Method for evaluating the quality of silicon wafers based on image features

Abstract

The present invention discloses a method for evaluating the quality of silicon wafers based on image features, which relates to the technical field of image processing, and solves the technical problems that it is difficult to extract features from surface images and quantify the features; secondly, it is difficult to select features based on feature parameters to generate feature vectors; then, it is difficult to analyze and identify silicon wafer defects based on feature vectors and difficult to detect silicon wafer deformation; then, it is difficult to analyze the dopant distribution in the ROI region and difficult to analyze the metal contamination of the silicon wafer; finally, it is difficult to comprehensively analyze the quality evaluation index of the silicon wafer based on various data. The present invention can identify the defects on the surface of the silicon wafer through image processing and feature extraction technologies, measure the deformation amount of the silicon wafer through deformation detection technologies; obtain the silicon wafer performance change rate by analyzing the dopant distribution in the ROI region, and obtain the metal contamination rate through metal contamination detection technologies; comprehensively evaluate the quality of the silicon wafer based on various data.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and specifically relates to a method for evaluating the quality of silicon wafers based on image features. Background Art

[0002] With the rapid development of image processing technology, the method for evaluating the quality of silicon wafers based on image features has gradually become a research hotspot. Silicon wafers are key materials for solar cells, semiconductor devices, etc., and their quality directly affects the performance and reliability of products. Therefore, accurate and rapid quality evaluation of silicon wafers is an important link to ensure product quality and production efficiency. Image processing technology can automatically extract the image features on the surface of silicon wafers and analyze and process these features through machine learning algorithms, so as to achieve accurate evaluation of the quality of silicon wafers.

[0003] The following problems exist in the prior art: First, it is difficult to extract features from the surface image and quantify the features; second, it is difficult to select features based on feature parameters to generate feature vectors; then, it is difficult to analyze and identify silicon wafer defects based on feature vectors and difficult to detect silicon wafer deformation; then, it is difficult to analyze the dopant distribution in the ROI region and difficult to analyze the metal contamination of silicon wafers; finally, it is difficult to comprehensively analyze the quality evaluation index of silicon wafers based on various data. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention provides a method for evaluating the quality of silicon wafers based on image features to solve the following technical problems:

[0005] First, it is difficult to extract features from the surface image and quantify the features; second, it is difficult to select features based on feature parameters to generate feature vectors; then, it is difficult to analyze and identify silicon wafer defects based on feature vectors and difficult to detect silicon wafer deformation; then, it is difficult to analyze the dopant distribution in the ROI region and difficult to analyze the metal contamination of silicon wafers; finally, it is difficult to comprehensively analyze the quality evaluation index of silicon wafers based on various data.

[0006] To solve the above problems, the first aspect of the present invention provides a method for evaluating the quality of silicon wafers based on image features, including the following steps:

[0007] S1: Collect the surface image and performance data of the silicon wafer after initial cleaning, where the performance data includes: silicon wafer thickness, carrier concentration and mobility, hole concentration and mobility, elementary charge and resistivity, and record the initial cleaning time length;

[0008] S2: Preprocess the collected surface image, including: filtering and denoising, image enhancement, and grayscale conversion; preprocess the collected performance data, including: cleaning and standardization;

[0009] S3: Extract features from the preprocessed surface image, quantify the features, and extract feature parameters; preprocess the extracted feature parameters, including: cleaning and standardization; perform feature selection based on the preprocessed feature parameters, and generate a feature vector according to the selection result;

[0010] S4: Analyze and identify wafer defects based on the feature vector, and calculate the defect value according to the analysis result; perform wafer deformation detection and analyze and calculate the deformation amount of the wafer according to the detection result; analyze the dopant distribution in the ROI area; analyze and calculate the change rate of wafer performance according to the dopant distribution result; analyze the metal contamination of the wafer and calculate the metal contamination rate according to the analysis result;

[0011] S5: Comprehensively analyze and evaluate the quality of the wafer based on the defect value, deformation amount, change rate of wafer performance, and metal contamination rate.

[0012] As a further solution of the present invention: In step S3, when extracting features from the preprocessed surface image, quantifying the features, and extracting feature parameters, the following steps are included:

[0013] Based on the preprocessed surface image, use an edge detection algorithm to extract the edge contour of the wafer; through morphological processing based on the edge contour of the wafer, extract the shape features of the surface image, and the feature parameters of the shape features include: the area within the edge contour area and the total length of the edge contour; through a feature detection algorithm, detect the ROI area from the preprocessed surface image, and the feature parameters extracted according to the detection result include: the coordinates of the four corners, the center point coordinates, the width, the height, and the area of the ROI area; extract texture features according to the gray-level co-occurrence matrix of the surface image, and the feature parameters of the texture features include: contrast, energy, homogeneity, entropy, and correlation.

[0014] As a further solution of the present invention: In step S3, when performing feature selection based on the preprocessed feature parameters and generating a feature vector according to the selection result, the following steps are included:

[0015] Generate a first feature vector according to the extracted feature parameters, and the first feature vector is expressed as ;

[0016] By analyzing the formula:

[0017]

[0018] Obtain the selected feature parameters ; where Y represents the selection threshold coefficient, represents the value of the i-th feature parameter in the first feature vector, and N represents the total number of feature parameters in the first feature vector;

[0019] Generate a second feature vector according to the selected characteristic parameters, and the second feature vector is expressed as , where M represents the total number of the selected characteristic parameters.

