A spectral library optimization method and system based on sensitivity analysis

Abstract

This invention discloses a method and system for optimizing a spectral library based on sensitivity analysis, belonging to the field of semiconductor optical measurement. The optimization method includes: acquiring simulated spectra; performing sensitivity analysis on preset library construction parameters based on the simulated spectra to obtain the sensitivity of each parameter; obtaining the value range of each parameter and, based on the sensitivity and value range, obtaining the partitioning weight of each parameter; calculating the number of grid divisions for each parameter based on its partitioning weight; obtaining grid points according to the number of grid divisions, with each grid point corresponding to a set of geometric parameters, thus obtaining a parameter mesh; and simulating the geometric parameters corresponding to each grid point in the parameter mesh using the Fourier modal method to generate an optimized spectral library. The spectral library obtained by this optimization method has a balanced sample distribution and small bias, improving the accuracy and stability of the neural network model.

Description

Technical Field

[0001] This invention relates to the field of semiconductor optical measurement, and in particular to a method and system for optimizing a spectral library based on sensitivity analysis. Background Technology

[0002] In high-density integrated circuit manufacturing, nanoscale metrology is crucial for ensuring device performance. Elliptic polarization spectroscopy, as an important nanoscale metrology technique, is widely used to characterize critical dimensions and optical constants of semiconductor chip structures due to its advantages such as surface sensitivity, non-destructiveness, and non-invasiveness. It is based on the change in polarization state of linearly polarized light after reflection from a sample. By establishing a model, the simulated spectrum is calculated using the Fourier mode method (FMM) and similar algorithms, and then fitted with the experimental spectrum to obtain the desired result.

[0003] Traditional optical geometric dimension measurement methods rely on complex electromagnetic calculations, which, while accurate, are computationally expensive and slow. To accelerate the measurement process, neural network models are used to replace traditional calculations, learning the mapping relationship between parameters and spectra through a pre-generated spectral library. However, the sampling distribution of the spectral library depends on the operator's experience, leading to uneven sample distribution and large biases, which in turn affects the accuracy and stability of the neural network model. Summary of the Invention

[0004] This solution aims to provide a spectral library optimization method and system based on sensitivity analysis to solve the above-mentioned technical problems, so as to make the optimized spectral library samples have a balanced distribution and small deviation, thereby improving the accuracy and stability of the neural network model.

[0005] To address the above problems, this invention provides a method and system for optimizing a spectral library based on sensitivity analysis, comprising the following steps:

[0006] Obtain simulated spectra;

[0007] Sensitivity analysis of preset library construction parameters is performed based on simulated spectra to obtain the sensitivity of each library construction parameter;

[0008] Obtain the value range of each database creation parameter, and based on the sensitivity and value range of each database creation parameter, obtain the partitioning weight of each database creation parameter;

[0009] The number of grid divisions for each database construction parameter is calculated based on the division weight of each database construction parameter.

[0010] Grid points are obtained based on the number of grid divisions for each database construction parameter. Each grid point corresponds to a set of geometric parameters, thus obtaining the parameter grid.

[0011] An optimized spectral library is generated by simulating the geometric parameters corresponding to each grid point in the parameter network using the Fourier modal method.

[0012] In the above scheme, by acquiring simulated spectra, sensitivity analysis is performed on preset library construction parameters, i.e., the sensitivity of each library construction parameter to changes in spectral output is calculated. Based on the parameter value range and sensitivity of each library construction parameter, the partition weight of each parameter is obtained, thereby calculating the number of grid divisions for each parameter. Grid points are obtained based on the number of grid divisions for each parameter, where a set of geometric parameters corresponding to each grid point constitutes a parameter network. Finally, the geometric parameters corresponding to each grid point in the parameter network are simulated using the Fourier modal method to generate an optimized spectral library, which can then be used to train a neural network. The spectral images optimized by the above scheme have a balanced sampling distribution and small sample bias, greatly improving the accuracy and stability of the neural network model.

