Deep learning-based ep-type resonant ring metamaterial counter-design system
By using a deep learning-based EP-type resonant ring metamaterial inverse design system and training a neural network with coupled mode parameters, the problems of complex metamaterial design process and high computational resource requirements are solved, enabling a faster and more transparent design process and enhancing the interpretability of physical mechanisms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2024-06-24
- Publication Date
- 2026-07-31
AI Technical Summary
Existing metamaterial reverse design processes are computationally intensive, cumbersome, and lack interpretability. Traditional methods require significant computational resources and specialized skills, and the design process is opaque.
A deep learning-based EP-type resonant ring metamaterial inverse design system is adopted. The system trains a neural network through coupled mode parameters, generates transmission curves, and performs inverse design. Simulation software is used to accelerate the design process and increase interpretability.
It improves the speed of metamaterial design, simplifies the design process, increases the interpretability of the design process, enables better exploration of physical mechanisms, and meets specific design needs.
Smart Images
Figure CN118692605B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent applications of deep learning algorithms, and in particular relates to a deep learning-based EP-type resonant ring metamaterial inverse design system. Background Technology
[0002] Currently, reverse design of metamaterials primarily relies on continuous parameter scanning using high-speed computers or parameter searching using various global or local optimization algorithms. For a specific reverse design, multiple simulations and modeling processes are often required. This involves optimizing a cost function to search the full parameter space for each task, iterating frequently to achieve the design requirements step by step. Therefore, the number of simulations the computer can perform, and the duration of each simulation, significantly determines the total time of the metamaterial reverse design phase. As the requirements for metamaterial reverse design continue to increase, these traditional design methods demand ever-higher computing power, and the complex and tedious design process places increasingly higher demands on researchers' time and expertise.
[0003] In the reverse design of metamaterials, deep learning can be used to obtain the expected electromagnetic response by optimizing the design parameter space without requiring initial guesses about the physical mechanisms. This design approach is mainly based on a combination of advanced algorithms and electromagnetic simulation software, seeking a solution system that minimizes the target distance (or maximizes the design performance), directly focusing on the target to solve the design problem. Once trained, the model can almost immediately solve the design problem. However, excessive disregard for physical mechanisms also makes the reverse design process using neural networks extremely opaque, lacking interpretability of the physical meaning of the data and the design process. By training with the coupled mode parameters of this EP-type resonant ring metamaterial, and utilizing the one-to-one correspondence between the coupled mode parameters and the metamaterial's transmission curve and structure, the interpretability of the data and design process is improved, making it easier to explore and learn its potential physical mechanisms. For traditional metamaterial design teams, this invention introduces machine learning into the metamaterial design process, while also introducing coupled mode parameters. While accelerating the design speed, it is easier to uncover the physical meaning behind the material's electromagnetic response, which has strong practical significance for achieving on-demand design for specific design goals. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning-based inverse design system for EP-type resonant ring metamaterials. This system utilizes deep learning to solve the problems of slow progress, complex process, and the need for a lot of experience in metamaterial design. At the same time, it differs from the traditional deep learning design process, increases the interpretability of the design process, is more conducive to exploring and learning its potential physical mechanisms, and better meets the design requirements of most functions of this type of metamaterial.
[0005] To achieve the above-mentioned technical effects, the technical system of the present invention is as follows:
[0006] A deep learning-based EP-type resonant ring metamaterial inverse design system specifically includes the following steps:
[0007] S1: Based on the coupled-mode equation, randomly select coupled-mode parameters to generate the transmission curve;
[0008] S2: Preprocess the data of the transmission curve;
[0009] S3: Divide the dataset into a training set and a test set;
[0010] S4: Build and train a contrastive learning neural network architecture;
[0011] S5: Simulation of EP-type resonant ring metamaterial using simulation software;
[0012] S6: Preprocess the simulated metamaterial data;
[0013] S7: Use a trained neural network architecture to extract the coupling mode parameters of the simulated metamaterial data;
[0014] S8: Using this model, the corresponding EP-type resonant ring metamaterial structure is designed in reverse based on the required transmission curve.
