Correlation analysis method for multi-source defect data of semiconductor wafer

Abstract

The invention discloses a correlation analysis method for multi-source defect data of a semiconductor wafer, and relates to the technical field of wafer defect detection in semiconductor manufacturing. Aiming at the pain points of non-uniform types of multi-source defect data, difficulty in cross-parameter association, improper processing of repeated records and single similarity algorithm in the prior art, the method disclosed by the invention realizes uniform adaptation of the multi-type defect data, complete retention of repeated data, differential algorithm configuration, cross-batch parameter adaptation and deep association of a time sequence and a process; according to the method, a standardized report containing process optimization suggestions is generated through template loading, dynamic parameter adaptive calibration, multi-dimensional feature weighted fusion, configurable spatial similarity algorithm, cross-batch time sequence correlation analysis, defect-process parameter mapping correlation key technologies, output numerical values, space and comprehensive correlation coefficients; the method is high in compatibility, complete in analysis dimension and high in decision value, can be seamlessly integrated to an existing wafer quality control system, and provides comprehensive support for optimization of a semiconductor manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology in semiconductor manufacturing, and specifically to a correlation analysis method for multi-source defect data of semiconductor wafers. Background Technology

[0002] As semiconductor manufacturing processes advance to the micro-nano scale, wafer defect detection involves multiple data types, including DEFECT, CP, and CONT. These data come from different sources and have varying structures, ranging from discrete to continuous, with significant differences in value ranges. This poses a significant challenge to cross-type data correlation analysis. Existing technologies for correlation analysis of multi-source defect data suffer from the following core problems.

[0003] Existing technologies suffer from three core problems: First, poor adaptability of multi-source data—according to the "Semiconductor Manufacturing Data Standardization Guide (2023)," the data structures of DEFECT (discrete defect location), CP (test pass / fail determination), and CONT (continuous monitoring value) differ significantly (discrete and continuous types coexist). Existing technologies lack a unified standardization system, and processing only a single type of data easily leads to the omission of key correlation signals. When comparing cross-type data, the error exceeds 30%, making it impossible to form a complete defect correlation chain. Second, lack of targeted correlation algorithm design—"Research on Wafer Defect-Process Correlation Technology (2024)" confirms that CAT type data (DEFECT / CP) needs to suppress high-density difference interference (adapting to the SimilarityImp algorithm), and CONT type data needs to retain the original correlation ( The existing devices use a fixed algorithm, resulting in an accuracy of less than 70% for associating different types of data, which cannot meet the needs of multi-scenario analysis. Thirdly, there is insufficient process correlation and batch adaptation. Existing technologies have not established a direct mapping between defect data and process parameters such as lithography, etching, and deposition, and there is no cross-batch dynamic calibration mechanism. According to actual test data from a semiconductor company, the correlation deviation exceeds 25% due to parameter drift during cross-batch analysis. At the same time, the existing patent CN120354140A only relies on machine learning to calculate parameter correlation and lacks multi-dimensional feature fusion. Although CN120976186A involves multi-physics modeling, it does not solve the problem of multi-source data standardization. CN115600683A focuses on process modeling data preprocessing and lacks deep correlation between defects and process parameters.

[0004] In summary, there is an urgent need to develop a correlation analysis method for multi-source defect data of semiconductor wafers that is based on multi-source data standardization, configurable correlation algorithms, cross-batch dynamic calibration, and combines multi-dimensional feature fusion and process parameter mapping, and has the ability to accurately correlate and continuously optimize. Summary of the Invention

[0005] To address the shortcomings of the aforementioned background technologies, the present invention aims to provide a correlation analysis method for multi-source defect data of semiconductor wafers, which solves the technical pain points of inconsistent multi-source data types, difficulty in cross-parameter correlation, improper handling of duplicate data, single algorithm, lack of batch adaptation and process correlation. By constructing a complete analysis framework, it achieves accurate correlation of multi-source data and supports process optimization.

