A method for intelligent design of a multiplexing superstructure device for terahertz fingerprint sensing

Abstract

The application provides a multiplexing superstructure device intelligent design method for terahertz fingerprint sensing, comprising the following steps: obtaining expected resonance frequencies, covering a target fingerprint frequency range through equidistant sampling to obtain a series of expected resonance frequencies; setting expected structure parameters, setting expected structure parameters of a target superstructure device according to actual physical limitations and processing conditions; combining the obtained expected resonance frequencies and the set expected structure parameters to form an input vector of a reverse design network; and predicting target structure parameters, constructing and training a reverse design network based on deep learning to predict remaining structure parameters of a series of superstructure surface modules capable of realizing the target fingerprint frequency range. The reverse design of the multiplexing superstructure device is realized through the deep learning neural network, and the device structure meeting the target fingerprint requirements can be quickly and accurately derived. The design precision and efficiency are significantly improved, and the ability to respond to changes in sensing requirements in real time is possessed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of terahertz fingerprint sensing and artificial intelligence algorithm, and particularly relates to a multiplexing superstructure device intelligent design method for terahertz fingerprint sensing. BACKGROUND

[0002] The multiplexing superstructure device for terahertz fingerprint sensing is an advanced device based on super surface design, which is specially used for realizing efficient multifunctional sensing and substance identification in the terahertz frequency band. Terahertz fingerprint sensing uses electromagnetic waves in the terahertz band (frequency range of 0.1 THz to 10 THz) to identify the chemical composition and physical properties of a substance by detecting its characteristic absorption spectrum, i.e. "molecular fingerprint spectrum". This technology has wide applications in chemical analysis, biomedical imaging, security inspection, communication and material science, because it can provide high-precision detection capability under label-free and non-destructive conditions.

[0003] Through the multiplexing technology, a superstructure device is composed of a series of super surface modules capable of realizing high-quality factor resonance, which can simultaneously sense multiple terahertz "fingerprint" characteristics, improving the detection efficiency and accuracy of the system. The resonant modes (such as guided mode resonance, Fano resonance, dipole resonance, etc.) used in traditional super surface modules usually have low quality factor, high radiation and energy loss, and insufficient local field enhancement. In contrast, the q-BIC mode realizes ultra-high quality factor and more significant local field enhancement, has low loss, high efficiency and greater design flexibility, and is therefore increasingly favored in modern super surface design, suitable for high-performance optical devices, sensors and nonlinear optical applications.

[0004] According to the occurrence mode, BIC is usually divided into two types, namely symmetric protection type BIC and accidental type BIC. Among them, accidental type BIC occurs accidentally due to its dependence on specific geometric parameters and material properties, making it difficult to explore the structure of its excitation condition. This is mainly due to the high sensitivity to parameters, computational complexity, lack of general design criteria, and challenges in experimental implementation. Traditional design methods usually rely heavily on the designer's solid physical reserves and deep experience accumulation, and require constant trial and error and simulation during the design process, resulting in a complex and time-consuming design process, and it is difficult to achieve optimal performance. Although the method of searching for optimal design parameters through optimization algorithm can overcome the complexity and high time consumption of the design process to some extent, due to the complexity of super surface modules and the high-dimensional design space, it often faces the problem of convergence difficulty.

[0005] In recent years, the application of machine learning algorithm to assist in the design of metasurfaces has attracted widespread attention. It can directly predict the required structure parameters through neural networks according to the desired optical response. Researchers quickly explore the design space through efficient model training and verification, thereby reducing experimental costs and time, and significantly accelerating the design process of single-resonance electromagnetic metasurfaces.

[0006] However, in the current mainstream spectral-based inverse design paradigm, designers often have difficulty providing physically realistic target spectra, which prevents the neural network from inversely predicting the required structure. At the same time, the geometric structure designed by the designer is often completely determined by the neural network, which causes some predicted structure parameters to be inconsistent with the physical reality or instrument processing limits.

[0007] In addition, in the application of terahertz fingerprint sensing, designers usually focus on the frequency range where the fingerprint features are located, which will affect the accuracy of device fingerprint sensing. Therefore, there is an urgent need for an inverse design paradigm based on the target fingerprint frequency range, which will greatly improve the design convenience of multiplexing superstructure devices for terahertz fingerprint sensing. At the same time, the predicted structure customization of the inverse design network will greatly improve the problem of designing structures that do not conform to the physical reality or instrument processing limits. SUMMARY

[0008] The purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide an intelligent design method for a multiplexing superstructure device for terahertz fingerprint sensing, which is used to predict the remaining structure parameters of the target superstructure device according to the target fingerprint frequency range and the desired structure parameters.