[0013] Furthermore, the sensitivity analysis of the preset library construction parameters based on simulated spectra to obtain the sensitivity of each library construction parameter includes:

[0014] The sensitivity of each database construction parameter is calculated using the partial derivative finite difference method, and is expressed as follows:

[0015]

[0016] Where F_λ represents the simulated spectrum at the λ-th wavelength, p_i represents the expected process value of the ith library construction parameter, Δp_i represents the parameter perturbation step size, N represents the wavelength index, and paraSense[i] represents the sensitivity of the ith library construction parameter.

[0017] In the above scheme, the influence of changes in the collection parameters on the optical measurement results is obtained by calculating the finite difference method using partial derivatives. Specifically, the sensitivity paraSense[i] of each collection parameter is obtained by simulating the output change of the spectrum, providing an important basis for subsequent mesh generation. Besides calculating the maximum value of the output change of the simulated spectrum as in the above scheme, the sensitivity of each collection parameter can also be obtained by calculating the average value or other variations. Compared to the existing method of manually setting the sampling interval, the above scheme integrates sensitivity analysis into automatic mesh generation, achieving dynamic adjustment of sampling density and improving accuracy.

[0018] Further, the process of obtaining the value range of each database creation parameter, and based on the sensitivity and value range of each database creation parameter, obtaining the partitioning weight of each database creation parameter, specifically includes:

[0019] Based on the database creation parameters, calculate the value range paraRange[i] for each database creation parameter, which represents the value range of the i-th database creation parameter;

[0020] The sensitivity of each database creation parameter is multiplied by the range of values ​​for each database creation parameter to obtain the partition weight of each database creation parameter, specifically expressed as: paraWeight[i] = paraSense[i] × paraRange[i];

[0021] In the formula, paraWeight[i] represents the partition weight of the i-th database creation parameter, paraSense[i] represents the sensitivity of the i-th database creation parameter, and paraRange[i] represents the parameter value range of the i-th database creation parameter.

[0022] The above scheme calculates the partition weight of each database construction parameter, assigning higher weights to more representative database construction parameter data, so that the data can play a greater role in the following steps, thereby improving the accuracy of the neural network training set.

[0023] Furthermore, the calculation of the number of grid divisions for each database construction parameter based on the division weight of each database construction parameter specifically includes:

[0024] Normalize each database creation parameter;

[0025] The maximum and minimum weight values, based on the database construction parameters, are normalized and represented as follows:

[0026]

[0027] Where normParaWeight[i] is the partition weight after normalization of the i-th database creation parameter, paraWeightMax represents the maximum value among all database creation parameter partition weights, and paraWeightMin represents the minimum value among all database creation parameter partition weights.

[0028] The above scheme combines sensitivity analysis with weighted normalization, assigning higher weights to highly sensitive library construction parameters, making them more important in subsequent calculations. This provides higher accuracy and stability for subsequent automatic mesh generation and dynamic adjustment of sampling density. Furthermore, this scheme is particularly suitable for elliptically polarized light measurement techniques, commonly used to characterize critical dimensions and optical constants of semiconductor chips.

[0029] Furthermore, the calculation of the number of grid divisions for each database construction parameter based on the division weight of each database construction parameter specifically includes:

[0030] The base value used for initial grid number allocation is calculated as follows:

[0031]

[0032] Where LibCountUser represents the number of preset spectral libraries, and m represents the number of library creation parameters;

[0033] The number of grid divisions for each database construction parameter is calculated based on the normalized partition weights and base values, specifically as follows:

[0034] Grid[i]=normParaWeight[i]×base;

[0035] Where Grid[i] represents the number of grid divisions for the i-th database creation parameter.

