Intelligent design method of multiplexing super-structure device for terahertz fingerprint sensing

Abstract

The invention provides a terahertz fingerprint sensing-oriented multiplexing meta-structure device intelligent design method, which comprises the steps of obtaining expected resonant frequencies, and covering a target fingerprint frequency range through equal-interval sampling to obtain a series of expected resonant frequencies; setting expected structure parameters, and setting the expected structure parameters of the target super-structure device according to actual physical limits and processing conditions; combining the obtained expected resonant frequency with 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, and predicting other structure parameters of a series of meta-structure surface modules capable of realizing a target fingerprint frequency range. The reverse design of the multiplexing super-structure device is realized through the deep learning neural network, and the device structure meeting the target fingerprint requirement can be quickly and accurately deduced. The design precision and efficiency are remarkably improved, and the capability of responding to sensing demand changes in real time is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of terahertz fingerprint sensing and artificial intelligence algorithms, and specifically relates to an intelligent design method for a multiplexed meta-device for terahertz fingerprint sensing. Background Art

[0002] The multiplexed metadevice for terahertz fingerprint sensing is an advanced device based on metasurface design, which is specially designed to achieve efficient multifunctional sensing and material identification in the terahertz band. Terahertz fingerprint sensing uses electromagnetic waves in the terahertz band (frequency range of 0.1THz to 10THz) to identify the chemical composition and physical properties of substances by detecting their characteristic absorption spectrum, namely "molecular fingerprint spectrum". This technology is widely used in chemical analysis, biomedical imaging, security inspection, communications and materials science because it can provide high-precision detection capabilities under label-free and non-destructive conditions.

[0003] Through multiplexing technology, a meta-device is composed of a series of metasurface modules that can achieve high-quality factor resonance, which can simultaneously sense multiple terahertz "fingerprint" features and improve the detection efficiency and accuracy of the system. The resonant modes used by traditional metasurface modules (such as guided mode resonance, Fano resonance, dipole resonance, etc.) usually have problems such as low quality factor, high radiation and energy loss, and insufficient local field enhancement. In contrast, the q-BIC mode achieves ultra-high quality factor and more significant local field enhancement, with low loss, high efficiency and greater design flexibility. Therefore, it is increasingly favored in modern metasurface design and is suitable for high-performance optical devices, sensors and nonlinear optical applications.

[0004] According to the different occurrence modes, BICs are usually divided into two types, namely symmetry-protected BICs and sporadic BICs. Among them, sporadic BICs appear occasionally because they rely on specific geometric parameters and material properties, making it very difficult to explore the structure of their excitation conditions. This is mainly due to the high sensitivity to parameters, computational complexity, lack of general design guidelines, and challenges in experimental implementation. Traditional design methods usually rely heavily on the solid physical reserves and deep experience accumulation of designers, and the design process requires continuous trial and error and simulation, 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 algorithms can overcome the complexity and high time consumption of the design process to a certain extent, it often faces the problem of convergence difficulties due to the complexity of the metasurface module and the high-dimensional design space.

[0005] In recent years, the application of machine learning algorithms to assist in the design of metasurfaces has received widespread attention. Neural networks can directly predict the required structural parameters based on the desired optical response. Researchers can 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 reverse design paradigm, designers often find it difficult to provide a target spectrum that conforms to physical reality, resulting in the inability of the neural network to reversely predict the desired structure. At the same time, the geometric structure of the designer's reverse design is often completely determined by the neural network, resulting in some predicted structural parameters not conforming to physical reality or instrument processing limits.

[0007] In addition, in applications for terahertz fingerprint sensing, designers usually focus on the frequency range where the fingerprint features are located, which will be related to the accuracy of the device's fingerprint sensing. Therefore, there is an urgent need for a reverse design paradigm based on the target fingerprint frequency range, which will greatly improve the design convenience of multiplexed metadevices for terahertz fingerprint sensing. At the same time, the reverse design network with customizable predicted structures will also greatly improve the problem that the designed structure does not conform to the actual physical situation or the instrument processing limit. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to provide an intelligent design method for a multiplexed metadevice for terahertz fingerprint sensing, which is used to predict the remaining structural parameters of the target metadevice based on the target fingerprint frequency range and the expected structural parameters.

