A method and system for generating test patterns for curve layout based on machine learning

Abstract

The present application relates to the technical field of integrated circuit layout detection, and particularly relates to a curve layout test pattern generation method and system based on machine learning, comprising: obtaining historical curve layout data and preprocessing to obtain an effective data set; extracting geometric and defect correlation features, and obtaining a feature matrix through correlation analysis and dimension reduction; generating a defect mode label by K-means clustering combined with contour coefficient correction clustering number; inputting the feature matrix and the label into a classification model to output parameters to generate an initial test pattern, and obtaining an optimized test pattern through redundancy and coverage optimization; importing a test system for verification, and if the verification does not meet the standard, the model is retrained, and if the verification meets the standard, the final pattern is output. The present application realizes accurate identification of defect modes with the aid of machine learning, balances detection accuracy and efficiency, improves the applicability of test patterns, and is suitable for integrated circuit curve layout detection.

Description

A Machine Learning-Based Method and System for Generating Curve Layout Test Graphics Technical Field

[0001] This invention relates to the field of integrated circuit layout inspection technology, and more specifically, to a method and system for generating curve layout test graphics based on machine learning. Background Technology

[0002] In the field of semiconductor integrated circuit manufacturing, the design of test patterns for curve layouts directly affects the accuracy and efficiency of inspection due to their complex geometric contours and diverse defect types (such as contour distortion and local damage). Traditional test pattern generation relies heavily on manual experience or fixed rules, which has significant limitations: on the one hand, manual design requires repeated debugging, resulting in low efficiency and difficulty in covering potential defect patterns in complex curves; on the other hand, test patterns generated by fixed rules are prone to having too many redundant test points and insufficient coverage of high-frequency defect areas, leading to excessively long testing times or high false negative rates. Existing generation methods based on simple algorithms can achieve preliminary automation, but they cannot accurately extract the geometric features and defect-related features of the curve layout, and lack effective clustering and classification of defect patterns. This makes the test patterns insufficiently targeted, making it difficult to balance inspection accuracy and efficiency. With the increasing complexity of integrated circuit layouts, there is an urgent need for a technical solution that can combine historical data and machine learning to achieve accurate feature selection, intelligent defect pattern recognition, and test pattern optimization to meet the requirements of high-precision and high-efficiency curve layout inspection. Summary of the Invention

[0003] In view of this, the present invention addresses the shortcomings of the prior art by proposing a method and system for generating curve layout test graphics based on machine learning, aiming to solve at least one of the problems mentioned in the background art.

[0004] In a first aspect, the present invention provides a method for generating curve layout test graphics based on machine learning, comprising the following steps:

[0005] Acquire historical curve layout data, clean and complete the historical curve layout data, and output a valid dataset containing normal layout data and marked defect layout data;

[0006] Geometric features and defect-related features of the curve layout in the effective dataset are extracted, and highly correlated features are screened through correlation analysis and then dimensionality reduction is performed to obtain the feature matrix.

[0007] The feature matrix is ​​input into the K-means algorithm. An initial cluster number k value is preset. After clustering, the cluster number is judged and corrected based on the silhouette coefficient to obtain the defect pattern label corresponding to the highly correlated feature.

[0008] Using the feature matrix and defect pattern labels as input, the core parameters of the test pattern are output through the classification model and an initial test pattern is generated. The test points are adjusted based on redundancy. Then, for the high-frequency defect area determined based on historical defect distribution, the test point density is optimized according to the coverage and an optimized test pattern is output.

[0009] The optimized test graph is imported into the layout testing system for verification. If the verification fails, feedback is sent to the clustering step to correct the k value or adjust the classification model parameters and retrain. If the verification succeeds, the final test graph is output.

[0010] In some embodiments, the step of acquiring historical curve map data, cleaning and completing the historical curve map data, and outputting a valid dataset containing normal map data and map data with marked defects includes:

[0011] The layout data for annotating defects includes information on defect type, defect location, and defect size;

[0012] Read the geometric parameters of the historical curve layout data;

[0013] When one or two parameters are missing in the historical curve map data, the missing parameters are filled in based on the existing contour fitting.

[0014] If three or more parameters are missing, the sample is filtered out.

[0015] Check the defect annotation information. If two or more annotation fields are missing, the annotation is deemed invalid and the sample is filtered out.

[0016] When a field is missing one item, it is filled in based on the statistical mean of similar defects.

[0017] In some embodiments, the step of extracting the geometric features and defect correlation features of the curve pattern in the effective dataset, filtering highly correlated features through correlation analysis, and obtaining a feature matrix through dimensionality reduction includes:

[0018] Based on the map contour fitting results, determine whether the sample is a regular curve or an irregular curve. When the sample is a regular curve, extract the core geometric parameters.

[0019] When the sample is an irregular curve, equal-interval sampling is used to extract the curvature sequence of the sampling points and the slope difference between adjacent sampling points to form a feature vector.

[0020] The correlation between geometric features and defect types is calculated based on the Pearson correlation coefficient, and a correlation threshold is preset to retain only high correlation features whose correlation between geometric features and defect types is greater than or equal to the correlation threshold.

[0021] When the number of highly correlated features is greater than or equal to a preset threshold for the number of highly correlated features, principal component analysis is used to reduce dimensionality and determine the initial dimensionality reduction dimension.

[0022] When the number of highly correlated features is less than a preset threshold for the number of highly correlated features, process features are extracted. After extraction, if the number of highly correlated features is greater than or equal to the preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension.

[0023] In some embodiments, when the number of highly correlated features is greater than or equal to a preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension; or when the number of highly correlated features is less than the preset threshold for the number of highly correlated features, process features are extracted, and after extraction, if the number of highly correlated features is greater than or equal to the preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension, including:

[0024] Calculate the cumulative variance contribution rate after dimensionality reduction. When the cumulative variance contribution rate is less than a preset contribution rate threshold, reduce the initial dimensionality reduction dimension.

[0025] When the cumulative variance contribution rate is greater than or equal to the preset contribution rate threshold, the initial dimensionality reduction dimension is not changed.

[0026] In some embodiments, the feature matrix is ​​input into a K-means algorithm, an initial cluster number k is preset, and the cluster number is determined and corrected based on the silhouette coefficient after clustering to obtain defect pattern labels corresponding to highly correlated features, including:

[0027] When the contour coefficient is less than a preset coefficient threshold, the initial cluster number K is increased to obtain the defect pattern label;

[0028] When the contour coefficient is greater than or equal to a preset coefficient threshold, the initial cluster number K value is not changed, and the defect pattern label is obtained.

[0029] When the defect pattern label is obtained, a preset number of clustered samples are randomly selected for manual verification. When the accuracy rate is less than the preset accuracy threshold, the clustering distance is adjusted.

[0030] When the accuracy is greater than or equal to the preset accuracy threshold, the cluster distance is not adjusted.