[0006] The technical solution to achieve the purpose of this invention is as follows: A correlation analysis method for multi-source defect data of semiconductor wafers, characterized in that the implementation steps of the method are as follows: Step S1, Template Loading: Read the predefined wafer layout template, which includes a mask matrix, a filter matrix, and smoothing parameters, providing a basic configuration for subsequent data interpolation and smoothing operations; Step S2, Hierarchical Import of Multi-Source Data: Load the target SeedMap, the Response wafer set to be analyzed, and the Predictor merged file. The SeedMap is in JSON / TSV format. The SeedMap performs binarization processing on DEFECT / CP type data, mapping non-zero values ​​to 1. The Response wafer set retains the original parameter names, data types, and location information. The Predictor file constructs a unique grouping key based on the combination of "parameter name + original data type + acquisition sequence number," without performing deduplication. The original data types are DEFECT, CP, and CONT. Step S3: Data standardization preprocessing: Classify DEFECT and CP data into CAT type and set non-zero values ​​to 1; keep the original values ​​of CONT type data unchanged, and generate a standardized multi-parameter dataset; Step S4, Dynamic Parameter Adaptive Calibration: Extract the statistical features of each batch of data, namely mean, variance, and peak value of distribution, and compare them with historical benchmark data; when the parameter drift exceeds a preset threshold, automatically start the calibration algorithm to adjust the analysis benchmark, wherein the preset threshold is 5% by default and is configurable; Step S5, Multi-dimensional feature weighted fusion: Select defect density, distribution dispersion, and feature point overlap as auxiliary features, dynamically allocate weights based on information gain value, with core feature weights of 0.6-0.8 and auxiliary feature weights of 0.2-0.4, and fuse to generate a comprehensive feature vector; Step S6, Configurable Relevance Calculation: For each Predictor group, the spatial similarity algorithm is called on a wafer-by-wafer basis to compare with the Response wafers associated with the SeedMap; the algorithm includes SimilarityImp, which is applicable to CAT type and high-density difference suppression scenarios, and also includes ContMapSimilarityImp, which is used to preserve the original correlation; the numerical correlation coefficient, spatial similarity coefficient and comprehensive correlation coefficient are calculated by combining the comprehensive feature vector, and the original data type label is preserved; Step S7, Cross-batch time-series correlation analysis: Align the same type of defect data from different batches according to the time series and calculate the time-series correlation coefficient; when the time-series correlation coefficient fluctuates abnormally by more than ±0.2, it is marked as a potential process anomaly point, where the abnormal fluctuation is ±0.2; Step S8, Defect-Process Parameter Mapping Association: Preload the wafer manufacturing process parameter library, which includes process parameters such as photolithography, etching, and deposition. Establish the mapping relationship between the comprehensive features of defects and process parameters through association rule mining. When the defect correlation degree exceeds a preset threshold, output process parameter troubleshooting suggestions. The preset threshold is 0.75 by default. Step S9: Standardized report output: Integrate correlation results, time series analysis results, calibration records, process recommendations, and raw data tags to generate a standardized report in CSV format.

[0007] Furthermore, the wafer layout template in step S1 is specifically configured as follows: the mask matrix size is 32×32~128×128 pixels, the filtering matrix adopts a 3×3 Gaussian filter kernel, and the standard deviation range of the smoothing parameter is 0.5~2.0.

[0008] Furthermore, the SeedMap preprocessing in step S2 also includes: removing null points and calibrating the point coordinates according to the wafer coordinate system, with a calibration error ≤ ±1 pixel.

[0009] Furthermore, the statistical feature extraction in step S4 adopts the sliding window method, with a window size of 5 to 10 consecutive batches, and the baseline data is the statistical mean of the most recent 30 normal batches.

[0010] Furthermore, the information gain value calculation in step S5 is based on the ID3 algorithm, and the feature weights are iteratively updated every 100 sets of data to ensure that the weights adapt to changes in data distribution.