[0015] Furthermore, the specific process of step S1 is as follows:
[0016] Step S1.1: Using the coupled-mode equation, a series of transmission curves are generated by substituting random coupled-mode parameters, including γ. x γ y , Γ x , Γ y , κ. Its coupled mode equation is:
[0017]
[0018] Furthermore, the specific process of step S2 is as follows:
[0019] Step S2.1: The data extraction involves extracting the transmission curve as the raw data;
[0020] Step 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:
[0021]
[0022] Where max is the maximum value of the sample and min is the minimum value of the sample.
[0023] Step S2.3: The data dimensionality reduction involves removing 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 transmission curve data and coupling mode parameter data of different types into training set and test set according to a 9:1 ratio;
[0026] Step S3.2: Divide the training set into 5 equal parts for 5-fold cross-validation.
[0027] Furthermore, the specific process of step S4 is as follows:
[0028] Step S4.1: The coupled mode parameter data encoder is a neural network specifically built for pre-trained coupled mode parameter data, and its input data includes γ. x γ y , Γ x , Γ y κ. The output is a vector of size 1×1024;
[0029] Step S4.2: The transmission curve data encoder is a universal neural network built for all transmission curve data. Its input is spectral data of size 1×1×512 with a unit length width, single channel and length of 512, and the output is a vector of size 1×1024.
[0030] Step S4.3: The training of the neural network refers to using the neural network to make the corresponding coupling mode parameter data most similar to the transmission curve data.
[0031] Furthermore, the specific process of step S5 is as follows:
[0032] Step S5.1: The material selection involves choosing polyimide (PI) as the metamaterial intermediate layer and titanium and gold as the materials for the two resonant rings to obtain high loss contrast.
[0033] Step S5.2: The frequency selection is to select a broadband frequency of 2.0-2.8THz for simulation;
[0034] Step S5.3: The pattern setting refers to the arrangement of two resonant rings orthogonally to form a single structural unit, and a series of such structural units are arranged cyclically to form a rectangular surface;
[0035] Step S5.4: The size setting refers to setting the thickness h of the PI intermediate layer, the outer diameters d1 and d2 of the two resonant rings, the arc width w, and the gap sizes g1 and g2 of the openings of the two resonant rings. The outer diameters d1 and d2 of the two resonant rings should be the same, and the same resonant frequency should be achieved by using different gap sizes g1 and g2.
[0036] Step S5.5: The extraction of the transmission spectrum response of the metamaterial obtained from the simulation and the expression form of the simulated metamaterial refers to using the CST microwave studio to simulate EP-type resonant ring metamaterials with different structural parameters, and extracting the transmission curve of the simulated metamaterial and the expression form of the simulated metamaterial structure.
[0037] Furthermore, the specific process of step S6 is as follows:
[0038] Step S6.1: The data extraction involves extracting the simulated transmission curve as the raw data;
[0039] Step S6.1: 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:
[0040]
[0041] Step S6.1: Where max is the maximum value of the sample and min is the minimum value of the sample;
[0042] Step S6.1: The data dimensionality reduction involves removing equally spaced points from the complete data, so that all variable-length data are reduced to the same dimension.
[0043] Furthermore, the specific process of step S7 is as follows:
[0044] Step S7.1: Use the trained neural network architecture to extract the coupling mode parameters of the simulated metamaterial data through the transmission curve.
[0045] Furthermore, the specific process of step S8 is as follows:
[0046] Step S8.1: By using the one-to-one correspondence between the coupling mode parameters and the metamaterial structure, the structure of the EP-type resonant ring metamaterial is designed in reverse using the required transmission curve. Attached Figure Description
[0047] Figure 1 This is a system flowchart of the present invention.
[0048] Figure 2 This is a diagram of the neural network architecture of the present invention.
