Multi-modal deep learning model construction method for metasurface electromagnetic characteristic estimation, medium, equipment and application

Abstract

The invention belongs to the technical field of electronic information, and discloses a multi-modal deep learning model construction method, medium, equipment and application for metasurface electromagnetic characteristic estimation, and the method comprises the steps: carrying out the binaryzation of metasurface patterns corresponding to all structure parameter vectors in a metasurface parameter space; constructing a multi-modal deep learning model MMN and initializing trainable parameters; calculating an accurate electromagnetic response corresponding to the structural parameter vector by using a finite element method, taking the structural parameter vector and the binarization matrix as input, taking the electromagnetic response as output, and pairing to form a training sample set; and training an MMN model by using a gradient descent method according to the loss function and the training sample set. The MMN model obtained by the invention can accurately estimate the electromagnetic characteristics of the metasurface, and the size of the required training sample set is 39% of the size of the sample set required by a single-mode model.

Description

Technical Field

[0001] The present invention belongs to the field of electronic information technology, and in particular relates to a multimodal deep learning model construction method, medium, equipment and application for estimating electromagnetic properties of metasurfaces. Background Art

[0002] Currently: Using deep learning to estimate the electromagnetic properties of metasurfaces is a widely studied problem, involving fields such as wireless communications and electromagnetic simulation. A major research trend in using deep learning to estimate the electromagnetic properties of metasurfaces is to improve the accuracy and efficiency of the electromagnetic property estimation process. For improving the accuracy and efficiency of the electromagnetic property estimation process, the use of multimodal deep learning models has been less studied.

[0003] Generally speaking, the accuracy and efficiency of estimating the electromagnetic properties of metasurfaces using deep learning depends on two aspects. The first aspect is the scale of the training sample set required to build and train the deep learning model, and the second aspect is the ability of the deep learning model itself to extract data features from the training sample set. The larger the scale and the stronger the feature extraction ability, the higher the accuracy and efficiency, and vice versa. After obtaining the training sample set, existing methods extract features from the data in the training sample set by constructing a single-modal deep learning model, such as using a fully connected layer and a matching activation layer, a normalization layer, etc. to extract features from the metasurface parameter vector. At this time, the deep learning model only uses data from the metasurface parameter vector, or uses a convolutional layer and a matching normalization layer, etc. to extract features from the metasurface pattern shape. At this time, the deep learning model only uses data from the metasurface pattern shape, which is a single modality.

[0004] There are currently many methods for optimizing single-modal deep learning models to improve the ability of deep learning models to extract data features. For example, optimizing the structure of single-modal deep learning models, mathematically optimizing the optimal number of each network layer in the deep learning model, the optimal number of neurons in each layer, etc. Or optimizing the training method of single-modal deep learning models, optimizing the loss function used when training the model, etc. In actual application scenarios, optimizing the structure of single-modal deep learning models is usually not worth the cost, because the larger the model, the stronger the feature extraction ability. Therefore, in general, for a given set of training samples, you only need to expand the scale of the model to obtain better electromagnetic response estimation results. However, such methods are usually limited by the computing power of the hardware, and the excessively large model scale will cause the amount of calculation of the model to expand sharply during the training process.

[0005] For a given metasurface structure, the structure can be accurately characterized using either the structural parameter vector or the pattern shape, so the structural parameter vector and the pattern shape are data of different modes of the same metasurface structure. Obviously, the structural parameter vector and the pattern shape (usually represented in matrix form) have completely different probability distributions, so a multimodal deep learning model can be constructed to simultaneously extract the features of the metasurface structure from the structural parameter vector and the pattern shape, thereby improving the accuracy and efficiency of the electromagnetic property estimation process.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows: the single-modal deep learning model can only use a single mode, such as the structural parameters of the metasurface structure in vector form or the pattern shape of the metasurface structure in matrix form. Using only a single mode does not fully explore the ability of the deep learning model to extract features, which reduces the accuracy and efficiency of the deep learning model in the process of electromagnetic property estimation.