[0009] In a first aspect, the present application provides an intelligent design method for a multiplexing superstructure device for terahertz fingerprint sensing, which comprises the following steps:

[0010] Obtain the desired resonance frequency, cover the target fingerprint frequency range by equidistant sampling, and obtain a series of desired resonance frequencies;

[0011] Set the desired structure parameters, set the desired structure parameters of the target superstructure device according to the actual physical limitations and processing conditions;

[0012] Form an input vector, combine the obtained desired resonance frequencies and the set desired structure parameters to form an input vector of the inverse design network;

[0013] Predict the target structure parameters, construct and train the inverse design network based on deep learning, and predict the remaining structure parameters of a series of superstructure surface modules that can achieve the target fingerprint frequency range.

[0014] Preferably, the multiplexed metasurface device is composed of an artificial designed metasurface module, which adopts a metal split ring resonator (SRR) array structure and takes quasi-bound states in the continuum (q-BIC) as the resonant mode.

[0015] Preferably, the structure and substrate material of the multiplexed metasurface device are 200 nm thick gold (Au) and 200 μm thick silicon dioxide (SiO2), respectively, and the dimensions of the remaining metasurface unit structure parameters p and w are 120 μm and 8 μm, respectively.

[0016] In a second aspect, the embodiments of the present application provide a training method of a structure customizable metasurface inverse design network, which is used in the prediction of target structure parameters in the intelligent design method of the terahertz fingerprint sensing oriented multiplexed metasurface device of the first aspect. The training method comprises the following steps:

[0017] An electromagnetic simulation software is used to model a given metasurface module, calculate the corresponding spectral response, and obtain an original data set;

[0018] The original data set is preprocessed, including resonance position calibration, data cleaning, normalization, and structure parameter splitting, to obtain a data set for training the inverse design network and a data set for the forward prediction network;

[0019] The forward prediction network is constructed, and the obtained data set of the forward prediction network is applied for training and verification; and

[0020] The inverse design network is constructed, connected in series with the trained forward prediction network, and the obtained data set of the inverse design network is applied for training and verification.

[0021] Further preferably, the preprocessing of the original data set further comprises:

[0022] The single resonance position of the spectral response in the original data set is calibrated using an intelligent recognition algorithm and / or manual calibration method, and samples that do not meet the single resonance characteristic are removed;

[0023] The corresponding structure parameters are scaled according to a certain scale, and are linearly scaled to 0 and 1 for normalization processing;

[0024] The structure parameters are split into expected structure parameters and remaining structure parameters, and the splitting is repeated until all cases are obtained;

[0025] The padding value (pad) representing the remaining structure parameters is combined with the expected structure parameters to form an expected structure vector, and the expected structure vector with fixed dimensions is obtained for inputting into a neural network.

[0026] Further preferably, the data set samples of the forward prediction network to be trained are obtained according to the resonance positions and the structure parameters corresponding to the data set samples after preprocessing;

[0027] The data set samples of the inverse design network to be trained are obtained according to the resonance positions and the expected structure vectors corresponding to the data set samples after preprocessing.

[0028] Further preferably, the forward prediction network comprises a fully connected neural network, a convolutional neural network or a residual neural network;

[0029] The data set samples of the forward prediction network are divided into a training set and a validation set according to a proportion, the structure parameters in the training set are input into the forward prediction network to be trained to predict the resonance positions, and the corresponding predicted resonance positions are output;

[0030] The resonance loss is calculated according to the difference between the predicted resonance positions and the resonance positions corresponding to the training set, the forward prediction network corresponding to the resonance loss is back propagated and the network parameters are updated, and the forward prediction network is converged through multiple iterations.

[0031] Further preferably, the training set of the forward prediction network is input into the corresponding forward prediction network for training, and the specific training method is as follows:

[0032] The initial parameters are set, the batch size is set to 16, the total number of learning times is set to 820, the initial learning rate is set to 0.0001, the learning rate decay strategy is set to decrease the learning rate by 20% every 40 epochs, Adam is used as the gradient descent optimizer, mean square error (MSE) is used as the loss function, and ReLU function is used as the activation function of the entire network model.

[0033] Preferably, the inverse design network comprises a fully connected neural network, a convolutional neural network or a residual neural network;

[0034] The inverse design network is connected in series with the trained forward prediction network, and the network parameters of the forward prediction network are fixed;

[0035] The expected resonance positions and the expected structure vectors in the training set of the inverse design network are input into the inverse design network to be trained to predict the structure parameters, and the corresponding predicted structure vectors are output;

[0036] The expected structure parameters in the expected structure vector are replaced with the corresponding parameters in the predicted structure vector to obtain a target structure vector, and the target structure vector is input into the forward prediction network with fixed network parameters to obtain corresponding predicted resonance positions;

[0037] A resonance loss is calculated according to the predicted resonance position and the resonance position corresponding to the training set; the corresponding inverse design network is back propagated and the network parameters are updated according to the resonance loss, and the inverse design network is made to converge through multiple iterations.