[0020] As a further solution of the present invention: in step S4, analyzing and identifying wafer defects according to the feature vector and calculating the defect value according to the analysis result includes the following steps:

[0021] According to the selected second feature vector, use a convolutional neural network model to analyze whether there are defects on the wafer and identify the defect types;

[0022] Label the second feature vector and the wafer defect types, where the wafer defect types include: cracks, scratches, dents, chipping, and bubbles; input the labeled second feature vector and the wafer defect types into the convolutional neural network model for training;

[0023] According to the trained convolutional neural network model, through the surface image of the wafer collected and preprocessed in real time, repeat step S3, input the feature vector extracted from the surface image of the wafer collected in real time into the trained convolutional neural network model for identification to determine whether there are defects on the wafer, and output the defect types;

[0024] When it is identified that there are defects on the wafer, the characteristic parameters of the defects extracted include: defect area and defect depth;

[0025] Through the analysis formula:

[0026] Obtain the defect value ; where represents the identified defect area, A represents the area within the edge contour region, represents the identified defect depth, and W represents the wafer thickness;

[0027] When it is identified that there are no defects on the wafer, perform wafer deformation detection.

[0028] As a further solution of the present invention: in step S4, performing wafer deformation detection and analyzing and calculating the deformation amount of the wafer according to the detection result includes the following steps:

[0029] When it is identified that there are no defects on the wafer, collect and preprocess the image of the undeformed standard wafer, extract the characteristic parameters of the image of the undeformed standard wafer and perform preprocessing of cleaning and standardization, and generate a standard feature vector through feature selection ;

[0030] According to the second feature vector and the standard feature vector, perform comparison and analysis, through the analysis formula:

[0031] Obtain the comparison similarity S; where, represents the second feature vector, represents the standard feature vector, represents the modulus of the second feature vector, represents the modulus of the standard feature vector;

[0032] Set a threshold according to the calculation result of the comparison similarity, and analyze and judge whether the silicon wafer is deformed; when , the silicon wafer is not deformed; when , the silicon wafer is deformed;

[0033] When the silicon wafer is deformed, through the analysis formula:

[0034] Obtain the deformation amount B; where, represents the i-th feature parameter of the second feature vector, represents the i-th feature parameter of the standard feature vector, and M represents the total number of selected feature parameters;

[0035] When the silicon wafer is not deformed, analyze the dopant distribution in the ROI region.

[0036] As a further solution of the present invention: Analyzing the dopant distribution in the ROI region in step S4 includes the following steps:

[0037] When the silicon wafer is not deformed, analyze the dopant distribution in the ROI region according to the extracted feature parameters of the ROI region using a deep learning model;

[0038] Label the feature parameters of the ROI region and the feature parameters of the ROI region in the silicon wafer surface image containing the dopant distribution; the dopant distribution includes: uniform distribution, local aggregation, and sparse distribution; input the labeled feature parameters of the ROI region and the feature parameters of the ROI region in the silicon wafer surface image containing the dopant distribution into the deep learning model for training;

[0039] According to the trained deep learning model, through the preprocessed silicon wafer surface image collected in real time, extract the feature parameters of the ROI region in the real-time silicon wafer surface image, input the extracted feature parameters of the ROI region into the trained deep learning model for analyzing and identifying the dopant distribution in the ROI region, and output the result of the dopant distribution in the ROI region.

[0040] As a further solution of the present invention: Analyzing and calculating the silicon wafer performance change rate according to the dopant distribution result includes the following steps:

[0041] Through the analysis formula:

[0042] Obtain the conductivity of the silicon wafer , where represents the carrier concentration, represents the elementary charge, represents the carrier mobility, represents the hole concentration, represents the hole mobility;

[0043] According to the characteristic parameters of the ROI region of the silicon wafer surface image collected in real time, when it is recognized that the dopant distribution in the ROI region is locally aggregated or sparsely distributed, through the analysis formula:

[0044] Obtain the silicon wafer performance change rate , where represents the conductivity of the silicon wafer when the dopant distribution in the ROI region is uniformly distributed, represents the conductivity of the silicon wafer when the dopant distribution in the ROI region is locally aggregated or sparsely distributed.

[0045] As a further solution of the present invention: In step S4, analyze the metal contamination of the silicon wafer and calculate the metal contamination rate according to the analysis result, including the following steps:

[0046] Obtain a silicon wafer surface image sample containing metal contamination and preprocess the silicon wafer surface image sample;

[0047] Extract the metal features in the preprocessed silicon wafer surface image sample, including: color features and texture features; the color features include: RGB value, HSV value and grayscale value; the texture features include: contrast, energy, homogeneity, entropy and correlation;

[0048] Label the metal features in the preprocessed silicon wafer surface image sample, and input the metal features in the silicon wafer surface image sample into a deep learning model for training;

[0049] According to the trained deep learning model, through the silicon wafer surface image collected and preprocessed in real time, extract the metal features in the real-time silicon wafer surface image, input the extracted metal features into the trained deep learning model for analyzing and identifying whether there is metal contamination in the real-time silicon wafer surface image, and output the identification result;

[0050] When it is recognized that there is metal contamination, through the analysis formula:

[0051] Obtain the metal contamination rate G; where A represents the area within the edge contour region, represents the area of the metal contamination region.