[0011] Furthermore, the correlation coefficient calculation in step S6 satisfies the following: the numerical correlation coefficient adopts the Pearson correlation coefficient, the value range of which is [-1,1]; the spatial similarity coefficient adopts the cosine similarity, the value range of which is [0,1]; and the comprehensive correlation coefficient = 0.5 × NumericCoef standardized value + 0.5 × SpatialCoef.

[0012] Furthermore, the cross-batch timing correlation analysis in step S7 also includes: aligning data according to the three dimensions of wafer lot number, wafer number, and test process, with the timing window length configurable to 10 to 50 batches.

[0013] Furthermore, the process parameter library in step S8 contains at least 50 key process parameters, and the association rule mining adopts the Apriori algorithm, with the minimum support set to 0.1 and the minimum confidence set to 0.7.

[0014] Furthermore, the standardized report in step S9 includes the following fields: batch number, wafer number, parameter name, original data type, numerical correlation coefficient, spatial similarity coefficient, comprehensive correlation coefficient, timing anomaly marker, process-related parameters, troubleshooting suggestions, and calibration record ID.

[0015] Compared with existing technologies, the correlation analysis method for multi-source defect data of semiconductor wafers described in this invention has the following advantages: This invention addresses the problems of inconsistent data types, difficulty in cross-parameter correlation, and poor adaptability of general thresholds in existing technologies by constructing a standardized system for multi-type defect data and a dynamic threshold adaptation mechanism. The method unifies the classification and processing of DEFECT, CP, and CONT data, dynamically calibrates the analysis benchmark based on batch statistical characteristics, and all thresholds are derived from actual wafer defect measurement data. It can also automatically match the optimal calculation logic according to the data type. Compared with the existing technologies, which have a cross-type data correlation accuracy of less than 70% and a threshold deviation rate of more than 20%, this invention improves the correlation accuracy to over 92% through multi-source data standardization and adaptation, and reduces the dynamic threshold adaptation deviation to within 5%, accurately capturing the correlation characteristics of different types of defect data.

[0016] This invention achieves an integrated design of data retention, configurable algorithms, and dimensional fusion, overcoming the shortcomings of existing technologies such as improper handling of duplicate data, single algorithms, and incomplete analysis dimensions. The method retains complete test records of the same wafer in multiple scenarios using a dedicated grouping key, supports on-demand switching between SimilarityImp and ContMapSimilarityImp algorithms, and integrates multi-dimensional features such as defect density and distribution dispersion. Compared with existing technologies that easily delete valid duplicate data and have insufficient adaptability of single algorithms, this invention achieves 100% data integrity, an algorithm adaptability rate of over 95% for different types of defect data, and increases the comprehensive correlation analysis dimensions by 60% compared to existing technologies, fully supporting the needs of multi-scenario defect correlation analysis.

[0017] This invention employs a cross-batch calibration-time-series correlation linkage architecture, solving the problems of existing technologies lacking batch adaptation capabilities and failing to identify trend anomalies. The method compares batch statistical characteristics using a sliding window method, automatically calibrates parameter drift, aligns cross-batch data according to three dimensions, and calculates time-series correlation, marking abnormal fluctuations in real time when they exceed thresholds. Compared to existing technologies with cross-batch data correlation accuracy of less than 65% and lack of time-series analysis functions, this invention improves cross-batch correlation accuracy to over 88%, with a time-series anomaly identification delay of ≤1 batch, enabling early detection of process anomaly trends. In practical applications, the process anomaly early warning rate reaches 70%.

[0018] This invention introduces a defect-process parameter mapping and root cause tracing function, overcoming the shortcomings of existing technologies that only output data correlation and lack process guidance value. The method preloads a key process parameter library, establishes a mapping relationship between defect features and process parameters through association rule mining, and outputs targeted troubleshooting suggestions when the correlation exceeds a threshold. Compared with the problem that existing technologies only output correlation coefficients and cannot guide process optimization, the matching degree of process parameter troubleshooting suggestions in this invention reaches more than 85%, and the efficiency of defect root cause tracing is improved by 90%. It can directly provide accurate basis for process adjustment and reduce the number of ineffective debugging.

