Wafer defect detection data processing method and system

Abstract

The invention relates to the technical field of semiconductor manufacturing and detection, and discloses a wafer defect detection data processing method and system.The wafer defect detection data processing method comprises the steps that a structured random measurement matrix is constructed based on wafer surface characteristics, and compression measurement data is collected; identifying local manifold structures corresponding to different defect types by using a spectral clustering algorithm; constructing a multi-scale image Laplacian operator, and extracting defect geometric characteristics; performing differential reconstruction processing according to the defect type, and solving an optimization target containing a nuclear norm and a graph Laplacian regularization item; dynamically adjusting a sampling strategy based on reconstruction quality evaluation; according to the method, high-precision defect reconstruction under a low sampling rate is realized, the data processing efficiency is improved, the complex defect reconstruction quality is improved, the method adapts to diversified defect processing, the storage and transmission burden is reduced, and the computing resource allocation is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor manufacturing inspection, and more specifically, it relates to a method and system for processing wafer defect detection data. Background Art

[0002] In the process of semiconductor integrated circuit manufacturing, wafer surface defect detection is a key link to ensure product quality and improve the yield. With the continuous development of integrated circuit manufacturing technology, the chip feature size continues to shrink, and the requirements for the accuracy and efficiency of wafer surface defect detection are also constantly increasing.

[0003] The existing wafer defect detection technologies mainly have the following technical problems: Traditional detection methods need to perform high-density full-scan on the wafer surface, collect a large number of data points to ensure detection accuracy, resulting in time-consuming data acquisition and heavy processing burden; The existing reconstruction algorithms have poor effects when dealing with complex structure defects and cannot effectively distinguish and process different types of defects; Traditional methods usually assume that defect data is located on a single global manifold, ignoring the characteristic that different types of defects may be distributed on multiple local manifolds, resulting in uneven reconstruction quality and inability to adapt to the diversity of wafer defect morphologies.

[0004] With the continuous reduction of semiconductor process nodes and the increase of wafer size, traditional detection methods can no longer meet the needs of modern semiconductor manufacturing in terms of data processing efficiency and reconstruction quality, and it is necessary to develop efficient and accurate wafer defect detection data processing technologies. Summary of the Invention

[0005] The present invention provides a method and system for processing wafer defect detection data, which solves the technical problems of low data processing efficiency, low reconstruction quality of complex defects, and inability to adapt to defect diversity in related technologies.

[0006] The present invention provides a method for processing wafer defect detection data, including: constructing a structured random measurement matrix based on wafer surface characteristics, and collecting compressed measurement data of wafer surface defects; Using the spectral clustering algorithm to analyze the compressed measurement data, automatically identify the local manifold structures corresponding to different defect types, and establish the correspondence between defect types and local manifolds; Constructing a multi-scale graph Laplacian operator for each type of defect, and extracting the geometric characteristics of defects at different scales; Performing differential reconstruction processing according to the defect type, solving the optimization objective including the nuclear norm and the graph Laplacian regularization term for each type of defect, and obtaining the reconstruction result; Based on the quality evaluation of the reconstruction result, dynamically adjust the distribution strategy of sampling points, and increase the sampling density in the areas with poor reconstruction quality.

[0007] In a preferred embodiment, the structured random measurement matrix is constructed as follows: ; wherein, is the structured random measurement matrix, is a randomly selected matrix for selecting some sampling points from the full sampling space, is a Fourier transform matrix for converting the spatial domain signal to the frequency domain.

[0008] In a preferred embodiment, the steps of analyzing the compressed measurement data by using the spectral clustering algorithm include: Constructing a defect data affinity matrix , and the calculation formula for its elements is: ; wherein, represents the affinity between the th data point and the th data point, , are respectively the th and the th measurement data points, represents the square of the Euclidean distance between two data points, is the Gaussian kernel parameter, represents the exponential function; Calculating the Laplacian matrix: ; wherein, is the Laplacian matrix, is a diagonal matrix, is the defect data affinity matrix; Calculating the eigenvalues and eigenvectors of , and selecting the eigenvectors corresponding to the smallest non-zero eigenvalues to form an eigenmatrix; Normalizing each row in the eigenmatrix and applying the K-means clustering algorithm to obtain defect type clusters.

