A method for establishing a surface quality identification model based on a scattering matrix and application thereof

Abstract

The application provides a surface quality recognition model based on a scattering matrix and an application thereof, and belongs to the field of optical detection.The method comprises the following steps: classifying the surface quality of a plurality of samples and obtaining the scattering light intensity and electric field intensity corresponding to each sample; performing normalization processing on the scattering light intensity of each sample to obtain the corresponding scattering matrix; performing visual processing on each scattering matrix according to the relationship between the scattering light intensity and the electric field intensity, thereby obtaining a plurality of updated scattering matrices; marking each updated scattering matrix according to the surface quality category of the sample, thereby obtaining a data set; and training a deep learning model using the data set to obtain a surface quality recognition model.The application can obtain the spatial distribution results of multiple incident angles and scattering angles, and can avoid the problem that the order of magnitude difference between the input data and the output data of the scattering light intensity is large, thereby causing a large prediction error, and finally obtain a surface quality recognition model with high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of optical detection, and more specifically, relates to a method for establishing a surface quality identification model based on a scattering matrix and its application. Background Technology

[0002] Optical components are widely used in equipment such as astronomical telescopes, laser-guided systems, and ultra-intense lasers and EUV (Extreme Ultraviolet) lithography equipment. These devices have extremely high requirements for the surface quality of optical components. However, rapidly characterizing the surface quality of large optical components during manufacturing is a challenging problem. Common inspection methods such as white light interferometers and atomic force microscopes are all offline inspections, which cannot effectively guide the manufacturing process, thus limiting processing efficiency and yield.

[0003] To accelerate the inspection efficiency of optical components, the surface quality of the processed optical components can be detected using light scattering methods. This method, employing a non-contact inspection approach, is easy to integrate, enables online inspection, and ensures processing efficiency and accuracy, showing promising application prospects. However, current technologies can only identify surface quality at a fixed incident angle. For full-incident-angle detection technology, the collected information is too abundant to be directly applied. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method for establishing a surface quality identification model based on the scattering matrix and its application, aiming to solve the problem that the existing identification methods cannot be applied to the full incident angle detection technology.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for establishing a surface quality identification model based on a scattering matrix is ​​provided, the method comprising the following steps:

[0006] S1 classifies the surface quality of several standard samples and obtains the scattered light intensity and electric field intensity of each standard sample.

[0007] S2 normalizes the scattered light intensity of each of the standard samples to obtain their corresponding scattering matrix.

[0008] S3 visualizes each of the scattering matrices based on the relationship between the scattered light intensity and the electric field intensity, thereby obtaining several updated scattering matrices.

[0009] S4 labels each of the updated scattering matrices according to the surface quality category of the standard sample to obtain the dataset;

[0010] S5 uses the dataset to train a deep learning model to obtain a surface quality recognition model.

[0011] As a further preferred embodiment, the method further includes a surface quality refinement process, specifically comprising the following steps:

[0012] S6. Establish several test samples, each of which has a different surface quality category from the standard sample. Then, obtain the scattered light intensity and light field intensity corresponding to a certain test sample, and use the current surface quality recognition model obtained in step S5 to identify it. Perform equidistant classification on the two surface quality categories with the highest confidence to obtain a new surface quality category.

[0013] S7 classifies the surface quality of each test sample according to the new surface quality category and repeats steps S1 to S5 to update the surface quality identification model to obtain the next surface quality identification model.

[0014] S8 updates the test sample and repeats steps S6 and S7 until the accuracy of the surface quality identification model reaches the threshold, thereby obtaining the final surface quality identification model.

[0015] As a further preferred embodiment, the surface quality is one of roughness, defects, or contaminants.

[0016] As a further preferred embodiment, in step S1, the intensity of scattered light and the electric field intensity are obtained by using finite element simulation.

[0017] As a further preferred embodiment, in step S2, normalization is performed using the following formula:

[0018]

[0019] In the formula, Let I be the normalized scattered light intensity corresponding to incident angle i and scattering angle j. ij Let I be the unnormalized scattered light intensity corresponding to incident angle i and scattering angle j. μ I represents the average intensity of the scattered light. σ The scattering matrix, representing the variance of the scattered light intensity, is established as follows:

[0020]

[0021] As a further preferred embodiment, in step S3, the following formula is used to calculate... and replace For visualization purposes:

[0022]

[0023] In the formula, E after normalization ij E ij Let E be the electric field intensity corresponding to the incident angle i and the scattering angle j. μ E is the mean value of the electric field strength.σ Let be the variance of the electric field strength.

