A ring-shaped metamaterial counter-design system based on contrast learning

Abstract

The application discloses a kind of circular ring metamaterial counter-design systems based on contrast learning, the system aims at solving the problems such as slow progress of metamaterial design, large demand data volume, need a lot of experience on metamaterial design, use contrast learning method to speed up its metamaterial design process, comprising: constructing metamaterial, obtaining metamaterial simulation data;Metamaterial simulation data is reduced dimension and normalized;Divide data set into training set and test set;Reflectance curve encoder, circular ring structure encoder and similarity calculation matrix are respectively constructed using training set;For the contrast learning architecture built, use 5-fold cross-validation to distinguish its specific performance;Model parameters are modified using test set, and the optimal counter-design result is obtained;The model is exported and used as a metamaterial counter-design system;The method can accurately design metamaterial without parameterization scanning and prior knowledge, and speeds up the design process of metamaterial.

Description

Technical Field

[0001] This invention belongs to the field of intelligent applications of deep learning algorithms, and particularly relates to a toroidal metamaterial inverse design system based on contrastive learning. Background Technology

[0002] Currently, metamaterial design mainly relies on computational electromagnetics methods such as finite-difference time-domain, finite element method, finite integral method, and method of moments under simulation platforms such as HFSS, CST, and COMSOL to directly solve Maxwell's equations numerically. Despite having different application scenarios and functional requirements, metamaterial research still relies heavily on direct inspiration from physics. That is, researchers formulate specific electromagnetic property indicators such as amplitude, bandwidth, phase, and polarization based on actual application scenarios and functional requirements. With their own design experience or understanding of physical properties, they continuously search, design, and iteratively optimize to find a relatively optimal design scheme.

[0003] Deep learning, a branch of machine learning, is an algorithm that uses artificial neural networks as its architecture to learn representations of data. Currently, deep learning has received widespread attention from both academia and business. As researchers delve deeper into the study of deep learning algorithms, it has begun to penetrate various industries, including human-computer interaction, image segmentation and recognition, autonomous driving, text and speech translation, advertising recommendation, data mining, financial analysis, medical diagnosis, logistics and transportation, scientific exploration, video games, and artistic creation—a wide range of engineering applications. In recent years, deep learning has also achieved numerous significant results in scientific research, such as using neural networks to predict protein structures and using deep learning to help discover mathematical conjectures and theorems. Summary of the Invention

[0004] The purpose of this invention is to provide a toroidal metamaterial inverse design system based on comparative learning. This system solves the problems of slow progress, complicated process, and the need for a lot of experience in metamaterial design, and meets the design requirements of most functions of toroidal metamaterials.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0006] A toroidal metamaterial inverse design system based on contrastive learning includes the following steps:

[0007] S1: Simulate the toroidal metamaterial using simulation software;

[0008] S2: Preprocess the raw data;

[0009] S3: Divide the data into a training set and a test set;

[0010] S4: Build a contrastive learning neural network architecture;

[0011] S5: Based on the model training results, test the model and adjust the model architecture to obtain the relatively optimal reverse design results;

[0012] S6: Use the trained model to reverse engineer metamaterials.

[0013] Furthermore, the specific process of step S1 is as follows:

[0014] Step S1.1: Select simulation materials. Select copper for both the surface pattern material and the backplate material of the metamaterial, and select FR-4 standard material for the substrate material.

[0015] Step S1.2: Select the simulation frequency. Select a broadband frequency of 2 to 30 GHz for simulation.

[0016] Step S1.3: Material size settings: Set the thickness of the base material to 1.5mm, and the length and width to 9mm; set the thickness of the metal back plate to 0.018mm, and the length and width to 9mm; set the thickness of the surface pattern material to 0.018mm.

[0017] Step S1.4: Construct a metamaterial surface pattern by dividing the radius from 0 to 4 mm into rings at intervals of 0.25 mm. Each ring has an equal probability of being selected. Randomly select any number of rings to form a set of concentric ring metamaterial surface patterns.