[0009] In a preferred embodiment, the construction of the multi-scale graph Laplacian operator includes: Constructing an adjacency graph corresponding to the defect type, wherein the adjacency relationship is determined based on a scale parameter; Calculating the weight matrix corresponding to the adjacency graph, wherein the weight value reflects the similarity between data points; Calculating the diagonal matrix: ; Among them, is the diagonal element of the th row and th column of the matrix, which is equal to the sum of all elements in the th row of the weight matrix . represents the similarity or connection strength between the th data point and the th data point, represents the sum of the connection strengths between the th data point and all other points; Calculate the normalized Laplacian matrix: ; Among them, represents the normalized Laplacian matrix of the nd type of defect at the th scale, is the identity matrix, is the weight matrix, is the degree matrix, which is a diagonal matrix.

[0010] In a preferred embodiment, the step of performing differential reconstruction processing according to the defect type adopts the augmented Lagrangian multiplier method, including: Introduce an auxiliary variable , and transform the optimization problem into: ; Construct the augmented Lagrangian function: ; Among them, represents the data matrix to be reconstructed for the st type of defect, is the introduced auxiliary variable, represents the compressed measurement data for the th type of defect, represents the measurement matrix for the th type of defect, is the nuclear norm regularization parameter, which is used to promote the low-rank property of the reconstruction result, is the graph Laplacian regularization parameter, is the weight coefficient for the th scale, is the graph Laplacian operator for the th type of defect at the th scale, represents the regularization term for the graph Laplacian operator, represents the augmented Lagrangian function, is the Lagrange multiplier, is the Lagrange multiplier term, is the penalty parameter, denotes the square of the Frobenius norm between and alternately optimize , and , until convergence.

[0011] In a preferred embodiment, the step of performing differential reconstruction processing according to the defect type includes a reconstruction quality evaluation function , and the calculation method is: ; wherein, denotes the quality evaluation function of the reconstructed data matrix , , , respectively denote the reconstruction quality of the -th, -th, -th data points, denotes the total number of data points; The calculation formula for the reconstruction quality of each data point is: ; wherein, denotes the reconstruction quality of the -th data point, is the reconstruction error, is the reconstruction uncertainty, is the weight coefficient, used to balance the influence of the reconstruction error and uncertainty.

[0012] In a preferred embodiment, the step of generating an adaptive sampling strategy based on the reconstruction quality in the quality evaluation step based on the reconstruction result includes: Dividing the defect area into multiple sub-areas: ; wherein, , , respectively denote the -th, -th, -th sub-areas, denotes the total number of sub-areas; Calculating the average reconstruction quality of each sub-area: ; wherein, denotes the The average reconstruction quality score of a sub-region, indicating the number of data points contained in the th sub-region, indicating the reconstruction quality score of the th data point, indicating the sum of the reconstruction quality scores of all data points within the th sub-region; Calculating the sampling priority according to the reconstruction quality: ; wherein, indicating the sampling priority of the th sub-region, being a monotonically decreasing function; Allocating sampling resources according to the sampling priority.

[0013] In a preferred embodiment, the process of implementing adaptive sampling in the quality assessment step based on the reconstruction result includes: Updating the measurement matrix according to the sampling strategy ; Collecting new compressed measurement data ; Combining with the original data for reconstruction: ; wherein, indicating the updated reconstruction result matrix, being the reconstruction function, indicating the newly collected compressed measurement data, indicating the updated measurement matrix, indicating the original reconstruction result; Iterating the above process in a loop until the reconstruction quality meets the preset threshold.

