Construction method and application of design model of metamaterial sensing chip based on deep learning

Abstract

The invention provides a construction method and application of a design model of a metamaterial sensing chip based on deep learning, and the method comprises the steps: inputting a transmission curve collection point set into an input layer of the design model, and carrying out the normalization processing, thereby obtaining a one-dimensional sequence; the one-dimensional sequence is respectively input into a plurality of convolution branches of the parallel convolution layer for convolution processing to obtain convolution features, the plurality of convolution features are input into the linear layer for splicing to obtain a spliced sequence, the spliced sequence is input into the output layer for processing, a structure parameter group is output, and the structure parameter group can be used for design of a metamaterial sensing chip.

Description

Technical Field

[0001] The present invention relates to the field of sensor chip design, and in particular to a method for constructing a design model of a metamaterial sensor chip based on deep learning and its application. Background Art

[0002] The rapid development and breakthroughs of deep learning and artificial neural networks have led to the continuous expansion of their application areas. In addition to the traditional field of computer science, they are also used in physics, chemistry, materials science, etc.

[0003] Metamaterial structure design often requires a fairly complex design process and design cycle. The traditional calculation method starts with certain initial conditions and boundary conditions, and uses computational electromagnetic simulation to solve the discrete Maxwell equations in space and time. By setting sufficient grids and iterative steps, the optical properties of a given structure can be accurately calculated. It is often necessary to fine-tune the geometry and iteratively perform simulations to gradually approach the target response. This process relies heavily on past experience with design templates, and due to limitations in simulation capabilities and time, only limited design parameters are adjusted when searching for the best structure.

[0004] Therefore, deep learning methods can be used to discover the implicit relationship between the physical structure parameters of metamaterials and their electromagnetic responses, thereby helping designers to quickly design metamaterial sensors on demand. However, the speed of designing electromagnetically induced transparent metamaterials using deep learning models is still not fast enough, and the accuracy is still not high enough. Summary of the invention

[0005] The purpose of the present invention is to provide a method for constructing a design model of a metamaterial sensor chip based on deep learning and its application, so as to construct a design model of a metamaterial sensor chip that can be reversely designed to simultaneously generate three transparent window resonance peaks, so as to improve the design efficiency of metamaterial sensors.

[0006] To achieve the purpose, the present technical solution provides a method for constructing a design model of a metamaterial sensor chip based on deep learning, comprising the following steps: A set of transmission curve collection points corresponding to the metamaterial sensor chip with different structural parameter groups is obtained as training data, wherein the upper structure of the metamaterial sensor chip includes a π-type resonance unit and a rectangular resonance unit arranged at intervals, and the π-type opening of the π-type resonance unit is arranged opposite to the rectangular resonance unit, wherein the structural parameter group includes the horizontal side length, the vertical bar length, the width of the π-type opening, the long side length and the short side length of the rectangular resonance unit; Input the training data into the design framework for training until the training conditions are met, wherein the design framework includes an input layer, a parallel convolution layer, a linear layer, and an output layer connected in sequence, and the parallel convolution layer includes multiple convolution branches, each convolution branch includes convolution layers of different convolution dimensions, and each convolution layer includes a ReLU activation function and a Pool operation; The transmission curve collection point set is input into the input layer for normalization to obtain a one-dimensional sequence. The one-dimensional sequence is input into multiple convolution branches of the parallel convolution layer for convolution to obtain convolution features. Multiple convolution features are input into the linear layer for splicing to obtain a spliced ​​sequence. The spliced ​​sequence is input into the output layer for processing and outputs a structural parameter group.

[0007] In the second aspect, the present scheme provides a design method for a metamaterial sensor chip, inputs the transmission curve collection points containing three transparent window resonance peaks into the design model constructed by the method for constructing a design model of a metamaterial sensor chip based on deep learning, and outputs a structural parameter group corresponding to the metamaterial sensor chip.