[0024] According to another aspect of the present invention, a surface quality identification model obtained using the above-described method is provided.

[0025] According to another aspect of the present invention, a method for identifying the surface quality of optical components using the above-described surface quality identification model is provided, the method specifically comprising:

[0026] (1) Obtain the scattered light intensity and electric field intensity of the optical element under test, and then normalize and visualize them to obtain the updated scattering matrix of the optical element under test.

[0027] (2) Substitute the updated scattering matrix into the surface quality identification model for identification, thereby obtaining the surface quality of the optical element under test.

[0028] As a further preferred embodiment, the surface quality identification method further includes determining an optimal filter before step (1), specifically:

[0029] (a) Add different types of noise to the scattering matrix of a standard sample, remove each type of noise using multiple filters, and then perform normalization and visualization to obtain several denoised updated scattering matrices.

[0030] (b) Substitute each of the denoised scattering new emission matrices into the surface quality recognition model for recognition and compare it with the surface quality of the standard sample to obtain its corresponding accuracy. Select the filter with the highest accuracy corresponding to different noise to obtain the optimal filter corresponding to different noise.

[0031] (c) After obtaining the scattered light intensity and electric field intensity of the optical element under test, first determine the type of noise and process it using the corresponding optimal filter before proceeding to steps (1) and (2).

[0032] As a further preferred embodiment, in step (a), a Gaussian filter, a median filter, and a mean filter are used to remove noise for each category, respectively.

[0033] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0034] 1. This invention proposes to establish a dataset using a scattering matrix, which can obtain the spatial distribution results of multiple incident angles and scattering angles, enriching the roughness scattering information. The uniformization process can avoid the problem of large prediction errors caused by the large difference in the order of magnitude between the input and output data of scattered light intensity, which is convenient for subsequent deep learning. At the same time, through visualization processing, the scattered light intensity can be visualized, and the relationship between scattered light intensity and surface quality can be intuitively understood, ultimately obtaining a surface quality recognition model with high accuracy.

[0035] 2. In particular, the present invention first uses surface quality category standard samples of a large classification for deep learning, and then uses test samples to further refine the surface quality categories and train them. This can effectively reduce the problem of increased training costs caused by large amounts of data and useless roughness categories in the deep learning process, effectively improve training efficiency, and obtain more accurate roughness categories after training.

[0036] 3. Meanwhile, considering that noise during data acquisition can lead to a decrease in prediction accuracy, this invention proposes to use the established quality identification model to select the optimal filter corresponding to different noise levels, which can further improve the accuracy of quality identification. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the establishment of an optical element surface quality identification model based on a scattering matrix, as provided in an embodiment of the present invention.

[0038] Figure 2 This is the roughness scattering simulation model provided in the embodiments of the present invention;

[0039] Figure 3 These are the updated scattering matrices of different samples provided in the embodiments of the present invention, where (a) is a smooth plane and (b) is a plane with RMS0.1λ and corr2λ.

[0040] Figure 4 This is a flowchart of the deep learning process provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] like Figure 1 As shown, this invention provides a method for establishing a surface quality identification model based on a scattering matrix. This method includes the following steps:

[0043] S1 classifies the surface quality of several standard samples and obtains the scattered light intensity and electric field intensity of each standard sample.

[0044] Since the intensity of the scattered light collected by S2 approximates a Gaussian distribution, it can be normalized using zero-mean standardization. The corresponding formula is:

[0045]

[0046] In the formula, Let I be the normalized scattered light intensity corresponding to incident angle i and scattering angle j. ij Let I be the unnormalized scattered light intensity corresponding to incident angle i and scattering angle j. μ I represents the average intensity of the scattered light. σ The scattering matrix, representing the variance of the scattered light intensity, is established as follows:

[0047]