[0052] As a further solution of the present invention: Step S5 includes the following steps:

[0053] By analyzing the formula:

[0054]

[0055] Obtain the quality evaluation index Q; where D represents the defect value, B represents the deformation amount, R represents the change rate of the silicon wafer performance, G represents the metal contamination rate, 、 、 and respectively represent the weight coefficients of the defect value, the deformation amount, the change rate of the silicon wafer performance, and the metal contamination;

[0056] Set the threshold H. When the quality evaluation index is greater than or equal to the threshold H, the quality of the silicon wafer is unqualified; when the quality evaluation index is less than the threshold H, the quality of the silicon wafer is qualified;

[0057] Analyze whether the quality of the silicon wafer after initial cleaning is qualified through the surface image and performance data of the silicon wafer collected in real time after initial cleaning; when the quality of the silicon wafer is unqualified, clean the silicon wafer again, increase the time of the Etch etching step by 12 seconds and increase the time length of the Coat coating step to 60 seconds, and judge whether the quality of the silicon wafer after re - cleaning is qualified; when it is judged that the quality of the silicon wafer after re - cleaning is unqualified, re - clean and detect the silicon wafer.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] Through image processing and feature extraction technologies, the present invention can accurately identify and quantify the defects on the surface of the silicon wafer. Through deformation detection technology, it can accurately measure the deformation amount of the silicon wafer. By analyzing the dopant distribution and the change rate of the silicon wafer performance in the ROI region, it can evaluate the performance change of the silicon wafer in the doping and other process steps. Through metal contamination detection technology, it can accurately measure the metal contamination rate on the surface of the silicon wafer;

[0060] By performing feature selection on the extracted feature parameters, the present invention screens out the most critical features for silicon wafer quality evaluation, which can reduce the interference of redundant information on the evaluation results and improve the accuracy and reliability of the evaluation; by using a deep learning model to perform intelligent analysis on the surface image of the silicon wafer, this intelligent analysis method not only reduces the possibility of manual intervention and misjudgment, but also improves the efficiency and accuracy of the analysis, realizing the rapid and accurate evaluation of the silicon wafer quality;

[0061] By comprehensively analyzing the surface image and various performance data of the silicon wafer, the present invention comprehensively evaluates the quality of the silicon wafer, can more accurately reflect the overall quality status of the silicon wafer, and improves the accuracy of evaluating the quality of the silicon wafer; for the silicon wafers initially judged to be unqualified, the time of the Etch etching step is increased and the time of the Coat coating step is extended for re-cleaning, improving the quality of the silicon wafer. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0063] Figure 1 It is a flowchart of the system method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0065] Please refer to Figure 1 , an embodiment of the first aspect of the present invention provides a method for evaluating the quality of a silicon wafer based on image features, including the following steps:

[0066] S1: Collect the surface image and performance data of the silicon wafer after initial cleaning. The performance data includes: silicon wafer thickness, carrier concentration and mobility, hole concentration and mobility, elementary charge, and resistivity, and record the length of the initial cleaning time;

[0067] S2: Perform preprocessing on the collected surface image, including: filtering and denoising, image enhancement, and grayscale conversion; perform preprocessing on the collected performance data, including: cleaning and standardization;

[0068] S3: Extract features and quantify features from the preprocessed surface image, and extract feature parameters; perform preprocessing on the extracted feature parameters, including: cleaning and standardization; perform feature selection based on the preprocessed feature parameters, and generate a feature vector according to the selection result;

[0069] S4: Identify the wafer defects based on the eigenvector analysis and calculate the defect value according to the analysis results; detect the wafer deformation and calculate the amount of wafer deformation according to the detection results; analyze the dopant distribution in the ROI region; analyze and calculate the change rate of the wafer performance according to the dopant distribution results; analyze the metal contamination of the wafer and calculate the metal contamination rate according to the analysis results.

[0070] S5: Comprehensively analyze and evaluate the quality of the wafer based on the defect value, the amount of deformation, the change rate of the wafer performance, and the metal contamination rate.

[0071] Specifically, obtain the surface image of the wafer after initial cleaning, and use the corresponding test instruments to measure the performance data of the wafer, such as thickness, carrier concentration, carrier mobility, and resistivity, according to the standard test methods. Use algorithms such as Gaussian filtering and mean filtering to remove the noise in the image, and improve the clarity of the image through contrast enhancement and sharpening algorithms. Convert the color image into a grayscale image to reduce the data volume and improve the processing speed. Remove the outliers or invalid data in the performance data, and convert the performance data into the same dimension to eliminate the dimension difference between different parameters. Extract the features from the preprocessed surface image, and clean and standardize the extracted feature parameters. Perform feature selection on the feature parameters through formulas and generate eigenvectors. Identify the defects on the wafer surface based on the eigenvector analysis; detect the deformation of the wafer. In the ROI region, use the image processing algorithm to extract the dopant distribution information; analyze and identify the metal contamination on the wafer surface. Perform comprehensive calculations on the various data to obtain a quality evaluation index, and evaluate the quality of the wafer according to the size of the quality evaluation index.

[0072] In one embodiment of the present invention, in step S3, the feature extraction and quantization are performed according to the preprocessed surface image, and the feature parameters are extracted, including the following steps:

[0073] According to the preprocessed surface image, use the edge detection algorithm to extract the edge contour of the wafer; through morphological processing based on the edge contour of the wafer, extract the shape features of the surface image, and the feature parameters of the shape features include: the area within the edge contour region and the total length of the edge contour; through the feature detection algorithm, detect the ROI region from the preprocessed surface image, and extract the feature parameters of the ROI region according to the detection results, including: the coordinates of the four corners, the center point coordinates, the width, the height, and the area of the ROI region; extract the texture features according to the gray-level co-occurrence matrix of the surface image, and the feature parameters of the texture features include: contrast, energy, homogeneity, entropy, and correlation.