[0014] In a preferred embodiment, when solving the optimization problem in the step of performing differential reconstruction processing according to the defect type, different regularization parameters and are respectively set for different types of defects, where: For regular geometric shape defects, increase the weight of to maintain the low-rank property; For complex shape defects, increase the weight of to maintain the local manifold structure.

[0015] In a preferred embodiment, a wafer defect detection data processing system for performing a wafer defect detection data processing method includes: A compression measurement module, which is used to construct a structured random measurement matrix based on the characteristics of the wafer surface and collect the compression measurement data of the wafer surface defects; A defect classification module, which is used to analyze the compression measurement data by using the spectral clustering algorithm, automatically identify the local manifold structures corresponding to different defect types, and establish the correspondence between the defect types and the local manifolds; A feature extraction module, which is used to construct a multi-scale graph Laplacian operator for each type of defect and extract the defect geometric characteristics at different scales; An adaptive reconstruction module, which is used to perform differential reconstruction processing according to the defect type, solve the optimization objective containing the nuclear norm and the graph Laplacian regularization term for each type of defect, and obtain the reconstruction result; A sampling optimization module, which is used to dynamically adjust the distribution strategy of the sampling points based on the quality evaluation of the reconstruction result and increase the sampling density in the areas with poor reconstruction quality.

[0016] The beneficial effects of the present invention are as follows: The wafer defect detection data processing method, system and device provided by the present invention solve the technical problems of low data processing efficiency, poor reconstruction quality of complex defects and inability to adapt to defect diversity in the field of wafer defect detection, and provide a more efficient and accurate defect detection solution for the semiconductor manufacturing industry. Description of the Drawings

[0017] Figure 1 is a flowchart of a wafer defect detection data processing method of the present invention; Figure 2 is a detailed flowchart of collecting the compression measurement data of the wafer surface defects of the present invention; Figure 3 is a detailed flowchart of establishing the correspondence between the defect types and the local manifolds of the present invention; Figure 4 is a detailed flowchart of extracting the defect geometric characteristics at different scales of the present invention; Figure 5 is a detailed flowchart of obtaining the reconstruction result of the present invention; Figure 6 is a detailed flowchart of increasing the sampling density in the areas with poor reconstruction quality of the present invention. Detailed Embodiments

[0018] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and that changes can be made to the functions and arrangements of the elements discussed without departing from the scope of protection of the content of this specification. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0019] In at least one embodiment of the present invention, a method for processing wafer defect detection data is disclosed, as Figures 1 to 6 shown, which includes the following steps: Step 1, construct a structured random measurement matrix based on the surface characteristics of the wafer, and collect compressed measurement data of the wafer surface defects; The specific steps include: 1.1, construct a structured random measurement matrix; Construct a structured random measurement matrix based on the surface characteristics of the wafer , where represents the set of real numbers, is the number of sampling points, is the number of complete data points, and . This matrix satisfies the Restricted Isometry Property (RIP), ensuring accurate reconstruction of complete information at low sampling rates.

[0020] The construction method of the structured random measurement matrix is as follows: ; where is the structured random measurement matrix, is a random selection matrix for selecting some sampling points from the full sampling space; is the Fourier transform matrix for converting the spatial domain signal to the frequency domain. By combining the structured characteristics of the wafer surface, is optimized to make it more suitable for capturing the key information of the wafer surface defects.

[0021] 1.2, execute an optimized sampling strategy; By optimizing the sampling strategy, only about 30% of the key sampling points are selected for data acquisition, and the sampling process is expressed as: ; where are the measured values obtained by sampling, is the complete data of the wafer surface.

[0022] 1.3, obtain compressed measurement data; The measured values obtained by sampling are used as compressed measurement data for subsequent defect analysis and reconstruction processing. Compared with the traditional full-sampling method, this step significantly reduces the data acquisition time and storage space requirements.