[0048] S3 To visualize the spatial distribution of the collected scattered light intensity, the scattering matrix needs to be visualized to analyze the scattered light intensity distribution of different surface qualities and different contours corresponding to the same surface quality. This is because the scattered light intensity I is related to |E(r)| 2 They are directly proportional, where E(r) is the electric field strength. It can be calculated using the relationship between scattered light intensity and electric field strength, using the following formula. and replace By visualizing each scattering matrix, several updated scattering matrices can be obtained:

[0049]

[0050] In the formula, E after normalization ij E ij Let E be the electric field intensity corresponding to the incident angle i and the scattering angle j. μ E is the mean value of the electric field strength. σ Let V be the variance of the electric field strength;

[0051] S4 labels each updated scattering matrix according to the surface quality category of the standard sample to obtain the dataset, and divides the labeled dataset into training set and test set;

[0052] S5 uses a dataset to train a deep learning model to obtain a surface quality recognition model.

[0053] This invention utilizes a scattering matrix to establish a dataset, enabling the acquisition of spatial distribution results with multiple incident and scattering angles. This enriches the roughness scattering information, solves the problem of decreased classification accuracy caused by spatial distribution images of light scattering, and avoids the influence of the 256 values ​​of image pixels on the classification results, thereby further improving classification accuracy. The uniformization process avoids the problem of excessive prediction errors caused by large differences in the magnitude of the input and output data of scattered light intensity, facilitating subsequent deep learning. Furthermore, visualization processing allows for the visualization of scattered light intensity, providing an intuitive understanding of the relationship between scattered light intensity and surface quality, ultimately resulting in a surface quality recognition model with high accuracy.

[0054] Furthermore, the method also includes a surface quality category refinement process, specifically comprising the following steps:

[0055] S6 establishes several test samples, each with a different surface quality category from the standard sample. Then, the scattered light intensity and light field intensity corresponding to a certain test sample are obtained, and the current surface quality identification model obtained in step S5 is used to identify it. The surface quality categories with the highest confidence are classified equally to obtain a new surface quality category.

[0056] S7 classifies the surface quality of each test sample according to the new surface quality category and repeats steps S1 to S5 to update the surface quality identification model to obtain the next surface quality identification model.

[0057] S8 updates the test sample and repeats steps S6 and S7 until the accuracy of the surface quality identification model reaches the threshold, thereby obtaining the final surface quality identification model.

[0058] This invention continuously refines the surface quality classification by performing equidistant classification on the top two surface quality categories with the highest confidence levels. At the same time, through training, it can further improve the accuracy of the surface quality recognition model, thereby reducing the problem of increased training costs caused by large amounts of useless roughness categories in the learning process and reducing the training time of deep learning.

[0059] Furthermore, surface quality can be one of roughness, defects, or contaminants. For example, during roughness measurement, the surface roughness profiles corresponding to the same roughness level can differ. To reduce the difficulty of preparing samples with different roughness profiles corresponding to the same roughness level, finite element simulation can be used to obtain the scattered light intensity and electric field intensity. When surface quality is roughness, it can be classified using (surface roughness) RMS and correlation length (corr), and several roughness levels can be obtained by arranging and combining RMS and corr.

[0060] Furthermore, to avoid information loss caused by using images as a dataset for deep learning training, we can analyze the information corresponding to the directly acquired electric field intensity E(r)|. 2 Zero-mean normalization is performed to avoid the influence of the collected scattering amplitude on model training. At the same time, normalization of the scattering amplitude can prevent gradient vanishing during training caused by excessively small scattering amplitude.

[0061] Furthermore, the ResNet50 network architecture was selected as the deep learning classification network model. The first convolutional layer of ResNet was set to have 1 input channel and a stride of 1. Image features were extracted using 7*7 convolutional kernels and repeated residual blocks were used for feature extraction. The Adam optimization algorithm was selected as the optimizer. During model training, the accuracy of the trained models was compared, and the training model with the highest accuracy was selected. The training effect of the model was verified using a test set, and the accuracy of the trained model was finally obtained.

[0062] According to another aspect of the present invention, a surface quality identification model established using the above method is provided.

[0063] According to another aspect of the present invention, a method for identifying the surface quality of optical components using the above-described surface quality identification model is provided, the method specifically comprising:

[0064] (1) Obtain the scattered light intensity and electric field intensity of the optical element under test, and then normalize and visualize them to obtain the updated scattering matrix of the optical element under test.