[0009] In a first aspect, the present invention proposes a method for intelligent design of a multiplexed metadevice for terahertz fingerprint sensing, the method comprising the following steps:

[0010] Obtain the expected resonance frequency, cover the target fingerprint frequency range by sampling at equal intervals, and obtain a series of expected resonance frequencies;

[0011] Setting expected structural parameters: According to actual physical limitations and processing conditions, setting expected structural parameters of the target meta-device;

[0012] Forming an input vector, combining the acquired desired resonance frequency with the set desired structural parameters to form an input vector of an inverse design network;

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

[0014] Preferably, the multiplexed metadevice is composed of a metasurface module based on artificial design, and the metasurface module adopts a metal open ring resonator SRR array structure with a quasi-continuous domain bound state q-BIC as the resonance mode.

[0015] Preferably, the materials of the structure and substrate of the multiplexed meta-device are 200 nm thick gold Au and 200 μm thick silicon dioxide SiO2, respectively, and the sizes of the other meta-unit structural parameters p and w are 120 μm and 8 μm, respectively.

[0016] In a second aspect, an embodiment of the present invention provides a training method for predicting a structure-customizable metasurface reverse design network, the method being used in the step of predicting target structural parameters of the intelligent design method for multiplexed metadevices for terahertz fingerprint sensing described in the first aspect, the training method comprising the following steps:

[0017] Using electromagnetic simulation software, modeling is performed based on a given metasurface module, and the corresponding spectral response is calculated to obtain the original data set;

[0018] Preprocessing the original data set, including resonance position calibration, data cleaning, normalization and structural parameter splitting, to obtain a data set for training a reverse design network and a data set for a forward prediction network;

[0019] Constructing the forward prediction network, and applying the obtained data set of the forward prediction network for training and verification; and

[0020] 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 used for training and verification.

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

[0022] Using an intelligent recognition algorithm and / or a manual calibration method to calibrate the single resonance position of the spectral response in the original data set, and remove samples that do not meet the single resonance characteristics;

[0023] The corresponding structural parameters are scaled according to a certain scale, so that they are linearly scaled to between 0 and 1 for normalization;

[0024] Split the structural parameters into expected structural parameters and other structural parameters, and repeat the splitting until all situations are obtained;

[0025] The padding value pad representing the remaining structural parameters is combined with the expected structural parameters into an expected structural vector, and the expected structural vector with a fixed dimension is obtained for input into the neural network.

[0026] Further preferably, according to the resonance positions and structural parameters corresponding to the preprocessed data set samples, a data set sample of the forward prediction network to be trained is obtained;

[0027] According to the resonance positions and expected structure vectors corresponding to the preprocessed data set samples, the data set samples of the inverse design network to be trained are obtained.

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

[0029] Dividing the data set samples of the forward prediction network into a training set and a validation set according to a ratio, inputting the structural parameters in the training set into the predicted resonance position of the forward prediction network to be trained, and outputting the corresponding predicted resonance position;

[0030] The resonance loss is calculated according to the difference between the predicted resonance position and the resonance position corresponding to the training set, and the forward prediction network corresponding to the resonance loss is back-propagated and the network parameters are updated, and the forward prediction network converges after multiple rounds of iterations.

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

[0032] Initialize the parameters, set the batch size batch_size to 16, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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.

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

[0034] Connecting the reverse design network in series with the trained forward prediction network, and fixing the network parameters of the forward prediction network;

[0035] Inputting the expected resonance position and the expected structure vector in the reverse design network training set into the reverse design network to be trained, predicting the structure parameters, and outputting the corresponding predicted structure vector;

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

[0037] 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 converges after multiple rounds of iterations.

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

[0039] Initialize the parameters, set the batch size batch_size to 32, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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.

[0040] Compared with the prior art, the beneficial results of the present invention are:

[0041] (1) The inverse design paradigm based on the target fingerprint frequency range will greatly improve the design convenience of multiplexed metadevices for terahertz fingerprint sensing. It overcomes the problem in the mainstream spectrum-based inverse design paradigm that designers often find it difficult to provide a target spectrum that conforms to physical reality, resulting in the inability of the neural network to reversely predict the required structure.

[0042] (2) The inverse design paradigm of predictive structure customization enables researchers to customize some of the structural parameters in the predicted structure, overcoming the problem in the traditional nanophotonics inverse design paradigm that the structural parameters of the designer's reverse design are often completely determined by the neural network, resulting in some of the predicted structural parameters not conforming to the actual physical situation or the instrument processing limits. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and are used together with the description to explain the principles of the present invention. It will be easy to recognize other embodiments and many expected advantages of the embodiments because they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with each other. The same reference numerals refer to corresponding similar parts.