[0036] The above scheme obtains a base value for initially allocating the number of grid cells by using the preset number of spectral libraries and the number of library construction parameters specified by the user. Then, it calculates the number of grid cells for each library construction parameter based on the normalized partitioning weights and the base value. By comprehensively considering the sensitivity, value range, and partitioning weight of each library construction parameter, the automatically calculated optimal number of grid cells avoids the problem of unreasonable grid partitioning caused by manual settings, thereby effectively improving computational efficiency and accuracy.

[0037] Furthermore, the step of obtaining grid points based on the number of grid divisions for each database construction parameter, with each grid point corresponding to a set of geometric parameters, to obtain a parameter mesh, specifically includes:

[0038] Calculate the corresponding grid points based on the number of grid divisions for each parameter. Each grid point corresponds to a set of geometric parameters, represented as follows:

[0039]

[0040] In the formula, paraMin[i] represents the minimum value of the i-th database construction parameter, Grid[i] represents the number of grid divisions for the i-th database construction parameter, j represents any positive integer less than Grid[i], and m represents the number of database construction parameters.

[0041] The geometric parameters corresponding to each grid point in the above scheme can form a parameter network. Then, based on the sampling of the parameter network, an optimized neural network training set is generated by simulating the spectral library of the geometric parameters corresponding to the grid point using the Fourier modality method. This realizes an automatic optimization scheme based on sensitivity analysis, ensuring that the sampling points are concentrated in the effective area and improving accuracy.

[0042] This invention also provides a spectral library optimization system based on sensitivity analysis, comprising: a measurement result acquisition module for acquiring simulated spectra; a sensitivity analysis module for performing sensitivity analysis on preset library construction parameters based on the simulated spectra to acquire the sensitivity of each library construction parameter; a weight partitioning module for acquiring the parameter value range of each library construction parameter and acquiring the partitioning weight of each library construction parameter based on the sensitivity and parameter value range of each library construction parameter; a grid partitioning module for calculating the number of grid partitions for each library construction parameter based on the partitioning weight of each library construction parameter; a parameter mesh acquisition module for acquiring grid points according to the number of grid partitions for each library construction parameter, with each grid point corresponding to a set of geometric parameters to obtain a parameter mesh; and a spectral library optimization module for simulating the geometric parameters corresponding to each grid point in the parameter mesh using the Fourier modal method to generate an optimized spectral library.

[0043] The proposed system architecture is simple and easy to implement, and can be used to implement a sensitivity analysis-based spectral library optimization method and system as described above. The system obtains simulated spectra through a measurement result acquisition module; performs sensitivity analysis on preset library construction parameters based on the simulated spectra through a sensitivity analysis module to obtain the sensitivity of each parameter; assigns a weight to each parameter based on its sensitivity and value range through a weight partitioning module; calculates the number of grid divisions for each parameter based on its partition weight; obtains grid points based on the number of grid divisions for each parameter through a parameter mesh acquisition module, with each grid point corresponding to a set of geometric parameters, thus generating the parameter mesh; and finally, uses a spectral library optimization module to simulate the geometric parameters corresponding to each grid point in the parameter mesh using the Fourier mode method to generate an optimized spectral library.

[0044] Furthermore, the sensitivity analysis module is used to perform sensitivity analysis on preset library construction parameters based on simulated spectra, and to obtain the sensitivity of each library construction parameter, including:

[0045] The sensitivity of each database construction parameter is calculated using the partial derivative finite difference method, and is expressed as follows:

[0046]

[0047] Where F_λ represents the simulated spectrum at the λ-th wavelength, p_i represents the expected process value of the ith library construction parameter, Δp_i represents the parameter perturbation step size, N represents the wavelength index, and paraSense[i] represents the sensitivity of the ith library construction parameter.

[0048] The above scheme uses partial derivatives to calculate the finite difference method to obtain the impact of changes in the library construction parameters on the optical measurement results. Specifically, it obtains the sensitivity paraSense[i] of each library construction parameter by simulating the output change of the spectrum, providing an important basis for subsequent mesh generation. Besides calculating the maximum value of the output change of the simulated spectrum as described above, the sensitivity of each library construction parameter can also be obtained by calculating the average value or other variations. Compared to the existing method of manually setting the sampling interval, the above scheme integrates sensitivity analysis into automatic mesh generation, enabling dynamic adjustment of the sampling density and improving accuracy.