[0065] (2) Substitute the updated scattering matrix into the surface quality identification model for identification, thereby obtaining the surface quality of the optical element under test.

[0066] Furthermore, the surface quality identification method also includes determining the optimal filter before step (1), specifically:

[0067] (a) Add different types of noise to the scattering matrix of a standard sample, remove each type of noise using multiple filters, and then perform normalization and visualization to obtain several denoised updated scattering matrices.

[0068] (b) Substitute each denoised scattering update matrix into the surface quality recognition model for recognition and compare it with the surface quality of the standard sample to obtain its corresponding accuracy. Select the filter with the highest accuracy for different noise levels to obtain the optimal filter for different noise levels.

[0069] (c) After obtaining the scattered light intensity and electric field intensity of the optical element under test, first determine the type of noise and process it using the corresponding optimal filter before proceeding to steps (1) and (2).

[0070] Furthermore, in step (a), the added noise includes Gaussian noise, Poisson noise, and check noise, and the filters used include Gaussian filters, median filters, and mean filters.

[0071] The technical solution provided by the present invention will be further described below with reference to specific embodiments.

[0072] Step 1: Establishing a Defect Scattering Simulation Data Model

[0073] Roughness modeling involves two related parameters: the correlation length (CORR) and the surface roughness (RMS). The commonly used type of surface roughness is Gaussian roughness, and the corresponding surface power spectral density function is:

[0074]

[0075] To avoid the tedious and time-consuming process of preparing datasets through experiments, finite element simulation was used to expand the database, and the accuracy of the simulation data was verified experimentally. Parallel light was selected as the incident light, incident on the surface of the sample at different angles, and the distribution of scattered light intensity was monitored using a monitor. The established roughness scattering simulation model is as follows: Figure 2 As shown, the distribution of scattered light intensity corresponding to different incident angles (-80° to 80°, with an interval of 2°) is obtained for each scattering angle (-90° to 90° hemisphere).

[0076] Step 2: Expanding the Scattering Matrix Database

[0077] To obtain the scattering matrices corresponding to different roughness levels, nine different roughness levels were constructed. The selected RMS values ​​were 0.05λ, 0.1λ, and 0.2λ, and the selected corr values ​​were 0.5λ, 1.0λ, and 2.0λ, where λ is the incident light wavelength. By arranging and combining the RMS and corr values, nine roughness levels were obtained. Furthermore, the seed value in the roughness surface construction process was varied, resulting in 190 different rough surfaces corresponding to each roughness level.

[0078] Step 3: Preprocessing of the scattering matrix database

[0079] The light scattering matrix data is extracted from scattering angles between -80° and 80°, with an angle interval of 2°, corresponding to 81 data points. The incident angle range is (-80° to 80°), with an angle interval of 2°, also corresponding to 81 data points. An 81*81 scattering matrix was established, and n training data points were selected, resulting in an n*81*81 dataset. The training data was then normalized, with the corresponding light scattering amplitude range being [0,1]. To facilitate data import for network training, the pre-trained scattering matrix data was reshaped to obtain an n*6561 matrix. The selected batch size for data import was 64, and the dataset was divided into training and test sets in a ratio of 7.5:2.5. The data preprocessing workflow is as follows: Figure 4 As shown.

[0080] Step 4: Deep Learning Model Construction and Training

[0081] Deep learning was trained using the collected dataset, with the ResNet50 network architecture selected as the deep learning classification network model. The first convolutional layer of ResNet had 1 input channel, and image features were extracted using 7x7 convolutional kernels, with repeated residual blocks used for feature extraction. The learning rate for the scattering matrix dataset was set to 0.003, and the learning rate for single-incident-angle scattering was set to 0.002. The batch size for both datasets was 64, and the Adam optimizer was selected as the model optimizer.

[0082] Step 5: Scattering Matrix Noise Filtering

[0083] In this study, Gaussian noise, Poisson noise, and salt-and-pepper noise were added to the scattering matrix, and the noise was analyzed. Common Gaussian, median, and mean filters were used to remove noise present in the scattering matrix acquisition process, and a trained deep learning database was used for prediction. The corresponding filters showed good performance for Gaussian and Poisson noise, improving classification accuracy by up to 4%.