[0023] Step 2: Use the spectral clustering algorithm to analyze the compressed measurement data, automatically identify the local manifold structures corresponding to different defect types, and establish the correspondence between defect types and local manifolds; The specific steps include: 2.1, Construct the defect data affinity matrix; Based on the compressed measurement data , construct the defect data affinity matrix , and its elements are calculated as follows: ; where represents the affinity between the -th data point and the -th data point, , are the -th and -th measurement data points respectively, represents the square of the Euclidean distance between two data points, is the Gaussian kernel parameter, represents the exponential function.

[0024] 2.2, Execute the spectral clustering algorithm to identify the local manifold structure; Based on the affinity matrix , perform the following steps for spectral clustering: Calculate the Laplacian matrix: ; where is the Laplacian matrix, is the diagonal matrix, is the defect data affinity matrix; where is the diagonal matrix: ; where represents the degree of the -th node, represents the affinity between the -th data point and the -th data point; Calculate 's eigenvalues and eigenvectors; Select the eigenvectors corresponding to the smallest non-zero eigenvalues to form the eigenmatrix ; Normalize each row in the feature matrix ; Apply the K-means clustering algorithm to the normalized matrix to obtain defect type clusters

[0025] Through this process, the defect data is automatically divided into categories, each category corresponding to a local manifold structure , thus forming a manifold set , where represents the set of all local manifolds, , , respectively represent the local manifold structures corresponding to the , , th class of defects, represents the preset number of defect types

[0026] 2.3, Establish the correspondence between defect types and local manifolds Associate each defect data point with its corresponding manifold category to establish a mapping relationship: ; where represents the mapping function, represents the th compressed measurement data point, represents the local manifold structure corresponding to the th class of defects This mapping relationship will be used for subsequent differentiation processing to ensure that different types of defects can be processed with the most suitable methods for their characteristics

[0027] Step 3, Construct a multi-scale graph Laplacian for each class of defects and extract the defect geometric characteristics at different scales The specific steps include 3.1, Construct a multi-scale graph Laplacian For each identified defect type (local manifold ), construct a set of multi-scale graph Laplacians: ; where represents the set of multi-scale graph Laplacians constructed for the th class of defects, , , respectively represent at , , The Laplacian operator constructed under the scale, where is the number of scales.

[0028] For the th scale of the th defect type, the calculation method of the Laplacian operator is as follows: Construct the adjacency graph of this type of defect data , where the adjacency relationship is determined based on the scale parameter ; Calculate the weight matrix corresponding to the adjacency graph , where the weight value reflects the similarity between data points; Calculate the diagonal matrix , where: ; where, is the diagonal element of the th row and th column of this matrix, equal to the sum of all elements in the th row of the weight matrix , represents the similarity or connection strength between the th data point and the th data point, represents the total connection strength of the th data point with all other points; Calculate the normalized Laplacian matrix: ; where, represents the normalized Laplacian matrix of the th type of defect under the th scale, is the identity matrix, is the weight matrix, is the degree matrix, which is a diagonal matrix.

[0029] 3.2, Extract multi-scale geometric features; Use the constructed multi-scale Laplacian operator to extract the geometric features of the defect data. For the th defect type, its feature extraction process is expressed as: ; where, represents the multi-scale feature set of the th type of defect, , , respectively represent the features extracted for the -th type of defect at the , , -th scale. represents the total number of selected scales, and its calculation method is: ; where represents the feature extracted for the -th type of defect at the -th scale, is the feature extraction function, which can be methods such as eigenvalue decomposition, spectral analysis, etc., represents the normalized Laplacian matrix for the -th type of defect at the -th scale, is the compressive measurement data belonging to the -th defect type.

[0030] 3.3 to form a multi-scale feature representation; Fuse the geometric features extracted at different scales to form a multi-scale feature representation for each defect type . These feature representations capture the structural information of the defects at different scales and provide a basis for subsequent differential reconstruction.