[0044] Figure 1 A schematic diagram of a flow chart of an intelligent design method for a multiplexed meta-device for terahertz fingerprint sensing according to an embodiment of the present invention;

[0045] Figure 2 A flowchart of the multiplexed meta-device intelligent design method according to an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of the structure of a multiplexed meta-device for terahertz fingerprint sensing and an open resonant ring meta-surface based on a quasi-continuous domain bound state according to an embodiment of the present invention;

[0047] Figure 4 A flowchart of a training method for a reverse design network of a predictive structure-customizable metasurface according to an embodiment of the present invention;

[0048] Figure 5 Flow chart of the network training method for reverse design of the metasurface according to an embodiment of the present invention;

[0049] Figure 6 A flowchart of preprocessing an original data set by the training method of an embodiment of the present invention;

[0050] Figure 7 A conceptual schematic diagram of the network training method for inverse design of a metasurface according to an embodiment of the present invention;

[0051] Figure 8 (a) is a comparison diagram of the simulated spectrum resonance position and the expected resonance position corresponding to the target structural parameters predicted when the structural parameters are not fixed in the embodiment of the present invention;

[0052] Figure 8 (b) is an embodiment of the present invention in which the fixed structural parameters d1 is 6 μm, d2 is 16 μm, and g x Comparison of the simulated spectrum resonance position and the expected resonance position corresponding to the target structure parameters predicted for 6μm;

[0053] Fig. 9 A conceptual schematic diagram of a multiplexed meta-device intelligent design method according to an embodiment of the present invention;

[0054] Fig.10 (a) is a schematic diagram of the intelligent design results of a multiplexed meta-device in which the starting frequency of the target fingerprint sensing range is set to 0.9 THz, the cut-off frequency is set to 1.1 THz, and the number of target meta-modules is set to 8 without fixing the structural parameters in an embodiment of the present invention;

[0055] Fig.10 (b) is an embodiment of the present invention in which the fixed structural parameters d1 is 6 μm, d2 is 16 μm, and g x Schematic diagram of the intelligent design results of a multiplexed meta-device with a target fingerprint sensing range of 0.9-1.2THz and a target number of meta-modules of 10 when the sensor size is 6μm. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0057] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0058] In a first aspect, an embodiment of the present invention discloses an intelligent design method for a multiplexed meta-device for terahertz fingerprint sensing, such as Figure 1 As shown, the method comprises the following steps:

[0059] S1. Obtain the expected resonance frequency, cover the target fingerprint frequency range by sampling at equal intervals, and obtain a series of expected resonance frequencies;

[0060] S2. Setting expected structural parameters: According to actual physical limitations and processing conditions, setting expected structural parameters of the target meta-device;

[0061] S3, forming an input vector, combining the acquired desired resonance frequency with the set desired structural parameters to form an input vector of the inverse design network;

[0062] S4. Predict the target structural parameters, build and train the inverse design network based on deep learning, and predict the remaining structural parameters of a series of metasurface modules that can achieve the target fingerprint frequency range.

[0063] Specifically, the present invention proposes an intelligent design method for multiplexed metadevices for terahertz fingerprint sensing, which is a deep learning method used to predict the remaining structural parameters of the target metadevice based on the target fingerprint frequency range and expected structural parameters.

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

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

[0066] S201, setting expected structural parameters of the target meta-device according to actual physical limitations and processing conditions;

[0067] S301, combining the desired resonance frequency and the desired structural parameters to form an input vector of the inverse design network;

[0068] S401, construct and train a reverse design network based on deep learning to predict the remaining structural parameters of a series of metasurface modules that can achieve the target fingerprint frequency range.

[0069] Specifically, the multiplexed metadevice is a sensor device used in the terahertz band, which can realize multi-functional or multi-channel terahertz fingerprint detection on the same device through multiplexing. It consists of a series of metasurface modules based on artificial design.

[0070] In this embodiment, the metasurface module adopts a metal open resonant ring (SRR) array structure with a quasi-continuous domain bound state (q-BIC) as the resonant mode. The SRR structure is an annular or square metal ring with one or more slits (openings) that destroy the structural integrity of the metal ring, thereby introducing specific electromagnetic resonance characteristics. The resonant frequency of the SRR structure can be adjusted by changing the geometric dimensions 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. The resonant state of the electromagnetic structure is transformed from the ideal continuous domain bound state through symmetry breaking or parameter tuning. It has a very high quality factor and can significantly enhance the local electric and magnetic fields of the electromagnetic structure.

