Narrow-band filter device design method and device based on machine learning

Abstract

The invention relates to the field of computer science and technology and micro-nano optics, in particular to a narrow-band filter device design method based on machine learning, which comprises the following steps: firstly, selecting a design carrier of a narrow-band filter, and then obtaining a simulation spectrum corresponding to structural parameters through simulation calculation. And fitting the simulation spectrum to obtain excitation peak parameters, thereby establishing a corresponding relation table between the simulation spectrum data and the excitation peak parameters. And based on the corresponding relation table, after an optimization target is set, a candidate parameter set is obtained by utilizing machine learning, and then a representative parameter set for guiding the production of the device is obtained by utilizing multi-target optimization. The invention also comprises a device generated by using the design method. According to the invention, a machine learning intelligent algorithm is integrated into the design process of a photonic device, a data driving and automation method is formed, the problems of low efficiency and poor precision and expandability in the design of a high-performance narrow-band filter device are solved, and a powerful technical tool is provided for the rapid development of integrated and miniaturized optical devices.

Description

Technical Field

[0001] This invention relates to the fields of computer science and technology and micro-nano optics, and in particular to a method and device for designing narrowband filter devices based on machine learning. Background Technology

[0002] Narrowband filtering devices play an irreplaceable role in key areas of modern technology, such as improving communication capacity, performing high-precision detection, analyzing microscopic information, and ensuring information security. In recent years, many narrowband filtering devices based on metamaterials have emerged. Continuous bound states (BICs) can achieve extremely high quality factors (Q-factors) and strong optical field localization in open systems, providing a novel physical mechanism for breakthroughs in photonic device performance. Since they can be realized through the design of periodic unit structures on optical metasurfaces, they have become a major research hotspot in micro-nano photonics in recent years, easily achieving narrowband filtering functions. However, traditional quasi-BIC designs rely on intuitive parameter scanning with symmetry breaking, which suffers from bottlenecks such as high resource consumption and low optimization accuracy, making it difficult to meet the directional requirements of multi-dimensional indicators such as excitation peak position, linewidth, and peak value in industrial applications.

[0003] In existing research, optimization of continuous bound state structures mostly relies on large-scale parameter scans and experience. Traditional optimization methods, such as gradient descent and genetic algorithms, also face two major challenges when dealing with high-dimensional parameter spaces: First, the time complexity of electromagnetic simulation calculations increases exponentially with the parameter dimension, leading to difficulties in estimating time costs. , (For parameter dimension); secondly, the topological protection characteristic of BIC leads to a large number of local extrema in the objective function, approximately This severely restricts the research and development efficiency of high-performance metasurface devices. Therefore, how to achieve rapid structural optimization to realize high-throughput optical output has become a challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a design method and device for narrowband filtering devices based on machine learning, which mainly solves the problems existing in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by this invention is to provide a narrowband filter device design method based on machine learning, characterized by comprising the following steps:

[0006] Step S100: Select a design carrier for the narrowband filter; the design carrier defines the selectable range of structural parameters for the narrowband filter.

[0007] Step S200: Within the selectable range, select different structural parameters and obtain the first simulation spectral dataset corresponding to the structural parameters through simulation calculation;

[0008] Step S300: Based on the first simulated spectrum dataset, the excitation peak parameters of the spectrum are obtained by fitting; the excitation peak parameters include linewidth parameters and peak position parameters;

[0009] Step S400: Combine the structural parameters with the excitation peak parameters obtained from the first simulation spectrum dataset to establish a corresponding relationship table;

[0010] Step S500: Set optimization target according to the excitation peak parameter; perform machine learning based on the correspondence table to obtain a candidate parameter set composed of the structural parameters; the spectrum formed by the values ​​of the structural parameters in the candidate parameter set corresponds to the excitation peak parameter, and the spectrum formed by the values ​​of the structural parameters that are locally better than other values ​​of the structural parameters corresponds to the excitation peak parameter.

[0011] Step S600: Based on the candidate parameter set, perform multi-objective optimization to form a representative parameter set; the values ​​of the structural parameters in the representative parameter set form the spectrum corresponding to the excitation peak parameter, and the values ​​of the structural parameters that are globally superior to the values ​​of other structural parameters form the spectrum corresponding to the excitation peak parameter.