[0038] Further preferably, the inverse design network training set is input into the corresponding inverse design network for training, and the specific training method is as follows:

[0039] The initial parameters are set, the batch size is set to 32, the total learning times are set to 820, the initial learning rate is set to 0.0001, the learning rate decay strategy is set to 20% reduction of the learning rate every 40 epochs, Adam is used as the gradient descent optimizer, mean square error (MSE) is used as the loss function, and ReLU function is used as the activation function of the entire network model.

[0040] Compared with the prior art, the beneficial results of the present application are as follows:

[0041] (1) The inverse design paradigm based on the target fingerprint frequency range will greatly improve the design convenience of the multiplexing metasurface device for terahertz fingerprint sensing, and overcome the problem that in the mainstream optical spectrum-based inverse design paradigm, the designer often has difficulty in providing a target spectrum that conforms to the physical reality, so that the neural network cannot inversely predict the required structure.

[0042] (2) The inverse design paradigm of the predicted structure can be customized, which enables researchers to customize part of the structure parameters in the predicted structure, and overcomes the problem that in the traditional nanophotonic inverse design paradigm, the structure parameters designed by the designer are often completely determined by the neural network, so that some of the predicted structure parameters do not conform to the physical reality or the instrument processing limit. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and serve to explain principles of the present application. Other embodiments and many of the intended advantages of the present application will be readily appreciated as the same becomes better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale as some embodiments have been simplified for purposes of this disclosure. Like reference numerals designate corresponding similar parts throughout the various drawings.

[0044] Figure 1 A flowchart of a terahertz fingerprint sensing-oriented multiplexing metasurface intelligent design method according to an embodiment of the present application;

[0045] Figure 2 A flowchart of the multiplexing metasurface intelligent design method according to an embodiment of the present application;

[0046] Figure 3 Structure diagram of a terahertz fingerprint sensing multiplexing super-structure device and an open resonant ring super-structure surface based on quasi-continuum bound states for an embodiment of the present application;

[0047] Figure 4 Flowchart of a training method of a super-structure surface inverse design network for an embodiment of the present application;

[0048] Figure 5 Flowchart of the training method of the super-structure surface inverse design network for an embodiment of the present application;

[0049] Figure 6 Flowchart of the training method for an embodiment of the present application for preprocessing of an original data set;

[0050] Figure 7 Conceptual diagram of the training method of the super-structure surface inverse design network for an embodiment of the present application;

[0051] Figure 8 (a) Comparison diagram of simulated spectral resonance positions corresponding to target structure parameters predicted without fixing structure parameters and expected resonance positions for an embodiment of the present application;

[0052] Figure 8 (b) Comparison diagram of simulated spectral resonance positions corresponding to target structure parameters predicted with fixing structure parameters d1 as 6 μm and d2 as 16 μm and g x as 6 μm for an embodiment of the present application;

[0053] Figure 9 Conceptual diagram of the intelligent design method of the multiplexing super-structure device for an embodiment of the present application;

[0054] Figure 10 (a) Intelligent design result diagram of the multiplexing super-structure device for an embodiment of the present application with fixing structure parameters d1 as 6 μm and d2 as 16 μm and g

[0055] Figure 10 (b) Intelligent design result diagram of the multiplexing super-structure device for an embodiment of the present application with fixing structure parameters d1 as 6 μm and d2 as 16 μm and g x as 6 μm and setting target fingerprint sensing range as 0.9-1.2 THz and target super-structure module number as 10. DETAILED DESCRIPTION

[0056] The application will be described in further detail below with reference to the drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application. In addition, it should be noted that only parts related to the application are shown in the drawings for ease of description.

[0057] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict. The application will be described in further detail below with reference to the drawings and embodiments.

[0058] In a first aspect, embodiments of the application disclose an intelligent design method for a terahertz fingerprint sensing multiplexing superstructure device, as shown in Figure 1 The method comprises the following steps:

[0059] S1, obtaining expected resonance frequencies, covering a target fingerprint frequency range by equidistant sampling to obtain a series of expected resonance frequencies;

[0060] S2, setting expected structure parameters, setting expected structure parameters of a target superstructure device according to actual physical limitations and processing conditions;

[0061] S3, forming an input vector, combining the obtained expected resonance frequencies and the set expected structure parameters to form an input vector of a reverse design network;

[0062] S4, predicting target structure parameters, constructing and training the reverse design network based on deep learning to predict the remaining structure parameters of a series of superstructure surface modules capable of realizing the target fingerprint frequency range.

[0063] Specifically, the application proposes an intelligent design method for a terahertz fingerprint sensing multiplexing superstructure device, which is a deep learning method for predicting target superstructure device remaining structure parameters according to target fingerprint frequency range and expected structure parameters.

