Reverse design method of metal material metasurface based on transfer learning

Abstract

The invention discloses a reverse design method of a metal material metasurface based on transfer learning. The reverse design method comprises the following steps: 1, acquiring a metasurface data set of a source task; 2, constructing a network model of a series architecture, which comprises a forward prediction network module and a reverse design network module, and is used for processing P and A to obtain a prediction value, constructing a loss function and obtaining a trained series network model; 3, acquiring a metasurface data set of a migration task, performing migration learning, migrating a pre-trained forward prediction network module and a trained series network model, processing # imgabs0 # and # imgabs1 # to obtain predicted values for constructing a loss function, and obtaining a pre-trained forward migration model and a migration series network model; and 4, inputting the target spectrum to the migration series network model to obtain reasonable metasurface unit parameters. According to the method, the efficiency and precision of spectrum prediction and structure design can be improved, and the design process of the metasurface can be accelerated remarkably, so that the design period is shortened.

Description

Technical Field

[0001] The present invention belongs to the technical field of novel artificial composite material supersurfaces, and specifically is a reverse design method of metal material supersurfaces based on transfer learning. Background Art

[0002] With the widespread application of metasurfaces in optics, optoelectronics, wireless communications and other fields, the design of metasurfaces has become a hot topic in current research. The design of metasurfaces usually relies on precise electromagnetic calculations and simulations, involving a large amount of data processing and computing resources, especially in the design of metasurfaces made of metal materials. The spectral characteristics of different metal materials have an important influence on the performance of metasurfaces, so accurately predicting these spectral characteristics and performing structural design has become a technical challenge. Traditional design methods rely on experience and a large number of parameter scans. This process is not only computationally expensive, but also time-consuming, which limits the design efficiency and optimization capabilities.

[0003] Although the deep learning methods in the prior art have made significant progress in the fields of image recognition and natural language processing, their application in metasurface design still faces some difficulties. Deep learning models usually require large-scale data sets for training in order to effectively learn the optical properties of specific materials and structures. However, the design data sets of metasurfaces are often difficult to obtain, and training deep learning models requires a lot of computing resources and time, resulting in high costs in the training process. Therefore, how to reduce dependence on large data sets and improve the efficiency and accuracy of metasurface design has become a key issue that needs to be solved in this field. Summary of the invention

[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and proposes a reverse design method for metal material metasurfaces based on transfer learning, in order to improve the efficiency and accuracy of spectral prediction and structural design, and to significantly accelerate the design process of the metasurface, thereby shortening the design cycle.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0006] The inverse design method of a metal material super surface based on transfer learning of the present invention is characterized in that it comprises the following steps:

[0007] Step 1: Obtain the hypersurface dataset of the source task;

[0008] Step 1.1: Get the metal material parameters represented by the Drude model { } and randomly generated structure parameters { } and constitute the parameters of the metasurface unit , ,in, They represent the plasma resonance frequency and plasma collision frequency characterized by the Drude model, represent the period, thickness, length and width of the metasurface unit respectively;

[0009] Step 1.2: Obtain the spectral data corresponding to the parameter P of the metasurface unit ,in, represents the spectral value at the i-th sampling point; N represents the total number of sampling points;

[0010] Step 2: Build a network model TNN with a tandem architecture, including: forward prediction network module FPN and inverse design network module IDN, and use it to and Processing is performed to obtain the predicted value, which is used to construct the loss function L TNN , thereby training TNN and obtaining the trained series network model TNN_1;

[0011] Step 3: Obtain the hypersurface dataset of the migration task and perform migration learning to obtain the migration pre-trained forward prediction network module FPN_1 and the trained tandem network model TNN_1 for and Processing is performed to obtain the corresponding predicted value, thereby constructing the loss function And the loss function L TNN-Transfer , which are used to train the transfer models FPN-Transfer and TNN-Transfer respectively, and obtain the pre-trained forward transfer model FPN_2 and the transfer series network model TNN_2;

[0012] Step 4: Sample the target spectral data Input into the trained migration series network model TNN_2, and get The corresponding predicted parameters of the hypersurface unit as well as Spectral verification data after FPN_2 processing in TNN_2 ,in, represents the j-th parameter prediction value of the hypersurface unit parameter output by the trained migration series network model TNN_2, represents the spectral prediction value of the i-th sampling point output by the migration series network model TNN_2, As the super surface unit parameters that meet the target requirements, and using test consistency of goals.