[0012] Further, in step S100, the design carrier is a square lattice metasurface resonator, which is composed of nanopillar structures within a periodic lattice; the structural parameters include the lattice period and structural size of the nanopillar junction.

[0013] Furthermore, step S300 includes the following sub-steps:

[0014] Step S301: Use an adaptive background fitting function to fit a first background set based on the first simulated spectral dataset;

[0015] Step S302: Based on the first background set, select the points where significant peaks are located from the first simulated spectrum dataset and record them as the first significant peak data point set;

[0016] Step S303: For each significant peak in the first simulated spectral dataset, replace the data point corresponding to the significant peak and the N adjacent data points with null values ​​to form a second simulated spectral dataset.

[0017] Step S304: Based on the smooth spline background fitting function, fit a second background set with the marked background based on the second simulated spectral dataset;

[0018] Step S305: Subtract the first simulated spectral dataset from the second background dataset, detect the points where the significant peaks are located again, and record them as the second significant peak data point set;

[0019] Step S306: For each significant peak in the second significant peak data point set, select the data point corresponding to the significant peak and N adjacent data points, and perform peak fitting on it to obtain the peak fitting result corresponding to each significant peak.

[0020] Step S307: Iterate through each of the peak fitting results to obtain the linewidth parameter and peak position parameter of the significant peak, and record them as the excitation peak parameter;

[0021] Step S308: Traverse each of the peak fitting results, and for multiple peak fitting results with an overlap greater than the overlap threshold, retain the peak fitting result with the largest peak and the corresponding excitation peak parameter.

[0022] Further, in step S301, the adaptive background fitting function is a polynomial fitting; the polynomial fitting starts from order 0 and increases until the polynomial fitting residual satisfies the fitting condition, and finally outputs the final residual, the optimal polynomial coefficients, and the optimal polynomial order.

[0023] Further, in step S304, the smoothed spline background fitting function includes a smoothing factor; when the smoothing factor is 0, the fitted curve passes through all data points; when the smoothing factor is greater than 0, the fitted curve balances the fit and smoothness; the smoothed spline background fitting function generates new independent and dependent variables based on effective data points, and finally outputs the fitted value.

[0024] Further, in step S306, the peak fitting uses the Fano linear fitting formula, expressed as:

[0025] .

[0026] Furthermore, step S500 includes sub-steps.

[0027] Step S501: Set the optimization target as the excitation peak position and the excitation peak linewidth, requiring the excitation peak position to be as close as possible to the preset peak position, and the excitation peak structure linewidth to be as small as possible;

[0028] Step S502: Divide the parameter space of the structural parameters in the correspondence table into blocks to form multiple parameter subspaces;

[0029] Step S503: Traverse each parameter subspace and apply the Gaussian process regression algorithm to perform a global search within each parameter subspace to find candidate solutions for each parameter subspace.

[0030] Step S504: All the candidate subspace sets constitute the candidate parameter set.

[0031] Further, in step S600, a pessimistic Pareto optimization algorithm is used for multi-objective optimization, which includes the following sub-steps:

[0032] Step S601: Group the candidate parameter set to obtain candidate parameter subsets;

[0033] Step S602: In each subset of candidate parameters, select one or more values ​​of the structural parameters that perform well in multiple factors of the optimization objective as the optimal parameter combination of the subset of candidate parameters.

[0034] Step S603: Based on obtaining the optimal parameter combination using all the candidate parameter subsets, use the pessimistic Pareto optimization algorithm to find the optimal parameter combination that is not completely surpassed by other optimal parameter combinations in all factors of the optimization objective, and use it as the frontier output.

[0035] Step S604: Use a sample selection algorithm to select the optimal parameter combination from the frontier output to form the representative parameter set.

[0036] The present invention also discloses a narrowband filter device implemented using the above-mentioned machine learning-based narrowband filter device design method, characterized in that it is a square lattice metasurface resonator with a multi-layer stacked system; the substrate layer of the resonator is located at the bottom layer, and the periodic lattice layer is located above the substrate layer and arranged in the vertical direction; each lattice layer contains a nanopillar structure arranged in a three-dimensional periodic manner.