[0031] In some embodiments, the feature matrix is ​​input into a K-means algorithm, an initial cluster number k is preset, and the cluster number is determined and corrected based on the silhouette coefficient after clustering to obtain defect pattern labels corresponding to highly correlated features. The algorithm further includes:

[0032] The dataset containing the dimensionality reduction features, defect pattern labels, and corresponding test graph core parameters is divided into a training set and a validation set according to a certain ratio. The defect pattern labels are used to establish the correlation between the dimensionality reduction features and the test graph core parameters. A random forest model is trained based on the training set, and the model performance is evaluated using the validation set. The prediction accuracy of the validation set is calculated.

[0033] When the prediction accuracy of the validation set is greater than or equal to a preset accuracy threshold, it is deemed qualified;

[0034] When the prediction accuracy of the validation set is less than a preset accuracy threshold, the cumulative percentage of the top three features in the model output is calculated:

[0035] If the cumulative percentage is less than a preset percentage threshold, it is determined that the feature discrimination is insufficient, and high discrimination features are added. The high discrimination features include the wire density around the defect.

[0036] If the cumulative proportion is greater than or equal to the preset proportion threshold, the feature discrimination is deemed qualified, the model hyperparameters are adjusted, and the model is retrained until the prediction accuracy of the validation set meets the standard.

[0037] In some embodiments, the process of taking the feature matrix and defect pattern labels as input, outputting core parameters of the test pattern through a classification model and generating an initial test pattern, adjusting test points based on redundancy, and then optimizing the test point density based on coverage for high-frequency defect areas determined based on historical defect distribution, and outputting an optimized test pattern includes:

[0038] When the redundancy of the initial test pattern is greater than a preset first threshold, the test points in the initial test pattern are adjusted in combination with the characteristics of the corresponding defect pattern. When the redundancy of the initial test pattern is less than or equal to the preset first threshold, the coverage of the initial test pattern to the high-frequency defect region corresponding to each defect pattern is calculated. The high-frequency defect region is determined based on the historical defect distribution associated with the defect pattern label.

[0039] The defect pattern label is used to mark the priority of test points corresponding to different defect types. When the redundancy of the initial test pattern is greater than the preset first threshold, the test points are sorted from high to low according to the number of overlaps of the test points, and the key test points corresponding to the defect pattern label are retained first, while non-key test points with more than 2 overlaps are deleted in turn.

[0040] After deleting non-critical test points with more than 2 overlaps, the redundancy is recalculated. When the redundancy is less than or equal to the preset first threshold after recalculation, the distribution area of ​​each defect type is statistically analyzed based on the defect pattern label. The area where the frequency of defect occurrence is greater than the preset frequency threshold is determined as the high-frequency defect area, and the coverage of the initial test map of the high-frequency defect area is calculated.

[0041] After deleting non-critical test points with more than 2 overlaps, the redundancy is recalculated. If the redundancy is greater than a preset first threshold after recalculation, the above deletion and calculation process continues until the redundancy is less than or equal to the preset first threshold. Then, the coverage of the initial test pattern to the high-frequency area of ​​the defect is calculated, and the density of the original test points in the high-frequency area of ​​the defect is increased based on the coverage. Finally, an optimized test pattern is output.

[0042] In some embodiments, the step of taking the feature matrix and defect pattern labels as input, outputting core parameters of the test pattern through a classification model and generating an initial test pattern, adjusting test points based on redundancy, and then optimizing the test point density based on coverage for high-frequency defect areas determined based on historical defect distribution, and outputting an optimized test pattern, further includes:

[0043] The coverage is the ratio of the area covered by the test pattern in the high-frequency defect region to the total area of ​​that region;

[0044] When the coverage is less than the preset coverage threshold, it is determined that the risk of missed detection is high. Test points in high-frequency defect areas are densified. The coverage is recalculated after each densification until the coverage is greater than or equal to the preset coverage threshold.

[0045] When the coverage is greater than or equal to a preset coverage threshold, it is determined that the coverage is qualified, and an optimized test pattern is generated. The test pattern includes test point coordinates, scanning path and detection parameters.

[0046] In some embodiments, the process of importing the optimized test graph into the layout testing system for verification, and if the verification fails, feeding back to the clustering step to correct the k value or adjust the classification model parameters and retrain, and if the verification succeeds, outputting the final test graph, includes:

[0047] When the false negative rate exceeds the preset false negative rate threshold, the test image is deemed substandard, and the sample weight of this type of false negative defect is increased to retrain the model.

[0048] When the false negative rate is less than or equal to the preset false negative rate threshold, it is determined that the false negative risk is qualified and the test time judgment process begins.

[0049] When the actual test time exceeds the preset time threshold, it is determined that the test efficiency is low. Under the premise of ensuring that the coverage of the high-frequency area of ​​the defect is greater than or equal to the preset coverage threshold, the test point density in the non-high-frequency area is reduced.

[0050] When the actual test time is less than the preset time threshold, the test pattern is deemed qualified and the final test pattern is output.

[0051] Secondly, the present invention provides a curve layout test graph generation system based on machine learning, comprising:

[0052] The acquisition module is configured to acquire historical curve map data, clean and complete the historical curve map data, and output a valid dataset containing normal map data and map data with marked defects.

[0053] An extraction module is electrically connected to the acquisition module. The extraction module is configured to extract the geometric features and defect correlation features of the curve map in the effective data set, filter highly correlated features through correlation analysis, and obtain a feature matrix through dimensionality reduction.

[0054] The modeling module is electrically connected to the extraction module. The modeling module is configured to input the feature matrix into the K-means algorithm, preset the initial number of clusters k, and after clustering, judge and correct the number of clusters based on the silhouette coefficient to obtain the defect pattern label corresponding to the highly correlated feature.

[0055] The testing module is electrically connected to the modeling module. The testing module is configured to take the feature matrix and defect pattern labels as input, output the core parameters of the test pattern through the classification model and generate an initial test pattern, adjust the test points based on redundancy, and then optimize the test point density according to the coverage for the high-frequency defect area determined based on the historical defect distribution, and output an optimized test pattern.

[0056] An optimization module is electrically connected to the testing module. The optimization module is configured to import the optimized test graph into the layout testing system for verification. If the verification fails, it will feed back to the clustering step to correct the k value or adjust the classification model parameters and retrain. If the verification succeeds, it will output the final test graph.

[0057] Compared with existing technologies, the advantages of this invention are as follows: By cleaning and completing data and accurately extracting features, combined with correlation analysis and dimensionality reduction, highly correlated features are effectively screened, providing high-quality data support for subsequent modeling and improving the accuracy of defect pattern recognition. K-means clustering is used, combined with silhouette coefficient correction to adjust the number of clusters, and manual verification ensures the reliability of defect pattern labels, making the generated test graphics more closely match the actual defect types and enhancing detection targeting. Redundancy optimization reduces overlapping test points, and coverage adjustment densifies high-frequency defect areas, reducing test redundancy, shortening test time, and lowering the risk of missed detections, achieving a balance between detection accuracy and efficiency. A model verification and iteration mechanism is introduced, dynamically adjusting the model and test graphics based on the missed detection rate and test time, continuously optimizing performance, and ensuring that the final output test graphics adapt to the needs of complex curve layout detection. Compared with traditional manual or fixed rule generation methods, this significantly improves the applicability and reliability of test graphics, providing an efficient solution for integrated circuit curve layout detection.