[0013] The inverse design method of a metal material supersurface based on transfer learning described in the present invention is also characterized in that the step 2 includes the following steps:

[0014] Step 2.1: Construct the forward prediction network module FPN, including: normalization layer, residual block, fully connected layer and activation function ReLu, and Process and output Corresponding spectral prediction data ,in, Represents the spectral prediction value of the i-th sampling point;

[0015] Step 2.2: Based on and Construct the forward loss function L A , used to train the forward prediction network module FPN to obtain the pre-trained forward prediction network module FPN_1;

[0016] Step 2.3: Construct a reverse design network module IDN with the same structure as the forward prediction network module FPN, freeze the network weights of the pre-trained forward prediction network module FPN_1, and then connect FPN_1 and the reverse design network module IDN in series to obtain a series network TNN;

[0017] The tandem network TNN Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN, respectively. and equal;

[0018] Will Input into FPN_1 of the series network TNN for processing, and get Corresponding spectral prediction data ,in, Represents the final spectral prediction value of the i-th sampling point output by TNN;

[0019] Step 2.4: Based on as well as Construct the loss function L of the cascade network TNN , and used to train TNN to obtain the trained series network model TNN_1;

[0020] LTNN (1)

[0021] In formula (1), represents the jth parameter of the hypersurface unit parameter, represents the j-th parameter prediction value of the hypersurface unit parameter output by the series network; .

[0022] Furthermore, the step three includes the following steps:

[0023] Step 3.1: After changing the metal material used by the hypersurface unit, follow the process of step 1 to obtain the hypersurface data set of the migration task, including: the parameters of the hypersurface unit of the migration task as well as Corresponding spectral data ,in, They represent the plasma resonance frequency and plasma collision frequency of the metal material used by the metasurface unit of the migration task characterized by the Drude model, represent the period, thickness, length and width of the metasurface unit of the migration task respectively, represents the spectral value of the i-th sampling point of the migration task;

[0024] Step 3.2: Freeze the network weights of the first m layers of the pre-trained forward prediction network FPN_1 and use it as the forward prediction model FPN-Transfer for the migration task. Process and output Corresponding spectral prediction data ,in, Represents the spectral prediction value of the i-th sampling point of the migration task;

[0025] Step 3.3: Based on and Constructing the loss function of the forward transfer model FPN-Transfer , and used to train FPN-Transfer to obtain the pre-trained forward transfer model FPN_2;

[0026] (2)

[0027] Step 3.4: Replace FPN_1 in the trained series network model TNN_1 with FPN_2 to obtain the series network model TNN-Transfer for the migration task, and Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN-Transfer, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN-Transfer, respectively. and equal;

[0028] Will Input into FPN_2 of TNN_1 and process it to get Corresponding spectral prediction data , Represents the spectral prediction value of the i-th sampling point output by TNN-Transfer;

[0029] Step 3.5: Based on as well as Construct the loss function L TNN-Transfer , and train TNN-Transfer to obtain the trained transfer series network model TNN_2;

[0030] (3)

[0031] In formula (3), The jth parameter representing the hypersurface unit parameter of the migration task, The jth parameter prediction value of the hypersurface unit parameter of the migration task representing the output of the cascaded network; .

[0032] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the reverse design method, and the processor is configured to execute the program stored in the memory.

[0033] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the reverse design method when the computer program is executed by a processor.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. In the process of acquiring the data set, the present invention not only considers the structural parameters of the hypersurface, but also uses the Drude model to parametrically characterize the metal materials involved in the hypersurface structure; in the process of transfer learning, the material properties of metals are different but obey the same material model, which enhances the commonality between tasks.

[0036] 2. The present invention reduces the dependence of deep learning on training data, saving data and computing costs. Traditional metasurface design methods usually rely on a large amount of experimental data or electromagnetic simulation calculations, which is not only time-consuming but also computationally expensive. The present invention uses transfer learning technology to apply the experience learned by the network model from the metasurface data of known metal materials to the design of different metal materials, greatly reducing the dependence on large-scale data sets, thereby reducing the time and cost of data collection and processing.