[0037] Furthermore, the substrate layer is made of aluminum oxide; the lattice layer is made of silicon.

[0038] In view of the above technical features, the present invention provides a machine learning-based narrowband filter device design method and device, which integrates machine learning intelligent algorithms into the design process of photonic devices, successfully transforming the traditional process that relies on experience and trial and error into a data-driven and automated method. Compared with the prior art, its significant advantages are: solving the problems of low efficiency, poor accuracy and scalability in the design of high-performance narrowband filter devices, and providing a powerful technical tool for the rapid development of integrated and miniaturized optical devices. Attached Figure Description

[0039] Figure 1 This is a flowchart of a preferred embodiment of the machine learning-based narrowband filter device design method of the present invention;

[0040] Figure 2 yes Figure 1 The flowchart shows the method for background removal and excitation peak fitting.

[0041] Figure 3 yes Figure 1 The flowchart below illustrates the method of machine learning using the Gaussian regression function.

[0042] Figure 4 yes Figure 1 The flowchart below illustrates a method for multi-objective optimization of machine learning results.

[0043] Figure 5 This is a schematic diagram of a preferred embodiment of a narrowband filter device designed using a machine learning-based narrowband filter device design method according to the present invention.

[0044] Figure 6 This is a three-dimensional schematic diagram of a preferred embodiment of a narrowband filter device designed using a machine learning-based narrowband filter device design method according to the present invention.

[0045] In the figure: 1-substrate layer, 2-periodic lattice layer. Detailed Implementation

[0046] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0047] Please see Figure 1 This invention discloses a narrowband filter device design method based on machine learning. As shown in the figure, a preferred embodiment includes the following steps:

[0048] Step S1: Select the design carrier for the narrowband filter device.

[0049] Narrowband filters have various basic structures, and one of them is chosen as the design framework. The design framework defines the range of possible structural parameters for the narrowband filter, thus limiting subsequent design to optimizing these parameters using machine learning, thereby simplifying the design process.

[0050] In this embodiment, a square lattice metasurface resonator is selected as the design carrier for the narrowband filter. The design object is a nanopillar junction within a periodic square lattice. The lattice period and structural dimensions of the nanopillar junction constitute the structural parameters of this type of narrowband filter.

[0051] Step S2: Obtain the first simulated spectral dataset.

[0052] Within the selectable range of structural parameters, multiple sets of different values ​​are chosen for each structural parameter. For each set of values, calculations are performed using electromagnetic simulation software based on the principle of rigorous coupled-wave analysis to obtain the corresponding simulated spectral dataset. After iterating through all selected values, all simulated spectral data are merged to form the first simulated spectral dataset. The first simulated spectral dataset describes the simulated spectra corresponding to different values ​​of the structural parameters.

[0053] Step S3: Fit the excitation peak parameters of the first simulated spectral dataset.

[0054] Based on the first simulated spectral dataset, the excitation peak parameters in the spectrum were obtained by fitting using the Fano peak parameter fitting algorithm. These parameters include two types: peak position parameters and linewidth parameters. Peak position parameters describe the location of the excitation peak, while linewidth parameters describe the shape (width) of the excitation peak.

[0055] In the Fano peak fitting results, the highest peak value after fitting is selected as the final significant peak fitting result, and the final significant fitting result is recorded. The fitting results are then visualized and plotted.

[0056] Step S4: Establish a correspondence table between structural parameters and excitation peak parameters.

[0057] Structural parameters determine device characteristics, and their spectral characteristics are contained in the first simulated spectral dataset. When designing narrowband filters, the main focus is on the peak position and linewidth parameters of the excitation peak, aiming for accurate peak position (consistent with the expected filtering peak) and small linewidth (meeting the narrowband requirement). Simply put, structural parameters are the specific metrics that need to be selected and optimized in the design, while excitation peak parameters are the evaluation metrics. After establishing a correspondence between the two, the results can be submitted to machine learning for optimization.

[0058] Step S5: Use machine learning to obtain a set of candidate parameters.

[0059] After each structural parameter is divided within its value range according to a set step size, the value vectors formed by all structural parameters together form the value space of the structural parameters.