[0064] The expected structure parameters are structure parameters that need to be fixed for the target superstructure device, which are used to meet actual physical limitations and processing conditions; the remaining structure parameters are structure parameters that need to be predicted according to a series of superstructure surface modules designed according to the target fingerprint frequency range and the expected structure parameters. Figure 2 The flowchart of the intelligent design method comprises the following steps:

[0065] S101, covering a target fingerprint frequency range by equidistant sampling to obtain a series of expected resonance frequencies;

[0066] S201, setting expected structure parameters of a target superstructure device according to actual physical limitations and processing conditions;

[0067] S301, combine the desired resonance frequency with the desired structure parameters to form an input vector of the inverse design network;

[0068] S401, construct and train the inverse design network based on deep learning to predict the remaining structure parameters of a series of metasurface modules capable of achieving the target fingerprint frequency range.

[0069] Specifically, the multiplexed metadevice is a sensor device for the terahertz band, which can realize multi-functional or multi-channel terahertz fingerprint detection on the same device through multiplexing, and is composed of a series of artificially designed metasurface modules.

[0070] In this embodiment, the metasurface module adopts a metal split ring resonator (SRR) array structure, and the quasi-continuous domain bound state (q-BIC) is used as the resonance mode. The SRR structure is a ring-shaped or square-shaped metal ring with one or more slits (openings), which breaks the structural integrity of the metal ring and introduces specific electromagnetic resonance characteristics. The resonance frequency of the SRR structure can be adjusted by changing the geometric size of the ring, the size of the opening, and the characteristics of the medium. The q-BIC is a state close to the continuous domain bound state, which is transformed from the resonance state of the electromagnetic structure through symmetry breaking or parameter tuning, and has a very high quality factor, which can significantly enhance the local electric field and magnetic field of the electromagnetic structure.

[0071] In one specific embodiment, the multiplexed metadevice adopts an array structure of metasurface modules based on split ring resonators, and realizes terahertz fingerprint sensing by multiplexing the structure parameters of the metasurface modules. The metasurface module adopts an open ring structure based on quasi-continuous domain bound state (q-BIC). Figure 3 The structure of the multiplexed metadevice and the metasurface module is shown in the figure, and the materials of the structure and the substrate are gold (Au, 200 nm thick) and silicon dioxide (SiO2, 200 μm thick), respectively. The dimensions of the remaining supercell structure parameters p and w are 120 μm and 8 μm, respectively. In particular, d1, d2, gx and gy are four adjustable parameters. The q-BIC mode can be excited by adjusting the size of the four adjustable parameters.

[0072] In a second aspect, the embodiments of the present application also disclose a training method of a structure-customizable metasurface inverse design network, which is used in the prediction of target structure parameters in the intelligent design method of the multiplexed metadevice for terahertz fingerprint sensing of the first aspect, i.e. in step S4, as shown in the figure, the training method comprises the following steps: Figure 4

[0073] ​S41, modeling according to a given metasurface module, calculating the corresponding spectral response to obtain an original data set;

[0074] S42, preprocessing the original data set, including resonance position calibration, data cleaning, normalization and structure parameter splitting, to obtain a data set for training the inverse design network and a data set for the forward prediction network;

[0075] S43, constructing the forward prediction network, and applying the obtained data set of the forward prediction network for training and verification; and

[0076] S44, constructing the inverse design network, connecting the trained forward prediction network, and applying the obtained data set of the inverse design network for training and verification.

[0077] Specifically, Figure 5 The specific flowchart of the training method includes the following steps:

[0078] A1, modeling according to a given metasurface module, calculating the corresponding spectral response to obtain an original data set;

[0079] A2, preprocessing the original data set, including resonance position calibration, data cleaning, normalization and structure parameter splitting, further obtaining the corresponding data set of the inverse design network to be trained and the forward prediction network connected to assist the training;

[0080] A3, constructing the forward prediction network, applying the forward prediction network data set to train and verify the forward prediction network;

[0081] A4, constructing the inverse design network, connecting the trained forward prediction network, and applying the inverse design network data set to train and verify the inverse design network.

[0082] In one specific embodiment, the commercial electromagnetic simulation software CST Studio Suite frequency domain solver is used to perform full-wave electromagnetic simulation on the open resonant ring metasurface and calculate the transmittance spectrum in the frequency range of 0.8 THz-1.3 THz. In the simulation, the adjustable parameters d1 are set in the range of 1-12 μm, d2 in the range of 6-18 μm, g xThe value range of d is 6-16 pm, and the value range of gy is 14-34 pm; the periodic Floquet boundary condition is used in the x-y direction, and the open boundary condition is used in the z direction; the electric field direction of the terahertz wave under vertical incidence is excited along the y direction, that is, the TE wave is incident; the dielectric constant of the SiO2 substrate is 3.9. In order to better meet the actual situation, the loss characteristics of the metal material are retained. The grid is set to be less than 1 / 10 of the minimum critical wavelength to obtain a convergent result. Finally, 18018 sample groups are generated by parameter scanning to form an original data set.