[0037] 3. The present invention improves the efficiency and accuracy of spectral prediction and structural design by transferring the knowledge of source tasks based on the transfer learning model, and significantly shortens the design cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention;

[0039] Figure 2 This is a residual block structure diagram of the present invention.

[0040] Figure 3 It is a diagram of the series network structure of the present invention. DETAILED DESCRIPTION

[0041] In this embodiment, a reverse design method of a metal material super surface based on transfer learning is provided. The specific process is as follows: Figure 1 As shown, the following steps are included:

[0042] Step 1: Obtain the hypersurface dataset of the source task;

[0043] Step 1.1: Get the metal material parameters represented by the Drude model { } and randomly generated structure parameters { } and constitute the parameters of the super surface unit , ,in, They represent the plasma resonance frequency and plasma collision frequency characterized by the Drude model, They respectively represent the period, thickness, length and width of the metasurface unit; in this embodiment, the metal material used is aluminum, the period range of the metasurface unit is 0.78~0.88um, the thickness range is 0.07~0.17um, the length range is 0.72~0.84um, and the width range is 0.36~0.42um. The parameters of the metasurface unit are randomly combined to obtain 5682 sets of data.

[0044] Step 1.2: Obtain the spectral data corresponding to the parameter P of the metasurface unit ,in, Represents the spectral value at the i-th sampling point; N represents the total number of sampling points; in this embodiment, Python-FDTD is used for joint simulation, and the simulation wavelength is set to 400~800nm, that is, the visible light band. The absorption spectrum corresponding to each set of metasurface unit parameters is collected, and the absorption spectrum data is uniformly sampled at 200 points to obtain the corresponding 5682 sets of spectral data.

[0045] Step 2: Build a network model TNN with a tandem architecture, including: forward prediction network module FPN and inverse design network module IDN, and use it to and Processing is performed to obtain the predicted value, which is used to construct the loss function, so as to train the TNN and obtain the trained series network model TNN_1;

[0046] Step 2.1: Construct the forward prediction network module FPN, including: normalization layer, residual block, fully connected layer and activation function ReLu, and Process and output Corresponding spectral prediction data ,in, represents the spectral prediction value of the i-th sampling point; in this embodiment, the network consists of 7 layers of full connection interspersed with 2 residual blocks, the loss function of the first 6 layers is ReLu, the loss function of the output layer is Sigmoid, and the number of neurons in the fully connected layer is 32, 64, 128, 256, 256, 256, and 200 respectively; the structure of the residual block is as follows Figure 2 shown.

[0047] Step 2.2: Based on and Construct the forward loss function L A , used to train the forward prediction network module FPN to obtain the pre-trained forward prediction network module FPN_1; in this embodiment, the Adam optimizer is used, the learning rate is set to 0.00003, the average value of the loss function of 5682 groups of data is calculated, the training iteration is 500 times, and the loss function threshold is 0.001.

[0048] Step 2.3: Construct a reverse design network module IDN with the same structure as the forward prediction network module FPN, freeze the network weights of the pre-trained forward prediction network module FPN_1, and then connect FPN_1 and the reverse design network module IDN in series to obtain a series network TNN; in this embodiment, the loss function of the reverse design network module IDN all uses ReLu, and the structure of the series network is as follows: Figure 3 shown.

[0049] The tandem network TNN Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN, respectively. and equal;

[0050] Will Input into FPN_1 of the series network TNN for processing, and get Corresponding spectral prediction data ,in, Represents the final spectral prediction value of the i-th sampling point output by TNN.

[0051] Step 2.4: Based on as well as Construct the loss function L of the cascade network TNN , and used to train TNN to obtain the trained series network model TNN_1;

[0052] L TNN (1)

[0053] In formula (1), represents the jth parameter of the hypersurface unit parameter, represents the j-th parameter prediction value of the hypersurface unit parameter output by the series network; .