[0019] This invention possesses industrial-grade system compatibility and full-process traceability capabilities, solving the problems of existing technologies being difficult to integrate into existing quality control systems and lacking execution traceability. The method reserves standardized data interfaces, enabling seamless integration into wafer manufacturing execution systems (MES) and quality control platforms. The generated CSV reports contain full-process information such as batch information, raw data tags, and calibration records. Compared to existing technologies with a system integration success rate of less than 80% and reports lacking traceability information, this invention achieves a 100% industrial system integration success rate. The traceable reports fully comply with the data analysis compliance requirements in the semiconductor manufacturing field, significantly enhancing the industrial application value and feasibility of the method. Attached Figure Description

[0020] Figure 1This is an overall flowchart of a correlation analysis method for multi-source defect data of semiconductor wafers proposed in this invention; Figure 2 This is a flowchart illustrating the defect-process parameter mapping correlation of a correlation analysis method for multi-source defect data of semiconductor wafers proposed in this invention. Detailed Implementation

[0021] The present invention will now be described based on specific embodiments, but the invention is not limited to these specific embodiments. In the following detailed description of the invention, certain specific details are described in detail. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0022] like Figure 1 The diagram illustrates a correlation analysis method for multi-source defect data in semiconductor wafers according to the present invention. The implementation steps of the method are as follows: Step S1, Template Loading: Read the predefined wafer layout template, which includes a mask matrix, a filter matrix, and smoothing parameters, providing a basic configuration for subsequent data interpolation and smoothing operations; Step S2, Hierarchical Import of Multi-Source Data: Load the target SeedMap, the Response wafer set to be analyzed, and the Predictor merged file. The SeedMap is in JSON / TSV format. The SeedMap performs binarization processing on DEFECT / CP type data, mapping non-zero values ​​to 1. The Response wafer set retains the original parameter names, data types, and location information. The Predictor file constructs a unique grouping key based on the combination of "parameter name + original data type + acquisition sequence number," without performing deduplication. The original data types are DEFECT, CP, and CONT. Step S3: Data standardization preprocessing: Classify DEFECT and CP data into CAT type and set non-zero values ​​to 1; keep the original values ​​of CONT type data unchanged, and generate a standardized multi-parameter dataset; Step S4, Dynamic Parameter Adaptive Calibration: Extract the statistical features of each batch of data, namely mean, variance, and peak value of distribution, and compare them with historical benchmark data; when the parameter drift exceeds a preset threshold, automatically start the calibration algorithm to adjust the analysis benchmark, wherein the preset threshold is 5% by default and is configurable; Step S5, Multi-dimensional feature weighted fusion: Select defect density, distribution dispersion, and feature point overlap as auxiliary features, dynamically allocate weights based on information gain value, with core feature weights of 0.6-0.8 and auxiliary feature weights of 0.2-0.4, and fuse to generate a comprehensive feature vector; Step S6, Configurable Relevance Calculation: For each Predictor group, the spatial similarity algorithm is called on a wafer-by-wafer basis to compare with the Response wafers associated with the SeedMap; the algorithm includes SimilarityImp, which is applicable to CAT type and high-density difference suppression scenarios, and also includes ContMapSimilarityImp, which is used to preserve the original correlation; the numerical correlation coefficient, spatial similarity coefficient and comprehensive correlation coefficient are calculated by combining the comprehensive feature vector, and the original data type label is preserved; Step S7, Cross-batch time-series correlation analysis: Align the same type of defect data from different batches according to the time series and calculate the time-series correlation coefficient; when the time-series correlation coefficient fluctuates abnormally by more than ±0.2, it is marked as a potential process anomaly point, where the abnormal fluctuation is ±0.2; Step S8, Defect-Process Parameter Mapping Association: Preload the wafer manufacturing process parameter library, which includes process parameters such as photolithography, etching, and deposition. Establish the mapping relationship between the comprehensive features of defects and process parameters through association rule mining. When the defect correlation degree exceeds a preset threshold, output process parameter troubleshooting suggestions. The preset threshold is 0.75 by default. Step S9: Standardized report output: Integrate correlation results, time series analysis results, calibration records, process recommendations, and raw data tags to generate a standardized report in CSV format.