[0007] The difficulty of solving the above problems and defects is: in order to simultaneously utilize the structural parameter data in vector form and the pattern shape data in matrix form, it is necessary to construct and train a deep learning model that can simultaneously input vector form data and matrix form data. The model needs to be able to simultaneously extract features from the structural parameter vector and the pattern shape and fuse the features. The fused features will be used to estimate the electromagnetic response of the metasurface.

[0008] The significance of solving the above problems and defects is: the multimodal deep learning model constructed by the present invention simultaneously utilizes two different modes of the metasurface, which can improve the accuracy of the estimation of the electromagnetic response of the metasurface while keeping the size of the training sample set unchanged; or in order to achieve specific electromagnetic response estimation performance, it can reduce the demand for the size of the training sample set, thereby improving the estimation efficiency. Summary of the invention

[0009] In response to the problems existing in the prior art, the present invention provides a multimodal deep learning model construction method, medium, device and application for estimating the electromagnetic properties of metasurfaces.

[0010] The present invention is implemented as follows: a method for constructing a multimodal deep learning model for estimating electromagnetic properties of a metasurface, the method comprising:

[0011] Initialize the structural parameter vector of the hypersurface in parameter space;

[0012] Binarizing the metasurface pattern corresponding to each structural parameter vector;

[0013] Construct a multimodal deep learning model MMN and initialize trainable parameters;

[0014] The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarization matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set;

[0015] According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

[0016] Furthermore, the multimodal deep learning model construction method for metasurface electromagnetic property estimation specifically includes the following steps:

[0017] In the first step, a uniform sampling method is used to sample a hypersurface structure of size N in a given parameter space. The parameter vector corresponding to the hypersurface structure can be represented by a set X = {x1, x2, ..., x N}express;

[0018] The second step is to use 3D modeling software to model the hypersurface and extract its surface pattern for any parameter vector x in the set X;

[0019] The third step is to binarize the metasurface pattern, where the area covered by the metal or resistor film is binarized to 1, and the other parts are binarized to 0. The binarization matrix corresponding to each metasurface structure in the parameter space can be represented by a set P = {p1, p2, ..., p N}express;

[0020] The fourth step is to construct an MMN model using a fully connected layer, a one-dimensional batch normalization layer, a convolutional layer, a two-dimensional batch normalization layer, an activation layer, and a regularization layer. The MMN model has two input nodes and one output node, and the parameter set of the neurons included in the model is W.

[0021] The fifth step is to use the finite element method to calculate the electromagnetic characteristic vector of the metasurface corresponding to each vector in the set X, and obtain the set Y = {y1, y2, ..., y N};

[0022] The sixth step is to take the structural parameter vector set X and the corresponding binary matrix set P as input, and the electromagnetic characteristic set Y as output, and construct a training sample set Z = {z1, z2, …, z N}={(X,P,Y)}={(x1,p1,y1),(x2,p2,y2),…,(x N ,p N ,y N )};

[0023] Step 7: Take turns to add x with index i in sample set Z i and p i As the input of MMN, the MMN model is used to perform forward propagation operations on the samples.i and p i obtain the floating-point value vector y generated by the output node neurons i ';

[0024] In the eighth step, according to the floating-point value vector y i ' and y in the original training sample set i , use the mean square error loss function to calculate the loss function value L of the model MMN i ;

[0025] In the ninth step, according to the loss function value L i backward calculate the gradient G corresponding to the trainable parameters in the model MMN i , and according to the gradient G i use the gradient descent method to update the parameters W of the MMN model;

[0026] In the tenth step, repeat steps seven to nine to complete the training of the MMN model.