[0054] Step 3: Obtain the hypersurface dataset of the migration task and perform migration learning, migrate the pre-trained forward prediction network module FPN_1 and the trained tandem network model TNN_1, and Processing is performed to obtain the predicted value, which is used to construct the loss function, and the transfer models FPN-Transfer and TNN-Transfer are trained respectively to obtain the pre-trained forward transfer model FPN_2 and the transfer series network model TNN_2;

[0055] Step 3.1: After changing the metal material used by the hypersurface unit, follow the process of step 1 to obtain the hypersurface data set of the migration task, including: the parameters of the hypersurface unit of the migration task as well as Corresponding spectral data ,in, They represent the plasma resonance frequency and plasma collision frequency of the metal material used by the metasurface unit of the migration task characterized by the Drude model, represent the period, thickness, length and width of the metasurface unit of the migration task respectively, Represents the spectral value of the i-th sampling point of the migration task; in this embodiment, the metal material used is changed to gold, the ranges of period, thickness, length, and width are consistent with step one, and the parameters of the metasurface unit are randomly combined to obtain 5682 sets of data; the simulation wavelength setting remains unchanged, and the sampling data of the absorption spectrum corresponding to the metasurface unit of the migration task is obtained, and the sampling point is 200.

[0056] Step 3.2: Freeze the network weights of the first m layers of the pre-trained forward prediction network FPN_1. In this embodiment, m=5, and use it as the forward prediction model FPN-Transfer for the migration task. Process and output Corresponding spectral prediction data ,in, Represents the spectral prediction value of the i-th sampling point of the migration task; In this embodiment, in order to verify the role of transfer learning in reducing the amount of training data, the relationship between the amount of training data and the performance of the network model is discussed, and the Pearson correlation coefficient is used for characterization. It is finally determined that the amount of training data used for the migration task is 1000 sets of data, which can achieve the prediction effect of direct learning with 5682 sets of data.

[0057] Step 3.3: Based on and Construct the loss function of the forward transfer model FPN-Transfer and use it to train FPN-Transfer to obtain the pre-trained forward transfer model FPN_2;

[0058] (2)

[0059] Step 3.4: Replace FPN_1 in the trained series network model TNN_1 with FPN_2 to obtain the series network model TNN-Transfer for the migration task, and Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN-Transfer, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN-Transfer, respectively. and equal.

[0060] Will Input into FPN_2 of TNN_1 and process it to get Corresponding spectral prediction data , Represents the spectral prediction value of the i-th sampling point output by TNN-Transfer.

[0061] Step 3.5: Based on as well as Construct the loss function L TNN-Transfer , and train TNN-Transfer to obtain the trained transfer series network model TNN_2;

[0062] L TNN-Transfer (3)

[0063] In formula (3), The jth parameter representing the hypersurface unit parameter of the migration task, The jth parameter prediction value of the hypersurface unit parameter of the migration task representing the output of the cascaded network; .

[0064] Step 4: Sample the target spectral data Input into the trained migration series network model TNN_2, and get The corresponding predicted parameters of the hypersurface unit as well as Spectral verification data after FPN_2 processing in TNN_2 ,in, represents the j-th parameter prediction value of the hypersurface unit parameter output by the trained migration series network model TNN_2, represents the spectral prediction value of the i-th sampling point output by the migration series network model TNN_2, As the super surface unit parameters that meet the target requirements, and using test In this embodiment, the target spectrum is the data sampling of the custom spectrum that achieves 60% absorption at 550nm, and the sampling points are 200. The reasonable hypersurface unit parameters are obtained through the trained migration series network model TNN_2, and the predicted spectrum data corresponding to the hypersurface unit parameters output by the network are compared with the target spectrum data for verification, proving that the hypersurface unit parameters given by the network meet the requirements for achieving the target spectrum.

[0065] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0066] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A reverse design method for metal material supersurface based on transfer learning, characterized in that: The following steps are involved: Step 1: Obtain the hypersurface dataset of the source task; Step 1.1: Get the metal material parameters represented by the Drude model { } and randomly generated structure parameters { } and constitute the parameters of the metasurface unit , ,in, They represent the plasma resonance frequency and plasma collision frequency characterized by the Drude model, represent the period, thickness, length and width of the metasurface unit respectively; Step 1.2: Obtain the spectral data corresponding to the parameter P of the metasurface unit ,in, represents the spectral value at the i-th sampling point; N represents the total number of sampling points; Step 2: Build a network model TNN with a tandem architecture, including the forward prediction network module FPN and the inverse design network module IDN, and use it to and Processing is performed to obtain the predicted value, which is used to construct the loss function L TNN , thereby training TNN and obtaining the trained series network model TNN_1; Step 3: Obtain the hypersurface dataset of the migration task and perform migration learning to obtain the migration pre-trained forward prediction network module FPN_1 and the trained tandem network model TNN_1 for and Processing is performed to obtain the corresponding predicted value, thereby constructing the loss function And the loss function L TNN-Transfer , which are used to train the transfer models FPN-Transfer and TNN-Transfer respectively, and obtain the pre-trained forward transfer model FPN_2 and the transfer series network model TNN_2; Step 4: Sample the target spectral data Input into the trained migration series network model TNN_2, and get The corresponding predicted parameters of the hypersurface unit as well as Spectral verification data after FPN_2 processing in TNN_2 ,in, represents the j-th parameter prediction value of the hypersurface unit parameter output by the trained migration series network model TNN_2, represents the spectral prediction value of the i-th sampling point output by the migration series network model TNN_2, As the super surface unit parameters that meet the target requirements, and using test consistency of goals.