[0058] The above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0059] Other features and aspects of this disclosure will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0060] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0061] Figure 1 is a flowchart of the curve layout test graphic generation method based on machine learning provided in an embodiment of the present invention;

[0062] Figure 2 is a functional block diagram of the curve layout test graph generation system based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0064] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0065] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0066] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0067] Referring to Figures 1 and 2, in the first embodiment, a method for generating curve layout test graphics based on machine learning according to an embodiment of this application includes the following steps:

[0068] S100. Obtain historical curve layout data, clean and complete the historical curve layout data, and output a valid dataset containing normal layout data and marked defect layout data.

[0069] S200. Extract the geometric features and defect correlation features of the curve layout in the effective dataset, filter the highly correlated features through correlation analysis, and obtain the feature matrix through dimensionality reduction.

[0070] S300. Input the feature matrix into the K-means algorithm, preset the initial number of clusters k, and after clustering, judge and correct the number of clusters based on the silhouette coefficient to obtain the defect pattern label corresponding to the highly correlated feature.

[0071] S400. Using the feature matrix and defect pattern labels as input, the core parameters of the test pattern are output through the classification model and an initial test pattern is generated. The test points are adjusted based on redundancy. Then, for the high-frequency defect area determined based on historical defect distribution, the test point density is optimized according to the coverage and an optimized test pattern is output.

[0072] S500. Import the optimized test graph into the layout test system for verification. If the verification fails, feed back to the clustering step to correct the k value or adjust the classification model parameters and retrain. If the verification succeeds, output the final test graph.

[0073] It should be understood that the input is historical curve map data (including normal maps and maps with defects marked, and the defect markings must include type / location / size); the data is cleaned and completed through "parameter missing judgment + defect marking validity judgment" (if 1-2 parameters are missing, contour fitting is used to complete them, and ≥3 missing parameters are filtered out), and the output is a noise-free and missing effective dataset; to avoid poor data interfering with the accuracy of subsequent modeling;

[0074] First, extract the "geometric features" of the curves (center / radius for regular curves, curvature sequence for irregular curves) and "defect correlation features". Then, use the Pearson correlation coefficient to screen for highly correlated features (correlation threshold r≥0.5). Finally, use PCA to reduce dimensionality (cumulative variance contribution rate≥85%) to obtain the feature matrix. Focus on the core features that are strongly correlated with defects to reduce the curse of dimensionality and improve modeling efficiency.

[0075] The K-means algorithm is used to cluster the feature matrix, with an initial K value of 3. The K value is adjusted by the silhouette coefficient (threshold ≥ 0.6) to finally obtain 3-5 defect pattern labels (such as "curvature change zone edge deviation" and "arc connection short circuit"). Using "feature matrix + defect pattern labels" as input, a random forest model is used to output the core parameters of the test image (test point density: 2 points / mm). 2 (Scanning path: clockwise; Detection voltage: 3.3V) This transforms abstract defect patterns into quantifiable test graphic parameters, enabling intelligent parameter generation.

[0076] Calculate the initial test pattern redundancy (area of ​​overlapping coverage area / total test area), with a preset first threshold of 15%; if the redundancy > 15%, remove test points with more than 2 overlaps according to "test point coverage overlap count from high to low" until the redundancy ≤ 15%;

[0077] For the "high-frequency defect regions" obtained from clustering (e.g., curvature abrupt change areas accounting for 60% of the total defects), calculate the coverage (test coverage area / total area of ​​high-frequency regions), with a preset coverage threshold of 95%; if the coverage is <95%, increase the density by "original test point density × 1.5" (e.g., original density 2 points / mm). 2 →3 points / mm 2 Continue until the coverage is ≥95%, then output the optimized test graph.

[0078] The optimized test graphics are imported into layout testing systems such as AOI (Automated Optical Inspection). If the verification fails (e.g., false negative rate > 3% or test time > 5 minutes), the process is returned to the modeling / graphics generation stage for reprocessing. If the requirements are met (false negative rate ≤ 3% and test time ≤ 5 minutes), the final graphics are output. Iterative corrections ensure that the test graphics meet mass production requirements.

[0079] In some specific embodiments, the step of acquiring historical curve map data, cleaning and completing the historical curve map data, and outputting a valid dataset containing normal map data and map data with marked defects includes:

[0080] The layout data for annotating defects includes information on defect type, defect location, and defect size;

[0081] Read the geometric parameters of the historical curve layout data;

[0082] When one or two parameters are missing in the historical curve map data, the missing parameters are filled in based on the existing contour fitting.

[0083] If three or more parameters are missing, the sample is filtered out.

[0084] Check the defect annotation information. If two or more annotation fields are missing, the annotation is deemed invalid and the sample is filtered out.

[0085] When a field is missing one item, it is filled in based on the statistical mean of similar defects.

[0086] It should be understood that data preprocessing is the foundation for subsequent modeling and graphics generation. This step uses a dual screening logic of "parameter integrity judgment + defect annotation validity judgment" to remove noisy data and supplement key information, ensuring the "accuracy" and "completeness" of the effective dataset and avoiding modeling deviations due to missing or incorrect data.

[0087] The "map data with defects marked" must include three core pieces of information: defect type (short circuit / open circuit / edge deviation), defect location (coordinate range, such as (10,20)-(12,22) mm), and defect size (length / width, such as 0.1 mm). If any of these pieces of information are missing, further judgment is required.

[0088] Read the geometric parameters of the layout (such as the center coordinates and radius of a circle, and the curvature of sampling points for irregular curves); if 1-2 parameters are missing: use CAD tools such as AutoCAD to fit and complete the layout based on the existing outline (e.g., if the two sampling points and radius of a circle are known, deduce the center coordinates); if ≥3 parameters are missing: determine it as an "invalid sample" and filter it directly (e.g., if the center, radius, and sampling points of a circle are all missing, it cannot be fitted and completed).

[0089] Check the labeled fields (type / location / size). If ≥2 items are missing (e.g., only the defect type is labeled, but both the location and size are missing): the label is deemed invalid and the sample is filtered out. If 1 item is missing (e.g., only the defect size is missing): the label is filled in based on the statistical mean of similar defects (e.g., the mean size of an edge deviation defect is 0.1±0.02mm, and it is filled in as 0.1mm). After processing according to the above rules, a valid dataset with no missing parameters and no invalid labels is obtained, which can be directly used for subsequent feature extraction.