[0083] Figure 6 A flowchart for preprocessing the original data set generated in step A2 for electromagnetic simulation software, including steps of:

[0084] A21: Resonance position calibration is performed on the original data set samples, an intelligent recognition algorithm and / or manual calibration method is used to calibrate the single resonance position of the corresponding spectral response in the original data set, and the data set samples in the original data set that do not meet the single resonance characteristic are removed;

[0085] A22: Normalization processing is performed on the original data set samples, and the corresponding structure parameters are scaled according to a certain scale, so that they are linearly scaled to between 0 and 1;

[0086] A23: The structure parameters are split to obtain corresponding expected structure parameters and remaining structure parameters;

[0087] A24: The filling value representing the remaining structure parameters and the expected structure parameters are combined to obtain an expected structure vector.

[0088] The splitting in step A23 is repeatedly performed multiple times to obtain all splitting conditions of the structure parameters.

[0089] In step A24, since the dimensions of the expected structure parameters and the remaining structure parameters obtained by splitting are not fixed, and the input dimension of the structure prediction network is fixed, in order to enable the data to be input into the neural network for training, the filling value (pad) representing the remaining structure parameters and the expected structure parameters are combined to obtain an expected structure vector with fixed dimensions for input into the neural network.

[0090] In this embodiment, the pad value is -1, and Table 1 shows 16 kinds of splitting combinations and corresponding expected structure vectors for a group of four-dimensional structure parameters (d1, d2, g x , g y ).

[0091] Table 1 is all splitting combinations and corresponding expected structure vectors for splitting a group of four-dimensional structure parameters into expected structure parameters and remaining structure parameters in the embodiment.

[0092]

[0093] Table 1

[0094] Further, according to the resonance position and the structure parameter corresponding to the preprocessed data set sample, a data set sample corresponding to the to-be-trained forward prediction network is obtained; and according to the resonance position and the expected structure vector corresponding to the preprocessed data set sample, a data set sample corresponding to the to-be-trained structure prediction network is obtained.

[0095] In the embodiment, 18018 original data set samples are collected through three-dimensional simulation software. Through preprocessing operation, 14201 data set samples of the forward prediction network are obtained (the number of samples is reduced due to data cleaning operation), and 227216 data set samples of the reverse design network are obtained (the number of samples is increased by 16 times due to structure parameter splitting operation).

[0096] The forward prediction network in step A3 can be a neural network such as a fully connected neural network, a convolutional neural network, and a residual neural network.

[0097] In the embodiment, in order to achieve higher degeneracy and efficiency, a fully connected neural network is selected as the forward prediction network, which sequentially includes a plurality of linear layers and activation function layers.

[0098] The data set samples of the forward prediction network are divided according to a certain proportion to obtain a corresponding forward prediction network training set and a verification set.

[0099] In the embodiment, a division ratio of 9:1 is adopted to obtain 12781 training data set samples and 1420 verification data set samples corresponding to the to-be-trained forward prediction network.

[0100] The structure parameters in the forward prediction network training set are input into the to-be-trained forward prediction network to predict the resonance position, and the corresponding predicted resonance position is output; the resonance loss is calculated according to the predicted resonance position and the resonance position corresponding to the training set; the corresponding forward prediction network is subjected to back propagation and network parameter updating according to the resonance loss, and the forward prediction network is made to converge through multiple iterations.

[0101] In the embodiment, the forward prediction network training set is input into the corresponding forward prediction network for training, and the specific training method is as follows:

[0102] The initialization parameters are set as follows: the batch size is 16, the total number of learning epochs is 820, the initial learning rate is 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs, Adam is used as the gradient descent optimizer, mean square error (MSE) is used as the loss function, and the ReLU function is used as the activation function of the entire network model.

[0103] In this embodiment, the forward prediction network training process converges well, and the validation set loss reaches 3.5 x 10 -5 .

[0104] The reverse design network described in step A4 can be a fully connected neural network, a convolutional neural network, a residual neural network, or the like.

[0105] In this embodiment, in order to achieve high degeneracy and high efficiency, a fully connected neural network is selected as the reverse design network, which sequentially includes a plurality of linear layers, activation function layers, and sigmod layers.

[0106] The reverse design network dataset samples are divided according to a certain proportion to obtain the corresponding reverse design network training set and validation set.

[0107] In this embodiment, a 9:1 division ratio is used to obtain 204494 training data sets and 22722 validation data sets for the reverse design network to be trained.