[0074] Specifically, edge detection algorithms such as Canny, Sobel, and Laplacian are used to apply the edge detection algorithm to the preprocessed image, extract the wafer edge in the detected wafer surface image, and form the edge contour of the wafer. The edge contour is processed by morphological operations such as erosion, dilation, opening operation, and closing operation. These operations can remove small miscellaneous points on the edge contour, fill holes, smooth the edge, and extract the shape features of the surface image. The feature parameters of the shape features include: the area within the edge contour region and the total length of the edge contour. Algorithms such as color threshold, shape matching, and template matching are used to apply the ROI detection algorithm to locate the region of interest. The four corner coordinates, center point coordinates, width, height, and area of the ROI region are extracted. The four corner coordinates can be obtained from the vertices of the detected ROI bounding box. The center point coordinates can be obtained by calculating the centroid of the ROI region. The width and height can be directly obtained from the dimensions of the ROI bounding box. The area can be obtained by calculating the number of pixels within the ROI region. The gray-level co-occurrence matrix of the preprocessed image is calculated, and texture feature parameters such as contrast, energy, homogeneity, entropy, and correlation are extracted from the gray-level co-occurrence matrix. These feature parameters can be obtained by calculating the statistical characteristics of the gray-level co-occurrence matrix. The contrast is obtained by calculating the variance of the element values in the gray-level co-occurrence matrix. The energy is the sum of the squares of the element values in the gray-level co-occurrence matrix. The homogeneity is obtained by calculating the reciprocal of the sum of the squares of the differences between the element values in the gray-level co-occurrence matrix and the mean values of their corresponding rows or columns. The entropy is the sum of the negative logarithms of the probability distributions of the element values in the gray-level co-occurrence matrix. The correlation is obtained by calculating the correlation coefficient between the element values in the gray-level co-occurrence matrix and the mean values of their corresponding rows or columns.

[0075] In one embodiment of the present invention, in step S3, feature selection is performed according to the preprocessed feature parameters, and a feature vector is generated according to the selection result, including the following steps:

[0076] Generate a first feature vector according to the extracted feature parameters, and the first feature vector is expressed as ;

[0077] By analyzing the formula:

[0078]

[0079] Obtain the selected feature parameters ; where Y represents the selection threshold coefficient, represents the value of the i-th feature parameter in the first feature vector, and N represents the total number of feature parameters in the first feature vector;

[0080] Generate a second feature vector according to the selected feature parameters, and the second feature vector is expressed as , where M represents the total number of selected feature parameters.

[0081] Specifically, according to the feature extraction of edge detection, morphological processing, ROI region detection, and gray-level co-occurrence matrix texture, the feature parameters extracted from the silicon wafer surface image are arranged in order to form a feature vector. According to the analysis requirements, the selection threshold coefficient Y is set to 3. This coefficient is used to screen important feature parameters and is usually dynamically adjusted based on factors such as the significance, correlation, or contribution degree of the feature parameters. According to the results of the formula, the feature parameters that meet the conditions are screened out. These screened feature parameters constitute the second feature vector.

[0082] In one embodiment of the present invention, the step S4 of analyzing the feature vector to identify silicon wafer defects and calculating the defect value according to the analysis results includes the following steps:

[0083] According to the selected second feature vector, use a convolutional neural network model to analyze whether there are defects on the silicon wafer and identify the defect type;

[0084] Label the second feature vector and the silicon wafer defect type, where the silicon wafer defect type includes: cracks, scratches, dents, chipping, and bubbles; input the labeled second feature vector and the silicon wafer defect type into the convolutional neural network model for training;

[0085] According to the trained convolutional neural network model, through the preprocessed silicon wafer surface image collected in real time, repeat step S3, input the feature vector extracted from the silicon wafer surface image collected in real time into the trained convolutional neural network model for identification to determine whether there are defects on the silicon wafer, and output the defect type;

[0086] When it is identified that there are defects on the silicon wafer, the feature parameters of the defects extracted include: defect area and defect depth;

[0087] By analyzing the formula:

[0088] Obtain the defect value ; where, represents the identified defect area, A represents the area within the edge contour region, represents the identified defect depth, and W represents the silicon wafer thickness;

[0089] When it is identified that there are no defects on the silicon wafer, perform silicon wafer deformation detection.

[0090] Specifically, based on the second feature vector extracted from the silicon wafer surface image, the defect types of the silicon wafer are labeled. The defect types of the silicon wafer include, but are not limited to: cracks, scratches, dents, chipped edges, and bubbles. Create a data set, where each sample contains the second feature vector and its corresponding defect type label. Divide the data set into a training set, a validation set, and a test set for model training, validation, and testing. Build a convolutional neural network model and train the convolutional neural network model using the training set data. During the training process, adjust the model parameters through optimization methods such as backpropagation algorithm and gradient descent to minimize the loss function. Monitor metrics such as the loss value and accuracy during the training process to evaluate the training effect of the model. Use the validation set data to validate the trained model. Adjust the model structure or hyperparameters according to the validation results to improve the generalization ability of the model. Use an image acquisition device to collect silicon wafer surface images in real time, and preprocess the collected images, including denoising and enhancing contrast. Repeat the steps of feature parameter extraction and feature selection for the preprocessed silicon wafer surface images, extract the second feature vector of the real-time image, and input the extracted second feature vector into the trained convolutional neural network model for recognition. The model will output whether the silicon wafer has defects and the types of defects. When it is identified that the silicon wafer has defects, further extract the feature parameters of the defects, including the defect area and the defect depth. These parameters can be obtained through image processing techniques and geometric measurement methods. According to the extracted defect feature parameters and the relevant parameters of the silicon wafer, use an analysis formula to calculate the defect value. The defect value can be used to evaluate the severity and influence range of the defect. According to the actual application effect and feedback, continuously optimize and update the convolutional neural network model. More data can be collected to enhance the generalization ability of the model, or the model structure can be adjusted to improve the recognition accuracy and efficiency.