[0031] Step 4: Perform differential reconstruction processing according to the defect type. Solve the optimization objective containing the nuclear norm and the graph Laplacian regularization term for each type of defect to obtain the reconstruction result; The specific steps include: 4.1, Construct a classification adaptive reconstruction model; For the -th type of defect (corresponding to the manifold ), construct the following optimization objective: ; where is the compressive measurement data belonging to the -th type of defect; is the corresponding measurement matrix; is the complete defect data to be reconstructed; is the nuclear norm, which promotes the low-rank property of the solution; is the regularization term based on the graph Laplacian operator, which preserves the geometric properties of the defect at the -th scale; is the nuclear norm regularization parameter, is the graph Laplacian regularization parameter; is the The weight coefficients of each scale.

[0032] 4.2, solve the optimization problem; For each type of defect, the augmented Lagrangian multiplier method (ADMM) is used to solve the above optimization problem. The specific steps are as follows: Introduce auxiliary variables , and transform the original problem into: ; Construct the augmented Lagrangian function: ; Among them, represents the data matrix to be reconstructed for the th type of defect, is the introduced auxiliary variable, represents the compressed measurement data for the th type of defect, represents the measurement matrix for the th type of defect, is the nuclear norm regularization parameter, which is used to promote the low-rank property of the reconstruction result, is the graph Laplacian regularization parameter, is the weight coefficient of the th scale, is the th type of defect at the th scale of the graph Laplacian operator, represents the regularization term of the graph Laplacian operator, represents the augmented Lagrangian function, is the Lagrangian multiplier, is the Lagrangian multiplier term, is the penalty parameter, represents and the square of the Frobenius norm between; Alternately optimize , and until convergence.

[0033] 4.3, generate the reconstruction result; Take the obtained by the optimization solution as the reconstruction result of the th type of defect. Since the specific manifold structure and multi-scale characteristics of the defect type are considered, this reconstruction result can better retain the morphological details of the defect.

[0034] Merge the reconstruction results of all categories to obtain the complete reconstruction result of the wafer defect data , among which, Represents the reconstructed result of the complete wafer defect data after final merging, , , respectively represent the reconstructed results of the , , types of defects, represents the preset number of defect types.

[0035] Step 5, based on the quality assessment of the reconstructed result, dynamically adjust the distribution strategy of sampling points, and increase the sampling density in the areas with poor reconstruction quality; Specifically include: 5.1, Evaluate the reconstruction quality; For the reconstructed result , construct a reconstruction quality evaluation function: ; Among them, represents the quality evaluation function of the reconstructed data matrix , , , respectively represent the reconstruction quality of the , , th data points, represents the total number of data points; The evaluation function can be calculated using the following formula: ; Among them, represents the reconstruction quality evaluation value of the th data point, represents the actual measurement value of the th data point, represents the measurement matrix corresponding to the th data point, represents the reconstructed value of the th data point, is the scaling factor, is the characteristic variance of this type of defect.

[0036] 5.2, Generate an adaptive sampling strategy; Based on the reconstruction quality evaluation result, generate an optimized strategy for the next round of sampling: Calculate the sampling priority of each region , among which, represents the sampling priority of region , represents the mapping function from reconstruction quality to sampling priority, is the region Reconstruction quality of internal data points; Allocate sampling resources according to priority, and regions with higher priority obtain more sampling points; Generate a new structured random measurement matrix to make it have a higher sampling density in regions with lower reconstruction quality.

[0037] The priority calculation function can be defined as: ; where, represents the sampling priority of region , represents the number of data points in region , represents the average reconstruction quality of all data points in region , represents the reconstruction quality evaluation value of the -th data point in region .

[0038] 5.3, Implement adaptive sampling; Use the updated measurement matrix to perform a new round of data acquisition and obtain updated compressed measurement data . If necessary, repeat steps 1 to 5 until the reconstruction quality meets the preset threshold requirements.

[0039] Through this adaptive sampling strategy, the system can concentrate limited sampling resources on the most needed regions, further improving the overall reconstruction accuracy while maintaining a low overall sampling rate.