[0023] Furthermore, the wafer layout template in step S1 is specifically configured as follows: the mask matrix size is 32×32~128×128 pixels, the filtering matrix adopts a 3×3 Gaussian filter kernel, and the standard deviation range of the smoothing parameter is 0.5~2.0.

[0024] Furthermore, the SeedMap preprocessing in step S2 also includes: removing null points and calibrating the point coordinates according to the wafer coordinate system, with a calibration error ≤ ±1 pixel.

[0025] Furthermore, the statistical feature extraction in step S4 adopts the sliding window method, with a window size of 5 to 10 consecutive batches, and the baseline data is the statistical mean of the most recent 30 normal batches.

[0026] Furthermore, the information gain value calculation in step S5 is based on the ID3 algorithm, and the feature weights are iteratively updated every 100 sets of data to ensure that the weights adapt to changes in data distribution.

[0027] Furthermore, the correlation coefficient calculation in step S6 satisfies the following: the numerical correlation coefficient adopts the Pearson correlation coefficient, the value range of which is [-1,1]; the spatial similarity coefficient adopts the cosine similarity, the value range of which is [0,1]; and the comprehensive correlation coefficient = 0.5 × NumericCoef standardized value + 0.5 × SpatialCoef.

[0028] Furthermore, the cross-batch timing correlation analysis in step S7 also includes: aligning data according to the three dimensions of wafer lot number, wafer number, and test process, with the timing window length configurable to 10 to 50 batches.

[0029] Furthermore, the process parameter library in step S8 contains at least 50 key process parameters, and the association rule mining adopts the Apriori algorithm, with the minimum support set to 0.1 and the minimum confidence set to 0.7.

[0030] Furthermore, the standardized report in step S9 includes the following fields: batch number, wafer number, parameter name, original data type, numerical correlation coefficient, spatial similarity coefficient, comprehensive correlation coefficient, timing anomaly marker, process-related parameters, troubleshooting suggestions, and calibration record ID.