[0049] Furthermore, the weighting module is used to obtain the parameter value range of each database creation parameter, and based on the sensitivity of each database creation parameter and the parameter value range of each database creation parameter, to obtain the partitioning weight of each database creation parameter, including:

[0050] Based on the database creation parameters, calculate the value range paraRange[i] for each database creation parameter, which represents the value range of the i-th database creation parameter;

[0051] The sensitivity of each database creation parameter is multiplied by the range of values ​​for each database creation parameter to obtain the partition weight of each database creation parameter, specifically expressed as: paraWeight[i] = paraSense[i] × paraRange[i];

[0052] In the formula, paraWeight[i] represents the partition weight of the i-th database creation parameter, paraSense[i] represents the sensitivity of the i-th database creation parameter, and paraRange[i] represents the parameter value range of the i-th database creation parameter.

[0053] The above scheme calculates the partition weight of each database construction parameter, assigning higher weights to more representative database construction parameter data, so that the data can play a greater role in the following steps, thereby improving the accuracy of the neural network training set.

[0054] Furthermore, the parameter mesh acquisition module is used to acquire mesh points according to the number of mesh divisions for each database construction parameter, where each mesh point corresponds to a set of geometric parameters, including:

[0055] Calculate the corresponding grid points based on the number of grid divisions for each parameter. Each grid point corresponds to a set of geometric parameters, represented as follows:

[0056]

[0057] In the formula, paraMin[i] represents the minimum value of the i-th database construction parameter, Grid[i] represents the number of grid divisions for the i-th database construction parameter, j represents any positive integer less than Grid[i], and m represents the number of database construction parameters.

[0058] The geometric parameters corresponding to each grid point in the above scheme can form a parameter network. Then, based on the sampling of the parameter network, an optimized neural network training set is generated by simulating the spectral library of the geometric parameters corresponding to the grid point using the Fourier modality method. This realizes an automatic optimization scheme based on sensitivity analysis, ensuring that the sampling points are concentrated in the effective area and improving accuracy. Attached Figure Description

[0059] Figure 1 A schematic diagram of a spectral library optimization method based on sensitivity analysis is provided in an embodiment of the present invention;

[0060] Figure 2 This invention provides a semiconductor example illustrating a spectral library optimization method and system based on sensitivity analysis, as provided in an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of a spectral library optimization system architecture based on sensitivity analysis, provided as an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figure 1 This embodiment provides a spectral library optimization method based on sensitivity analysis, including the following steps:

[0064] Step S1: Obtain the simulated spectrum;

[0065] Step S2: Perform sensitivity analysis on the preset library construction parameters based on simulated spectra to obtain the sensitivity of each library construction parameter;

[0066] Step S3: Obtain the parameter value range of each database creation parameter, and based on the sensitivity of each database creation parameter and the parameter value range of each database creation parameter, obtain the partitioning weight of each database creation parameter;

[0067] Step S4: Calculate the number of grid divisions for each database construction parameter based on the division weight of each database construction parameter;

[0068] Step S5: Obtain grid points according to the number of grid divisions for each database construction parameter. Each grid point corresponds to a set of geometric parameters to obtain the parameter grid.

[0069] Step S6: Simulate the geometric parameters corresponding to each grid point in the parameter network using the Fourier modal method to generate an optimized spectral library.