[0018] Furthermore, the specific process of step S2 is as follows:

[0019] Step S2.1: Extract data, using the reflection curve from the extracted simulation curve as the raw data;

[0020] Step S2.2: Data normalization. Extract the maximum and minimum values ​​from the data and perform linear normalization to scale the data to the interval between 0 and 1. The transformation formula is as follows:

[0021]

[0022] Where max is the maximum value of the sample and min is the minimum value of the sample.

[0023] Step S2.3: Dimensionality reduction of data. Remove equally spaced points from the complete data so that all variable-length data are reduced to the same dimension.

[0024] Furthermore, the specific process of step S3 is as follows:

[0025] Step S3.1: Divide the dataset into a training set and a test set in a 9:1 ratio;

[0026] Step S3.2: Divide the training set into cross-training and validation sets according to the 5-fold cross-validation ratio.

[0027] Furthermore, the specific process of step S4 is as follows:

[0028] Step S4.1: The reflection curve encoder is constructed, mainly consisting of 6 parts. The first part is used to expand the data dimension from 1 dimension to 64 dimensions; the second part is used to expand the data dimension from 64 dimensions to 256 dimensions; the third part is the residual connection from 64 dimensions to 256 dimensions; the fourth part is used to expand the data dimension from 256 dimensions to 1024 dimensions; the fifth part is the residual connection from 256 dimensions to 1024 dimensions; and the sixth part is the dimensional linear mapping unit, which is used to expand the data dimension and remap it to a 1024-dimensional vector.

[0029] Step S4.2: The circular ring structure encoder is built, which mainly consists of 1 part, which directly maps the one-dimensional data of the circular ring structure into a 1024-dimensional vector;

[0030] Step S4.3: Constructing the similarity calculation matrix involves multiplying two 1024-dimensional vectors to obtain the similarity matrix.

[0031] Furthermore, the specific process of step S5 is as follows:

[0032] Step S5.1: Divide the data into data packets of size 512, and train on all data in batches;

[0033] Step S5.2: Use 5-fold cross-validation to average the 5-fold validation results and use this average to guide the adjustment of the model architecture.

[0034] Furthermore, the specific process of step S6 is as follows:

[0035] Step S6.1: Export the trained model as a ".pkl" file, which can be used to perform reverse engineering on given metamaterial reflection curve data.

[0036] Furthermore, the specific process of step S7 is as follows:

[0037] Step S7.1: Generate a full-space metamaterial pattern representation matrix with a radius of 0 to 4 mm;

[0038] Step S7.2: Input the expected performance curve and the full-space metamaterial pattern representation matrix into the trained model to obtain the reverse-designed metamaterial structure.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] The method of this invention can accurately design metamaterials without parametric scanning or prior knowledge, thus accelerating the metamaterial design process. Attached Figure Description

[0041] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0042] Figure 1 This invention relates to one type of metamaterial pattern and its digital representation method.

[0043] Figure 2 This is a flowchart of the present invention.

[0044] Figure 3 This is a diagram of the neural network architecture during training in this invention. Specific implementation methods

[0045] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0046] To better illustrate this embodiment, some steps in the accompanying drawings may be simplified and do not represent the actual detailed steps;

[0047] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] Part 1: Acquisition of Metamaterial Simulation Data.

[0050] Single metamaterial structures such as Figure 1 As shown, the entire structure consists of three layers with a total thickness of 1.536 mm. The bottom layer is a square metal backplate made of pure copper with a side length s of 9 mm. The middle layer is a dielectric layer made of FR-4 material, also with a side length s of 9 mm and a thickness h of 1.5 mm. The top layer is a pattern layer, whose pattern consists of a randomly generated 16-bit sequence. Each bit represents the presence or absence of a copper-made ring within a radius of 0.25 mm. The maximum diameter c of the ring is 8 mm. This layer has the same thickness as the bottom backplate.

[0051] Use a Python script to randomly generate a 16-bit sequence, and use this sequence to construct something similar to... Figure 1 The metamaterial structure is shown. The structure was simulated using the CST microwave studio, and the simulated spectral curve S11 was derived as the input value for the inverse design of the toroidal material. The corresponding 16-bit sequence of the structure was used as the target value for the inverse design of the toroidal metamaterial.

[0052] Repeat the above process 25,000 times to obtain a total of 25,000 pairs of spectral responses and structural sequences.