[0071] In a specific embodiment, the multiplexed meta-device adopts a metasurface module array structure based on an open resonant ring, and terahertz fingerprint sensing is realized by multiplexing the structural parameters of the metasurface module. The metasurface module adopts an open resonant ring structure based on a quasi-continuum bound state (q-BIC). Figure 3 The schematic diagram of the structure of the multiplexed metadevice and metasurface module is shown in Figure 1. The materials of the structure and substrate are gold (Au, 200nm thick) and silicon dioxide (SiO2, 200μm thick), respectively. The sizes of the remaining metaunit structural parameters p and w are 120μm and 8μm, 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, an embodiment of the present invention further discloses a training method for a reverse design network of a predictive structure-customizable metasurface, the method being used in the step of predicting target structural parameters of the intelligent design method for multiplexed metadevices for terahertz fingerprint sensing described in the first aspect, i.e., in step S4, as Figure 4 As shown, the training method includes the following steps:

[0073] S41. Using electromagnetic simulation software, modeling is performed according to a given metasurface module, and the corresponding spectral response is calculated to obtain an original data set;

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

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

[0076] S44, constructing the reverse design network, connecting the trained forward prediction network in series, and using the obtained data set of the reverse design network for training and verification.

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

[0078] A1, using electromagnetic simulation software, modeling is performed according to a given metasurface module, the corresponding spectral response is calculated, and the original data set is obtained;

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

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

[0081] A4, constructing the reverse design network, connecting the trained forward prediction network in series, and applying the reverse design network dataset to train and verify the reverse design network.

[0082] In a specific embodiment, a full-wave electromagnetic simulation is performed on the split ring resonator metasurface using the commercial electromagnetic simulation software CST Studio Suite frequency domain solver, and the transmittance spectrum in the frequency range of 0.8THz-1.3THz is calculated. In the simulation, the adjustable parameter d1 is set to a value range of 1-12μm, the value range of d2 is 6-18μm, and g is set to xThe value range of is 6-16μm, and the value range of gy is 14-34μm; the periodic Floquet boundary condition is used in the xy 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 conform to 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 convergence results. Finally, 18018 samples are generated by parameter scanning to form the original data set.

[0083] Figure 6 A schematic diagram of a process for preprocessing the original data set generated by the electromagnetic simulation software in step A2, comprising the steps of:

[0084] A21: calibrate the resonance position of the original data set samples, use an intelligent recognition algorithm and / or a manual calibration method to calibrate the single resonance position of the corresponding spectral response in the original data set, and remove the data set samples in the original data set that do not meet the single resonance characteristics;

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

[0086] A23: Splitting the structural parameters to obtain corresponding expected structural parameters and other structural parameters;

[0087] A24: Combine the fill values ​​representing the remaining structural parameters and the expected structural parameters to obtain the expected structural vector.

[0088] In step A23, the splitting is repeated multiple times to obtain all splitting conditions of the structural parameters.

[0089] In step A24, since the dimensions of the expected structural parameters and the remaining structural 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 padding value (pad) representing the remaining structural parameters and the expected structural parameters are combined to obtain an expected structure vector with a fixed dimension for input into the neural network.

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

[0091] Table 1 shows all the split combinations of a set of four-dimensional structural parameters into expected structural parameters and other structural parameters and the corresponding expected structural vectors in the embodiment.

[0092]

[0093] Table 1

[0094] Furthermore, according to the resonance positions and structural parameters corresponding to the preprocessed data set samples, the data set samples corresponding to the forward prediction network to be trained are obtained; according to the resonance positions and expected structural vectors corresponding to the preprocessed data set samples, the data set samples corresponding to the structure prediction network to be trained are obtained.

[0095] In this embodiment, a total of 18018 original data set samples were collected by 3D simulation software. Through preprocessing operations, 14201 forward prediction network data sets were obtained (data cleaning operations resulted in a reduction in the number of samples) and 227216 reverse design network data sets were obtained (structure parameter splitting operations resulted in a 16-fold increase in the number of samples).

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

[0097] In this embodiment, for higher degeneracy and efficiency, a fully connected neural network is selected as the forward prediction network, which includes several linear layers and activation function layers in sequence.