[0108] The reverse design network dataset is divided according to a certain proportion to obtain the corresponding reverse design network training set and validation set.

[0109] The reverse design network is connected in series with the trained forward prediction network, and the network parameters of the forward prediction network are fixed. The expected resonance position and expected structure vector in the reverse design network training set are input into the reverse design network to be trained to predict the structure parameters, and the corresponding predicted structure vector is output. The expected structure parameters in the expected structure vector are replaced with the corresponding parameters in the predicted structure vector to obtain a target structure vector. The target structure vector is input into the forward prediction network with fixed network parameters to obtain the corresponding predicted resonance position. The resonance loss is calculated according to the predicted resonance position and the resonance position corresponding to the training set. The reverse design network is subjected to back propagation and network parameter updating according to the resonance loss, and the reverse design network is converged after multiple iterations.

[0110] Figure 7 The framework diagram of the training method in this embodiment is as follows:

[0111] Firstly, the four-dimensional expected structure vector is spliced with a one-dimensional expected resonance position into a five-dimensional input vector, which is input into the inverse design network to predict the structure parameters, and a four-dimensional predicted structure vector is obtained; secondly, the expected structure parameters in the four-dimensional expected structure vector are replaced with the corresponding parameters in the four-dimensional predicted structure vector to obtain a four-dimensional target structure vector; then, the four-dimensional target structure vector is input into the trained forward prediction network to obtain the corresponding predicted resonance position; finally, the resonance loss is calculated according to the predicted resonance position and the resonance position corresponding to the training set; the corresponding inverse design network is back propagated and the network parameters are updated according to the resonance loss, and the inverse design network is converged through multiple iterations.

[0112] In this embodiment, the inverse design network training set is input into the corresponding inverse design network for training, and the specific training method is as follows:

[0113] The initial parameters are set, the batch size batch_size is 32, the total learning times epoch is 820, the initial learning rate is 0.0001, the learning rate decay strategy is that the learning rate is reduced by 20% every 40 epochs, Adam is used as the gradient descent optimizer, mean square error (MSE) is used as the loss function, and ReLU function is used as the activation function of the entire network model.

[0114] In this embodiment, the inverse design network training process converges well, and the validation set loss reaches 1.02x10 -5 .

[0115] From the above training results, it can be seen that the prediction structure customizable metasurface inverse design network training method proposed in the present application can effectively train the inverse design network.

[0116] In this embodiment, the method for predicting the trained prediction structure customizable metasurface inverse design network is as follows: firstly, normalize the structure parameters to be fixed, and combine them with the padding values representing the structure parameters to be predicted to obtain a four-dimensional expected structure vector; then, the four-dimensional expected structure vector and the one-dimensional expected resonance position are spliced into a five-dimensional input vector, which is input into the trained inverse design network to predict the structure parameters, and a four-dimensional predicted structure vector is obtained; finally, the expected structure parameters in the four-dimensional expected structure vector are replaced with the corresponding parameters in the four-dimensional predicted structure vector to obtain a four-dimensional target structure vector, and the final four-dimensional target structure parameters are obtained through denormalization.

[0117] Figure 8(a) is a comparison of the simulated spectral resonance position and the expected resonance position corresponding to the predicted target structural parameters in this embodiment without fixing the structural parameters. The input vector is the expected structural vector [pad, pad, pad, pad] and the normalized spliced ​​vector of the expected resonance position of 1.0500 THz. The denormalized target structural parameters predicted by the inverse design network are [9.9 μm, 19.9 μm, 3.6 μm, 22.8 μm]. The resonance position corresponding to the spectral response obtained by modeling and simulating the target structural parameters is 1.0480 THz, with an error of 0.0020 THz.

[0118] Figure 8 (b) In this embodiment, the fixed structural parameters d1 are 6 μm, d2 are 16 μm, and g x The image shows a comparison between the simulated spectral resonance positions and the expected resonance positions corresponding to the predicted target structural parameters at a resolution of 6 μm. The input vector consists of the expected structural vector [6 μm, 16 μm, 6 μm, pad] and a normalized spliced ​​vector of the expected resonance position at 0.9500 THz. The denormalized target structural parameters predicted by the inverse design network are [6 μm, 16 μm, 6 μm, 18.9 μm]. The resonance position corresponding to the spectral response obtained from modeling and simulating the target structural parameters is 0.9475 THz, with an error of 0.0025 THz.

[0119] Table 2 shows the results for this embodiment with and without fixed structural parameters, with fixed structural parameter d2 of 16 μm, and with fixed structural parameter d2 of 16 μm. x The diameter is 6 μm and the structural parameters are fixed at d1 = 6 μm, d2 = 16 μm, and g x A table comparing the simulated spectral resonance positions with the expected resonance positions for the predicted target structure parameters in four cases with a diameter of 6 μm. The average error of the resonance positions for the 12 examples in the table is 1.1875 × 10⁻⁶. -3 THz.