[0091] In one embodiment of the present invention, in step S4, the silicon wafer deformation is detected and the deformation amount of the silicon wafer is analyzed and calculated according to the detection result, including the following steps:

[0092] When it is identified that the silicon wafer has no defects, collect and preprocess the image of the undeformed standard silicon wafer, extract the feature parameters of the undeformed standard silicon wafer image, and perform preprocessing of cleaning and standardization, and generate a standard feature vector through feature selection ;

[0093] According to the second feature vector and the standard feature vector, perform comparison and analysis through the analysis formula:

[0094] Obtain the comparison similarity S; where, represents the second feature vector, represents the standard feature vector, represents the modulus of the second feature vector, represents the modulus of the standard feature vector;

[0095] Set a threshold according to the calculation result of the comparison similarity, and analyze and judge whether the silicon wafer is deformed; when it is the case, the silicon wafer is not deformed; when it is the case, the silicon wafer is deformed;

[0096] When the silicon wafer is deformed, by analyzing the formula:

[0097] the deformation amount B is obtained; where represents the i-th characteristic parameter of the second characteristic vector, represents the i-th characteristic parameter of the standard characteristic vector, and M represents the total number of selected characteristic parameters;

[0098] When the silicon wafer is not deformed, analyze the dopant distribution in the ROI region.

[0099] Specifically, when it is identified that there are no defects on the silicon wafer, an image of an undeformed standard silicon wafer is collected. The collected standard silicon wafer image is preprocessed, including steps such as denoising and enhancing the contrast, to ensure the image quality. For the preprocessed standard silicon wafer image, the feature extraction step is repeated to extract feature parameters. The extracted feature parameters are cleaned to remove invalid or redundant information. The feature parameters are standardized to ensure the numerical comparability of different feature parameters. Through the feature selection method, the most representative features are selected from the cleaned and standardized feature parameters to generate a standard feature vector. For the surface image of the silicon wafer collected in real time and preprocessed, a second feature vector is extracted. According to the second feature vector and the standard feature vector, a comparison analysis is performed. The comparison similarity is calculated using the analysis formula. According to the calculation result of the comparison similarity, the threshold is set to , and the threshold is dynamically adjusted according to the actual situation. When the comparison similarity is greater than or equal to the threshold, it is considered that the silicon wafer is not deformed; when the comparison similarity is less than the threshold, it is considered that the silicon wafer is deformed. When the silicon wafer is deformed, the deformation amount is calculated using the analysis formula. When the silicon wafer is not deformed, the dopant distribution in the ROI region is further analyzed.

[0100] In one embodiment of the present invention, the analysis of the dopant distribution in the ROI region in step S4 includes the following steps:

[0101] When the silicon wafer is not deformed, analyze the dopant distribution in the ROI region using a deep learning model based on the extracted feature parameters of the ROI region;

[0102] The characteristic parameters of the ROI region and the characteristic parameters of the ROI region in the silicon wafer surface image containing the dopant distribution. The dopant distribution includes: uniform distribution, local aggregation, and sparse distribution. Input the characteristic parameters of the labeled ROI region and the characteristic parameters of the ROI region in the silicon wafer surface image containing the dopant distribution into a deep learning model for training.

[0103] According to the trained deep learning model, through the preprocessed silicon wafer surface image collected in real time, extract the characteristic parameters of the ROI region in the real-time silicon wafer surface image. Input the extracted characteristic parameters of the ROI region into the trained deep learning model to analyze and identify the dopant distribution in the ROI region, and output the result of the dopant distribution in the ROI region.

[0104] Specifically, collect the image of the silicon wafer surface to ensure that the image is clear, without blur or distortion for subsequent processing and analysis. From the collected silicon wafer surface image, extract the region of interest ROI. The ROI region should cover the part of the silicon wafer that may contain dopants. For the extracted ROI region, label its characteristic parameters, which specifically depend on the manifestation form of the dopant on the silicon wafer. According to professional knowledge or actual observation, label the ROI region in the silicon wafer surface image containing the dopant distribution. The dopant distribution includes but is not limited to: uniform distribution, local aggregation, and sparse distribution. Combine the labeled ROI region characteristic parameters and the dopant distribution into a data set. The data set should contain multiple samples, and each sample includes the characteristic parameters of an ROI region and the corresponding dopant distribution label. Select a deep learning model according to the complexity of the problem and the scale of the data set. Use the labeled data set to train the deep learning model. During the training process, adjust the model parameters through optimization methods such as backpropagation algorithm and gradient descent to minimize the loss function. Monitor indicators such as the loss value and accuracy during the training process to evaluate the training effect of the model. Use the validation set data to validate the trained model. Adjust the model structure or hyperparameters according to the validation results to improve the generalization ability of the model. Collect and preprocess the silicon wafer surface image in real time. From the preprocessed silicon wafer surface image, extract the characteristic parameters of the ROI region. Input the extracted characteristic parameters of the ROI region into the trained deep learning model for analysis. In practical applications, collect the feedback data from users, including but not limited to: indicators such as the accuracy of the result, false alarm rate, and missed alarm rate. According to the feedback data, optimize and improve the deep learning model. The model structure, hyperparameters can be adjusted or more training data can be added to improve the performance of the model. Replace the current used model with the optimized model. Ensure that the new model has better performance and accuracy in practical applications.