[0098] The forward prediction network data set samples are divided according to a certain ratio to obtain corresponding forward prediction network training sets and validation sets.

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

[0100] The structural parameters in the forward prediction network training set are input into the forward prediction network to be trained, the resonance position is predicted, 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 back-propagated and the network parameters are updated according to the resonance loss, and the forward prediction network converges after multiple rounds of iterations.

[0101] In this embodiment, the forward prediction network training set is input into the corresponding forward prediction network for training. The specific training method is:

[0102] Initialize the parameters, set the batch size batch_size to 16, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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.

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

[0104] In step A4, the reverse design network can be a neural network such as a fully connected neural network, a convolutional neural network, a residual neural network, etc.

[0105] In this embodiment, for higher degeneracy and efficiency, a fully connected neural network is selected as the inverse design network, which includes a number of linear layers, activation function layers and sigmoid layers in sequence.

[0106] For the reverse design network data set samples, the reverse design network data set samples are divided according to a certain ratio to obtain corresponding reverse design network training sets and verification sets;

[0107] In this embodiment, a division ratio of 9:1 is adopted to obtain 204,494 training data sets and 22,722 verification data sets corresponding to the inverse design network to be trained.

[0108] The reverse design network data set is divided according to a certain ratio to obtain a corresponding reverse design network training set and verification 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 the expected structure vector in the reverse design network training set are input into the reverse design network to be trained, the structure parameters are predicted, 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 the 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 corresponding reverse design network is back-propagated and the network parameters are updated according to the resonance loss, and the reverse design network converges after multiple rounds of iterations.

[0110] Figure 7 This is a schematic diagram of the framework of the training method described in this embodiment:

[0111] Firstly, a four-dimensional expected structure vector and a one-dimensional expected resonance position are concatenated into a five-dimensional input vector, which is input into a reverse design network to predict the structure parameters, thereby obtaining a four-dimensional predicted structure vector. 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. According to the resonance loss, the corresponding reverse design network is back-propagated and the network parameters are updated, and the reverse design network converges after multiple rounds of iterations.

[0112] In this embodiment, the reverse design network training set is input into the corresponding reverse design network for training. The specific training method is:

[0113] Initialize the parameters, set the batch size batch_size to 32, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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.

[0114] In this embodiment, the reverse design network training process converges well, and the validation set loss reaches 1.02×10 -5 .

[0115] It can be seen from the above training results that the reverse design network training method for predicting customizable metasurface structures proposed in the present invention can effectively train the reverse design network.

[0116] In this embodiment, the method for predicting the trained predictive structure customizable metasurface reverse design network is as follows: first, the structural parameters that need to be fixed are normalized, and combined with the filling values ​​representing the structural parameters that need 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 reverse design network to predict the structural parameters to obtain a four-dimensional predicted structure vector; finally, the expected structural 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 by denormalization.

[0117] Figure 8(a) is a comparison diagram of the simulated spectral resonance position and the expected resonance position corresponding to the target structural parameters predicted in this embodiment without fixing the structural parameters. The input vector is the normalized concatenation vector of the expected structural vector [pad, pad, pad, pad] and the expected resonance position 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 is 6 μm, d2 is 16 μm, and g x The comparison chart of the simulated spectral resonance position and the expected resonance position corresponding to the predicted target structural parameters when the target structure is 6μm, the input vector is the expected structure vector [6μm, 16μm, 6μm, pad] and the normalized concatenation vector of the expected resonance position 0.9500THz, and the denormalized target structural parameters predicted by the inverse design network are [6μm, 16μm, 6μm, 18.9μm]. The resonance position corresponding to the spectral response obtained by modeling and simulating the target structural parameters is 0.9475THz, with an error of 0.0025THz.

[0119] Table 2 shows the non-fixed structural parameters, fixed structural parameters d2 is 16μm, fixed structural parameters d2 is 16μm, g x 6μm and fixed structural parameters d1 is 6μm, d2 is 16μm, g x The comparison table of the simulated spectral resonance positions and expected resonance positions corresponding to the target structural parameters predicted in four cases of 6μm. The average error of the corresponding resonance positions of the 12 examples in the table is 1.1875×10 -3 THz.

[0120]

[0121] Table 2

[0122] It can be seen from the above prediction results that the reverse design network training method for predicting customizable metasurface structures proposed in the present invention can effectively train the reverse design network, so that the reverse design network can effectively predict the remaining structural parameters of the metasurface module based on the customer-customized expected structural parameters and expected resonance positions.