[0040] Technical effects of this embodiment: The wafer defect detection data processing method provided in this embodiment has the following technical effects: Significantly improve data processing efficiency: Through the structured random measurement matrix and compressed sensing technology, a reconstruction accuracy of 95% can be achieved with only 30% of the sampling rate, the detection speed is increased by 2 times, and the detection rate remains unchanged.

[0041] Improve the reconstruction quality of complex defects: Through the classification adaptive reconstruction strategy, different processing methods are adopted for different types of defects, and the reconstruction accuracy is increased by 15%. In particular, the reconstruction quality of complex-shaped defects is significantly improved.

[0042] Adapt to the processing of diverse defects: By automatically identifying and modeling multiple local manifold structures, the problem of uneven reconstruction quality for different defect types is solved, enabling the system to adapt to various complex-shaped wafer defects.

[0043] Reduce storage and transmission burdens: Through the low-rank reconstruction model, while maintaining the reconstruction accuracy, the data storage space is reduced by 75%, and the data transmission efficiency is increased by 3 times.

[0044] Optimize the allocation of computing resources: Combining with the adaptive sampling strategy, the system can concentrate computing resources on obtaining key information, reduce 90% of redundant calculations, and lower the overall processing cost.

[0045] This embodiment is particularly applicable to the high-precision wafer defect detection scenario, effectively solving the technical problems in processing efficiency and reconstruction quality of traditional methods, and providing a more efficient and accurate defect detection solution for the semiconductor manufacturing industry.

[0046] Real application examples of this embodiment: This example is applied to the surface defect detection process of 12-inch wafers. There are various types of defects on the wafer surface, including micro-particles (about 0.1 - 5μm), scratches (length 0.5 - 20μm, width 0.1 - 2μm), metal residues (size 0.2 - 10μm), and uneven oxide layer regions, etc. Traditional detection methods require high-density sampling of the entire wafer surface, with about 10,000 data points to be collected per square millimeter area, resulting in a detection time of 5 - 8 minutes for a single wafer and a data processing time of 10 - 15 minutes.

[0047] Implementation process of this method: Construction and application of the structured random measurement matrix: In this example, the structured random measurement matrix is constructed in the following way: First, according to the geometric characteristics and defect distribution rules of the wafer surface, the sampling point distribution of the randomly selected matrix is matched with the wafer structure; then it is combined with the Fourier transform matrix to form the final measurement matrix.

[0048] The construction parameters of the measurement matrix are shown in Table 1: Table 1: Parameter configuration of the structured random measurement matrix;

[0049] Sampling is carried out through this matrix to obtain compressed measurement data , and the data volume is only 30% of that of traditional methods.

[0050] Parameter configuration of the spectral clustering algorithm and defect classification results: Apply the spectral clustering algorithm to the collected compressed measurement data to identify the local manifold structure of the defects. The algorithm parameter configuration and results are shown in Table 2: Table 2: Parameters and results of the spectral clustering algorithm;

[0051] This algorithm successfully classifies the defect data into 5 categories, with each category corresponding to a local manifold structure , and the classification results are highly consistent with the defect types manually labeled

[0052] Construction and feature extraction of the multi-scale graph Laplacian operator: For each type of defect, 3 graph Laplacian operators with different scales are constructed, and the parameters are shown in Table 3: Table 3: Parameters of the multi-scale graph Laplacian operator;

[0053] The Laplacian operator of each scale is used to capture the geometric characteristics of defects at different spatial scales, and they are combined to form a multi-scale feature representation

[0054] Parameter setting and solution of the classification adaptive reconstruction: For different types of defects, the parameter settings of the classification adaptive reconstruction algorithm are shown in Table 4: Table 4: Parameter settings of the classification adaptive reconstruction;

[0055] Iterative parameter setting of the ADMM algorithm: Penalty parameter , maximum number of iterations 100, convergence threshold .

[0056] Parameters and effects of the adaptive sampling optimization: Parameter setting of the reconstruction quality evaluation function: Scaling factor .