[0105] In one embodiment of the present invention, in step S4, analyzing and calculating the change rate of the silicon wafer performance according to the dopant distribution result includes the following steps:

[0106] By analyzing the formula:

[0107] the conductivity of the silicon wafer is obtained , where represents the carrier concentration, represents the elementary charge, represents the carrier mobility, represents the hole concentration, represents the hole mobility;

[0108] According to the characteristic parameters of the ROI region of the silicon wafer surface image collected in real time, when it is identified that the dopant distribution in the ROI region is locally aggregated or sparsely distributed, by analyzing the formula:

[0109] the change rate of the silicon wafer performance is obtained , where represents the conductivity of the silicon wafer when the dopant distribution in the ROI region is uniformly distributed, represents the conductivity of the silicon wafer when the dopant distribution in the ROI region is locally aggregated or sparsely distributed.

[0110] Specifically, through the ROI region of the silicon wafer surface image collected in real time, the carrier concentration, hole concentration, elementary charge, carrier mobility, and hole mobility of the ROI region are obtained. Substitute the collected data into the above formula to calculate the conductivity of the silicon wafer. First, it is necessary to determine the conductivity of the silicon wafer when the dopant distribution in the ROI region is uniformly distributed. This can be obtained by testing a silicon wafer with uniformly distributed dopants under the same conditions. Through the characteristic parameters of the ROI region of the silicon wafer surface image collected in real time, using a deep learning model or other image recognition technologies, it is identified whether the dopant distribution in the ROI region is locally aggregated or sparsely distributed. After identifying the dopant distribution, use the above formula again to calculate the conductivity of the current silicon wafer, that is, when the dopant is locally aggregated or sparsely distributed. According to the calculated conductivity and change rate of the silicon wafer performance, analyze the influence of the dopant distribution on the silicon wafer performance. If the change rate is large, it indicates that the dopant distribution has a significant impact on the silicon wafer performance, and the doping process needs to be further optimized.

[0111] In one embodiment of the present invention, in step S4, analyzing the metal contamination of the silicon wafer and calculating the metal contamination rate according to the analysis result includes the following steps:

[0112] Obtain a silicon wafer surface image sample containing metal contamination and preprocess the silicon wafer surface image sample;

[0113] Extract the metal features in the preprocessed silicon wafer surface image sample, including: color features and texture features; the color features include: RGB values, HSV values, and grayscale values; the texture features include: contrast, energy, homogeneity, entropy, and correlation;

[0114] Label the metal features in the preprocessed silicon wafer surface image sample, and input the metal features in the silicon wafer surface image sample into a deep learning model for training;

[0115] According to the trained deep learning model, extract the metal features in the real-time silicon wafer surface image through the real-time collected and preprocessed silicon wafer surface image, input the extracted metal features into the trained deep learning model to analyze and identify whether there is metal contamination in the real-time silicon wafer surface image, and output the identification result;

[0116] When it is identified that there is metal contamination, through the analysis formula:

[0117] Obtain the metal contamination rate G; where A represents the area within the edge contour region, represents the area of the metal contamination region.

[0118] Specifically, obtain surface image samples of silicon wafers containing metal contamination through detection equipment in the laboratory or on the production line, ensuring that the samples are representative and can cover different types of metal contamination. Denoise the images to eliminate noise and interference in the images. Perform image enhancement to improve the contrast and clarity of the images, making the metal contamination areas more obvious. Perform size normalization on the images to ensure that all images have the same size and resolution. Use image processing software or libraries to extract the RGB values, HSV values, and grayscale values of the images. Use the gray-level co-occurrence matrix of the texture analysis algorithm to extract texture features such as contrast, energy, homogeneity, entropy, and correlation of the images. Manually annotate the metal features in the preprocessed silicon wafer surface image samples to ensure the accuracy and integrity of the annotation. Divide the annotated image samples into a training set and a test set for training a deep learning model. Input the annotated training set image samples into the deep learning model for training and optimization. By adjusting hyperparameters such as model parameters and learning rates, improve the recognition accuracy and generalization ability of the model. Use real-time acquisition of silicon wafer surface images, preprocess the acquired images, and keep them consistent with those during training. Use the trained deep learning model to extract metal features in the real-time silicon wafer surface images. Compare and analyze the extracted metal features with those extracted during training. Judge whether there is metal contamination in the real-time silicon wafer surface images according to the output results of the deep learning model. Use the edge detection algorithm of the image processing algorithm to determine the edge contour area of the silicon wafer surface, segment and measure the identified metal contamination areas, calculate the area of the metal contamination areas, and obtain the metal contamination rate through a formula.