[0008] Compared with the prior art, this technical solution has the following characteristics and beneficial effects: This technical solution provides a method for constructing and applying a design model for a metamaterial sensor chip based on deep learning. The design model can adapt to the complexity of metamaterial device design and has good generalization ability. The design of metamaterial sensors involves the interaction of many physical parameters and electromagnetic properties, which is a highly complex task. The parallel convolution layer of the 1D architecture used in the design model can accurately simulate the complex input-output relationship through parallel learning of different data features, that is, the mapping from the input spectral sequence to the output structural parameter sequence. In addition, the nonlinear mapping ability of the design model originates from the activation function after each convolution layer, and can fit complex nonlinear functional relationships. In the field of metamaterial sensor design, many physical phenomena are nonlinear. This characteristic allows the design model to accurately capture hidden physical laws. Thanks to the diversified feature extraction and fusion of the parallel convolution branches, the design model can cope with various changes in sensor chip design, such as different materials and structures, so that the design model can be well used for the design of metamaterial sensor chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Schematic diagram of a metamaterial sensor chip that can generate three transparent window resonance peaks.

[0010] Figure 2 It is a schematic diagram of a sensing unit of a metamaterial sensing chip that can generate three transparent window resonance peaks.

[0011] Figure 3It is a transmission curve diagram of various parts of the metamaterial sensor chip that can produce three transparent window resonance peaks.

[0012] Figure 4 This is the electric field distribution diagram of the metamaterial sensor chip that can produce three transparent window resonance peaks.

[0013] Figure 5 It is a schematic diagram of the network framework of the design model.

[0014] Figure 6 is the loss of the test set and training set during the construction process of the designed model.

[0015] Figure 7 It is the transmission curve of the metamaterial sensor chip designed by the design model and the target transmission curve.

[0016] Figure 8 This is the result graph of the metamaterial sensor chip's response to the test objects with different refractive indices.

[0017] Fig. 9 The fitting result graph is obtained by linearly fitting the refractive index change of each transparent window relative to the surrounding test object.

[0018] In the figure: 10-π-type resonant unit, 11-horizontal bar, 12-first vertical bar, 13-second vertical bar, 100-π-type opening, 20-rectangular resonant closed loop. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0020] Those skilled in the art should understand that, in the disclosure of the present invention, the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the above terms should not be understood as limiting the present invention.

[0021] Embodiment 1 This scheme provides a method for constructing a design model of a metamaterial sensor chip based on deep learning. The design model can be used to design the structural parameters of the metamaterial sensor chip, and the metamaterial sensor chip can generate three transparent window resonance peaks to achieve simultaneous detection of multiple antibiotics.

[0022] Specifically, the method for constructing a design model of a metamaterial sensor chip based on deep learning provided in this solution includes the following steps: A set of transmission curve collection points corresponding to the metamaterial sensor chip with different structural parameter groups is obtained as training data, wherein the upper structure of the metamaterial sensor chip includes a π-type resonance unit and a rectangular resonance unit arranged at intervals, and the π-type opening of the π-type resonance unit is arranged opposite to the rectangular resonance unit, wherein the structural parameter group includes the horizontal side length, the vertical bar length, the width of the π-type opening, the long side length and the short side length of the rectangular resonance unit; Input the training data into the design framework for training until the training conditions are met, wherein the design framework includes an input layer, a parallel convolution layer, a linear layer, and an output layer connected in sequence, and the parallel convolution layer includes multiple convolution branches, each convolution branch includes convolution layers of different convolution dimensions, and each convolution layer includes a ReLU activation function and a Pool operation; The transmission curve collection point set is input into the input layer for normalization to obtain a one-dimensional sequence. The one-dimensional sequence is input into multiple convolution branches of the parallel convolution layer for convolution to obtain convolution features. Multiple convolution features are input into the linear layer for splicing to obtain a spliced ​​sequence. The spliced ​​sequence is input into the output layer for processing and outputs a structural parameter group.