[0027] Furthermore, the construction method of the MMN model in the multi-modal deep learning model construction method for estimating the electromagnetic characteristics of the metasurface includes:

[0028] Step 1, construct M fully connected layers FC1, FC2,..., FC M , M - 1 one-dimensional batch normalization layers BN1D1, BN1D2,..., BN1D M-1 , M - 1 activation layers RE1, RE2,..., RE M-1 , T convolutional layers CONV1, CONV2,..., CONV T , T two-dimensional batch normalization layers BN2D1, BN2D2,..., BN2D T and an output layer OC;

[0029] Step 2, set FC1 as the first input node of MMN, and FC1 is responsible for receiving x in the training samples; set CONV1 as the second input node of MMN, and CONV1 is responsible for receiving p in the training samples;

[0030] Step 3, connect the first Q (Q < M) fully connected layers, the first Q one-dimensional batch normalization layers, and the first Q activation layers in the order of {FC1, BN1D1, RE1, FC2,..., FC Q , BN1D Q , RE Q}, and the output of the previous layer network is used as the input of the next layer network;

[0031] Step 4, connect the T convolutional layers and the T two-dimensional batch normalization layers in the order of {CONV1, BN2D1, CONV2,..., CONV T,BN2D T} are connected in sequence, and the output of the previous network layer is used as the input of the next network layer;

[0032] Step 5: RE Q The output vector of BN2D T The output vectors of are added to obtain vector I;

[0033] Step 6: The remaining fully connected layers, one-dimensional batch normalization layers, activation layers, and output layers are arranged according to {FC Q+1 ,BN1D Q+1 ,RE Q+1 ,…,FC M ,OC} are connected in sequence, and the output of the previous layer of network is used as the input of the next layer of network, FC Q+1 The input of the layer is the vector I; the output of the OC layer neurons is the output of the MMN model.

[0034] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0035] Initialize the structural parameter vector of the hypersurface in parameter space;

[0036] Binarizing the metasurface pattern corresponding to each structural parameter vector;

[0037] Construct a multimodal deep learning model MMN and initialize trainable parameters;

[0038] The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarization matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set;

[0039] According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

[0040] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0041] Initialize the structural parameter vector of the hypersurface in parameter space;

[0042] Binarizing the metasurface pattern corresponding to each structural parameter vector;

[0043] Construct a multimodal deep learning model MMN and initialize trainable parameters;

[0044] The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarized matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set;

[0045] According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

[0046] Another object of the present invention is to provide an information data processing terminal, which is used to implement the multimodal deep learning model construction method for estimating the electromagnetic properties of metasurfaces. The information data processing terminal includes: a wireless communication terminal and an electromagnetic simulation terminal.

[0047] The multimodal deep learning model construction method was tested using a double-layer metasurface absorption-through integrated structure. Compared with the single-modal deep learning model trained with a complete training sample set, the multimodal deep learning model can achieve the same estimation accuracy using only 39% of the training samples. The training sample set size and the mean square error performance of the corresponding single-modal or multimodal deep learning model are shown in the attached figure. Figure 3 The dotted line in the figure represents the mean square error performance of the single-modal deep learning model, and the solid line represents the mean square error performance of the multi-modal deep learning model. The lower the mean square error value, the better the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 It is a flow chart of a method for constructing a multimodal deep learning model for estimating electromagnetic properties of a metasurface provided in an embodiment of the present invention.

[0050] Figure 2 is a flow chart of the MMN model construction method provided by an embodiment of the present invention;

[0051] Figure 3 is a mean square error performance curve diagram that can be obtained when the single-modality and multi-modality deep learning models provided in the embodiment of the present invention are trained using the same percentage of training sample set data;

[0052] Figure 4 It is a histogram of the size of the training sample set required for a unimodal or multimodal model to achieve the desired mean square error performance, provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] In response to the problems existing in the prior art, the present invention provides a multimodal deep learning model construction method, medium, device and application for metasurface electromagnetic property estimation. The present invention is described in detail below in conjunction with the accompanying drawings.

[0055] like Figure 1 As shown, the multimodal deep learning model construction method for metasurface electromagnetic property estimation provided by the present invention comprises the following steps:

[0056] S101: Initialize the structural parameter vector of the hypersurface in the parameter space;

[0057] S102: Binarizing the supersurface pattern corresponding to each structural parameter vector;

[0058] S103: Construct a multimodal deep learning model MMN and initialize trainable parameters;

[0059] S104: using the finite element method to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and considering the structural parameter vector and the binarization matrix as input, and the electromagnetic response as output, and pairing them to form a training sample set;

[0060] S105: According to the loss function and the training sample set, the MMN model is trained using the gradient descent method.