[0123] Fig. 9 Schematic diagram of the concept of the intelligent design method of the multiplexed meta-device described in the embodiment.

[0124] Firstly, the starting and ending frequencies of the target fingerprint frequency range are set, and the target fingerprint frequency range is covered by equally spaced sampling to obtain a series of expected resonance frequencies; secondly, according to the actual physical limitations and processing conditions, the four-dimensional expected structural parameter vector of the target meta-device is set; then, the series of expected resonance frequencies are combined with the expected structural parameter vectors in turn to form a series of five-dimensional input vectors of the inverse design network; finally, based on the trained inverse design network, predictions are performed in turn to obtain a series of meta-sensing modules of the terahertz fingerprint sensing meta-device that can achieve the target fingerprint frequency range.

[0125] Fig.10 (a) is a schematic diagram of the intelligent design results of a multiplexed metasurface device with a target fingerprint sensing range start frequency of 0.9 THz, a cutoff frequency of 1.1 THz, and a target number of metasurface modules of 8 when the structural parameters are not fixed in the embodiment. Table 3 is a comparison table of the predicted resonance frequency, design module parameters, and simulated resonance frequency corresponding to 8 metasurface modules:

[0126]

[0127] Table 3

[0128] The predicted resonance frequency is the resonance frequency predicted by the forward prediction network corresponding to the metasurface module designed based on the reverse design network.

[0129] The average error of the resonant frequency of the above design results is 1.56×10 -3 THz, the simulated resonance frequencies of the corresponding 8 metasurface modules cover our expected target fingerprint frequency range, which shows that the intelligent design method of multiplexed metasurface devices for terahertz fingerprint sensing proposed in the present invention can accurately design the design results of multiplexed metasurface modules according to the user-customized target fingerprint frequency range.

[0130] Fig.10 (b) In the embodiment, the fixed structural parameters d1 is 6 μm, d2 is 16 μm, and g x The schematic diagram of the intelligent design results of the multiplexed meta-device is shown in the figure below, with the target fingerprint sensing range of 0.9-1.2THz and the target number of meta-modules being 10 when the surface area is 6μm; Table 4 is a comparison table of the predicted resonance frequency, design module parameters and simulated resonance frequency corresponding to 10 meta-surface modules:

[0131]

[0132] Table 4

[0133] The average error of the resonant frequency of the above design results is 1.04×10 -3THz, the simulated resonance frequencies of the corresponding 10 metasurface modules cover our expected target fingerprint frequency range, and the design results meet the customization. This shows that the intelligent design method of multiplexed metasurface devices for terahertz fingerprint sensing proposed in the present invention can not only accurately design the design results of the multiplexed metasurface modules according to the user-customized target fingerprint frequency range, but also customize the design results according to the actual physical conditions and instrument processing limitations.

[0134] In summary, the intelligent design method of multiplexed metadevice for terahertz fingerprint sensing proposed in the present invention includes at least the following beneficial technical effects:

[0135] 1. The reverse design paradigm based on the target fingerprint frequency range will greatly improve the design convenience of multiplexed metadevices for terahertz fingerprint sensing, overcoming the problem in the mainstream spectrum-based reverse design paradigm that designers often find it difficult to provide a target spectrum that conforms to physical reality, resulting in the inability of the neural network to reversely predict the required structure;

[0136] 2. The inverse design paradigm of customizable predicted structures enables researchers to customize some structural parameters in the predicted structure, overcoming the problem in the traditional nanophotonics reverse design paradigm that the structural parameters of the designer's reverse design are often completely determined by neural networks, resulting in some predicted structural parameters not conforming to the actual physical situation or the instrument processing limits.

[0137] The above description is only a preferred embodiment of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present invention is not limited to the technical solution formed by a specific combination of the above technical features, but 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 above features are replaced with the technical features with similar functions disclosed in the present invention (but not limited to) to form a technical solution.

Claims

1. An intelligent design method for multiplexed metadevices for terahertz fingerprint sensing, characterized in that: The method comprises the following steps: Obtain the expected resonance frequency, cover the target fingerprint frequency range by sampling at equal intervals, and obtain a series of expected resonance frequencies; Setting expected structural parameters: According to actual physical limitations and processing conditions, setting expected structural parameters of the target meta-device; Forming an input vector, combining the acquired desired resonance frequency with the set desired structural parameters to form an input vector of an inverse design network; Predict the target structural parameters, build and train the inverse design network based on deep learning, and predict the remaining structural parameters of a series of metasurface modules that can achieve the target fingerprint frequency range.