[0120]

[0121] Table 2

[0122] As can be seen from the above prediction results, the training method for the metasurface reverse design network with customizable prediction structure proposed in this invention can effectively train the reverse design network, enabling the reverse design network to effectively predict the remaining structural parameters of the metasurface module based on the customer-customized desired structural parameters and desired resonance positions.

[0123] Figure 9 This is a conceptual schematic diagram of the intelligent design method for multiplexed meta-devices described in the embodiments.

[0124] Firstly, the start and end frequencies of the target fingerprint frequency range are set, the target fingerprint frequency range is covered by equidistant sampling to obtain a series of expected resonance frequencies; secondly, the four-dimensional expected structure parameter vector of the target super device is set according to the actual physical restrictions and processing conditions; then, the series of expected resonance frequencies are combined with the expected structure parameter vector in turn to form a series of five-dimensional input vectors of the inverse design network; finally, based on the trained inverse design network, the prediction is carried out in turn to obtain a series of super-structure sensing modules of the terahertz fingerprint sensing super-structure device capable of realizing the target fingerprint frequency range.

[0125] Figure 10 (a) The intelligent design result schematic diagram of the multiplexing super device with the target fingerprint sensing range start frequency of 0.9 THz, the end frequency of 1.1 THz, and the target super module number of 8 in the embodiment without fixed structure parameters. Table 3 is a comparison table of the predicted resonance frequencies, the design module parameters and the simulation resonance frequencies of the corresponding 8 super surface modules:

[0126]

[0127] Table 3

[0128] The predicted resonance frequency is the resonance frequency predicted by the forward prediction network based on the super surface module designed by the inverse design network.

[0129] The average error of the resonance frequencies of the above design result is 1.56x10 -3 The simulation resonance frequencies of the corresponding 8 super surface modules cover the expected target fingerprint frequency range, which shows that the intelligent design method of the multiplexing super device for terahertz fingerprint sensing proposed in the application can accurately design the design result of the multiplexing super surface module according to the target fingerprint frequency range customized by the user.

[0130] Figure 10 (b) The intelligent design result schematic diagram of the multiplexing super device with the target fingerprint sensing range of 0.9-1.2 THz and the target super module number of 10 in the embodiment with the fixed structure parameters d1 of 6 μm and d2 of 16 μm and g x 6 μm; Table 4 is a comparison table of the predicted resonance frequencies, the design module parameters and the simulation resonance frequencies of the corresponding 10 super surface modules:

[0131]

[0132] Table 4

[0133] The average error of the resonance frequencies of the above design result is 1.04x10 -3THz, the simulation resonance frequency of the corresponding 10 metasurface modules covers the target fingerprint frequency range we expect, and the design result meets the customization, which shows that the intelligent design method of the multiplexing metasurface device for terahertz fingerprint sensing proposed in the application can not only accurately design the design result of the multiplexing metasurface module according to the target fingerprint frequency range customized by the user, but also customize the design result according to the physical actual situation and instrument processing limit.

[0134] In summary, the intelligent design method of the multiplexing metasurface device for terahertz fingerprint sensing proposed in the application at least has the following beneficial technical effects:

[0135] 1. The inverse design paradigm based on the target fingerprint frequency range will greatly improve the design convenience of the multiplexing metasurface device for terahertz fingerprint sensing, and overcome the problem that in the mainstream spectrum-based inverse design paradigm, the designer often has difficulty in providing a target spectrum that conforms to the physical reality, so that the neural network cannot inversely predict the required structure;

[0136] 2. The inverse design paradigm with customizable predicted structure enables researchers to customize part of the structure parameters in the predicted structure, and overcomes the problem that in the traditional nanophotonics inverse design paradigm, the structure parameters inversely designed by the designer are often completely determined by the neural network, so that some of the predicted structure parameters do not conform to the physical reality or instrument processing limit.

[0137] The above description is only the preferred embodiment of the application and the explanation of the applied technical principles. Those skilled in the art should understand that the application range involved in the application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features and the technical features disclosed in the application (but not limited to) having similar functions.