2. According to the inverse design method of a metal material super surface based on transfer learning in claim 1, it is characterized in that: The step 2 comprises the following steps: Step 2.1: Construct the forward prediction network module FPN, including: normalization layer, residual block, fully connected layer and activation function ReLu, and Process and output Corresponding spectral prediction data ,in, Represents the spectral prediction value of the i-th sampling point; Step 2.2: Based on and Construct the forward loss function L A , used to train the forward prediction network module FPN to obtain the pre-trained forward prediction network module FPN_1; Step 2.3: Construct a reverse design network module IDN with the same structure as the forward prediction network module FPN, freeze the network weights of the pre-trained forward prediction network module FPN_1, and then connect FPN_1 and the reverse design network module IDN in series to obtain a series network TNN; The tandem network TNN Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN, respectively. and equal; Will Input into FPN_1 of the series network TNN for processing, and get Corresponding spectral prediction data ,in, Represents the final spectral prediction value of the i-th sampling point output by TNN; Step 2.4: Based on as well as Construct the loss function L of the cascade network TNN , and used to train TNN to obtain the trained series network model TNN_1; L TNN (1) In formula (1), represents the jth parameter of the hypersurface unit parameter, represents the j-th parameter prediction value of the hypersurface unit parameter output by the series network; .

3. The inverse design method of a metal material supersurface based on transfer learning according to claim 2 is characterized in that: The step three comprises the following steps: Step 3.1: After changing the metal material used by the hypersurface unit, follow the process of step 1 to obtain the hypersurface data set of the migration task, including: the parameters of the hypersurface unit of the migration task as well as Corresponding spectral data ,in, They represent the plasma resonance frequency and plasma collision frequency of the metal material used by the metasurface unit of the migration task characterized by the Drude model, represent the period, thickness, length and width of the metasurface unit of the migration task respectively, Represents the spectral value of the i-th sampling point of the migration task; Step 3.2: Freeze the network weights of the first m layers of the pre-trained forward prediction network FPN_1 and use it as the forward prediction model FPN-Transfer for the migration task. Process and output Corresponding spectral prediction data ,in, Represents the spectral prediction value of the i-th sampling point of the migration task; Step 3.3: Based on and Constructing the loss function of the forward transfer model FPN-Transfer , and used to train FPN-Transfer to obtain the pre-trained forward transfer model FPN_2; (2) Step 3.4: Replace FPN_1 in the trained series network model TNN_1 with FPN_2 to obtain the series network model TNN-Transfer for the migration task, and Process and output The corresponding predicted parameters of the hypersurface unit ,in, They represent the predicted values ​​of plasma resonance frequency and plasma collision frequency output by the inverse design module IDN in TNN-Transfer, respectively. They represent the predicted period, thickness, length and width of the hypersurface unit output by the inverse design module IDN in TNN-Transfer, respectively. and equal; Will Input into FPN_2 of TNN_1 and process it to get Corresponding spectral prediction data , Represents the spectral prediction value of the i-th sampling point output by TNN-Transfer; Step 3.5: Based on as well as Construct the loss function L TNN-Transfer , and train TNN-Transfer to obtain the trained transfer series network model TNN_2; (3) In formula (3), The jth parameter representing the hypersurface unit parameter of the migration task, The jth parameter prediction value of the hypersurface unit parameter of the migration task representing the output of the cascaded network; .

4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports a processor to execute the reverse design method described in any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the reverse design method according to any one of claims 1 to 3 are executed.

Images