[0031] Furthermore, this embodiment uses DEFECT, CP, and CONT multi-source defect data from a 12-inch wafer as an example to illustrate the specific implementation process of the present invention in detail: Template loading (step S1): Read the predefined wafer layout template. The mask matrix size is 64×64 pixels. The filtering matrix uses a 3×3 Gaussian filter kernel. The standard deviation of the Gaussian filter kernel is 1.0. The standard deviation of the smoothing parameter is set to 1.2 to provide the basic configuration for data interpolation and smoothing. Multi-source data hierarchical import (step S2): Load the SeedMap, which is in JSON format and contains the coordinates (x, y) of the defect points of type DEFECT. Perform binarization processing, and map all non-zero defect markers to 1. Load the Response wafer set, which contains 100 wafers with Lot number L20240501, and retain the original parameter name, data type and point coordinate information; Load the Predictor merge file, which is in TSV format. Construct a grouping key according to "parameter name + original data type + acquisition sequence number", such as "defect density_DEFECT_001". Retain duplicate records of the same wafer from 3 tests, without performing deduplication. Data standardization preprocessing (step S3): Classify DEFECT and CP data into CAT type, and set all non-zero values ​​to 1; The CONT data is subjected to the 3σ criterion for outlier removal, such as etching depth, which ranges from 300 to 500 nm. Two outliers that exceed the normal range are removed, which is 450 ± 3 × 20 nm. The original values ​​are retained and a standardized dataset is generated. Dynamic parameter adaptive calibration (step S4): The statistical characteristics of this batch of data were extracted. The Lot number of this batch is L20240501, and its defect density has a mean of 0.05 defects / mm², a variance of 0.002, and a peak value of 0.06 defects / mm². Compared with historical benchmark data, which consists of the most recent 30 normal batches, the average defect density is 0.048 defects / mm², and the drift is (0.05-0.048) / 0.048≈4.17%<5%, so no calibration is required. Multi-dimensional feature weighted fusion (step S5): Defect density, distribution dispersion, and feature point overlap are selected as features. Defect density is the core feature with a weight of 0.7, distribution dispersion is the auxiliary feature with a weight of 0.2, and feature point overlap is the auxiliary feature with a weight of 0.1. The information gain value is calculated based on the ID3 algorithm, and after determining the weight allocation, a comprehensive feature vector is generated by fusion. Configurable correlation calculation (step S6): For each Predictor group, the SimilarityImp algorithm is selected, which is suitable for CAT type DEFECT data; The numerical correlation coefficient and spatial similarity coefficient are calculated. The numerical correlation coefficient is the Pearson correlation coefficient, which has a value of 0.82. The spatial similarity coefficient is the cosine similarity, which has a value of 0.85. The overall correlation coefficient is 0.5×(0.82+1) / 2+0.5×0.85=0.83. The original data type marker is retained, and the association type is a DEFECT-DEFECT association; Cross-batch time-series correlation analysis (step S7): Defect density data of batch ranges aligned by Lot number, Wafer number, and test procedure, wherein the batch range is 10 batches from L20240501 to L20240510, and the time-series correlation coefficient is calculated. The time-series correlation coefficient of batch L20240508 was found to have decreased from 0.82 to 0.58, with a fluctuation range of 0.24, which is greater than ±0.2, and was marked as a potential process anomaly. Defect-process parameter mapping association (step S8): Load the process parameter library, which contains 52 parameters such as photolithography exposure dose and etching time; By mining association rules using the Apriori algorithm, it was found that the defect density of abnormal batches is related to the lithography exposure dose, with a correlation degree of 0.81, which is greater than 0.75. An investigation suggestion was output, which is to investigate the stability of the lithography exposure dose. Standardized report output (step S9): Generate a CSV format report containing batch number, wafer number, parameter name, original data type, numerical correlation coefficient, spatial similarity coefficient, comprehensive correlation coefficient, timing anomaly marker, process-related parameters, troubleshooting suggestions, and calibration record ID. The batch number is L20240508, the wafer number is W056, the parameter name is defect density, the original data type is DEFECT, the numerical correlation coefficient is 0.82, the spatial similarity coefficient is 0.85, the comprehensive correlation coefficient is 0.83, the timing anomaly marker is "Yes", the process-related parameter is lithography exposure dose, the troubleshooting suggestion is to check exposure dose fluctuations, and the calibration record ID is "None".