2. The intelligent design method of multiplexed metadevice for terahertz fingerprint sensing according to claim 1, characterized in that: The multiplexed meta-device is composed of a meta-surface module based on artificial design. The meta-surface module adopts a metal open resonant ring SRR array structure and takes a quasi-continuous domain bound state q-BIC as a resonance mode.

3. The intelligent design method of multiplexed metadevice for terahertz fingerprint sensing according to claim 1, characterized in that: The materials of the structure and substrate of the multiplexed meta-device are 200nm thick gold Au and 200μm thick silicon dioxide SiO2, respectively, and the sizes of the other meta-unit structural parameters p and w are 120μm and 8μm, respectively.

4. A training method for a reverse design network of a predictive structure-customizable metasurface, characterized in that: The method is used in the step of predicting target structural parameters of the intelligent design method for multiplexed metadevices 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 is performed based on a given metasurface module, and the corresponding spectral response is calculated to obtain the original data set; Preprocessing the original data set, including resonance position calibration, data cleaning, normalization and structural parameter splitting, to obtain a data set for training a reverse design network and a data set for a forward prediction network; Constructing the forward prediction network, and applying the obtained data set of the forward prediction network for training and verification; as well as 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 used for training and verification.

5. The training method of the reverse design network according to claim 4, characterized in that: The preprocessing of the original data set further comprises: Using an intelligent recognition algorithm and / or a manual calibration method to calibrate the single resonance position of the spectral response in the original data set, and remove samples that do not meet the single resonance characteristics; The corresponding structural parameters are scaled according to a certain scale, so that they are linearly scaled to between 0 and 1 for normalization; Split the structural parameters into expected structural parameters and other structural parameters, and repeat the splitting until all situations are obtained; The padding value pad representing the remaining structural parameters is combined with the expected structural parameters into an expected structural vector, and the expected structural vector with a fixed dimension is obtained for input into the neural network.

6. The training method of the reverse design network according to claim 5, characterized in that: Also includes: Obtaining a data set sample of the forward prediction network to be trained according to the resonance position and structural parameters corresponding to the preprocessed data set sample; According to the resonance positions and expected structure vectors corresponding to the preprocessed data set samples, the data set samples of the inverse design network to be trained are obtained.

7. The training method of the reverse design network according to claim 4, characterized in that: The forward prediction network includes a fully connected neural network, a convolutional neural network or a residual neural network; Dividing the data set samples of the forward prediction network into a training set and a validation set according to a ratio, inputting the structural parameters in the training set into the predicted resonance position of the forward prediction network to be trained, and outputting the corresponding predicted resonance position; The resonance loss is calculated according to the difference between the predicted resonance position and the resonance position corresponding to the training set, and the forward prediction network corresponding to the resonance loss is back-propagated and the network parameters are updated, and the forward prediction network converges after multiple rounds of iterations.

8. The training method of the reverse design network according to claim 7, characterized in that: The forward prediction network training set is input into the corresponding forward prediction network for training. The specific training method is: Initialize the parameters, set the batch size batch_size to 16, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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.

9. The training method of the reverse design network according to claim 4, characterized in that: The reverse design network includes a fully connected neural network, a convolutional neural network or a residual neural network; Connecting the reverse design network in series with the trained forward prediction network, and fixing the network parameters of the forward prediction network; Inputting the expected resonance position and the expected structure vector in the reverse design network training set into the reverse design network to be trained, predicting the structure parameters, and outputting the corresponding predicted structure vector; The expected structural parameters in the expected structural vector are replaced with the corresponding parameters in the predicted structural vector to obtain a target structural vector; the target structural vector is input into the forward prediction network with fixed network parameters to obtain a 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 corresponding inverse design network is back-propagated and the network parameters are updated according to the resonance loss, and the inverse design network converges after multiple rounds of iterations.

10. The training method of the reverse design network according to claim 9, characterized in that: The reverse design network training set is input into the corresponding reverse design network for training. The specific training method is: Initialize the parameters, set the batch size batch_size to 32, the total number of learning epochs to 820, the initial learning rate to 0.0001, the learning rate decay strategy is to reduce the learning rate by 20% every 40 epochs of training, 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