[0060] To reduce the difficulty of machine learning, the value space formed by the structural parameters is further divided into multiple smaller blocks. Within each block, the peak position and linewidth parameters of the excitation peak are set as optimization objectives, i.e., accurate peak position and minimum linewidth. Machine learning is then used to search within each block, and the search results are sorted according to the optimization objectives. The top-ranked solutions are selected as the candidate parameter set for that block. All candidate sets from different blocks are then merged to form the candidate parameter set for the entire structural parameters.

[0061] Each structural parameter value in the candidate parameter set is ranked first in a certain small block, and therefore is locally superior to those eliminated values ​​within that block. Specifically, "superior" means that the spectrum formed by this structural parameter corresponds to a more accurate peak position and a smaller linewidth in the excitation peak parameters.

[0062] Step S6: Multi-objective optimization yields a representative parameter set.

[0063] The candidate parameter set contains the values ​​of locally dominant structural parameters from each of the smaller blocks. Further multi-objective optimization is performed based on this candidate parameter set to obtain the values ​​of globally dominant structural parameters, which serve as the representative parameter set. The values ​​of the structural parameters in the representative parameter set, whose spectral excitation peak parameters are globally superior to those of other structural parameters, are used as the final output for actual device manufacturing.

[0064] Please see Figure 2 In step S3, the excitation peak parameters are fitted from the first simulated spectral dataset. This process involves multiple background removal and excitation peak fitting steps, and consists of the following sub-steps:

[0065] Step S31: Fit the first background set.

[0066] The data in the first simulated spectral dataset has no prior information; it comes directly from the simulation of structural parameters. Therefore, in the first round, an adaptive background fitting function that does not require prior information input is used to fit the first simulated spectral dataset to obtain the first background set.

[0067] In this embodiment, the adaptive background fitting function is fitted using a polynomial. The polynomial fitting starts from order 0 and increases until the polynomial fitting residuals satisfy the fitting conditions. Finally, the final residuals, the optimal polynomial coefficients, and the optimal polynomial order are output.

[0068] Specifically, the input parameters are x, y, max_order, and tol. Here, x is the independent variable of the data to be fitted, y is the dependent variable, max_order is the maximum order used for fitting, and tol is the tolerance threshold. The initial residual is set to infinity. The initial fitting configuration is as follows: the initial optimal polynomial order is set to 0, and the coefficients of the 0th-order polynomial are fitted as the initial values ​​of the optimal polynomial coefficients. Then, the fitting loop gradually increases the current polynomial order until max_order, fitting the polynomial coefficients of the current order and calculating the residual. If the absolute value of the difference between the calculated current residual and the residual from the previous step is less than tol, the loop ends. Otherwise, the solution is updated, and the loop continues. The function returns the residual after iteration, the optimal polynomial coefficients, and the optimal polynomial order.

[0069] Step S32: Screen out the set of data points with the first significant peak.

[0070] After obtaining the first background set, background data can be removed from the first simulated spectral dataset to identify the points containing significant peaks. The set of points corresponding to these initially identified significant peaks is the first significant peak data set. The first significant peak data set represents the result of the initial screening.

[0071] Step S33: Form the second simulation spectrum dataset.

[0072] Using the results of the initial screening, the first simulated spectral dataset is refined to prepare for the second screening. Specifically, based on the significant peak data points marked in the first significant peak data point set, their positions in the first simulated spectral dataset are located. Then, the data points corresponding to these significant peaks and their corresponding neighboring data points (N data points adjacent to the significant peak data points) are replaced with null values ​​from their original values. In other words, the significant peaks identified in the initial screening are removed from the first simulated spectral dataset to form the second simulated spectral dataset.

[0073] Step S34: Fit the second background set.

[0074] The significant peaks initially identified have been removed from the second simulated spectral dataset. Therefore, the second background fitting can be more aggressive without significantly affecting the already identified significant peaks. The second background fitting uses a smooth spline background fitting function to obtain the second background set with labeled background.