Claims

1. A smart design method for multiplexed meta-devices for terahertz fingerprint sensing, characterized in that, The method comprises the following steps: Obtaining expected resonance frequencies, covering the target fingerprint frequency range by equidistant sampling to obtain a series of expected resonance frequencies; Setting expected structure parameters, setting the expected structure parameters of the target metasurface device according to the actual physical limitations and processing conditions; Forming an input vector, combining the obtained expected resonance frequencies and the set expected structure parameters to form an input vector of the inverse design network; Predicting target structure parameters, constructing and training the inverse design network based on deep learning to predict the remaining structure parameters of a series of metasurface modules capable of achieving the target fingerprint frequency range; The remaining structure parameters are the structure parameters that need to be predicted corresponding to a series of metasurface modules designed according to the target fingerprint frequency range and the expected structure parameters; The inverse design network comprises a fully connected neural network, a convolutional neural network or a residual neural network: the inverse design network is connected in series with a trained forward prediction network, and the network parameters of the forward prediction network are fixed; the expected resonance positions and the expected structure vectors in the training set of the inverse design network are input into the inverse design network to be trained to predict the structure parameters, and the corresponding predicted structure vectors are output; the expected structure parameters in the expected structure vector are replaced with the corresponding parameters in the predicted structure vector to obtain a target structure vector; the target structure vector is input into the forward prediction network with fixed network parameters to obtain the corresponding predicted resonance positions; the resonance loss is calculated according to the predicted resonance positions and the resonance positions corresponding to the training set; the corresponding inverse design network is back propagated and the network parameters are updated according to the resonance loss, and the inverse design network is converged after multiple iterations; The forward prediction network comprises a fully connected neural network, a convolutional neural network or a residual neural network: the data set samples of the forward prediction network are divided into a training set and a validation set according to a proportion, the structure parameters in the training set are input into the forward prediction network to be trained to predict the resonance positions, and the corresponding predicted resonance positions are output; the resonance loss is calculated according to the difference between the predicted resonance positions and the resonance positions corresponding to the training set, the forward prediction network is back propagated and the network parameters are updated according to the resonance loss, and the forward prediction network is converged after multiple iterations.

2. The method of claim 1, wherein, The multiplexing metasurface device is composed of a metasurface module designed by artificial design, and the metasurface module adopts a metal open resonant ring SRR array structure with quasi-continuous domain bound state q-BIC as the resonance mode.

3. The method of claim 1, wherein, The structure and substrate material of the multiplexed superstructure device are 200 nm thick gold, Au, and 200 pm thick silicon dioxide, SiO2, respectively, with the remaining structure parameters p and w The dimensions are 120 pm and 8 pm, respectively.

4. A method for training a structure-customizable metasurface inverse design network, comprising: The method is used in the prediction of target structure parameters in the intelligent design method of the multiplexing metasurface device for terahertz fingerprint sensing according to any one of claims 1 to 3, and the training method comprises the following steps: Using electromagnetic simulation software, modeling according to a given metasurface module, calculating the corresponding spectral response to obtain an original data set; Pretreating the original data set, including resonance position calibration, data cleaning, normalization and structure parameter splitting, to obtain a data set for training the inverse design network and a data set for the forward prediction network; The forward prediction network is constructed, and the obtained data set of the forward prediction network is applied for training and verification; And The reverse design network is constructed, the trained forward prediction network is connected in series, and the obtained data set of the reverse design network is applied for training and verification.

5. The training method of claim 4, wherein, The preprocessing of the original data set further includes: The single resonance position of the spectral response in the original data set is labeled by using an intelligent identification algorithm and / or manual calibration method, and samples that do not meet the single resonance characteristic are removed; The corresponding structure parameters are scaled according to a certain scale, so that they are linearly scaled to 0 and 1 for normalization processing; The structure parameters are split into expected structure parameters and remaining structure parameters, and repeated splitting is performed until all cases are obtained; The padding value pad representing the remaining structure parameters is combined with the expected structure parameters to form an expected structure vector, and a fixed-dimension expected structure vector is obtained for inputting into a neural network.

6. The training method of claim 5, wherein, Further comprising: According to the resonance position and the structure parameters corresponding to the preprocessed data set samples, data set samples for training the forward prediction network are obtained; According to the resonance position and the expected structure vector corresponding to the preprocessed data set samples, data set samples for training the reverse design network are obtained.

7. The training method of claim 4, wherein, The forward prediction network training set is input into the corresponding forward prediction network for training, and the specific training method is: Initialize parameters, set the batch size batch_size to 16, the total number of learning epochs epoch to 820, the initial learning rate to 0.0001, the learning rate decay strategy to reduce the learning rate by 20% every 40 epochs, use Adam as the gradient descent optimizer, use mean square error MSE as the loss function, and use ReLU function as the activation function of the entire network model.

8. The training method of claim 4, wherein, The reverse design network training set is input into the corresponding reverse design network for training, and the specific training method is: Initialize parameters, set the batch size batch_size to 32, the total number of learning epochs epoch to 820, the initial learning rate to 0.0001, the learning rate decay strategy to reduce the learning rate by 20% every 40 epochs, use Adam as the gradient descent optimizer, use mean square error MSE as the loss function, and use ReLU function as the activation function of the entire network model.

Images