[0090] In this embodiment, the fitting process is as follows: First, professional CAD tools such as AutoCAD are used to identify and analyze the complete outline of the existing layout, determining the curve type (regular curves such as circles and ellipses, or irregular curves such as curves with varying curvature). Then, fitting calculations are performed based on the geometric characteristics of this type of curve and the correlation between known parameters. For example, if the data to be completed is a regular circle layout, with the coordinates of two sampling points and the radius parameter known, but the center coordinates are missing, then the geometric definition of a circle (the distance from any point on the circle to the center is equal to the radius) is used. An equation is established using the distance formula between two points and the radius constraint, and the x and y coordinates of the center are deduced. If it is an elliptical layout, with the major semi-axis and eccentricity known, but the minor semi-axis missing, then the inherent formulas for the eccentricity and the major and minor semi-axis of an ellipse (e=c / a, c) are used. 2 =a 2 -b 2 Where e is the eccentricity, a is the major semi-axis, b is the minor semi-axis, and c is the distance from the focus to the center, the missing minor semi-axis parameter is calculated by combining the fitting error control of the existing contour (usually the fitting error ≤ 0.01 mm). If it is an irregular curve, the curvature or slope information of some sampling points is known, but the curvature data of individual key sampling points is missing, then based on the characteristics of equal-interval sampling and the curvature change trend of adjacent sampling points, the missing curvature parameters are supplemented by linear interpolation or curve fitting algorithms to ensure that the supplemented parameters are consistent with the overall shape of the existing contour without significant deviation.

[0091] In some specific embodiments, the step of extracting the geometric features and defect correlation features of the curve pattern in the effective dataset, filtering highly correlated features through correlation analysis, and obtaining a feature matrix through dimensionality reduction includes:

[0092] Based on the map contour fitting results, determine whether the sample is a regular curve or an irregular curve. When the sample is a regular curve, extract the core geometric parameters.

[0093] When the sample is an irregular curve, equal-interval sampling is used to extract the curvature sequence of the sampling points and the slope difference between adjacent sampling points to form a feature vector.

[0094] The correlation between geometric features and defect types is calculated based on the Pearson correlation coefficient, and a correlation threshold is preset to retain only high correlation features whose correlation between geometric features and defect types is greater than or equal to the correlation threshold.

[0095] When the number of highly correlated features is greater than or equal to a preset threshold for the number of highly correlated features, principal component analysis is used to reduce dimensionality and determine the initial dimensionality reduction dimension.

[0096] When the number of highly correlated features is less than a preset threshold for the number of highly correlated features, process features are extracted. After extraction, if the number of highly correlated features is greater than or equal to the preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension.

[0097] It should be understood that features are the "core input" of machine learning modeling. This step, through the logic of "curve type classification → correlation filtering → feature quantity completion", prioritizes retaining features that are strongly related to defects, avoids redundant features from interfering with the model, and at the same time ensures that the feature dimensions are sufficient to support modeling, thus achieving a balance between "feature quality" and "dimensional rationality".

[0098] Classification criteria: Based on the results of the map outline fitting (e.g., if the least squares method is used for fitting, and the fitting error is ≤0.01mm, it is judged as a regular curve).

[0099] Regular curves (circles, ellipses, circular arcs): Extract core geometric parameters (circle: center coordinates (x, y), radius R; ellipse: major axis a, minor axis b, eccentricity e);

[0100] Irregular curves (variable curvature curves, free curves): Using equal-interval sampling (sampling interval 0.05mm), the curvature sequence of the sampling points (e.g., 100 sampling points correspond to 100 curvature values) and the slope difference between adjacent sampling points (e.g., the slope difference between the i-th and i+1-th sampling points) are extracted to form a feature vector;

[0101] Calculation method: Based on the Pearson correlation coefficient, the correlation between geometric features and defect types is quantified (e.g., the correlation coefficient between "number of curvature abrupt change points" and edge deviation defects is r=0.82, and the correlation coefficient between "center coordinates" and short-circuit defects is r=0.15).

[0102] Filtering threshold: The preset correlation threshold r≥0.5 retains only highly correlated features (such as retaining the number of curvature abrupt change points and the average slope difference, and removing the center coordinates of the circle).

[0103] The preset threshold for the number of highly correlated features is 5. If the number of highly correlated features after screening is ≥5 (such as the number of curvature abrupt change points, the average slope difference, the conductor width, the layer thickness, and the conductor density around the defect), the dimensionality reduction is directly entered. If it is <5 (such as only 3), process features such as "layer thickness" and "conductor width" are extracted until the number of features is ≥5.

[0104] In some specific embodiments, when the number of highly correlated features is greater than or equal to a preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension. Alternatively, when the number of highly correlated features is less than the preset threshold for the number of highly correlated features, process features are extracted to supplement the extraction. After the extraction, if the number of highly correlated features is greater than or equal to the preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension, including:

[0105] Calculate the cumulative variance contribution rate after dimensionality reduction. When the cumulative variance contribution rate is less than a preset contribution rate threshold, reduce the initial dimensionality reduction dimension.

[0106] When the cumulative variance contribution rate is greater than or equal to the preset contribution rate threshold, the initial dimensionality reduction dimension is not changed.

[0107] It should be understood that the core of PCA dimensionality reduction is to "reduce feature dimensions while retaining key information". This step judges the dimensionality reduction effect by "cumulative variance contribution rate" to avoid losing feature information due to excessive dimensionality reduction and to ensure that the features after dimensionality reduction can still support the accuracy of subsequent modeling.

[0108] For a set of highly correlated features (e.g., 5 features), the PCA algorithm is used for dimensionality reduction (e.g., initially reduced to 3 dimensions). The variance contribution rates of each principal component after dimensionality reduction are summed to obtain the cumulative variance contribution rate (e.g., the variance contribution rates of the 3-dimensional principal components are 45%, 30%, and 15%, respectively, totaling 90%).

[0109] The preset cumulative variance contribution rate threshold is 85%. If the cumulative variance contribution rate is <85% (e.g., 80% when reducing to 3 dimensions), it indicates that too much information has been lost. The dimensionality reduction dimension should be reduced (e.g., from 3 dimensions to 2 dimensions, and the cumulative variance contribution rate should be recalculated). If the cumulative variance contribution rate is ≥85% (e.g., 88% when reducing to 2 dimensions), the adjustment should be stopped, and the final dimensionality reduction dimension should be determined. The dimensionality-reduced feature matrix (e.g., a 2-dimensional feature matrix) has reduced dimensionality and retains ≥85% of the original feature information, and can be directly used for machine learning modeling.

[0110] In some specific embodiments, the feature matrix is ​​input into the K-means algorithm, an initial cluster number k is preset, and after clustering, the cluster number is determined and corrected based on the silhouette coefficient to obtain defect pattern labels corresponding to highly correlated features, including:

[0111] When the contour coefficient is less than a preset coefficient threshold, the initial cluster number K is increased to obtain the defect pattern label;

[0112] When the contour coefficient is greater than or equal to a preset coefficient threshold, the initial cluster number K value is not changed, and the defect pattern label is obtained.

[0113] When the defect pattern label is obtained, a preset number of clustered samples are randomly selected for manual verification. When the accuracy rate is less than the preset accuracy threshold, the clustering distance is adjusted.

[0114] When the accuracy is greater than or equal to the preset accuracy threshold, the cluster distance is not adjusted.

[0115] It should be understood that the purpose of K-means clustering is to "classify similar defects into the same pattern". Through the dual logic of "optimizing the number of clusters K by silhouette coefficient + ensuring clustering accuracy by manual verification", it avoids cluster overlap or classification confusion and obtains clear and reliable defect pattern labels, providing "defect type anchor points" for subsequent test parameter prediction.