[0070] In this embodiment, by acquiring simulated spectra, sensitivity analysis is performed on preset library construction parameters, i.e., the sensitivity of each library construction parameter is obtained by calculating the change of the spectral output to the library construction parameter. Based on the parameter value range and sensitivity of each library construction parameter, the partition weight of each parameter is obtained, thereby calculating the number of grid divisions for each parameter. Grid points are obtained based on the number of grid divisions for each parameter, where a set of geometric parameters corresponding to each grid point constitutes a parameter network. Finally, the geometric parameters corresponding to each grid point in the parameter network are simulated using the Fourier modal method to generate an optimized spectral library, which can then be used to train a neural network. The spectral image obtained through the above optimization method has a balanced sampling sample distribution and small sample bias, greatly improving the accuracy and stability of the neural network model. Furthermore, this invention is applicable to multiple optical measurement fields such as thin film thickness and optical geometry for generating spectral libraries in different parameter spaces, meeting the needs of various optical model training and prediction.

[0071] In one embodiment, the sensitivity analysis of preset library construction parameters based on simulated spectroscopy to obtain the sensitivity of each library construction parameter includes:

[0072] The sensitivity of each database construction parameter is calculated using the partial derivative finite difference method, and is expressed as follows:

[0073]

[0074] Where F_λ represents the simulated spectrum at the λ-th wavelength, p_i represents the expected process value of the ith library construction parameter, Δp_i represents the parameter perturbation step size, N represents the wavelength index, and paraSense[i] represents the sensitivity of the ith library construction parameter.

[0075] In this embodiment, the influence of changes in the library construction parameters on the optical measurement results is obtained through the finite difference method calculated by partial derivatives. Specifically, the sensitivity paraSense[i] of each library construction parameter is obtained by simulating the output change of the spectrum, providing an important basis for subsequent mesh generation. Besides calculating the maximum value of the output change of the simulated spectrum as described above, the sensitivity of each library construction parameter can also be obtained through other variations such as calculating the average value. Compared to the existing method of manually setting the sampling interval, the above scheme integrates sensitivity analysis into automatic mesh generation, achieving dynamic adjustment of sampling density and improving accuracy.

[0076] In one embodiment, obtaining the parameter value range of each database creation parameter, and obtaining the partitioning weight of each database creation parameter based on the sensitivity of each database creation parameter and the parameter value range of each database creation parameter, specifically includes:

[0077] Based on the database creation parameters, calculate the value range paraRange[i] for each database creation parameter, which represents the value range of the i-th database creation parameter;

[0078] The sensitivity of each database creation parameter is multiplied by the range of values ​​for each database creation parameter to obtain the partition weight of each database creation parameter, specifically expressed as: paraWeight[i] = paraSense[i] × paraRange[i];

[0079] In the formula, paraWeight[i] represents the partition weight of the i-th database creation parameter, paraSense[i] represents the sensitivity of the i-th database creation parameter, and paraRange[i] represents the parameter value range of the i-th database creation parameter.

[0080] In this embodiment, the partitioning weight of each database construction parameter is calculated during the implementation process. Higher weights are assigned to more representative database construction parameter data, so that the data can play a greater role in the following steps, thereby improving the accuracy of the neural network training set.

[0081] In one embodiment, calculating the number of grid divisions for each database construction parameter based on the partitioning weight of each database construction parameter specifically includes:

[0082] Normalize each database creation parameter;

[0083] The maximum and minimum weight values, based on the database construction parameters, are normalized and represented as follows:

[0084]

[0085] Where normParaWeight[i] is the partition weight after normalization of the i-th database creation parameter, paraWeightMax represents the maximum value among all database creation parameter partition weights, and paraWeightMin represents the minimum value among all database creation parameter partition weights.

[0086] This embodiment combines sensitivity analysis with weighted normalization, assigning higher weights to highly sensitive library construction parameters, making them more important in subsequent calculations. This provides higher accuracy and stability for subsequent automatic mesh generation and dynamic adjustment of sampling density. Furthermore, this approach is particularly suitable for elliptically polarized light measurement techniques, commonly used to characterize critical dimensions and optical constants of semiconductor chips.