[0023] The design model of the metamaterial sensor chip based on deep learning provided by this solution selects 1D architecture as the basic architecture, which is particularly suitable for processing one-dimensional sequence data, so as to capture local key features through convolution operations and mine nonlinear mapping relationships in one-dimensional sequence data. At the same time, this solution introduces parallel convolution layers based on the 1D architecture so that the design model can focus on multi-size features, thereby fully covering all kinds of key information in the design, making the design model have a more thorough understanding of complex mapping relationships.

[0024] Figure 5 This is the overall architecture diagram of the design framework provided by this solution, such as Figure 5 As shown, the transmission curve collection points are multiple spectral data points collected from the transmission curve, and the transmission curve collection points corresponding to the metamaterial sensor chip corresponding to each structural parameter group are input into the input layer for normalization processing to obtain a one-dimensional sequence. In the embodiment of this scheme, the metamaterial sensor chip corresponding to each structural parameter group collects 201 transmission curve collection points, and the input layer obtains a one-dimensional sequence after normalizing the transmission curve collection points.

[0025] In some embodiments, the parallel convolution layer includes three convolution branches of 1D architecture, each convolution branch includes three convolution layers of different convolution dimensions, and each layer includes a ReLU activation function and a Pool operation.

[0026] Specifically, the convolution dimensions of the three convolutional layers in each convolutional branch are 256, 128, and 32, respectively.

[0027] Specifically, the parallel convolution layer includes a first convolution branch, a second convolution branch and a third convolution branch in parallel, wherein the convolution kernels of the three convolution layers in the first convolution branch are 7, 5, and 3, respectively, the convolution kernels of the three convolution layers in the second convolution branch are 5, 5, and 5, respectively, and the convolution kernels of the three convolution layers in the third convolution branch are 3, 5, and 7, respectively.

[0028] It should be noted that the design model provided by this solution has efficient feature extraction capabilities. For one-dimensional sequences, the convolution kernel slides along the sequence dimension and can quickly focus on key local patterns. In the relevant spectral data of metamaterial sensor chip design, these local features may correspond to specific combinations of structural parameters, phased changes in electromagnetic response, etc. The introduction of parallel convolution layers in the design model further enhances this capability. Convolution branches of different branches are configured with convolution kernels of different sizes, each capturing local features of different scales, covering from small-scale subtle changes in the input data to trend features with larger spans, making the feature information extracted by the network more comprehensive and rich.

[0029] At the same time, the network architecture of this design model can reduce the number of parameters and computing costs. The convolution branches under the parallel convolution layer operate relatively independently, and then the convolution results are integrated through the linear layer to avoid redundant calculations. Different convolution branches focus on processing different types of features, and the allocation of computing resources is more reasonable, which improves computing efficiency without sacrificing performance. The convolution branches of the 1D architecture share receptive fields and weights, which greatly reduces the number of parameters that need to be trained.

[0030] In some embodiments, the linear layer includes a splicing linear layer and a fully connected linear layer. The convolution features output by the three convolution branches are all entered into the splicing linear layer for splicing to obtain splicing features. The splicing features are then mapped to dimension reduction through fully connected linear layers of different dimensions to obtain a spliced ​​sequence after dimensionality reduction.

[0031] In a specific embodiment, the convolution features of three convolution branches are concatenated in a linear layer to obtain a concatenated feature with a length of 96, and the concatenated feature is then mapped to a dimension reduction map in a fully connected linear layer with dimensions of 64 and 16 to obtain a concatenated sequence, and the concatenated sequence is input into the output layer to obtain a structural parameter group consisting of the horizontal side length of the π-type resonance unit, the vertical bar length, the width of the π-type opening, the length of the long side of the rectangular resonance unit, and the length of the short side.

[0032] About the collection of training data: This solution uses CST Studio Suite 2020 electromagnetic simulation software to model and simulate the metamaterial sensor chip to obtain the data set required for the experiment. In the process of obtaining the data set, the frequency domain solver in the software is used. Based on the proposed metamaterial sensor chip, the five structural parameters of the horizontal length of the π-type resonant unit, the vertical length, the width of the π-type opening, the length of the long side of the rectangular resonant unit, and the length of the short side are used as variables for simulation to complete the collection of transmission curve collection points. In the process of data collection, for the transmission curve obtained by device simulation, 201 points are evenly selected as transmission curve collection points to represent the transmission curve.