[0053] Dimensionality reduction is performed on spectral response data of different lengths by deleting data points from the spectral data at equal intervals, thereby reducing the length of the data to the same length.

[0054] All spectral response data are linearly normalized to scale them to the 0-1 range. The transformation formula is as follows:

[0055]

[0056] Where max is the maximum value of the sample and min is the minimum value of the sample.

[0057] The dataset was divided into training and testing sets in a 9:1 ratio; the training set was further divided into cross-training and validation sets according to a 5-fold cross-validation ratio.

[0058] Part Two: Building and Training Neural Network Architectures

[0059] The overall architecture of the neural network consists of three parts, the first of which is the spectral data encoder.

[0060] Because spectral data is complex and contains a wide variety of patterns, spectral data encoders employ complex methods, such as... Figure 3 Each conv module represents a convolutional layer, with the following numbers representing the kernel size and the number of output patterns for that convolutional layer, respectively; each BN module represents a batch normalization layer, with the following number representing the number of patterns it processes; and each ReLU module represents a rectified linear unit. After the initial convolutional layer extracts 64 primary patterns, each subsequent convolutional layer doubles the number of patterns.

[0061] What's special about this system is that after the first convolutional layer, a residual connection module is added every two convolutional layers. Each residual connection module also consists of a convolutional layer, a batch normalization layer, and a linear correction unit. It is used to establish skip connections between low-number-of-mode and high-number-of-mode systems.

[0062] Ultimately, the spectral data was processed into a vector of length 1024.

[0063] The second part is the structure encoder.

[0064] Because the structural data is relatively simple, the structural encoder consists of only three layers: the first layer is a linear fully connected layer, followed by a number indicating the size of its output vector; the second layer is a batch normalization layer; and the third layer is a linear correction unit.

[0065] Ultimately, the structured data was also processed into a vector of length 1024.

[0066] The third part is the similarity calculation module.

[0067] It multiplies the final vector of the structural data with the final vector of the spectral data to calculate their similarity, thus obtaining a similarity matrix.

[0068] Finally, the neural network uses the similarity calculation results to calculate the cross-entropy loss, which maximizes the diagonal elements of the similarity matrix and minimizes the remaining elements.

[0069] The previously segmented spectral response data and metamaterial structure sequence were randomly selected in batches. Each batch of data was input into the corresponding encoder, and the Adam optimization algorithm was selected to optimize the neural network.

[0070] We use 5-fold cross-validation to verify the optimality of the neural network structure.

[0071] The neural network architecture is continuously adjusted to minimize the loss value when the average of the 5-fold cross-validation results is taken, while also minimizing the loss value of the test data.

[0072] Part 3: Inverse design of toroidal metamaterials using neural networks.

[0073] Export the trained model as a ".pkl" file.

[0074] First, a 16-bit sequence is generated from 0 to 65536, which is 2 to the power of 16. This sequence represents the complete structural sequence of all space metamaterials corresponding to this invention.

[0075] Input the desired metamaterial performance curves into the ".pkl" file, along with the complete metamaterial structure sequence. The trained neural network will calculate the similarity between the two; the structure sequence with the highest similarity is the inverse design result of the toroidal metamaterial.

[0076] In cases where different metamaterial structures correspond to the same spectral response curve, this invention can calculate the metamaterial structures in the entire space based on similarity ranking and select multiple metamaterial structures that correspond to the desired spectral response curve.

[0077] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The toroidal metamaterial inverse design system based on comparative learning can accurately design metamaterials without parametric scanning or prior knowledge, thus accelerating the metamaterial design process.

[0078] The description of the embodiments of this invention is only for the purpose of helping to understand the method of this invention, and is merely a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention, and the content of this specification should not be construed as a limitation of this invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0079] In summary, this invention discloses a toroidal metamaterial inverse design system based on contrastive learning. This system aims to address the problems of slow metamaterial design progress, large data requirements, and the need for extensive experience in metamaterial design. It uses a contrastive learning method to accelerate the metamaterial design process, including: constructing the metamaterial and acquiring metamaterial simulation data; performing dimensionality reduction and normalization on the metamaterial simulation data; dividing the dataset into training and testing sets; constructing a reflection curve encoder, a toroidal structure encoder, and a similarity calculation matrix using the training set; using 5-fold cross-validation to differentiate the specific performance of the constructed contrastive learning architecture; modifying model parameters using the testing set to obtain the optimal inverse design result; and exporting the model for use as a metamaterial inverse design system. This method can accurately design metamaterials without parametric scanning or prior knowledge, thus accelerating the metamaterial design process.