[0116] Input the feature matrix, with a preset initial number of clusters K=3 (based on common defect types in curve layouts, such as edge deviation, short circuit, and open circuit); after training the K-means model, calculate the silhouette coefficient (to quantify the clustering effect, with a value range of [-1,1], the closer to 1 the better the clustering effect);

[0117] The preset silhouette coefficient threshold is 0.6. If the silhouette coefficient is <0.6 (e.g., 0.55 when K=3, indicating high cluster overlap), the K value is incremented by 1 (K=4), the clustering model is retrained and the silhouette coefficient is calculated again until the silhouette coefficient is ≥0.6 (e.g., 0.68 when K=4, indicating satisfactory clustering). If the silhouette coefficient is ≥0.6 (e.g., 0.62 when K=3), the K value is not adjusted.

[0118] Randomly select 10% of the clustered samples (e.g., 100 samples out of a total of 1000), and manually verify the accuracy of the defect pattern classification (e.g., determine whether all samples in the "curvature change zone edge deviation" category are of this defect); the preset accuracy threshold is 90%; if the accuracy is <90% (e.g., 85%), adjust the clustering distance metric (from Euclidean distance to Manhattan distance), re-cluster and verify; if the accuracy is ≥90% (e.g., 92%), do not adjust the distance; obtain 3-5 clear defect pattern labels (e.g., label 1: curvature change zone edge deviation, label 2: short circuit at arc junction, label 3: open circuit at free curve).

[0119] In some specific embodiments, the feature matrix is ​​input into the K-means algorithm, an initial cluster number k is preset, and the cluster number is determined and corrected based on the silhouette coefficient after clustering to obtain the defect pattern label corresponding to the highly correlated feature. The algorithm also includes:

[0120] The dataset containing the dimensionality reduction features, defect pattern labels, and corresponding test graph core parameters is divided into a training set and a validation set according to a certain ratio. The defect pattern labels are used to establish the correlation between the dimensionality reduction features and the test graph core parameters. A random forest model is trained based on the training set, and the model performance is evaluated using the validation set. The prediction accuracy of the validation set is calculated.

[0121] When the prediction accuracy of the validation set is greater than or equal to a preset accuracy threshold, it is deemed qualified;

[0122] When the prediction accuracy of the validation set is less than a preset accuracy threshold, the cumulative percentage of the top three features in the model output is calculated:

[0123] If the cumulative percentage is less than a preset percentage threshold, it is determined that the feature discrimination is insufficient, and high discrimination features are added. The high discrimination features include the wire density around the defect.

[0124] If the cumulative proportion is greater than or equal to the preset proportion threshold, the feature discrimination is deemed qualified, the model hyperparameters are adjusted, and the model is retrained until the prediction accuracy of the validation set meets the standard.

[0125] It should be understood that the random forest model is the core mapping tool of "feature matrix + defect pattern label → test graph parameters". Through the logic of "dataset partitioning → accuracy judgment → feature / hyperparameter optimization", it ensures the accuracy of model prediction and avoids parameter prediction deviation due to insufficient feature discrimination or inappropriate hyperparameters.

[0126] It includes dimensionality-reduced features, defect mode labels, and corresponding measured core parameters of test patterns (test point density, scan path, detection voltage); it is divided into a training set (70%, for model training) and a validation set (30%, for performance evaluation) according to 7:3; a random forest model is used (RandomForestRegressor in the Scikit-learn library is used and multi_output=True is set for multi-objective regression, and the initial hyperparameters are set as: the number of decision trees is 100, and the maximum depth is 10), and the prediction accuracy of the validation set is calculated, including: first, calculate the accuracy for each of the three output parameters. The accuracy of the test point density is 1 - min(1, |predicted value - true value| / true value), the accuracy of the scan path is 1 (when the predicted code is the same as the true code) or 0 (when they are different), and the accuracy of the detection voltage is 1 - min(1, |predicted value - true value| / 5V); the final overall accuracy is the weighted average of the three, with weights of 0.5, 0.3, and 0.2 in sequence; the preset prediction accuracy threshold for the validation set is 90%.

[0127] If the accuracy of the validation set ≥ 90% (such as 92%): Determine that the model is qualified and proceed to the subsequent pattern generation; if the accuracy of the validation set < 90% (such as 85%): Further analyze the reasons: Calculate the feature importance output by the model (such as the importance of "wire density around the defect" is 35%, the importance of "mean slope difference" is 25%, and the importance of "stacked thickness" is 15%), and count the cumulative proportion of the top 3 features. The preset cumulative proportion threshold is 60%; if the cumulative proportion < 60% (such as the top 3 features account for 55%): Determine that the feature discrimination is insufficient, and return to supplement high-discrimination features (such as "wire density around the defect" and "curve profile deviation"), and re-extract, reduce dimensions, and train after supplementation; if the cumulative proportion ≥ 60% (such as the top 3 features account for 65%): Determine that the hyperparameters are inappropriate, and use the grid search method to optimize within the preset parameter space (the number of decision trees: [100, 200, 300], the maximum depth: [10, 15, 20]), select the hyperparameter combination with the highest accuracy on the validation set, and retrain the model. This optimization process is iterated at most 5 times until the accuracy of the validation set ≥ 90%.

[0128] In some specific embodiments, taking the feature matrix and defect mode labels as inputs, outputting the core parameters of the test pattern through a classification model and generating an initial test pattern, adjusting the test points based on redundancy, and then optimizing the test point density according to the coverage for the defect high-frequency region determined based on the historical defect distribution, and outputting an optimized test pattern, including:

[0129] When the redundancy of the initial test pattern is greater than a preset first threshold, the test points in the initial test pattern are adjusted in combination with the characteristics of the corresponding defect pattern. When the redundancy of the initial test pattern is less than or equal to the preset first threshold, the coverage of the initial test pattern to the high-frequency defect region corresponding to each defect pattern is calculated. The high-frequency defect region is determined based on the historical defect distribution associated with the defect pattern label.

[0130] The defect pattern label is used to mark the priority of test points corresponding to different defect types. When the redundancy of the initial test pattern is greater than the preset first threshold, the test points are sorted from high to low according to the number of overlaps of the test points, and the key test points corresponding to the defect pattern label are retained first, while non-key test points with more than 2 overlaps are deleted in turn.

[0131] After deleting non-critical test points with more than 2 overlaps, the redundancy is recalculated. When the redundancy is less than or equal to the preset first threshold after recalculation, the distribution area of ​​each defect type is statistically analyzed based on the defect pattern label. The area where the frequency of defect occurrence is greater than the preset frequency threshold is determined as the high-frequency defect area, and the coverage of the initial test map of the high-frequency defect area is calculated.

[0132] After deleting non-critical test points with more than 2 overlaps, the redundancy is recalculated. If the redundancy is greater than a preset first threshold after recalculation, the above deletion and calculation process continues until the redundancy is less than or equal to the preset first threshold. Then, the coverage of the initial test pattern to the high-frequency area of ​​the defect is calculated, and the density of the original test points in the high-frequency area of ​​the defect is increased based on the coverage. Finally, an optimized test pattern is output.