[0033] In some specific embodiments, the total amount of data collected is 3125, 80% of which is used as a training set and 20% as a test set. A set of transmission curve collection points in the data set includes 201 sampling points of the transmission curve and five structural parameters: the horizontal side length, vertical bar length, width of the π-type opening, length of the long side of the rectangular resonance unit, and length of the short side.

[0034] About the training process: This solution trains the model in batches when the training set is input into the model, with 16 data in each batch. The training and validation losses of the designed model are as follows Figure 6 As shown, Figure 6 (a) in the figure corresponds to the training set loss. It can be seen that the training set loss converges to 0.018 after 400 cycles. Figure 6 (b) in the figure corresponds to the test set loss, which converges to 0.025 after 250 cycles. The model has reached a stable and low loss at 200 cycles of training. While having a relatively fast training speed, the MSE of the test set is as low as 0.029, indicating that the design model has a relatively good design capability for metamaterial sensor chips.

[0035] About the verification of the design model: After the MSE of the training set converges to 0.018, four transmission curves are randomly selected from the test set and input into the improved design model, and four corresponding structural parameter groups [the horizontal length of the π-type resonant unit, the vertical length, the width of the π-type opening, the length of the long side of the rectangular resonant unit, and the length of the short side] can be immediately designed. The four groups of structural parameters designed are: [279.165, 77.872, 131.098, 147.113, 84.065], [273.851, 86.069,125.673, 157.687,92.147], [279.423, 81.895,120.568,152.472,87.451],[269.421,85.732,131.734,146.673,88.356], in order to more intuitively compare the design effects, they are organized as shown in the following Table 1: Table 1 Designed structural parameter groups

[0036] The target parameters (a) / target parameters (b) / target parameters (c) and target parameters (d) in Table 1 correspond to the structural parameters of the metamaterial sensor chip corresponding to the four transmission curves selected from the test set, and the design parameters are the design values ​​output by the design model. This solution simulates the four sets of designed structural parameters and plots the obtained transmission curves together with the target curves input into the design model. Figure 7 It can be seen that the simulation results of the designed structural parameters are in good agreement with the target curve, and the design task of the target curve can be completed well.

[0037] The specific structure of the metamaterial sensor chip of this scheme is as follows Figure 1 and Figure 2 As shown, the metamaterial sensor chip includes: at least one sensor unit, wherein each sensor unit includes a substrate layer and a π-type resonance unit 10 and a rectangular resonance unit 20 disposed on the substrate layer and spaced apart, wherein the π-type resonance unit 10 includes a horizontally disposed horizontal bar 11 and two vertical bars disposed perpendicularly relative to the horizontal bar 11, the vertical bars include a first vertical bar 12 and a second vertical bar 13, a π-type opening 100 is formed between the first vertical bar 12 and the second vertical bar 13, wherein the rectangular resonance unit 20 is a rectangular closed-loop structure, and the opening direction of the π-type opening 100 is arranged directly opposite to the rectangular resonance unit 20; multi-mode coupling occurs between the π-type resonance unit 10 and the rectangular resonance unit 20, and interference destructive interaction generates three transparent window resonance peaks at three different frequencies. In this solution, the lengths of the first vertical bar 12 and the second vertical bar 12 are the same, and are both defined as the length of the vertical bar.

[0038] In some embodiments, the material of the π-type resonant unit 10 and the rectangular resonant unit 20 is copper, and the substrate layer is quartz. Specifically, the conductivity of copper is 5.96×10 7 S / m, and the relative dielectric constant of quartz is 1.9. The reason why this scheme chooses a material with a low dielectric constant as the material of the substrate layer is to enhance the sensing performance of the three-band electromagnetically induced transparent metamaterial sensor chip.