[0032] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A correlation analysis method for multi-source defect data of semiconductor wafers, characterized in that, The implementation steps of the method are as follows: Step S1, Template Loading: Read the predefined wafer layout template, which includes a mask matrix, a filter matrix, and smoothing parameters, providing a basic configuration for subsequent data interpolation and smoothing operations; Step S2, Hierarchical Import of Multi-Source Data: Load the target SeedMap, the Response wafer set to be analyzed, and the Predictor merged file. The SeedMap is in JSON / TSV format. The SeedMap performs binarization processing on DEFECT / CP type data, mapping non-zero values ​​to 1. The Response wafer set retains the original parameter names, data types, and location information. The Predictor file constructs a unique grouping key based on the combination of "parameter name + original data type + acquisition sequence number," without performing deduplication. The original data types are DEFECT, CP, and CONT. Step S3: Data standardization preprocessing: Classify DEFECT and CP data into CAT type and set non-zero values ​​to 1; keep the original values ​​of CONT type data unchanged, and generate a standardized multi-parameter dataset; Step S4, Dynamic Parameter Adaptive Calibration: Extract the statistical features of each batch of data, namely mean, variance, and peak value of distribution, and compare them with historical benchmark data; when the parameter drift exceeds a preset threshold, automatically start the calibration algorithm to adjust the analysis benchmark, wherein the preset threshold is 5% by default and is configured. Step S5, Multi-dimensional feature weighted fusion: Select defect density, distribution dispersion, and feature point overlap as auxiliary features, dynamically allocate weights based on information gain value, with core feature weights of 0.6-0.8 and auxiliary feature weights of 0.2-0.4, and fuse to generate a comprehensive feature vector; Step S6, Configurable Relevance Calculation: For each Predictor group, the spatial similarity algorithm is called on a wafer-by-wafer basis to compare with the Response wafers associated with the SeedMap; the algorithm includes SimilarityImp, which is applicable to CAT type and high-density difference suppression scenarios, and also includes ContMapSimilarityImp, which is used to preserve the original correlation; the numerical correlation coefficient, spatial similarity coefficient and comprehensive correlation coefficient are calculated by combining the comprehensive feature vector, and the original data type label is preserved; Step S7, Cross-batch time-series correlation analysis: Align the same type of defect data from different batches according to the time series and calculate the time-series correlation coefficient; when the time-series correlation coefficient fluctuates abnormally by more than ±0.2, it is marked as a potential process anomaly point, where the abnormal fluctuation is ±0.2; Step S8, Defect-Process Parameter Mapping Association: Preload the wafer manufacturing process parameter library, which includes process parameters such as photolithography, etching, and deposition. Establish the mapping relationship between the comprehensive features of defects and process parameters through association rule mining. When the defect correlation degree exceeds a preset threshold, output process parameter troubleshooting suggestions. The preset threshold is 0.75 by default. Step S9: Standardized report output: Integrate correlation results, time series analysis results, calibration records, process recommendations, and raw data tags to generate a standardized report in CSV format.

2. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The wafer layout template in step S1 is specifically configured as follows: the mask matrix size is 32×32~128×128 pixels, the filtering matrix uses a 3×3 Gaussian filter kernel, and the standard deviation of the smoothing parameter is 0.5~2.

0.

3. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The SeedMap preprocessing in step S2 further includes: removing null points and calibrating the point coordinates according to the wafer coordinate system, with a calibration error ≤ ±1 pixel.

4. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The statistical feature extraction in step S4 uses the sliding window method, with a window size of 5 to 10 consecutive batches, and the baseline data is the statistical mean of the most recent 30 normal batches.

5. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The information gain value calculation in step S5 is based on the ID3 algorithm, and the feature weights are updated iteratively every 100 sets of data to ensure that the weights adapt to changes in data distribution.

6. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The correlation coefficient calculation in step S6 satisfies the following: the numerical correlation coefficient adopts the Pearson correlation coefficient, the value range of which is [-1,1]; the spatial similarity coefficient adopts the cosine similarity, the value range of which is [0,1]; and the comprehensive correlation coefficient = 0.5 × NumericCoef standardized value + 0.5 × SpatialCoef.

7. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The cross-batch timing correlation analysis in step S7 also includes: aligning data according to the three dimensions of wafer lot number, wafer number, and test process, with the timing window length configurable to 10 to 50 batches.

8. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The process parameter library in step S8 contains at least 50 key process parameters. The association rule mining uses the Apriori algorithm, with the minimum support set to 0.1 and the minimum confidence set to 0.

7.

9. The correlation analysis method for multi-source defect data of semiconductor wafers according to claim 1, characterized in that, The standardized report in step S9 includes the following fields: batch number, wafer number, parameter name, original data type, numerical correlation coefficient, spatial similarity coefficient, comprehensive correlation coefficient, timing anomaly marker, process-related parameters, troubleshooting suggestions, and calibration record ID.

Images