[0087] In one embodiment, calculating the number of grid divisions for each database construction parameter based on the partitioning weight of each database construction parameter specifically includes:

[0088] The base value used for initial grid number allocation is calculated as follows:

[0089]

[0090] Where LibCountUser represents the number of preset spectral libraries, and m represents the number of library creation parameters;

[0091] The number of grid divisions for each database construction parameter is calculated based on the normalized partition weights and base values, specifically as follows:

[0092] Grid[i]=normParaWeight[i]×base;

[0093] Where Grid[i] represents the number of grid divisions for the i-th database creation parameter.

[0094] This embodiment obtains a base value for initially allocating the number of grid cells by using the preset number of spectral libraries and the number of library construction parameters specified by the user. Then, it calculates the number of grid cells for each library construction parameter based on the normalized partitioning weights and the base value. By comprehensively considering the sensitivity, value range, and partitioning weight of each library construction parameter, the automatically calculated optimal number of grid cells avoids the problem of unreasonable grid partitioning caused by manual settings, thereby effectively improving computational efficiency and accuracy.

[0095] In one embodiment, the step of obtaining grid points based on the number of grid divisions 'd' for each database construction parameter, with each grid point corresponding to a set of geometric parameters, to obtain a parameter grid, specifically includes:

[0096] Calculate the corresponding grid points based on the number of grid divisions for each parameter. Each grid point corresponds to a set of geometric parameters, represented as follows:

[0097]

[0098] In the formula, paraMin[i] represents the minimum value of the i-th database construction parameter, Grid[i] represents the number of grid divisions for the i-th database construction parameter, j represents any positive integer less than Grid[i], and m represents the number of database construction parameters.

[0099] In this embodiment, a parameter network can be constructed using the geometric parameters corresponding to each grid point. Then, based on the sampling of the parameter network, an optimized neural network training set is generated by simulating the spectral library of the geometric parameters corresponding to the grid point using the Fourier modality method. This achieves an automatic optimization scheme based on sensitivity analysis, ensuring that the sampling points are concentrated in the effective region and improving accuracy.

[0100] To more clearly illustrate the sensitivity analysis-based spectral library optimization method provided by the above scheme and highlight its technical advantages, this embodiment uses... Figure 2 The semiconductor example shown is used for illustration. Figure 2 The specific meanings involved are shown in Table 1 below:

[0101] Table 1. Detailed Explanation of Semiconductor Examples

[0102]

[0103] This embodiment compares the number of grid divisions for automatically generated (Auto) library construction parameters with the existing method of manually setting (SelfDefine). Specific data is shown in Table 1 below:

[0104] Table 2 Grid Allocation Table under Different Methods

[0105]

[0106]

[0107] As shown in Table 2, Auto represents the spectral library optimization method based on sensitivity analysis used in this embodiment; SelfDefine represents the manual setting method used in the prior art; L1_HT represents the thickness of the first thin film; G2_Depth represents the height of the trapezoid in the second layer;

[0108] Based on the Auto and SelfDefine datasets described in Table 2, neural network models were trained under the same conditions, resulting in two neural network models. Regression tests were then performed using simulated spectra (L1_HT:90, G2_Depth:40), and the results are as follows:

[0109] Table 3 Comparison of results under different methods

[0110]

[0111] Through the above steps, as shown in the relevant data results provided in Table 3, the spectral library optimization system based on sensitivity analysis applied in this embodiment automatically generates a reasonable distribution of sampling points in the specified geometric parameter space and dynamically adjusts the sampling density. This effectively avoids the blindness of manual operation and dependence on experience, improves the automation and intelligence level of the sampling process, realizes efficient and automatic grid division calculation, provides a scientific basis for parameter allocation in the library generation process, and effectively improves the accuracy of the final result.