[0039] In some embodiments, the rectangular resonance unit 20 includes two parallel long sides and two parallel short sides, and the long sides and the short sides are arranged vertically, and the length of the long side is greater than the length of the short side. Specifically, the long side of the rectangular resonance unit 20 is arranged parallel to the horizontal strip 11 of the π-type resonance unit 10, and the opening of the π-type opening 100 is arranged opposite to the long side of the rectangular resonance unit 20.

[0040] like Figure 2 As shown, the horizontal length of the π-type resonance unit of the metamaterial sensor chip is defined as c, the vertical length is d, the width of the π-type opening is g, the length of the long side of the rectangular resonance unit is l, and the length of the short side is w. It is these five structural parameters that constitute the structural parameter group that affect the position of the transmission window resonance peak generated by the metamaterial sensor chip.

[0041] This solution uses CST Studio Suite electromagnetic simulation software to complete the simulation of the three-band electromagnetically induced transparent metamaterial sensor chip. During the simulation, the boundary conditions of the three-band electromagnetically induced transparent metamaterial sensor chip unit in the x and y directions are set as unit cell boundaries, and the z direction is the open (add space) boundary, so as to simulate an infinite metamaterial periodic array, select the THz plane wave with the incident direction of z as the excitation source, and set the simulation frequency band to 0.3~1THz.

[0042] This scheme simulates the π-type resonant unit and the rectangular resonant unit respectively. The simulated transmission curve is shown in the figure below. Figure 3 As shown in Figure 2, the contribution of the three transparent windows is studied. Figure 3 Curve R1 represents the transmission curve when the π-type resonance unit is simulated alone, curve R2 represents the transmission curve when the rectangular resonance unit is simulated alone, and curve R3 represents the transmission curve of the overall sensor chip simulation. It can be seen that the three transparent windows are formed by coupling of multiple modes.

[0043] Next, we analyze its formation mode by combining the electric field distribution diagram and obtain the electric field distribution diagrams at seven key points with frequencies from low to high as shown in the figure below: Figure 4 As shown, Figure 4 (a) in the figure represents the first trough. Figure 4 (b) in the figure represents the first transparent window. Figure 4 (c) in the figure indicates the second trough. Figure 4 (d) in the figure indicates the second transparent window. Figure 4 (e) in the graph represents the third trough. Figure 4 (f) in the figure indicates the third transparent window. Figure 4 (g) in the figure indicates the fourth trough. The electric field distribution is divided into three parts for analysis, namely the horizontal bar P1 of the π-type resonance unit, the first and second vertical bars P2 of the π-type resonance unit, and the rectangular resonance unit as a whole P3. It can be seen that P2 and P3 are excited at the first trough, and only P1 is excited at the second trough. The three are all in an over-bright state before and after the first transparent window. The first transparent window is a bright-bright-bright coupling mode, and the electric field energy is gathered between the horizontal bar, the first and second vertical bars of the π-type resonance unit, and the rectangular resonance unit. Only P2 is strongly excited at the third trough, and P3 is not excited before and after the second transparent window. The second transmission window is a bright-dark-bright coupling mode, and the electric field energy is only gathered in the horizontal bar and two vertical bars of the π-type resonance unit. P1 is strongly excited at the fourth trough, and P3 is not excited before and after the third transparent window. The third transmission window is a bright-dark-bright coupling mode, and the electric field energy is only concentrated on the upper side and two arms of the π-type resonator. In summary, the three transparent windows are bright-bright-bright coupling, bright-dark-bright coupling and bright-dark-bright coupling modes respectively.

[0044] Embodiment 2 This solution provides an application method of the design model of a metamaterial sensor chip based on deep learning constructed according to Embodiment 1, that is, Embodiment 2 provides a design method of a metamaterial sensor chip: The transmission curve collection points containing three transparent window resonance peaks are input into the design model, and the structural parameter group of the corresponding metamaterial sensor chip is output.

[0045] The contents that are the same as those in the first embodiment will not be described in detail here.

[0046] It should be noted that, in a specific embodiment, the present solution utilizes the design method of the metamaterial sensor chip to design a metamaterial sensor chip that can be used for antibiotic detection.