[0133] It should be understood that excessive redundancy will lead to longer testing time and reduced mass production efficiency. This step reduces the number of repeated test areas without affecting coverage by using the logic of "redundancy calculation → overlapping test point reduction", thereby controlling redundancy within a reasonable range and balancing testing efficiency and reliability.

[0134] The core parameters of the output test pattern (e.g., test point density 2 points / mm) 2 The initial test graphic (including test point coordinates and scan path) is generated using CAD tools, with the input being the clockwise scan path.

[0135] Redundancy = Area of ​​overlapping coverage area / Total test area (e.g., the coverage areas of two test points overlap by 0.01mm) 2 The total test area is 1 mm². 2 (Then the redundancy is 1%).

[0136] First threshold: The preset redundancy first threshold is 15%.

[0137] If the initial redundancy is ≤15% (e.g., 12%): the redundancy is deemed sufficient, and the calculation of coverage proceeds directly.

[0138] If the initial redundancy is >15% (e.g., 18%): Use a greedy algorithm to optimize.

[0139] Sort all test points according to the number of overlaps in test point coverage from high to low (e.g., test point A overlaps 3 times, test point B overlaps 2 times, test point C overlaps 1 time, the sorting is A>B>C). Then, sequentially remove test points with more than 2 overlaps (e.g., remove A first, then check B). After removing each test point, recalculate the redundancy. If the recalculated redundancy is still greater than 15% (e.g., redundancy is 16% after removing A), continue removing the next test point with more than 2 overlaps (e.g., remove B). Stop removing test points when the redundancy is ≤15% (e.g., redundancy is 14% after removing B).

[0140] Output results: Test patterns with redundancy ≤15%, providing a foundation for subsequent coverage optimization.

[0141] In some specific embodiments, the process of taking the feature matrix and defect pattern labels as input, outputting core parameters of the test pattern through a classification model and generating an initial test pattern, adjusting test points based on redundancy, and then optimizing the test point density based on coverage for high-frequency defect areas determined based on historical defect distribution, and outputting an optimized test pattern, further includes:

[0142] The coverage is the ratio of the area covered by the test pattern in the high-frequency defect region to the total area of ​​that region;

[0143] When the coverage is less than the preset coverage threshold, it is determined that the risk of missed detection is high. Test points in high-frequency defect areas are densified. The coverage is recalculated after each densification until the coverage is greater than or equal to the preset coverage threshold.

[0144] When the coverage is greater than or equal to a preset coverage threshold, it is determined that the coverage is qualified, and an optimized test pattern is generated. The test pattern includes test point coordinates, scanning path and detection parameters.

[0145] It should be understood that, based on the clustering results, the defect distribution density of each region is statistically analyzed, and regions with a defect ratio of ≥50% are defined as "high-frequency defect regions" (e.g., curvature abrupt change regions account for 60% of all defects and are judged as high-frequency regions).

[0146] Calculation method: Coverage = Area of ​​the test pattern within the high-frequency region of the defect / Total area of ​​the high-frequency region of the defect (e.g., high-frequency region area 0.5mm) 2 The test pattern covers 0.46mm. 2 Then the coverage is 92%.

[0147] Coverage threshold: The preset coverage threshold is 95%.

[0148] If the coverage is ≥95% (e.g., 96%): the coverage is deemed qualified, and an optimized test pattern is generated (including test point coordinates, scan path, and detection voltage, e.g., detection voltage 3.3V).

[0149] If coverage is <95% (e.g., 92%): Increase the density of test points in high-frequency defect areas.

[0150] Encryption ratio: Encrypt by "original test point density × 1.5" (e.g., original density 2 points / mm). 2 After encryption, it is 3 points / mm 2 );

[0151] Iteration: After each encryption is completed (e.g., adding test points in high-frequency areas), the coverage is recalculated; if the coverage is still <95% (e.g., 93%) after encryption, encryption continues at the same ratio; encryption stops when the coverage is ≥95% (e.g., 95%).

[0152] Output results: Optimized test graphs with redundancy ≤15% and high-frequency area coverage ≥95%, providing input for subsequent verification.

[0153] In some specific embodiments, the optimized test graph is imported into the layout testing system for verification. If the verification fails, feedback is sent to the clustering step to correct the k-value or adjust the classification model parameters before retraining. If the verification succeeds, the final test graph is output, including:

[0154] When the false negative rate exceeds the preset false negative rate threshold, the test image is deemed substandard, and the sample weight of this type of false negative defect is increased to retrain the model.

[0155] When the false negative rate is less than or equal to the preset false negative rate threshold, it is determined that the false negative risk is qualified and the test time judgment process begins.

[0156] When the actual test time exceeds the preset time threshold, it is determined that the test efficiency is low. Under the premise of ensuring that the coverage of the high-frequency area of ​​the defect is greater than or equal to the preset coverage threshold, the test point density in the non-high-frequency area is reduced.

[0157] When the actual test time is less than the preset time threshold, the test pattern is deemed qualified and the final test pattern is output.

[0158] It should be understood that verification feedback is the "closed-loop core" of the method. Through the dual standards of "missed detection rate judgment (to ensure reliability) → test time judgment (to ensure efficiency)," it ensures that the optimized test graphics meet the mass production requirements. If they do not meet the standards, they are returned to the corresponding stage for iterative correction to prevent unqualified graphics from entering the mass production.

[0159] Optimize the import of test graphics into the layout testing system (such as the AOI system) to test curve layout samples containing known defects (such as 100 defective samples);

[0160] False negative rate calculation: False negative rate = number of undetected defects / total number of actual defects (e.g., if 4 out of 100 defects are not detected, the false negative rate is 4%); False negative rate threshold: The preset false negative rate threshold is 3%.

[0161] If the false negative rate is >3% (e.g., 4%): the test image is deemed substandard, and the system returns to the modeling module. The sample weight of this type of false negative defect is increased (e.g., from 1 to 1.5 to increase the model's attention to this defect), the model is retrained, and the image is generated.

[0162] If the false detection rate is ≤3% (e.g., 2%): the false detection risk is deemed acceptable, and the test time judgment begins;

[0163] Record the total actual testing time (e.g., 6 minutes) of the testing system for 100 layout samples.

[0164] Based on the efficiency requirements of mass production lines, the preset test time threshold is 5 minutes (i.e., the test time for 100 samples in a single batch is ≤ 5 minutes).

[0165] If the actual test time is greater than 5 minutes (e.g., 6 minutes): the test is deemed inefficient, and the system returns to the test pattern generation module. While ensuring ≥95% coverage of high-frequency defect areas, the test point density in non-high-frequency areas is reduced (e.g., from 2 points / mm). 2 Reduced to 1.5 points / mm 2 ), recalculate redundancy and coverage, and generate optimized graphics;

[0166] If the actual test time is ≤5 minutes (e.g., 4.5 minutes): the test pattern is deemed qualified;

[0167] Qualified test patterns with a false negative rate of ≤3% and a test time of ≤5 minutes can be directly used for mass production testing.

[0168] A second embodiment of a machine learning-based curve layout test graphics generation system according to an embodiment of this application includes:

[0169] The acquisition module is configured to acquire historical curve map data, clean and complete the historical curve map data, and output a valid dataset containing normal map data and map data with marked defects.