[0061] The multimodal deep learning model construction method for metasurface electromagnetic property estimation provided by the present invention can also be implemented by other steps by ordinary technicians in the industry. Figure 1 The multimodal deep learning model construction method for estimating the electromagnetic properties of metasurfaces provided by the present invention is only a specific embodiment.

[0062] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0063] The multimodal deep learning model construction method for metasurface electromagnetic property estimation provided by the present invention specifically includes the following steps:

[0064] In the first step, a uniform sampling method is used to sample a hypersurface structure of size N in a given parameter space. The parameter vector corresponding to the hypersurface structure can be represented by a set X = {x1, x2, ..., x N}express;

[0065] The second step is to use 3D modeling software to model the hypersurface and extract its surface pattern for any parameter vector x in the set X;

[0066] The third step is to binarize the metasurface pattern, where the area covered by the metal or resistor film is binarized to 1, and the other parts are binarized to 0. The binarization matrix corresponding to each metasurface structure in the parameter space can be represented by a set P = {p1, p2, ..., p N}express;

[0067] The fourth step is to construct an MMN model using a fully connected layer, a one-dimensional batch normalization layer, a convolutional layer, a two-dimensional batch normalization layer, an activation layer, and a regularization layer. The MMN model has two input nodes and one output node, and the parameter set of the neurons included in the model is W.

[0068] The fifth step is to use the finite element method to calculate the electromagnetic characteristic vector of the metasurface corresponding to each vector in the set X, and obtain the set Y = {y1, y2, ..., y N};

[0069] The sixth step is to take the structural parameter vector set X and the corresponding binary matrix set P as input, and the electromagnetic characteristic set Y as output, and construct a training sample set Z = {z1, z2, …, z N}={(X,P,Y)}={(x1,p1,y1),(x2,p2,y2),…,(x N ,p N ,y N )};

[0070] Step 7: Take turns to add x with index i in sample set Z i and p i As the input of MMN, the MMN model is used to perform forward propagation operations on the samples. i and p i Get the floating-point value vector y generated by the output node neuron i ′;

[0071] Step 8: According to the floating point value vector y i ′ and y in the original training sample set i , use the mean square error loss function to calculate the loss function value L of the model MMN i ;

[0072] The ninth step is to calculate the loss function value L according to the loss function value L. i Reverse calculation of the gradient G corresponding to the trainable parameters in the model MMN i , and according to the gradient G i Use gradient descent method to update the parameters W of the MMN model;

[0073] Step 10: Repeat steps 7 to 9 to complete the training of the MMN model.

[0074] like Figure 2 As shown, the MMN model construction method provided by the embodiment of the present invention includes:

[0075] Step 1: construct M fully connected layers FC1, FC2, ..., FC M , M-1 one-dimensional batch normalization layers BN1D1,BN1D2,…,BN1D M-1 , M-1 activation layers RE1,RE2,…,RE M-1 , T convolutional layers CONV1, CONV2, …, CONV1, T two-dimensional batch normalization layers BN2D1, BN2D2, …, BN2D T And an output layer OC;

[0076] Step 2: Set FC1 as the first input node of MMN, and FC1 is responsible for receiving x in the training sample; set CONV1 as the second input node of MMN, and CONV1 is responsible for receiving p in the training sample;

[0077] Step 3: The first Q fully connected layers, the first Q one-dimensional batch normalization layers, and the first Q activation layers are arranged according to {FC1, BN1D1, RE1, FC2, …, FC Q ,BN1D Q ,RE Q} are connected in sequence, and the output of the previous network layer is used as the input of the next network layer;

[0078] Step 4: Arrange T convolutional layers and T 2D batch normalization layers according to {Conv1, BN2D1, CONV2, …, CONV T ,BN2D T} are connected in sequence, and the output of the previous network layer is used as the input of the next network layer;

[0079] Step 5: RE Q The output vector of BN2D T The output vectors of are added to obtain vector I;

[0080] Step 6: The remaining fully connected layers, one-dimensional batch normalization layers, activation layers, and output layers are arranged according to {FC Q+1 ,BN1D Q+1 ,RE Q+1 ,…,FC M ,OC} are connected in sequence, and the output of the previous layer of network is used as the input of the next layer of network, FC Q+1 The input of the layer is the vector I; the output of the OC layer neurons is the output of the MMN model.