[0112] Please see Figure 3 This embodiment also provides a spectral library optimization system based on sensitivity analysis, including: a measurement result acquisition module for acquiring simulated spectra; a sensitivity analysis module for performing sensitivity analysis on preset library construction parameters based on simulated spectra to acquire the sensitivity of each library construction parameter; a weight partitioning module for acquiring the parameter value range of each library construction parameter and acquiring the partitioning weight of each library construction parameter based on the sensitivity and parameter value range of each library construction parameter; a grid partitioning module for calculating the number of grid partitions for each library construction parameter based on the partitioning weight of each library construction parameter; a parameter mesh acquisition module for acquiring grid points according to the number of grid partitions for each library construction parameter, with each grid point corresponding to a set of geometric parameters to obtain a parameter mesh; and a spectral library optimization module for simulating the geometric parameters corresponding to each grid point in the parameter mesh using the Fourier mode method to generate an optimized spectral library.

[0113] The system architecture proposed in this embodiment is simple and easy to implement, and can be used to implement the aforementioned sensitivity analysis-based spectral library optimization method and system. The system acquires simulated spectra through a measurement result acquisition module; performs sensitivity analysis on preset library construction parameters based on the simulated spectra through a sensitivity analysis module to obtain the sensitivity of each library construction parameter; obtains the partitioning weight of each library construction parameter based on its sensitivity and value range through a weight partitioning module; calculates the number of grid partitions for each library construction parameter based on its partitioning weight through a grid partitioning module; obtains grid points based on the number of grid partitions for each library construction parameter through a parameter mesh acquisition module, with each grid point corresponding to a set of geometric parameters, thus obtaining the parameter mesh; finally, uses the spectral library optimization module to simulate the geometric parameters corresponding to each grid point in the parameter mesh using the Fourier mode method to generate an optimized spectral library.

[0114] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for optimizing a spectral library based on sensitivity analysis, characterized in that, include: Obtain simulated spectra; Sensitivity analysis was performed on the preset library construction parameters based on simulated spectra to obtain the sensitivity of each parameter. This included calculating the sensitivity of each parameter using the partial derivative finite difference method, expressed as follows: ； ； Where F_λ represents the simulated spectrum at the λth wavelength, p_i represents the expected process value of the ith library construction parameter, Δp_i represents the parameter perturbation step size, N represents the wavelength index, and paraSense[i] represents the sensitivity of the ith library construction parameter. Obtain the value range of each database creation parameter, and based on the sensitivity and value range of each database creation parameter, obtain the partitioning weight of each database creation parameter; The number of grid divisions for each database construction parameter is calculated based on the division weight of each database construction parameter. Grid points are obtained based on the number of grid divisions for each database construction parameter. Each grid point corresponds to a set of geometric parameters, thus obtaining the parameter grid. An optimized spectral library is generated by simulating the geometric parameters corresponding to each grid point in the parameter network using the Fourier modal method.

2. The spectral library optimization method based on sensitivity analysis as described in claim 1, characterized in that, The process of obtaining the parameter value range for each database creation parameter, and obtaining the partitioning weight for each database creation parameter based on its sensitivity and value range, specifically includes: Based on the database creation parameters, calculate the value range paraRange[i] for each database creation parameter, which represents the value range of the i-th database creation parameter; The sensitivity of each database creation parameter is multiplied by the range of values ​​for that parameter to obtain the partitioning weight for each parameter, which is specifically expressed as follows: ; In the formula, paraWeight[i] represents the partition weight of the i-th database creation parameter, paraSense[i] represents the sensitivity of the i-th database creation parameter, and paraRange[i] represents the parameter value range of the i-th database creation parameter.

3. The spectral library optimization method based on sensitivity analysis as described in claim 2, characterized in that, The calculation of the number of grid divisions for each database construction parameter based on the partitioning weight of each database construction parameter specifically includes: Normalize each database creation parameter; The maximum and minimum weight values, based on the database construction parameters, are normalized and represented as follows: ; Where normParaWeight[i] is the partition weight after normalization of the i-th database creation parameter, paraWeightMax represents the maximum value among all database creation parameter partition weights, and paraWeightMin represents the minimum value among all database creation parameter partition weights.