[0119] In one embodiment of the present invention, step S5 includes the following steps:

[0120] By analyzing the formula:

[0121]

[0122] Obtain the quality evaluation index Q; where D represents the defect value, B represents the deformation amount, R represents the change rate of the silicon wafer performance, G represents the metal contamination rate, , , and respectively represent the weight coefficients of the defect value, deformation amount, change rate of the silicon wafer performance, and metal contamination;

[0123] Set a threshold H. When the quality evaluation index is greater than or equal to the threshold H, the quality of the silicon wafer is unqualified; when the quality evaluation index is less than the threshold H, the quality of the silicon wafer is qualified;

[0124] Analyze whether the quality of the initially cleaned silicon wafer is qualified based on the surface image and performance data of the silicon wafer collected in real time after initial cleaning; when the quality of the silicon wafer is unqualified, clean the silicon wafer again by increasing the time of the Etch etching step by 12 seconds and increasing the time length of the Coat coating step to 60 seconds, and determine whether the quality of the silicon wafer after re-cleaning is qualified; when it is determined that the quality of the silicon wafer after re-cleaning is unqualified, re-clean and detect the silicon wafer.

[0125] Specifically, according to the calculation results of the defect value, deformation amount, silicon wafer performance change rate, and metal contamination, the quality evaluation index is obtained through an analysis formula. Among them, the weight coefficient of the defect value is 0.3, the weight coefficient of the deformation amount is 0.2, the weight coefficient of the silicon wafer performance change rate is 0.2, and the weight coefficient of metal contamination is 0.3, ensuring that the sum of the weight coefficients is 1 to reflect the relative importance of each parameter in quality evaluation. According to the calculation result of the quality evaluation index, a threshold of 55% is set to evaluate whether the quality of the silicon wafer is qualified. The weight coefficients and thresholds are dynamically set according to the quality requirements of the silicon wafer and the actual application scenario. Based on the surface image and performance data of the silicon wafer collected in real time after initial cleaning, the quality evaluation index after initial cleaning can be calculated. Judge according to the quality evaluation index and the threshold after initial cleaning. When the quality of the silicon wafer is unqualified, re-cleaning is required, including: increasing the time of the Etch etching step by 12 seconds and increasing the time length of the Coat coating step to 60 seconds, and then calculating the quality evaluation index again. When it is determined that the quality of the silicon wafer after re-cleaning is unqualified, re-clean and detect the silicon wafer.

[0126] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for evaluating the quality of silicon wafers based on image features, characterized in that, Including the following steps: S1: Collect the surface images and performance data of the silicon wafers after initial cleaning. The performance data includes: silicon wafer thickness, carrier concentration and mobility, hole concentration and mobility, elementary charge, and resistivity. Record the length of the initial cleaning time; S2: Preprocess the collected surface images, including: filtering and denoising, image enhancement, and grayscale conversion; preprocess the collected performance data, including: cleaning and standardization; S3: Extract features from the preprocessed surface images and quantify the features to extract feature parameters; preprocess the extracted feature parameters, including: cleaning and standardization; perform feature selection based on the preprocessed feature parameters, and generate a feature vector according to the selection result; S4: Analyze and identify silicon wafer defects based on the feature vector and calculate the defect value according to the analysis result; perform silicon wafer deformation detection and analyze and calculate the deformation amount of the silicon wafer according to the detection result; analyze the dopant distribution in the ROI region; analyze and calculate the silicon wafer performance change rate according to the dopant distribution result; analyze the metal contamination of the silicon wafer and calculate the metal contamination rate according to the analysis result; S5: Comprehensively analyze and evaluate the quality of the silicon wafers based on the defect value, deformation amount, silicon wafer performance change rate, and metal contamination rate; Among them, in step S3, performing feature selection based on the preprocessed feature parameters and generating a feature vector according to the selection result includes the following steps: Generate a first feature vector according to the extracted feature parameters, and the first feature vector is expressed as ; By analyzing the formula: Obtain the selected characteristic parameters ; where Y represents the selection threshold coefficient, represents the value of the i-th characteristic parameter in the first eigenvector, and N represents the total number of characteristic parameters in the first eigenvector; According to the selected characteristic parameters, a second characteristic vector is generated, and the second characteristic vector is expressed as , where M represents the total number of the selected characteristic parameters; Among them, step S5 includes the following steps: By analyzing the formula: Obtain the quality evaluation index Q; where D represents the defect value, B represents the deformation amount, R represents the change rate of the silicon wafer performance, G represents the metal pollution rate, , , and respectively represent the weight coefficients of the defect value, the deformation amount, the change rate of the silicon wafer performance, and the metal pollution; Set a threshold H. When the quality evaluation index is greater than or equal to the threshold H, the quality of the silicon wafer is unqualified; when the quality evaluation index is less than the threshold H, the quality of the silicon wafer is qualified; Analyze whether the quality of the silicon wafers after initial cleaning is qualified through the real-time collected surface images and performance data of the silicon wafers after initial cleaning; when the quality of the silicon wafer is unqualified, perform secondary cleaning on the silicon wafer, increase the time of the Etch etching step by 12 seconds and increase the time length of the Coat coating step to 60 seconds, and determine whether the quality of the silicon wafer after secondary cleaning is qualified; when it is determined that the quality of the silicon wafer after secondary cleaning is unqualified, re-clean and detect the silicon wafer.

2. The method for evaluating the quality of silicon wafers based on image features according to claim 1, wherein, In step S3, extracting features from the preprocessed surface images and quantifying the features to extract feature parameters includes the following steps: Based on the preprocessed surface image, use an edge detection algorithm to extract the edge contour of the silicon wafer; through morphological processing based on the edge contour of the silicon wafer, extract the shape features of the surface image. The feature parameters of the shape features include: the area within the edge contour region and the total length of the edge contour; through a feature detection algorithm, detect the ROI region from the preprocessed surface image, and extract the feature parameters of the ROI region according to the detection result, including: the coordinates of the four corners of the ROI region, the center point coordinates, width, height, and area; extract texture features based on the gray-level co-occurrence matrix of the surface image. The feature parameters of the texture features include: contrast, energy, homogeneity, entropy, and correlation.