[0081] The technical effects of the present invention are described in detail below in conjunction with specific embodiments.

[0082] Embodiment 1:

[0083] Take a double-layer metasurface absorption and penetration integrated structure as an example. The bottom surface of the absorption and penetration integrated structure is metal copper, and the top surface is metal copper. The metal covering patterns of the bottom and top surfaces are different. A dielectric layer is filled between the metal layers of the bottom and top surfaces, and the relative dielectric constant is 1.1 and the relative magnetic permeability is 1. A data set of 2500 is constructed using a random sampling method, of which 2000 samples are used as training sample sets and 500 samples are used as test sample sets. All metasurfaces in the data set are characterized in two modes, including vectorized structural parameters and binary patterns in matrix form. A multimodal model is constructed using the multimodal deep learning model construction method for supersurface electromagnetic property estimation, which includes a total of 7 fully connected layers and 10 convolutional layers. For the fully connected layers, the first 5 are used to extract features from the structural parameter vector, and the last two are used to further extract features from the fused features; for the convolutional layers, 10 convolutional layers are used to extract features from the metal patterns on the bottom and top surfaces. For the unimodal model, there are a total of 7 fully connected layers, and their connection structure is consistent with the fully connected layer connection structure in the multimodal model. The difference between the two different deep learning models is that the multimodal model uses additional pattern shapes as the second modality data and fuses the features from the structural parameter vector and the pattern shape.

[0084] The 2000 training samples were regarded as 100% of the data, and different percentages of data were used to train the unimodal and multimodal models. The same 500 samples were used to test the trained deep learning models. Figure 3 The figure shows the mean square error performance that can be obtained by the single-modal and multi-modal deep learning models when using the same percentage of data, wherein the horizontal axis represents the percentage of data samples used in training in the total training data set samples, and the vertical axis represents the mean square error of the electromagnetic response estimation of the trained model on the test sample set. The figure shows that the multi-modal model can obtain the same performance as the single-modal model using only 39% of the training data, indicating that the multi-modal model has a higher estimation efficiency. Figure 4 The figure shows the size of the training sample set required for a unimodal or multimodal model to achieve the expected mean square error performance, where the horizontal axis is the expected mean square error performance of the unimodal or multimodal model, and the vertical axis is the size of the training sample set required to achieve the expected performance. The figure shows that the multimodal model always requires fewer training samples to achieve the expected performance, indicating that the multimodal deep learning model has a higher estimation accuracy.

[0085] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0086] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a multimodal deep learning model for estimating electromagnetic properties of a metasurface, characterized in that: The multimodal deep learning model construction method for metasurface electromagnetic property estimation includes: Initialize the structural parameter vector of the hypersurface in parameter space; Binarizing the metasurface pattern corresponding to each structural parameter vector; Construct a multimodal deep learning model MMN and initialize trainable parameters; The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarized matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set; According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