[0075] The input parameters for the smoothed spline background fitting function are x, y, and s. x represents the independent variable of the data to be fitted, y represents the dependent variable, and s is the smoothing factor. The smoothing factor can be adjusted as needed: when the smoothing factor is 0, the fitted curve passes through all data points, meaning it is not smooth. When the smoothing factor is greater than 0, the fitted curve balances goodness of fit and smoothness, and it will not pass through all data points. In this scenario, the data points passed through by the fitted curve are considered valid data points. Therefore, after completing the fitting, the smoothed spline background fitting function generates new independent and dependent variables based on the valid data points, and finally outputs the fitted value.

[0076] Step S35: Generate the second significant peak data point set.

[0077] The second background set contains relatively refined background information formed after quadratic fitting. Subtracting the first simulated spectral dataset from the second background set yields the points where significant peaks are located, which are recorded as the second significant peak data point set. This second significant peak data point set contains the positions of all significant peaks identified from the original spectrum.

[0078] Step S36: Perform peak fitting on the significant peaks.

[0079] For each significant peak in the second significant peak data point set, the Fano linear fitting formula is used to fit it. Specifically, the data point corresponding to the significant peak and its N neighboring data points are selected, and the peak is fitted using the Fano resonance function, expressed by the formula:

[0080] .

[0081] Input parameters , , , , , After the fitting is completed, for each significant peak in the second significant peak data point set, its peak fitting result is obtained, that is, the coefficient of the correlation fitting function of that peak is obtained.

[0082] Step S37: Obtain the excitation peak parameters from the peak fitting results.

[0083] The peak fitting provides a mathematical expression and coefficients for each significant peak. In the specific Fano linear fitting formula, each coefficient has a physical meaning. Therefore, the values ​​of the specific coefficients given in the peak fitting result can be used to derive the linewidth and peak position parameters of the significant peak, which are recorded as excitation peak parameters.

[0084] Step S38: Remove duplicate peak fitting results.

[0085] After obtaining the peak position parameters corresponding to each peak fitting result, it is possible to assess whether there is any overlap between these peak fitting results. For peak fitting results with an overlap degree greater than the overlap threshold, only the peak fitting result with the largest peak and its corresponding excitation peak parameters are retained. After removing duplicate peak fitting results, the excitation peak parameters corresponding to the retained peak fitting results constitute the final excitation peak parameters. In this embodiment, the overlap threshold is set to 90%. That is, if the overlap between two significant peaks is greater than 90%, the excitation peak parameters corresponding to the significant peak with the smaller peak value are deleted.

[0086] Please see Figure 3 In this embodiment, step S5 uses machine learning to obtain a set of candidate parameters. Specifically, it employs a grouped Gaussian regression function, which includes the following sub-steps:

[0087] Step S51: Set optimization goals.

[0088] In the design of narrowband filter devices, the optimization objectives are the excitation peak position and the excitation peak linewidth. The excitation peak position is required to be accurate and as close as possible to the preset peak position, and the excitation peak linewidth is required to be as small as possible.

[0089] Step S52: Group the structural parameters in the parameter space.

[0090] Machine learning requires exploring the parameter space, constructed from the space of all possible values ​​for the structural parameters, which generally contains a large number of optional values. To accelerate machine learning, avoid resource exhaustion, and ensure controllable performance, the parameter space, consisting of all possible values ​​of the structural parameters in the correspondence table, is first divided into multiple parameter subspaces. Specifically, this involves considering all structural parameters used in the simulation dataset. ,Pick The upper and lower limits, and the step size are set to By dividing the vectors within the upper and lower limits of each parameter according to the step size, m vectors are obtained. Let the space spanned by these m vectors be defined as the parameter space. Set the block size to batch, and divide the parameter space into multiple smaller blocks according to the batch size.

[0091] Step S53: Search for the optimal solution in the parameter subspace.

[0092] For a given parameter subspace, a Gaussian process regression algorithm is iteratively applied to perform a global search across the entire parameter subspace, finding candidate solutions for each parameter subspace. Within each parameter subspace, one or more candidate solutions are selected. The selected candidate solutions outperform the eliminated solutions in terms of the optimization objective.

[0093] Step S54: Merge the candidate sets of the subspaces to form a candidate parameter set.

[0094] The algorithm traverses all parameter subspaces, obtaining one or more candidate solutions in each subspace. All candidate solutions are then merged to form a candidate parameter set.