[0047] Specifically, the transmission curve collection points of the three transmission window resonance peaks including 0.76THz, 0.79THz and 0.85THz are input into the design model, and the structural parameter group of the corresponding metamaterial sensor chip is output. The metamaterial sensor chip can be used to detect chlortetracycline, tetracycline and sodium penicillin, wherein 0.76THz, 0.79THz and 0.85THz correspond to the fingerprint spectra of chlortetracycline, tetracycline and sodium penicillin, respectively.

[0048] Specifically, the structural parameter group of the metamaterial sensor chip obtained by this scheme is as follows: the length of the horizontal bar 11 of the π-type resonance unit 10 is 272µm~276µm, the length of the first vertical bar 12 and the second vertical bar 13 is 80µm~84µm, the width of the π-type opening 100 is 124µm~128µm, the length of the long side of the rectangular resonance unit 20 is 150µm~154µm, and the length of the short side is 86µm~90µm.

[0049] Preferably, the length of the horizontal bar 11 of the π-type resonance unit 10 is c is 274µm, the length of the first vertical bar 12 and the second vertical bar 13 d The width of the π-shaped opening is 82µm and 100 g The long side of the rectangular resonant unit 20 is 126µm. l The length is 152µm, the short side w The length of the substrate layer is 88µm. Px and Py 180µm.

[0050] In addition, the application team conducted a performance test on the preferred metamaterial sensor chip to study the response of the metamaterial sensor chip to the test objects with different refractive indices and obtained the following results: Figure 8 As shown, from Figure 8 It can be seen that when the refractive index of the object to be measured increases from 1.0 to 1.3 in steps of 0.1, the three transmission window resonance peaks all undergo frequency shifts to varying degrees. The resonance frequency of the first transmission window resonance peak moves from 0.755 THz to 0.706 THz, and the frequency shifts of the first transparent window resonance peak corresponding to each refractive index of the object to be measured are 164, 163, and 163 THz, resulting in a total redshift of 490 GHz; the resonance frequency of the second transmission window resonance peak moves from 0.798 THz to 0.770 THz, and the frequency shifts of the second transparent window resonance peak corresponding to each refractive index of the object to be measured are 95, 93, and 92 THz, resulting in a total redshift of 280 GHz; the resonance frequency of the third transmission window resonance peak moves from 0.848 THz to 0.802 THz, and the frequency shifts of the third transparent window resonance peak are 154, 154, and 152 THz, resulting in a total redshift of 460 GHz. By comparing with each other, it can be seen that the sensitivity of the three transmission window resonance peaks to the change of the refractive index of the object to be measured is, from low to high, the first transmission window resonance peak, the third transmission window resonance peak and the second transmission window resonance peak. The difference in sensitivity is caused by the different electromagnetic responses of the coupling fields of each transparent window to the object to be measured.

[0051] The linear fitting is performed on the refractive index changes of each transparent window relative to the surrounding test object to obtain the fitting results as follows: Fig. 9 As shown, S 1 , S2 and S 3 They represent the first, second and third transparent windows respectively. The slope of the fitted straight line is the sensitivity of the corresponding transparent window. The slopes of the three fitted straight lines are expressed by k 1 , k 2 and k 3 It means that k 1 =-0.163, k 2 =-0.093, k 3 =-0.155, its R 2 are 0.99977, 0.99931 and 0.9938 respectively, so the sensitivities of the three transparent windows are S 1 =163GHz / RIU, S 2 =93GHz / RIU, S 3 =155GHz / RIU. The center frequencies of the three transparent windows are 0.755THz, 0.795THz and 0.848THz, and the half-height widths are 0.051THz, 0.043THz and 0.032THz, respectively. The Q values ​​are 14.8, 18.5 and 26.5, and the FOMs are 3.2, 2.2 and 4.8, respectively. The sensing performances of the three transparent windows are somewhat different, and the specific sensing detection of chlortetracycline, tetracycline and penicillin sodium can be realized at the corresponding frequencies.