2. The method for constructing a multimodal deep learning model for estimating electromagnetic properties of a metasurface according to claim 1, characterized in that: The multimodal deep learning model construction method for estimating electromagnetic properties of metasurfaces is specifically The following steps are involved: In the first step, a uniform sampling method is used to sample a hypersurface structure of size N in a given parameter space. The parameter vector corresponding to the hypersurface structure can be represented by a set X = {x1, x2, ..., x N }express; The second step is to use 3D modeling software to model the hypersurface and extract its surface pattern for any parameter vector x in the set X; The third step is to binarize the metasurface pattern, where the area covered by the metal or resistor film is binarized to 1, and the other parts are binarized to 0. The binarization matrix corresponding to each metasurface structure in the parameter space can be represented by a set P = {p1, p2, ..., p N }express; The fourth step is to construct an MMN model using a fully connected layer, a one-dimensional batch normalization layer, a convolutional layer, a two-dimensional batch normalization layer, and an activation layer. The MMN model has two input nodes and one output node, and the parameter set of the neurons included in the model is W. The fifth step is to use the finite element method to calculate the electromagnetic characteristic vector of the metasurface corresponding to each vector in the set X, and obtain the set Y = {y1, y2, ..., y N }; The sixth step is to take the structural parameter vector set X and the corresponding binary matrix set P as input, and the electromagnetic characteristic set Y as output, and construct a training sample set Z = {z1, z2, …, z N }={(X,P,Y)}={(x1,p1,y1),(x2,p2,y2),…,(x N ,p N ,y N )}; Step 7: Take turns to add x with index i in sample set Z i and p i As the input of MMN, the MMN model is used to perform forward propagation operations on the samples. i and p i Get the floating-point value vector y generated by the output node neuron i ′; Step 8: According to the floating point value vector y i ′ and y in the original training sample set i , use the mean square error loss function to calculate the loss function value L of the model MMN i ; The ninth step is to calculate the loss function value L according to the loss function value L. i Calculate the gradient G corresponding to the trainable parameters in the model MMN i , and according to the gradient G i Use gradient descent method to update the parameters W of the MMN model; Step 10: Repeat steps 7 to 9 to complete the training of the MMN model.

3. The method for constructing a multimodal deep learning model for estimating electromagnetic properties of a metasurface according to claim 2, characterized in that: The MMN model construction method comprises: Step 1: construct M fully connected layers FC1, FC2, ..., FC M , M-1 one-dimensional batch normalization layers BN1D1,BN1D2,…,BN1D M-1 , M-1 activation layers RE1,RE2,…,RE M-1 , T convolutional layers CONV1, CONV2, ..., CONV1, T two-dimensional batch normalization layers BN2D1, BN2D2, ..., BN2D T And an output layer OC; Step 2: Set FC1 as the first input node of MMN, and FC1 is responsible for receiving x in the training sample; set CONV1 as the second input node of MMN, and CONV1 is responsible for receiving p in the training sample; Step 3: Connect the first Q (Q < M) fully connected layers, the first Q one-dimensional batch normalization layers, and the first Q activation layers in sequence according to the order of {FC1, BN1D1, RE1, FC2, …, FC Q , BN1D Q , RE Q}, with the output of the previous layer network serving as the input of the next layer network; Step 4: Arrange T convolutional layers and T 2D batch normalization layers according to {Conv1, BN2D1, CONV2, …, CONV T ,BN2D T } are connected in sequence, and the output of the previous network layer is used as the input of the next network layer; Step 5: RE Q The output vector of BN2D T The output vectors of are added to obtain vector I; Step 6: The remaining fully connected layers, one-dimensional batch normalization layers, activation layers, and output layers are arranged according to {FC Q+1 ,BN1D Q+1 ,RE Q+1 ,…,FC M ,OC} are connected in sequence, and the output of the previous layer of network is used as the input of the next layer of network, FC Q+1 The input of the layer is the vector I; the output of the OC layer neuron is the output of the MMN model.

4. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Initialize the structural parameter vector of the hypersurface in parameter space; Binarizing the metasurface pattern corresponding to each structural parameter vector; Construct a multimodal deep learning model MMN and initialize trainable parameters; The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarization matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set; According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

5. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps: Initialize the structural parameter vector of the hypersurface in parameter space; Binarizing the metasurface pattern corresponding to each structural parameter vector; Construct a multimodal deep learning model MMN and initialize trainable parameters; The finite element method is used to calculate the accurate electromagnetic response corresponding to the structural parameter vector, and the structural parameter vector and the binarized matrix are regarded as input, and the electromagnetic response is regarded as output, and they are paired to form a training sample set; According to the loss function and training sample set, the MMN model is trained using the gradient descent method.

6. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the multimodal deep learning model construction method for estimating the electromagnetic properties of the metasurface as described in any one of claims 1 to 3, and the information data processing terminal includes: a wireless communication terminal and an electromagnetic simulation terminal.

Images