[0057] A practical application case of the adaptive sampling strategy: After an initial 30% sampling of a certain wafer area, it is found through the reconstruction quality evaluation that the reconstruction quality of the linear scratch and the uneven annular oxidation areas is the lowest. The system automatically adjusts the sampling strategy, increases the sampling density of these two types of areas to 45%, and maintains the other areas at 25%. The overall sampling rate remains at about 30%

[0058] Verification of the technical effect: Improvement of data processing efficiency: The comparison of the data processing efficiency of this method in actual wafer detection with the traditional method is shown in Table 5: Table 5: Comparison of data processing efficiency;

[0059] Improvement of reconstruction quality: The comparison of the reconstruction quality of different types of defects is shown in Table 6: Table 6: Comparison of Defect Reconstruction Quality (Peak Signal-to-Noise Ratio PSNR, unit: dB);

[0060] As can be seen from Table 6, the proposed method has significantly improved the reconstruction quality of various types of defects, especially for linear scratches with complex shapes and uneven circular oxidation, which shows the most obvious improvement.

[0061] The above examples fully verify the effectiveness and practicality of this embodiment in actual wafer defect detection. It not only significantly improves the data processing efficiency but also improves the reconstruction quality of complex defects, which is particularly suitable for the requirements of high-efficiency and high-precision defect detection in modern semiconductor manufacturing.

[0062] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. A wafer defect detection data processing method, characterized in that: The following steps are involved: A structured random measurement matrix is ​​constructed based on the wafer surface characteristics to collect compressed measurement data of wafer surface defects; The spectral clustering algorithm is used to analyze the compressed measurement data, automatically identify the local manifold structure corresponding to different defect types, and establish the corresponding relationship between defect types and local manifolds; Construct a multi-scale graph Laplacian operator for each type of defect to extract the geometric characteristics of defects at different scales; Perform differentiated reconstruction processing according to defect types, solve the optimization objective including nuclear norm and graph Laplace regularization term for each type of defect, and obtain the reconstruction result; Based on the quality evaluation of the reconstruction results, the distribution strategy of the sampling points is dynamically adjusted to increase the sampling density in areas with poor reconstruction quality.

2. The method for processing wafer defect detection data according to claim 1, characterized in that: The structured random measurement matrix The construction method is: ； in, is the structured random measurement matrix, is a random selection matrix used to select some sampling points from the full sampling space. is the Fourier transform matrix, which is used to convert spatial domain signals into frequency domain.

3. The method for processing wafer defect detection data according to claim 1, characterized in that: The step of analyzing the compressed measurement data using the spectral clustering algorithm comprises: Constructing a Defect Data Affinity Matrix , the element calculation formula is: ； in, Indicates Data points and The affinity between data points, , Respectively and measurement data points, represents the square of the Euclidean distance between two data points, is the Gaussian kernel parameter, represents the exponential function; Compute the Laplacian matrix: ； in, is the Laplace matrix, is a diagonal matrix, is the defect data affinity matrix; calculate The eigenvalue and eigenvector of The eigenvectors corresponding to the non-zero eigenvalues ​​form the characteristic matrix; Normalize each row in the feature matrix and apply the K-means clustering algorithm to obtain defect type clusters.

4. The method for processing wafer defect detection data according to claim 1, characterized in that: The multiscale graph Laplacian The build includes: Construct an adjacency graph corresponding to the defect type, where the adjacency relationship is determined based on the scale parameter; Calculate the weight matrix corresponding to the adjacency graph, where the weight value reflects the similarity between data points; Compute the diagonal matrix: ； in, This matrix is Line The diagonal elements of the column are equal to the weight matrix No. The sum of all elements in a row, Indicates Data points and The similarity or connection strength between data points, Indicates the calculation The sum of the connection strengths between a data point and all other points; Compute the normalized Laplacian matrix: ； in, Indicates The defect in The normalized Laplacian matrix at each scale, is the identity matrix, is the weight matrix, is the degree matrix, which is a diagonal matrix.