[0052] Those skilled in the art should understand that the technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0053] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for constructing a design model of a metamaterial sensor chip based on deep learning, characterized in that: The following steps are involved: A set of transmission curve collection points corresponding to the metamaterial sensor chip with different structural parameter groups is obtained as training data, wherein the upper structure of the metamaterial sensor chip includes a π-type resonance unit and a rectangular resonance unit arranged at intervals, and the π-type opening of the π-type resonance unit is arranged opposite to the rectangular resonance unit, wherein the structural parameter group includes the horizontal side length, the vertical bar length, the width of the π-type opening, the long side length and the short side length of the rectangular resonance unit; Input the training data into the design framework for training until the training conditions are met, wherein the design framework includes an input layer, a parallel convolution layer, a linear layer, and an output layer connected in sequence, and the parallel convolution layer includes multiple convolution branches, each convolution branch includes convolution layers of different convolution dimensions, and each convolution layer includes a ReLU activation function and a Pool operation; The transmission curve collection point set is input into the input layer for normalization to obtain a one-dimensional sequence. The one-dimensional sequence is input into multiple convolution branches of the parallel convolution layer for convolution to obtain convolution features. Multiple convolution features are input into the linear layer for splicing to obtain a spliced ​​sequence. The spliced ​​sequence is input into the output layer for processing and outputs a structural parameter group.

2. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 1, characterized in that: The parallel convolution layer includes three 1D convolution branches, each of which includes three convolution layers with different convolution dimensions. The convolution dimensions of the three convolution layers in each convolution branch are 256, 128, and 32, respectively.

3. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 2, characterized in that: The parallel convolution layer includes a first convolution branch, a second convolution branch and a third convolution branch in parallel, wherein the convolution kernels of the three convolution layers in the first convolution branch are 7, 5, and 3 respectively, the convolution kernels of the three convolution layers in the second convolution branch are 5, 5, and 5 respectively, and the convolution kernels of the three convolution layers in the third convolution branch are 3, 5, and 7 respectively.

4. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 1, characterized in that: The linear layer includes a splicing linear layer and a fully connected linear layer. The convolution features output by the three convolution branches are all put into the splicing linear layer for splicing to obtain splicing features. The splicing features are then mapped to dimension reduction through fully connected linear layers of different dimensions to obtain the splicing sequence after dimensionality reduction.

5. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 4, characterized in that: The concatenated linear layer concatenates the convolutional features of the three convolutional branches to obtain a concatenated feature of length 96, which is then mapped to a dimension reduction in a fully connected linear layer with dimensions 64 and 16 to obtain a concatenated sequence.

6. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 1, characterized in that: The π-type resonance unit includes a horizontal bar and two vertical bars arranged perpendicular to the horizontal bar, the vertical bars include a first vertical bar and a second vertical bar arranged at intervals, and a π-type opening is formed between the first vertical bar and the second vertical bar, wherein the rectangular resonance unit is a rectangular closed-loop structure, multi-mode coupling occurs between the π-type resonance unit and the rectangular resonance unit, and interference destructive interaction generates three transparent window resonance peaks at three different frequencies.

7. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 6, characterized in that: The rectangular resonance unit comprises two parallel long sides and two parallel short sides, and the long sides and the short sides are vertically arranged, and the length of the long sides is greater than the length of the short sides.

8. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 7, characterized in that: The long side of the rectangular resonance unit is arranged parallel to the horizontal side of the π-type resonance unit, and the opening of the π-type opening is arranged opposite to the long side of the rectangular resonance unit.

9. The method for constructing a design model of a metamaterial sensor chip based on deep learning according to claim 1, characterized in that: The material of the π-type resonance unit and the rectangular resonance unit is metal copper, and the substrate layer is quartz.

10. A method for designing a metamaterial sensor chip, characterized in that: The transmission curve collection points containing three transparent window resonance peaks are input into the design model constructed according to the method for constructing a design model of a metamaterial sensor chip based on deep learning according to any one of claims 1 to 9, and a structural parameter group corresponding to the metamaterial sensor chip is output.

Images