[0049] Figure 3This is a structural diagram of the EP-type resonant ring metamaterial involved in the present invention. Specific implementation methods
[0050] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some steps in the drawings may be simplified and do not represent the actual detailed steps. It is understandable that some well-known structures and their descriptions may be omitted in the drawings for those skilled in the art.
[0051] The technical system of the present invention will be further described below with reference to the accompanying drawings.
[0052] Part 1: Generate transmission curves based on the coupled-mode equations and randomly selected coupled-mode parameters, and extract and preprocess the transmission curve data.
[0053] By substituting random coupled-mode parameters into the coupled-mode equation, a series of transmission curves can be generated. The coupled-mode equation is:
[0054]
[0055] Random coupling mode parameters include γ x γ y , Γ x , Γ y , κ.
[0056] Data extraction is performed on the transmission curve.
[0057] The extracted transmission curve data is normalized by extracting the maximum and minimum values. Linear normalization is then applied to scale the data to the 0-1 range. The transformation formula is as follows:
[0058]
[0059] Where max is the maximum value of the sample and min is the minimum value of the sample.
[0060] The normalized data is reduced in dimensionality by removing equally spaced points from the complete data, so that all data of variable length are reduced to the same dimension.
[0061] Part Two: Neural Network Architecture Setup and Training
[0062] according to Figure 2 The architecture of the training neural network consists of three parts: a coupled mode parameter data encoder, a transmission curve data encoder, and a neural network training module.
[0063] The coupled-mode parameter data encoder input is coupled-mode parameter data of size 1×1×512, with a unit width, single channel, and a length of 512. Its input data includes γ.x γ y , Γ x , Γ y The output is a vector of size 1×1024; the transmission curve data encoder takes transmission curve data of size 1×1×512 with a unit width, single channel, and length of 512 as input, and outputs a vector of size 1×1024.
[0064] The outputs of both the coupled-mode parameter data encoder and the transmission curve data encoder must be divided by their own norm to perform vector normalization.
[0065] The normalized vectors are multiplied to perform preliminary neural network training. The goal of training the neural network is to make the corresponding coupling mode parameter data most similar to the transmission curve data.
[0066] We use 5-fold cross-validation to verify the optimality of the neural network structure.
[0067] 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.
[0068] Export the trained model as a ".pkl" file.
[0069] Part Three: Acquisition of Metamaterial Simulation Data.
[0070] EP-type resonant ring metamaterial structure, such as Figure 3 As shown, this is an ultrathin bilayer structure composed of two resonant rings with nearly identical resonant frequencies but different openings and high loss contrast. These resonant rings are placed on the top and bottom surfaces of a polyimide (PI) interlayer with orthogonal gaps. The two resonant rings are made of different metals; for example, titanium and gold can be used to achieve high loss contrast. The period of the unit cell is p, and the thickness of the PI interlayer is h. The outer diameters of the two resonant rings are d1 and d2, respectively, with d1 and d2 being the same. The arc width is w, and the distance between the two outer arcs is s. By changing the above simulation structural parameters, a series of this type of metamaterial and its corresponding transmission curves can be obtained.
[0071] The simulated metamaterial structure and transmission curves were extracted as raw data;
[0072] The original data is normalized by extracting the mean values of features from different data types and subtracting the mean from each, thus centering all data to 0. The transformation formula is as follows:
[0073]
[0074] To reduce the dimensionality of the data, equally spaced points are removed from the complete dataset, reducing all variable-length data to the same dimension.
[0075] Part 4: Extracting the coupling mode parameters of the simulated metamaterial using a trained neural network for reverse engineering.
[0076] Input the simulated metamaterial transmission curve into the ".pkl" file, and simultaneously input the coupling mode parameter sequence into the ".pkl" file.
[0077] By leveraging the correspondence between the coupling mode parameters and the EP-type resonant ring metamaterial structure, the metamaterial structure corresponding to the required transmission curve can be designed in reverse based on the desired transmission curve.