3. A method for evaluating the quality of silicon wafers based on image features according to claim 1, characterized in that, In step S4, analyzing and identifying silicon wafer defects based on the feature vector and calculating the defect value according to the analysis result includes the following steps: According to the selected second feature vector, use a convolutional neural network model to analyze whether there are defects on the silicon wafer and identify the defect types. Label the second feature vector and the silicon wafer defect types. The silicon wafer defect types include: cracks, scratches, dents, chipping, and bubbles. Input the labeled second feature vector and the silicon wafer defect types into the convolutional neural network model for training. According to the trained convolutional neural network model, through the preprocessed silicon wafer surface image collected in real time, repeat step S3. Input the feature vector extracted from the silicon wafer surface image collected in real time into the trained convolutional neural network model for identification to determine whether there are defects on the silicon wafer and output the defect types. When it is identified that there are defects on the silicon wafer, extract the characteristic parameters of the defects, including: defect area and defect depth. By analyzing the formula: Obtain the defect value ; where represents the recognized defect area, A represents the area within the edge contour region, represents the recognized defect depth, and W represents the thickness of the silicon wafer; When it is identified that there are no defects on the silicon wafer, perform silicon wafer deformation detection.

4. The method for evaluating the quality of silicon wafers based on image features according to claim 3, wherein, In step S4, the silicon wafer deformation detection is performed and the deformation amount of the silicon wafer is analyzed and calculated according to the detection result, including the following steps: When no defect is identified on the silicon wafer, collect and preprocess the image of the undeformed standard silicon wafer, extract the characteristic parameters of the image of the undeformed standard silicon wafer, and perform preprocessing of cleaning and standardization, and generate a standard feature vector through feature selection ; According to the second eigenvector and the standard eigenvector, perform comparison and analysis through the analysis formula: Obtain the comparison similarity S; where, represents the second feature vector, represents the standard feature vector, represents the modulus of the second feature vector, represents the modulus of the standard feature vector; Set a threshold according to the calculation result of the comparison similarity, and analyze and judge whether the silicon wafer is deformed; when the silicon wafer is not deformed; when the silicon wafer is deformed; When the silicon wafer is deformed, by analyzing the formula: Obtain the deformation amount B; among them, represents the i-th eigenparameter of the second eigenvector, represents the i-th eigenparameter of the standard eigenvector, and M represents the total number of selected eigenparameters; When the silicon wafer does not deform, analyze the dopant distribution in the ROI region.

5. The method for evaluating the quality of silicon wafers based on image features according to claim 4, wherein, In step S4, the analysis of the dopant distribution in the ROI region includes the following steps: When the silicon wafer does not deform, use a deep learning model to analyze the dopant distribution in the ROI region according to the extracted characteristic parameters of the ROI region. Label the characteristic parameters of the ROI region and the characteristic parameters of the ROI region in the silicon wafer surface image containing the dopant distribution. The dopant distribution includes: uniform distribution, local aggregation, and sparse distribution. Input the labeled characteristic parameters of the ROI region and the characteristic parameters of the ROI region in the silicon wafer surface image containing the dopant distribution into the deep learning model for training. According to the trained deep learning model, through the preprocessed silicon wafer surface image collected in real time, extract the characteristic parameters of the ROI region in the real-time silicon wafer surface image. Input the extracted characteristic parameters of the ROI region into the trained deep learning model for analyzing and identifying the dopant distribution in the ROI region and output the result of the dopant distribution in the ROI region.

6. The method for evaluating the quality of silicon wafers based on image features according to claim 5, wherein, In step S4, the analysis and calculation of the silicon wafer performance change rate according to the dopant distribution result includes the following steps: By analyzing the formula: Obtain the conductivity of the silicon wafer , where represents the carrier concentration, represents the elementary charge, represents the carrier mobility, represents the hole concentration, represents the hole mobility; According to the characteristic parameters of the ROI region of the silicon wafer surface image collected in real time, when it is recognized that the dopant distribution in the ROI region is locally aggregated or sparsely distributed, by analyzing the formula: Obtain the silicon wafer performance change rate , where represents the silicon wafer conductivity when the dopant distribution in the ROI region is uniformly distributed, represents the silicon wafer conductivity when the dopant distribution in the ROI region is locally aggregated or sparsely distributed.

7. A method for evaluating the quality of silicon wafers based on image features according to claim 1, characterized in that, In step S4, the analysis of the metal contamination of the silicon wafer and the calculation of the metal contamination rate according to the analysis result include the following steps: Obtain the silicon wafer surface image sample containing metal contamination and preprocess the silicon wafer surface image sample. Extract the metal features in the preprocessed silicon wafer surface image sample, including: color features and texture features. The color features include: RGB value, HSV value, and grayscale value. The texture features include: contrast, energy, homogeneity, entropy, and correlation. Label the metal features in the preprocessed silicon wafer surface image sample and input the metal features in the silicon wafer surface image sample into the deep learning model for training. According to the trained deep learning model, by using the surface image of the silicon wafer after real-time acquisition and preprocessing, extract the metal features in the real-time silicon wafer surface image, input the extracted metal features into the trained deep learning model to analyze and identify whether there is metal contamination in the real-time silicon wafer surface image, and output the recognition result; When metal contamination is identified, by analyzing the formula: Obtain the metal pollution rate G; where A represents the area within the edge contour region, represents the area of the metal pollution region.

Images