[0170] An extraction module is electrically connected to the acquisition module. The extraction module is configured to extract the geometric features and defect correlation features of the curve map in the effective data set, filter highly correlated features through correlation analysis, and obtain a feature matrix through dimensionality reduction.

[0171] The modeling module is electrically connected to the extraction module. The modeling module is configured to input the feature matrix into the K-means algorithm, preset the initial number of clusters k, and after clustering, judge and correct the number of clusters based on the silhouette coefficient to obtain the defect pattern label corresponding to the highly correlated feature.

[0172] The testing module is electrically connected to the modeling module. The testing module is configured to take the feature matrix and defect pattern labels as input, output the core parameters of the test pattern through the classification model and generate an initial test pattern, adjust the test points based on redundancy, and then optimize the test point density according to the coverage for the high-frequency defect area determined based on the historical defect distribution, and output an optimized test pattern.

[0173] An optimization module is electrically connected to the testing module. The optimization module is configured to import the optimized test graph into the layout testing system for verification. If the verification fails, it will feed back to the clustering step to correct the k value or adjust the classification model parameters and retrain. If the verification succeeds, it will output the final test graph.

[0174] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating curve layout test graphics based on machine learning, characterized in that, The process includes the following steps: acquiring historical curve map data; cleaning and completing the historical curve map data to output a preprocessed valid dataset, wherein the historical curve map data includes normal map data and map data with defects; extracting geometric features and defect correlation features of the curve maps in the valid dataset, filtering highly correlated features through correlation analysis and obtaining a feature matrix through dimensionality reduction; using the feature matrix as input, setting an initial cluster number k value, clustering the feature matrix using the K-means algorithm, calculating the silhouette coefficient after clustering, and determining whether to correct the initial cluster number k value based on the silhouette coefficient; if it is determined that the initial cluster number k value should not be corrected, or if it is determined that the initial cluster number k value should be corrected and the correction is completed, obtaining the defect pattern label corresponding to the highly correlated feature; using the feature matrix and the defect pattern label as input, outputting a test result through a classification model. The test pattern is generated using the core parameters as input. The redundancy of the initial test pattern is calculated. When the redundancy exceeds a preset first threshold, the test points in the initial test pattern are adjusted based on the characteristics of the corresponding defect pattern. When the redundancy is less than or equal to the preset first threshold, the coverage of the initial test pattern over the high-frequency defect regions corresponding to each defect pattern is calculated. The high-frequency defect regions are determined based on the historical defect distribution associated with the defect pattern labels. The coverage is used to determine whether to increase the density of the original test points in the high-frequency defect regions based on the geometric features of the corresponding defect pattern, and an optimized test pattern is output. The optimized test pattern is then imported into the layout testing system for verification. If the verification fails, the verification results are fed back to the clustering step to correct the k-value or adjust the classification model parameters before retraining. If the verification succeeds, the final test pattern is output.

2. The method for generating curve layout test graphics based on machine learning according to claim 1, characterized in that, The process involves acquiring historical curve map data, cleaning and completing the historical curve map data, and outputting a preprocessed valid dataset. The historical curve map data includes normal map data and map data with annotated defects. Specifically, the map data with annotated defects includes defect type, defect location, and defect size information. Geometric parameters of the historical curve map data are read. When one or two parameters are missing in the historical curve map data, the missing parameters are completed based on existing contour fitting. When three or more parameters are missing, the current sample is filtered. Defect annotation information is checked; when two or more annotation fields are missing, the annotation is deemed invalid, and the sample is filtered. When one annotation field is missing, it is completed based on the statistical mean of similar defects.

3. The method for generating curve layout test graphics based on machine learning according to claim 2, characterized in that, The process of extracting geometric features and defect correlation features from the concentrated curve layout of the effective dataset, filtering highly correlated features through correlation analysis, and obtaining a feature matrix through dimensionality reduction includes: determining whether a sample is a regular or irregular curve based on the layout contour fitting results; extracting core geometric parameters when the sample is a regular curve; extracting the curvature sequence of sampling points and the slope difference between adjacent sampling points using equal-interval sampling when the sample is an irregular curve, forming a feature vector; calculating the correlation between geometric features and defect types based on the Pearson correlation coefficient, and setting a correlation threshold, retaining only highly correlated features whose correlation with the defect type is greater than or equal to the correlation threshold; when the number of highly correlated features is greater than or equal to the preset high-correlation feature number threshold, using principal component analysis for dimensionality reduction to determine the initial dimensionality reduction dimension; when the number of highly correlated features is less than the preset high-correlation feature number threshold, supplementing with extracted process features; and when the number of supplemented highly correlated features is greater than or equal to the preset high-correlation feature number threshold, using principal component analysis for dimensionality reduction to determine the initial dimensionality reduction dimension.

4. The method for generating curve layout test graphics based on machine learning according to claim 3, characterized in that, When the number of highly correlated features is greater than or equal to a preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension. Alternatively, when the number of highly correlated features is less than the preset threshold for the number of highly correlated features, process features are extracted and supplemented. If the number of highly correlated features after supplementation is greater than or equal to the preset threshold for the number of highly correlated features, principal component analysis is used for dimensionality reduction to determine the initial dimensionality reduction dimension. This includes: calculating the cumulative variance contribution rate after dimensionality reduction; when the cumulative variance contribution rate is less than a preset contribution rate threshold, the initial dimensionality reduction dimension is reduced; when the cumulative variance contribution rate is greater than or equal to the preset contribution rate threshold, the initial dimensionality reduction dimension remains unchanged.

5. The method for generating curve layout test graphics based on machine learning according to claim 4, characterized in that, The feature matrix is ​​used as input, and an initial cluster number k is preset. The feature matrix is ​​clustered using the K-means algorithm. After clustering, a silhouette coefficient is calculated. Based on the silhouette coefficient, it is determined whether to modify the initial cluster number k. If it is determined that the initial cluster number k should not be modified, or if it is determined that the initial cluster number k should be modified and the modification is completed, a defect pattern label corresponding to the highly correlated feature is obtained. This includes: when the silhouette coefficient is less than a preset coefficient threshold, increasing the initial cluster number k to obtain the defect pattern label; when the silhouette coefficient is greater than or equal to the preset coefficient threshold, not changing the initial cluster number k to obtain the defect pattern label; when the defect pattern label is obtained, a preset number of clustered samples are randomly selected for manual verification. When the accuracy is less than a preset accuracy threshold, the cluster distance is adjusted; when the accuracy is greater than or equal to the preset accuracy threshold, the cluster distance is not adjusted.