[0078] The descriptions of the embodiments herein are merely illustrative of the methods of the present invention and represent only a portion, not all, of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning based inverse design system for EP-type resonant ring metamaterials, characterized in that: Includes the following steps: S1: Based on the coupled-mode equation, randomly select coupled-mode parameters to generate the transmission curve; The coupled-mode equation is: The coupled mode parameters substituted include , , , , The horizontal axis of the transmission curve represents the incident frequency. The vertical axis represents transmittance; S2: Extract and preprocess the various data from the transmission curve; S3: Divide the dataset into a training set and a test set; S4: Build and train a contrastive learning neural network architecture; S4.1: The coupled mode parameter data encoder is a neural network specially built for pre-training coupled mode parameter data, the input data of which includes , , , , , and the output is a vector of size . S4.2: The aforementioned transmission curve data encoder is a universal neural network built for all transmission curve data, with an input of 512 units in width, single channel, and length. The size of the transmission spectrum data, output as A vector of size; S4.3: The training of the contrastive learning neural network is to make the corresponding coupled mode parameter data most similar to the transmission curve data; S5: Simulate the EP-type resonant ring metamaterial using simulation software; S6: Preprocess the simulated metamaterial data; S7: Use a trained neural network architecture to extract the coupling mode parameters of the simulated metamaterial data; S8: Using the coupling mode parameters, design the corresponding EP-type resonant ring metamaterial structure in reverse according to the required transmission curve.
2. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S2 includes data extraction, data normalization, and data dimensionality reduction of the transmission curve. S2.1: The data extraction mentioned above refers to extracting the transmission curve as the raw 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; S2.3: The data dimensionality reduction mentioned above is to remove equally spaced points from the complete data, so that all data of variable length are reduced to the same dimension.
3. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S3 specifically includes: S3.1: Divide the transmission curve data and coupling mode parameter data of different types into training set and test set according to a 9:1 ratio; S3.2: Divide the training set into 5 equal parts for 5-fold cross-validation.
4. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S5 specifically includes: material selection, frequency selection, pattern setting, size setting of the metamaterial, and extraction of the simulated metamaterial transmission curve and the simulated metamaterial representation. S5.1: The material selection is to choose polyimide as the metamaterial intermediate layer and titanium and gold as the materials of the two resonant rings to obtain high loss contrast. S5.2: The frequency selection is to select a broadband frequency of 2.0-2.8 THz for simulation; S5.3: The pattern setting refers to two resonant rings arranged orthogonally to form a single structural unit, and a series of such structural units are arranged cyclically to form a rectangular surface; S5.4: The size setting refers to setting the thickness h of the PI intermediate layer, the outer diameters d1 and d2 of the two resonant rings, the arc width w, and the gap sizes g1 and g2 of the openings of the two resonant rings. The outer diameters d1 and d2 of the two resonant rings should be the same, and the same resonant frequency can be achieved by setting different opening gap sizes g1 and g2. S5.5: The extraction of the metamaterial transmission spectrum response obtained from the simulation and the expression form of the simulated metamaterial refer to the use of CST microwave studio to simulate EP-type resonant ring metamaterials with different parameters, and to extract the metamaterial transmission curves and the expression form of the simulated metamaterial.
5. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S6 specifically includes: data extraction, data normalization, and data dimensionality reduction. S6.1: The data extraction mentioned above refers to using the extracted simulated transmission curve as the original data; S6.2: 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 S6.3 involves removing equally spaced points from the complete data, thereby reducing all variable-length data to the same dimension.
6. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S7 specifically includes: S7.1: Using a trained neural network architecture, the coupling mode parameters of the simulated metamaterial data are extracted through the transmission curve. These parameters include... , , , , This achieves a one-to-one correspondence between the transmission curve and the coupling mode parameters of the EP-type resonant ring metamaterial structure.
7. The deep learning based inverse design system of an EP-type resonant ring metamaterial according to claim 1, wherein: Step S8 specifically includes: S8.1: By using the one-to-one correspondence between the coupling mode parameters and the metamaterial structure, the required transmission curves are used to reverse-engineer the EP-type resonant ring metamaterial.