5. The method for processing wafer defect detection data according to claim 1, characterized in that: The step of performing differential reconstruction processing according to the defect type adopts an augmented Lagrange multiplier method, including: Introducing auxiliary variables , transforming the optimization problem into: ； Construct the augmented Lagrangian function: ； in, Indicates The data matrix to be reconstructed of the defect class, is the auxiliary variable introduced, Indicates Compressed measurement data of class defects, Indicates The measurement matrix of class defects, is the nuclear norm regularization parameter, It is used to make the reconstruction result have low rank characteristics. is the graph Laplace regularization parameter, For the The weight coefficient of each scale, For the The defect in The graph Laplacian operator at each scale, is the regularization term of the graph Laplacian operator, represents the augmented Lagrangian function, is the Lagrange multiplier, is the Lagrange multiplier term, is the penalty parameter, express and The square of the Frobenius norm between ; Alternating Optimization , and , until convergence.

6. The method for processing wafer defect detection data according to claim 1, characterized in that: The step of performing differentiated reconstruction according to defect type includes a reconstruction quality evaluation function , the calculation method is: ； in, Represents the reconstructed data matrix The quality assessment function is , , Respectively represent , , The reconstruction quality of each data point is Indicates the total number of data points; The reconstruction quality calculation formula for each data point is: ； in, Indicates The reconstruction quality of each data point is is the reconstruction error, To reconstruct uncertainty, is the weight coefficient, which is used to balance the impact of reconstruction error and uncertainty.

7. The method for processing wafer defect detection data according to claim 1, characterized in that: The step of generating an adaptive sampling strategy based on the reconstruction quality in the step of evaluating the quality of the reconstruction result comprises: Divide the defect area into multiple sub-areas: ； in, , , Respectively represent , , sub-regions, Indicates the total number of sub-regions; Calculate the average reconstruction quality for each sub-region: ； in, Indicates The average reconstruction quality score of sub-regions, Indicates The number of data points contained in each sub-region. Indicates The reconstruction quality score of each data point is Indicates The sum of the reconstruction quality scores of all data points in the sub-region; Calculate sampling priority based on reconstruction quality: ； in, Indicates The sampling priority of each sub-region is is a monotonically decreasing function; Sampling resources are allocated according to sampling priority.

8. The method for processing wafer defect detection data according to claim 1, characterized in that: The process of implementing adaptive sampling in the step of quality assessment based on reconstruction results includes: Update the measurement matrix according to the sampling strategy ; Collect new compression measurements ; Reconstruction based on the original data: ； in, represents the updated reconstruction result matrix, To reconstruct the function, represents newly acquired compressed measurement data, represents the updated measurement matrix, represents the original reconstruction result; The above process is iterated until the reconstruction quality meets the preset threshold.

9. The method for processing wafer defect detection data according to claim 1, characterized in that: In the step of performing differentiated reconstruction processing according to defect types, different regularization parameters are set for different types of defects when solving the optimization problem. and ,in: For regular geometric defects, increase The weights of , to maintain the low rank property; For complex morphological defects, increase to maintain the local manifold structure.

10. A wafer defect detection data processing system, characterized in that: A method for processing wafer defect detection data according to any one of claims 1 to 9, the system comprising: A compression measurement module is used to build a structured random measurement matrix based on the surface characteristics of the wafer and collect compression measurement data of wafer surface defects; Defect classification module, which is used to analyze compressed measurement data using spectral clustering algorithm, automatically identify the local manifold structure corresponding to different defect types, and establish the corresponding relationship between defect type and local manifold; Feature extraction module, which is used to construct a multi-scale graph Laplacian operator for each type of defect and extract the geometric characteristics of defects at different scales; An adaptive reconstruction module is used to perform differentiated reconstruction processing according to the defect type, and solve the optimization objective including the nuclear norm and graph Laplace regularization term for each type of defect to obtain the reconstruction result; The sampling optimization module is used to dynamically adjust the distribution strategy of sampling points based on the quality evaluation of the reconstruction results and increase the sampling density in areas with poor reconstruction quality.

Images