[0095] Please see Figure 4 In this embodiment, step S6 uses a pessimistic Pareto optimization algorithm for multi-objective optimization, which includes the following sub-steps:

[0096] Step S61: Group the candidate parameter sets.

[0097] The candidate parameter set contains a large number of candidate solutions in its subspace, so it is first grouped to improve the efficiency of multi-objective optimization. After grouping, the candidate parameter set yields multiple candidate parameter subsets, and the union of these subsets equals the original candidate parameter set.

[0098] Step S62: Select the optimal parameter combination for each candidate parameter subset.

[0099] The optimization objective includes multiple factors. For example, the first factor is to ensure the excitation peak position is accurate and as close as possible to the preset peak position, and the second factor is to minimize the excitation peak linewidth. In each subset of candidate parameters, considering each factor in the optimization objective, the value of one or more structural parameters that performs well for all factors is selected as the optimal parameter combination for that subset of candidate parameters.

[0100] Step S63: Use the pessimistic Pareto optimization algorithm to select the leading edge output.

[0101] The optimal parameter combinations selected from all candidate parameter subsets are merged, and multi-objective optimization is performed again. This optimization uses the pessimistic Pareto optimization algorithm, which seeks an optimal parameter combination that is not completely surpassed by other optimal parameter combinations in all factors included in the optimization objective, and uses this as the frontier output of the pessimistic Pareto optimization algorithm.

[0102] Step S64: Select a representative parameter set based on the leading edge output.

[0103] A representative parameter set is selected from the leading-edge output using a sample selection algorithm. One or more sets of data from the representative parameter set can be used in actual production. That is, by selecting one or more sets of data from the representative parameter set, arranging them in an array, drawing a layout, and conducting experimental processing, a device that meets the design requirements can be obtained.

[0104] Please see Figure 5 and Figure 6 This invention also discloses a narrowband filter device implemented using the aforementioned machine learning-based narrowband filter device design method. In a preferred embodiment, a square lattice metasurface resonator with a multi-layer stacked system is used to implement the narrowband filter device. This resonator includes a substrate layer 1 and a periodic lattice layer 2. The substrate layer 1 is located at the bottom and is made of aluminum oxide. Above the substrate layer 1 is the periodic lattice layer 2, made of silicon. It is located above the substrate layer and arranged vertically. The specific physical structure within the periodic lattice layer 2 is a nanopillar structure, with the nanopillar structures arranged in a three-dimensional periodic pattern.

[0105] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A machine learning-based method for designing narrowband filtering devices, characterized in that, Includes the following steps: Step S100: Select a design carrier for the narrowband filter; the design carrier defines the selectable range of structural parameters for the narrowband filter. Step S200: Within the selectable range, select different structural parameters and obtain the first simulation spectral dataset corresponding to the structural parameters through simulation calculation; Step S300: Based on the first simulated spectrum dataset, the excitation peak parameters of the spectrum are obtained by fitting; the excitation peak parameters include linewidth parameters and peak position parameters; Step S400: Combine the structural parameters with the excitation peak parameters obtained from the first simulation spectrum dataset to establish a corresponding relationship table; Step S500: Set the optimization target according to the excitation peak parameters; Machine learning is performed based on the correspondence table to obtain a candidate parameter set composed of the structural parameters; the spectrum formed by the values ​​of the structural parameters in the candidate parameter set corresponds to the excitation peak parameter, and the spectrum formed by the values ​​of the structural parameters that are locally better than other values ​​of the structural parameters corresponds to the excitation peak parameter. Step S600: Based on the candidate parameter set, perform multi-objective optimization to form a representative parameter set; The spectrum formed by the values ​​of the structural parameters in the representative parameter set corresponds to the excitation peak parameter, and the spectrum formed by the values ​​of the structural parameters that are globally superior to those of other structural parameters corresponds to the excitation peak parameter.

2. The machine learning-based narrowband filter device design method according to claim 1, characterized in that, In step S100, the design carrier is a square lattice metasurface resonator, which is composed of nanopillar structures within a periodic lattice; the structural parameters include the lattice period and structural size of the nanopillar junction.