Claims

1. A toroidal metamaterial inverse design system based on contrastive learning, characterized in that, Specifically, the steps include the following: S1: Simulate the toroidal metamaterial using simulation software; S2: Preprocess the raw data; S3: Divide the data into a training set and a test set; S4: Build a contrastive learning neural network architecture. Step S4.1: The reflection curve encoder is constructed, mainly consisting of 6 parts. The first part is used to expand the data dimension from 1 dimension to 64 dimensions; the second part is used to expand the data dimension from 64 dimensions to 256 dimensions; the third part is the residual connection from 64 dimensions to 256 dimensions; the fourth part is used to expand the data dimension from 256 dimensions to 1024 dimensions; the fifth part is the residual connection from 256 dimensions to 1024 dimensions; and the sixth part is the dimensional linear mapping unit, which is used to expand the data dimension and remap it to a 1024-dimensional vector. Step S4.2: The circular ring structure encoder is built, which mainly consists of 1 part, which directly maps the one-dimensional data of the circular ring structure into a 1024-dimensional vector; Step S4.3: Constructing the similarity calculation matrix involves multiplying two 1024-dimensional vectors to obtain the similarity matrix; S5: Based on the model training results, test the model and adjust the model architecture to obtain the relatively optimal reverse design results; S6: Use the trained model to reverse engineer metamaterials; S7: Generate full-space toroidal representation data with radii from 0 to 4 mm for metamaterial reverse design.

2. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S1 specifically includes the selection of metamaterial materials, frequency selection, size setting, and methods for constructing metamaterial surface patterns: S1.1: The material selection is to select copper for both the surface pattern material and the backplate material of the metamaterial, and to select FR-4 standard material for the substrate material; S1.2: The frequency selection is to select a broadband frequency of 2 to 30 GHz for simulation; S1.3: The aforementioned size settings are as follows: the thickness of the base material is set to 1.5mm, and both the length and width are 9mm; the thickness of the metal back plate is set to 0.018mm, and both the length and width are 9mm; the thickness of the surface pattern material is set to 0.018mm. S1.4: The method for constructing the metamaterial surface pattern involves dividing the radius from 0 to 4 mm into rings at intervals of 0.25 mm, with each ring having an equal probability of being selected, and randomly selecting any number of rings to form a set of concentric circle metamaterial surface patterns.

3. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S2 specifically includes data extraction, data normalization, and data dimensionality reduction: S2.1: The data extraction refers to extracting the reflection curve from the simulation curve as the original data; S2.2: The data normalization involves extracting the maximum and minimum values ​​from the data and performing linear normalization on the data, scaling it to the range of 0 to 1. The transformation formula is as follows: Where max is the maximum value of the sample and min is the minimum value of the sample; The data dimensionality reduction described in S2.3 involves removing equally spaced points from the complete data, thereby reducing all variable-length data to the same dimension.

4. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S3 specifically includes: Step S3.1: Divide the dataset into a training set and a test set in a 9:1 ratio; Step S3.2: Further divide the training set into cross-training and validation sets according to the 5-fold cross-validation ratio.

5. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S5 specifically includes: Step S5.1: Divide the data into data packets of size 512, and train on all data in batches; Step S5.2: Use 5-fold cross-validation to average the 5-fold validation results and use this average to guide the adjustment of the model architecture.

6. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S6 specifically includes: Step S6.1: Export the trained model as a ".pkl" file, which can be used to perform reverse engineering on given metamaterial reflection curve data.

7. The toroidal metamaterial inverse design system based on contrastive learning as described in claim 1, characterized in that, Step S7 specifically includes: Step S7.1: Generate a full-space metamaterial pattern representation matrix from 0 to 8 mm; Step S7.2: Input the expected performance curve and the full-space metamaterial pattern representation matrix into the trained model to obtain the reverse-designed metamaterial structure.

Images