∈R D It has D discrete variables a. i For each element ∈ R, i = 1, 2, ..., D, the encoder output is passed to a vector quantization layer. This layer transforms the element into a discrete vector using a nearest neighbor search and the discrete latent space, thus obtaining a discrete spatiotemporal coding matrix. For the 2-bit metasurface, we have Corresponding to the reflection phase The discrete latent space 'a' can also be adjusted according to the actual reflection coefficient of the coding element.

[0064] The decoder module of the physical-driven vector quantization autoencoder introduces a physical mechanism between the spatiotemporal coding matrix and harmonic scattering modes, and outputs a reconstructed harmonic scattering mode.

[0065] Step 3: Train the physical-driven vector quantization autoencoder in an unsupervised manner using the dataset;

[0066] The physics-driven vector quantization autoencoder is trained in an unsupervised manner. Its overall loss function consists of two terms: the first term is the reconstruction loss, which represents the difference between the reconstructed harmonic scattering pattern and the target scattering pattern; the second term is the encoding quantization loss, which represents the quantization difference between the continuous spatiotemporal encoding matrix and the discrete spatiotemporal encoding matrix.

[0067] Step 4: After the physical-driven vector quantization autoencoder model is trained, it takes the actual target harmonic scattering mode as input and quickly outputs the corresponding optimal discrete-time coding matrix.

[0068] The trained physical-driven vector quantization autoencoder has good generalization ability, so the actual target harmonic scattering mode does not need to be included in the dataset constructed in step 1. The target harmonic scattering mode can be directly constructed according to actual needs.

[0069] Figure 1 This is a schematic diagram of a spatiotemporally coded digital metasurface that uses a physically driven vector quantization autoencoder to rapidly acquire a spatiotemporally coded matrix and modulate electromagnetic waves. The physically driven vector quantization autoencoder takes the target harmonic scattering mode as input and generates a corresponding optimized discrete spatiotemporally coded matrix with high reliability and accuracy. Therefore, the spatiotemporally coded digital metasurface can modulate the incident wave into the desired harmonic mode in real time under the control of the generated spatiotemporally coded matrix.

[0070] Figure 2 This is a schematic diagram of a physically driven vector quantization autoencoder, consisting of an encoder module, a vector quantization layer, and a physically driven decoder module. The encoder module takes the target scattering pattern as input and generates the corresponding continuous spatiotemporal coding matrix. The vector quantization layer is used to obtain a discrete representation of the spatiotemporal coding matrix, and the decoder module outputs the reconstructed harmonic scattering pattern by introducing a physical computation mechanism between the spatiotemporal coding matrix and the harmonic scattering pattern.

[0071] Figure 3 (a) The target harmonic scattering pattern is used as the input to the encoder module; Figure 3 (b) is the continuous spatiotemporal coding matrix generated by the encoder module; Figure 3 (c) Discrete-space-time coding matrix after vector quantization layer; Figure 3 (d) The reconstructed harmonic scattering mode corresponding to the generated discrete spatiotemporal coding matrix is ​​used as the output of the decoder module.

[0072] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for generating a spatiotemporal coding matrix based on a physically driven vector quantization autoencoder, characterized in that, Includes the following steps: Step 1: Construct a dataset consisting of multiple sets of target harmonic scattering modes; Step 2: Based on the required spatiotemporal coding matrix size, construct an artificial neural network model. This model is a physical-driven vector quantization autoencoder, which includes an encoder module, a vector quantization layer, and a physical-driven decoder module. Step 3: Train the physical-driven vector quantization autoencoder in an unsupervised manner using the dataset constructed in Step 1; Step 4: After the physical-driven vector quantization autoencoder is trained, it takes the actual target harmonic scattering mode as input and quickly outputs the corresponding optimal discrete spatiotemporal coding matrix. In step 2, the physical drive vector quantization autoencoder takes the target harmonic scattering mode as input, first outputs a continuous spatiotemporal coding matrix through the encoder module, then obtains a discrete spatiotemporal coding matrix through the vector quantization layer, and then inputs it into the physical drive decoder module to output the reconstructed harmonic scattering mode. In step 2, the encoder module of the physical-driven vector quantization autoencoder is an artificial neural network model. The activation function of the output layer is tanh, and the output value is continuous, within a certain range. Between, corresponding to the reflection phase ; In step 2, the vector quantization layer of the physically driven vector quantization autoencoder contains a discrete latent space, which has D The encoder output is passed to a vector quantization layer, which uses nearest neighbor lookup to transform the discrete variables into discrete vectors using the discrete latent space, thus obtaining a discrete spatiotemporal coding matrix. In step 2, the decoder module of the physical drive vector quantization autoencoder introduces a physical mechanism between the spatiotemporal coding matrix and the harmonic scattering mode, and outputs a reconstructed harmonic scattering mode.

2. The spatiotemporal coding matrix generation method based on a physically driven vector quantization autoencoder according to claim 1, characterized in that, In step 1, each group of target harmonic scattering modes includes The scattering pattern of each harmonic frequency point, wherein the power of each harmonic frequency point is preserved or suppressed, and the beam direction of each harmonic pattern is randomly set.

3. The spatiotemporal coding matrix generation method based on a physically driven vector quantization autoencoder according to claim 1, characterized in that, In step 3, the physical-driven vector quantization autoencoder is trained in an unsupervised manner. Its overall loss function consists of two terms: the first term is the reconstruction loss, which represents the difference between the reconstructed harmonic scattering pattern and the target scattering pattern; the second term is the encoding quantization loss, which represents the quantization difference between the continuous spatiotemporal encoding matrix and the discrete spatiotemporal encoding matrix.

4. The spatiotemporal coding matrix generation method based on a physically driven vector quantization autoencoder according to claim 1, characterized in that, In step 4, the trained physical drive vector quantization autoencoder has generalization ability, so the actual target harmonic scattering mode does not need to be included in the dataset constructed in step 1. The target harmonic scattering mode can be directly constructed according to actual needs.

Images