4. The spectral library optimization method based on sensitivity analysis as described in claim 3, characterized in that, The calculation of the number of grid divisions for each database construction parameter based on the partitioning weight of each database construction parameter specifically includes: The base value used for initial grid number allocation is calculated as follows: ； Where LibCountUser represents the number of preset spectral libraries, and m represents the number of library creation parameters; The number of grid divisions for each database construction parameter is calculated based on the normalized partition weights and base values, specifically as follows: ； Where Grid[i] represents the number of grid divisions for the i-th database creation parameter.

5. The spectral library optimization method based on sensitivity analysis as described in claim 2, characterized in that, The process of obtaining grid points based on the number of grid divisions for each database construction parameter, with each grid point corresponding to a set of geometric parameters, to obtain a parameter grid, specifically includes: Calculate the corresponding grid points based on the number of grid divisions for each parameter. Each grid point corresponds to a set of geometric parameters, represented as follows: ； In the formula, paraMin[i] represents the minimum value of the i-th database construction parameter, Grid[i] represents the number of grid divisions for the i-th database construction parameter, j represents any positive integer less than Grid[i], and m represents the number of database construction parameters.

6. A spectral library optimization system based on sensitivity analysis, characterized in that, include: The measurement result acquisition module is used to acquire simulated spectra. The sensitivity analysis module is used to perform sensitivity analysis on preset library construction parameters based on simulated spectra, obtaining the sensitivity of each library construction parameter. This includes calculating the sensitivity of each library construction parameter using the partial derivative finite difference method, expressed as: ； ； Where F_λ represents the simulated spectrum at the λth wavelength, p_i represents the expected process value of the ith library construction parameter, Δp_i represents the parameter perturbation step size, N represents the wavelength index, and paraSense[i] represents the sensitivity of the ith library construction parameter. The weighting module is used to obtain the parameter value range of each database creation parameter, and to obtain the partitioning weight of each database creation parameter based on the sensitivity of each database creation parameter and the parameter value range of each database creation parameter. The grid partitioning module is used to calculate the number of grid partitions for each database construction parameter based on the partitioning weight of each database construction parameter. The parameter mesh acquisition module is used to obtain grid points based on the number of grid divisions for each database construction parameter. Each grid point corresponds to a set of geometric parameters, thus obtaining the parameter mesh. The spectral library optimization module is used to simulate the geometric parameters corresponding to each grid point in the parameter network using the Fourier modal method, and generate an optimized spectral library.

7. The spectral library optimization system based on sensitivity analysis as described in claim 6, characterized in that, The weighting module is used to obtain the parameter value range of each database creation parameter, and based on the sensitivity and parameter value range of each database creation parameter, to obtain the partitioning weight of each database creation parameter, including: Based on the database creation parameters, calculate the value range paraRange[i] for each database creation parameter, which represents the value range of the i-th database creation parameter; The sensitivity of each database creation parameter is multiplied by the range of values ​​for that parameter to obtain the partitioning weight for each parameter, which is specifically expressed as follows: ; In the formula, paraWeight[i] represents the partition weight of the i-th database creation parameter, paraSense[i] represents the sensitivity of the i-th database creation parameter, and paraRange[i] represents the parameter value range of the i-th database creation parameter.

8. The spectral library optimization system based on sensitivity analysis as described in claim 6, characterized in that, The parameter mesh acquisition module is used to acquire mesh points according to the number of mesh divisions for each database construction parameter. Each mesh point corresponds to a set of geometric parameters, including: Calculate the corresponding grid points based on the number of grid divisions for each parameter. Each grid point corresponds to a set of geometric parameters, represented as follows: ； In the formula, paraMin[i] represents the minimum value of the i-th database construction parameter, Grid[i] represents the number of grid divisions for the i-th database construction parameter, j represents any positive integer less than Grid[i], and m represents the number of database construction parameters.

Images