[0084] Step 6: Near-term roughness prediction

[0085] A roughness scattering matrix is ​​collected and classified using a trained surface quality recognition model. The confidence scores of the first two classifications are taken to determine the roughness categories nearest to the collected scattering matrix. To more accurately determine the roughness category, categories are selected at equal intervals between the first two categories, and deep learning is retrained. This method reduces the increase in training data volume caused by too many categories in the initial scattering matrix classification.

[0086] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for establishing a scattering matrix-based surface quality recognition model, characterized in that, The establishing method comprises the following steps: S1 classifying surface qualities of a plurality of samples and obtaining scattering light intensity and electric field intensity corresponding to each sample; S2 performing normalization processing on the scattering light intensity of each sample to obtain a corresponding scattering matrix, and performing normalization processing by using the following formula: wherein is the incident angle i、 is the scattering angle j is the corresponding unnormalized scattered light intensity, is the mean value of the scattered light intensity, is the variance of the scattered light intensity, the scattering matrix is established as: ； S3 visualizes each of the scattering matrices according to the relationship between the scattering light intensity and the electric field intensity, to obtain a plurality of updated scattering matrices, which are calculated using the following formula and replace to perform the visualizing processing: wherein is the processed visual representation , is the angle of incidence i、 is the angle of scattering j is the corresponding electric field strength, is the mean value of the electric field strength, is the variance of the electric field strength; S4 marking each updated scattering matrix according to the surface quality category of the sample to obtain a data set; S5 training a deep learning model by using the data set to obtain a surface quality recognition model.

2. The scatter matrix based surface quality identification model building method of claim 1, wherein, The establishing method further comprises a surface quality refinement process, specifically comprising the following steps: S6 establishing a plurality of test samples, each test sample having a different surface quality category from the samples, then obtaining scattering light intensity and electric field intensity corresponding to a test sample, and identifying the test sample by using the current surface quality recognition model obtained in step S5, and performing equidistant classification on the first two surface quality categories with respect to confidence to obtain a new surface quality category; S7 classifying the surface qualities of each test sample according to the new surface quality category, and repeating steps S1-S5 to update the surface quality recognition model to obtain a next surface quality recognition model; S8 updating the test samples and repeating steps S6 and S7 until the accuracy of the surface quality recognition model reaches a threshold, thereby obtaining a final surface quality recognition model.

3. The scatter matrix based surface quality identification model building method of claim 1, wherein, The surface quality is one of roughness, defects or contaminants.

4. The scatter matrix based surface quality identification model building method of claim 1, wherein, In step S1, the finite element simulation method is used to obtain the scattering light intensity and the electric field intensity.

5. A surface quality recognition model obtained by using the establishing method according to any one of claims 1-4.

6. A method for surface quality recognition of an optical element using the surface quality recognition model according to claim 5, characterized in that, The method specifically comprises: (1) obtaining scattering light intensity and electric field intensity of a to-be-tested optical element, then performing normalization processing and visualization processing to obtain an updated scattering matrix of the to-be-tested optical element; (2) substituting the updated scattering matrix into the surface quality recognition model for identification, thereby obtaining the surface quality of the to-be-tested optical element.

7. The surface quality recognition method according to claim 6, wherein The surface quality recognition method further comprises determining an optimal filter before step (1), specifically comprising: (a) adding different categories of noise to a scattering matrix of a sample, and using a plurality of filters to remove noise of each category, then performing normalization processing and visualization processing to obtain a plurality of denoised updated scattering matrices; (b) substituting each denoised updated scattering matrix into the surface quality recognition model for identification, and comparing the identification result with the surface quality of the sample to obtain an accuracy corresponding to each denoised updated scattering matrix, and selecting a filter corresponding to the highest accuracy of different noise, thereby obtaining an optimal filter corresponding to different noise; (c) after obtaining the scattering light intensity and the electric field intensity of the to-be-tested optical element, first determining the type of noise and then using the corresponding optimal filter to process, and then performing steps (1) and (2).

8. The surface quality recognition method of claim 7, wherein, In step (a), a Gaussian filter, a median filter and a mean filter are respectively used to remove noise of each category.

Images