3. The machine learning-based narrowband filter device design method according to claim 1, characterized in that, Step S300 includes the following sub-steps: Step S301: Use an adaptive background fitting function to fit a first background set based on the first simulated spectral dataset; Step S302: Based on the first background set, select the points where significant peaks are located from the first simulated spectrum dataset and record them as the first significant peak data point set; Step S303: For each significant peak in the first simulated spectral dataset, replace the data point corresponding to the significant peak and the N adjacent data points with null values ​​to form a second simulated spectral dataset. Step S304: Based on the smooth spline background fitting function, fit a second background set with the marked background based on the second simulated spectral dataset; Step S305: Subtract the first simulated spectral dataset from the second background dataset, detect the points where the significant peaks are located again, and record them as the second significant peak data point set; Step S306: For each significant peak in the second significant peak data point set, select the data point corresponding to the significant peak and N adjacent data points, and perform peak fitting on it to obtain the peak fitting result corresponding to each significant peak. Step S307: Iterate through each of the peak fitting results to obtain the linewidth parameter and peak position parameter of the significant peak, and record them as the excitation peak parameter; Step S308: Traverse each of the peak fitting results, and for multiple peak fitting results with an overlap greater than the overlap threshold, retain the peak fitting result with the largest peak and the corresponding excitation peak parameter.

4. The machine learning-based narrowband filter device design method according to claim 3, characterized in that, In step S301, the adaptive background fitting function is a polynomial fitting; the polynomial fitting starts from order 0 and increases until the polynomial fitting residual satisfies the fitting condition, and finally outputs the final residual, the optimal polynomial coefficients, and the optimal polynomial order.

5. The machine learning-based narrowband filter device design method according to claim 3, characterized in that, In step S304, the smoothed spline background fitting function includes a smoothing factor; when the smoothing factor is 0, the fitted curve passes through all data points; when the smoothing factor is greater than 0, the fitted curve balances the fit and smoothness; the smoothed spline background fitting function generates new independent and dependent variables based on effective data points, and finally outputs the fitted value.

6. The machine learning-based narrowband filter device design method according to claim 1, characterized in that, In step S306, the peak fitting uses the Fano linear fitting formula, expressed as: 。 7. The machine learning-based narrowband filter device design method according to claim 1, characterized in that, Step S500 includes sub-steps. Step S501: Set the optimization target as the excitation peak position and the excitation peak linewidth, requiring the excitation peak position to be as close as possible to the preset peak position, and the excitation peak structure linewidth to be as small as possible; Step S502: Divide the parameter space of the structural parameters in the correspondence table into blocks to form multiple parameter subspaces; Step S503: Traverse each parameter subspace and apply the Gaussian process regression algorithm to perform a global search within each parameter subspace to find candidate solutions for each parameter subspace. Step S504: All the candidate subspace sets constitute the candidate parameter set.

8. The machine learning-based narrowband filter device design method according to claim 1, characterized in that, In step S600, a pessimistic Pareto optimization algorithm is used for multi-objective optimization, which includes the following sub-steps: Step S601: Group the candidate parameter set to obtain candidate parameter subsets; Step S602: In each subset of candidate parameters, select one or more values ​​of the structural parameters that perform well in multiple factors of the optimization objective as the optimal parameter combination of the subset of candidate parameters. Step S603: Based on obtaining the optimal parameter combination using all the candidate parameter subsets, use the pessimistic Pareto optimization algorithm to find the optimal parameter combination that is not completely surpassed by other optimal parameter combinations in all factors of the optimization objective, and use it as the frontier output. Step S604: Use a sample selection algorithm to select the optimal parameter combination from the frontier output to form the representative parameter set.

9. A narrowband filter device implemented using the machine learning-based narrowband filter device design method as described in claim 1, characterized in that, The resonator is a square lattice metasurface resonator employing a multi-layer stacked system; the substrate layer of the resonator is located at the bottom layer, and the periodic lattice layers are located above the substrate layer and arranged in the vertical direction; each lattice layer contains nanopillar structures arranged in a three-dimensional periodic manner.

10. The machine learning-based narrowband filter device design method and device according to claim 9, characterized in that, The substrate layer is made of aluminum oxide; the lattice layer is made of silicon.

Images