6. The method for generating curve layout test graphics based on machine learning according to claim 5, characterized in that, Using the feature matrix as input, preset the initial number of clusters k, perform clustering on the feature matrix through the K-means algorithm, calculate the silhouette coefficient after clustering, and determine whether to correct the initial number of clusters k based on the silhouette coefficient. When it is determined not to correct the initial number of clusters k, or when it is determined to correct the initial number of clusters k and the correction is completed, obtain the defect mode label corresponding to the highly correlated feature. It further includes: dividing the dataset containing the dimensionality-reduced features, defect mode labels, and corresponding core parameters of the test pattern into a training set and a validation set in proportion, where the defect mode label is used to establish the correlation between the dimensionality-reduced features and the core parameters of the test pattern, training a random forest model based on the training set, evaluating the model performance using the validation set, and calculating the prediction accuracy of the validation set; when the prediction accuracy of the validation set is greater than or equal to the preset accuracy threshold, it is determined to be qualified; when the prediction accuracy of the validation set is less than the preset accuracy threshold, calculate the cumulative proportion of the top 3 items in the feature importance output by the model: if the cumulative proportion is less than the preset proportion threshold, it is determined that the feature discrimination is insufficient, and high-discrimination features are supplemented, and the high-discrimination features include the wire density around the defect; if the cumulative proportion is greater than or equal to the preset proportion threshold, it is determined that the feature discrimination is qualified, adjust the model hyperparameters, and retrain the model after adjustment until the prediction accuracy of the validation set reaches the standard.

7. The method for generating curve layout test graphics based on machine learning according to claim 6, characterized in that, The process involves taking the feature matrix and defect pattern labels as input, outputting core parameters of the test pattern through a classification model, generating an initial test pattern using these core parameters as input, calculating the redundancy of the initial test pattern, adjusting the test points in the initial test pattern based on the features of the corresponding defect pattern when the redundancy of the initial test pattern is greater than a preset first threshold, and calculating the coverage of the high-frequency defect regions corresponding to each defect pattern based on the coverage of the initial test pattern. The high-frequency defect regions are determined based on the historical defect distribution associated with the defect pattern labels, and the density of the original test points in the high-frequency defect regions is increased based on the geometric features of the corresponding defect patterns, according to the coverage. An optimized test pattern is then output. This includes: the defect pattern labels are used to mark the priority of test points corresponding to different defect types; when the redundancy of the initial test pattern is greater than the preset first threshold, the test points are adjusted according to the test points. Test points are sorted from highest to lowest overlap count, with priority given to retaining critical test points corresponding to defect pattern labels. Non-critical test points with more than two overlap counts are then sequentially removed. After removing non-critical test points with more than two overlap counts, redundancy is recalculated. If the redundancy is less than or equal to a preset first threshold after recalculation, the distribution areas of each defect type are statistically analyzed based on the defect pattern labels. Areas where the frequency of defect occurrence is greater than a preset frequency threshold are identified as high-frequency defect areas, and the coverage of the initial test pattern for these high-frequency defect areas is calculated. After removing non-critical test points with more than two overlap counts, redundancy is recalculated. If the redundancy is greater than the preset first threshold after recalculation, the above removal and calculation process continues until the redundancy is less than or equal to the preset first threshold. At this point, the coverage of the initial test pattern for the high-frequency defect areas is calculated, and the original test point density in the high-frequency defect areas is increased based on the coverage. An optimized test pattern is then output.

8. The method for generating curve layout test graphics based on machine learning according to claim 7, characterized in that, The process involves taking the feature matrix and the defect pattern labels as input, outputting core parameters of the test pattern through a classification model, generating an initial test pattern using these core parameters as input, and calculating the redundancy of the initial test pattern. If the redundancy of the initial test pattern exceeds a preset first threshold, the test points in the initial test pattern are adjusted based on the features of the corresponding defect pattern. If the redundancy of the initial test pattern is less than or equal to the preset first threshold, the coverage of the initial test pattern for the high-frequency defect regions corresponding to each defect pattern is calculated. These high-frequency defect regions are determined based on the historical defect distribution associated with the defect pattern labels, and the coverage is calculated accordingly. The method for determining whether to increase the density of original test points in the high-frequency region of a defect based on the geometric features of the corresponding defect pattern and outputting an optimized test pattern also includes: the coverage is the ratio of the area covered by the test pattern in the high-frequency region of the defect to the total area of ​​that region; when the coverage is less than a preset coverage threshold, it is determined that the risk of missed detection is high, and the test points in the high-frequency region of the defect are densified. After each densification, the coverage is recalculated until the coverage is greater than or equal to the preset coverage threshold; when the coverage is greater than or equal to the preset coverage threshold, it is determined that the coverage is qualified, and an optimized test pattern is generated. The test pattern includes test point coordinates, scanning path, and detection parameters.

9. The method for generating curve layout test graphics based on machine learning according to claim 8, characterized in that, The optimized test pattern is imported into the layout testing system for verification. If the verification fails, the k-value is corrected or the classification model parameters are adjusted based on the verification results, and the model is retrained. If the verification succeeds, the final test pattern is output. This includes: when the false negative rate is greater than a preset false negative rate threshold, the test pattern is deemed unqualified, and the sample weights of the current class of false negative defects are increased and the model is retrained; when the false negative rate is less than or equal to the preset false negative rate threshold, the false negative risk is deemed acceptable, and the test time judgment process begins; when the actual test time is greater than a preset time threshold, the test efficiency is deemed low, and the test point density in non-high frequency areas is reduced while ensuring that the coverage of the high frequency areas of the defect is greater than or equal to the preset coverage threshold; when the actual test time is less than the preset time threshold, the test pattern is deemed acceptable, and the final test pattern is output.

10. A curve layout test graph generation system based on machine learning, characterized in that, The method for generating test graphics of curve layout based on machine learning according to any one of claims 1 to 9 includes: a collection module configured to acquire historical curve layout data, clean and complete the historical curve layout data, and output a preprocessed effective dataset, wherein the historical curve layout data includes normal layout data and layout data with defects marked; an extraction module electrically connected to the collection module, configured to extract geometric features and defect correlation features of the curve layout in the effective dataset, filter highly correlated features through correlation analysis, and obtain a feature matrix through dimensionality reduction; a modeling module electrically connected to the extraction module, configured to take the feature matrix as input, preset an initial cluster number k, cluster the feature matrix using the K-means algorithm, calculate the silhouette coefficient after clustering, determine whether to correct the initial cluster number k based on the silhouette coefficient, and obtain a defect pattern label corresponding to the highly correlated feature when it is determined that the initial cluster number k should not be corrected, or when it is determined that the initial cluster number k should be corrected and the correction is completed; and a testing module, connected to the modeling module. The testing module is electrically connected to the feature matrix and the defect pattern label as input. It outputs core parameters of the test pattern through a classification model, generates an initial test pattern using these core parameters as input, calculates the redundancy of the initial test pattern, and adjusts the test points in the initial test pattern based on the features of the corresponding defect pattern when the redundancy is greater than a preset first threshold. When the redundancy is less than or equal to the preset first threshold, it calculates the coverage of the high-frequency defect regions corresponding to each defect pattern. The high-frequency defect regions are determined based on the historical defect distribution associated with the defect pattern label. The module then determines whether to increase the density of the original test points in the high-frequency defect regions based on the geometric features of the corresponding defect pattern, and outputs an optimized test pattern. An optimization module, electrically connected to the testing module, is configured to import the optimized test pattern into a layout testing system for verification. If the verification fails, the module feeds back the verification results to the clustering step to correct the k-value or adjust the classification model parameters before retraining. If the verification succeeds, the final test pattern is output.

Images