CN-122023262-A - Crystal arc-shaped defect detection method based on statistical inspection and multi-feature fusion

CN122023262ACN 122023262 ACN122023262 ACN 122023262ACN-122023262-A

Abstract

The invention discloses a wafer arc defect detection method based on statistical inspection and multi-feature fusion, and belongs to the technical field of wafer defect detection. The method comprises the steps of constructing a four-layer framework of defect classification, multi-dimensional inspection, interference elimination and mode output, selecting a curvature value, defect length, defect continuity and direction angle change rate as characteristics, realizing preliminary classification of defects through a hierarchical clustering algorithm, performing differentiated multi-dimensional special inspection on various arc defects after pre-classification, eliminating interference by combining a double-threshold screening and spatial separation mechanism, and outputting a complete defect mode report after algorithm optimization adaptation. The special inspection scheme is designed aiming at different types of defects such as arc scratches and particle clusters, the detection accuracy is improved through dynamic threshold and algorithm optimization, noise interference is effectively eliminated, an output report contains classification statistics, key parameters and process suggestions, reliable support is provided for wafer production process optimization, and pertinence and effectiveness of complex arc defect detection are improved.

Inventors

LI YUNHAO
LI KAILI
GUAN JIAN

Assignees

无锡智现未来科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (10)

1. The crystal circular arc defect detection method based on statistical inspection and multi-feature fusion is characterized by comprising the following steps of: s1, obtaining a wafer defect sample containing arc scratches, arc particle clusters, annular defect bands and segmented arcs, constructing defect classification, multi-dimensional inspection, interference elimination and mode output four-layer frames, and selecting curvature values, defect lengths, defect continuity and direction angle change rates as characteristics of the wafer arc defects; s2, carrying out preliminary classification on wafer defects through a hierarchical clustering algorithm, calculating curvature in a cluster after hierarchical clustering, and obtaining a pre-classified defect result by including parameters of curvature standard deviation and average curvature; Step S3, performing differential multidimensional special inspection on the pre-classified defects, and performing optimized distribution verification and statistical inspection on the defects; S4, combining a double-threshold screening and spatial separation mechanism to perform interference elimination on the detected defect data; And S5, outputting a complete defect mode report after algorithm optimization and adaptation processing.
2. The method for detecting the arc-shaped defects of the wafer based on the statistical test and the multi-feature fusion according to claim 1, wherein the hierarchical clustering algorithm in the step S2 is performed as follows: S21, dynamically classifying the types according to the historical defect data of the wafer through a K-means algorithm, selecting the peak change rate of the super threshold value according to the peak change rate, and determining the size of K according to the original defect characteristic mode, wherein K is the number of preset clusters, namely the total number of the types into which the target data need to be classified; s22, subdividing the large-class data output by the K-means by a Ward method, wherein the Ward method adopts a mode of combining from bottom to top and minimizing variance increment, carries out local feature enhancement on the large-class data output by the K-means, supplements local density and sub-region marking features, standardizes the features, and eliminates isolated points; S23, initializing defect points in a large class as independent sub-clusters, calculating variance increment after sub-cluster combination, and repeatedly combining the sub-clusters with the minimum cost; And S24, carrying out scale screening, space continuity check and repeated combination on the combined sub-clusters, removing invalid small clusters, distributing isolated points to obtain fine sub-clusters, and simultaneously calculating the in-class feature similarity to verify the classification validity and outputting a classification result.
3. The method for detecting the crystal circular arc defects based on the statistical test and the multi-feature fusion is characterized in that the differential multi-dimensional special test of the arc scratches in the step S3 is a multi-section fitting and curvature consistency test, a long arc is divided according to grids and then is fitted by a least square method, a fitting model comprises an arc, an elliptic arc and a parabolic arc, and when the deviation of each section of curvature and the overall average curvature is smaller than a preset threshold value, the continuous arc with consistent parameters is judged.
4. The method for detecting the arc-shaped defect of the crystal circle based on the statistical test and the multi-feature fusion according to claim 1, wherein in the step S3, the differentiated multi-dimensional special test of the arc-shaped particle cluster is a particle density and curvature association test, and when the particle distribution density is greater than a preset density threshold and the curvature variation coefficient after the particle coordinate fitting is less than a preset variation coefficient threshold, the high-density arc-shaped particle cluster is determined.
5. The method for detecting the circular arc defect of the crystal based on the statistical test and the multi-feature fusion according to claim 1, wherein in the step S3, the differentiated multi-dimensional special test of the annular defect band is a closure and radius consistency test, and when the ratio of the distances of the end points at two ends of the arc to the arc length reaches a preset closure threshold value and the ratio of the standard deviation of the annular radius to the average radius is smaller than a preset radius deviation threshold value, the complete annular defect band is judged.
6. The method for detecting the arc-shaped defects of the crystal circle based on the statistical test and the multi-feature fusion according to claim 1, wherein in the step S3, the differentiated multi-dimensional special test of the segmented arcs is a segment-segment correlation test, and when the cosine similarity of the direction angles of two adjacent segments of arcs is greater than a preset similarity threshold value and the absolute deviation of the curvature is less than a preset curvature deviation threshold value, the segmented arcs of the same source are determined.
7. The method for detecting the crystal circular arc defect based on the statistical test and the multi-feature fusion according to claim 1, wherein the optimized distribution verification and statistical test in the step S3 comprises the steps of continuously testing the newly added direction angle of the arc scratch, optimizing the arc global weighted chi-square test on the arc particle cluster, testing the distribution Krueskal-Wolis test on the defect on the inner side and the outer side of the newly added ring of the ring-shaped defect band, and testing the Bezier curve fit of the newly added segmented arc path of the segmented arc.
8. The method for detecting wafer circular arc defects based on statistical inspection and multi-feature fusion as claimed in claim 1, wherein the dual-threshold screening and spatial separation mechanism in step S4 is implemented by constructing a dual-dimensional dynamic threshold model based on the statistical distribution characteristics of wafer history normal data, wherein the threshold is adaptively adjusted along with the process parameters of wafer batch and detection area, and the threshold value of single-segment arc length is The formula is: In the middle of As the length average of the arc-shaped structure in the historical normal data, Is the standard deviation of the length of the tube, Is an adaptive coefficient; Goodness-of-fit threshold for arc The formula is: In the middle of The mean value of goodness of fit for the history normal arc, In order to fit the standard deviation of the goodness of fit, The process adaptation coefficients are used; when the detected arc length of a single segment And goodness of fit Or total number of arc-shaped defects in the same wafer area When the structure is judged to be noise and is directly filtered, the For a minimum effective number threshold based on historical defect density statistics, , In order to minimize the effective defect density, Is the detection area; The space separation mechanism adopts improved Hough transformation to extract the edge contour of the defect, introduces curvature constraint factors to optimize the arc line segment identification precision, and firstly discretizes the edge contour into a pixel point set Calculating curvature of adjacent pixel points Wherein At the level of the minimum value of the total number of the components, Is the serial number of the pixel point, represents the position identification of a single pixel point in the outline of the defect edge, takes the value as a positive integer, Is the total number of pixel points in the outline of the defect edge and is a subscript The maximum value of (2) is a positive integer, and the curvature threshold value is set Wherein Self-adaptive adjustment along with defect size, when the curvature average value of continuous pixel points And when the arc section is determined, And finally separating the straight line segment from the arc segment, and executing subsequent inspection on the arc segment only.
9. The method for detecting the crystal circular arc defects based on the statistical test and the multi-feature fusion according to claim 1, wherein the algorithm optimization adaptation in the step S5 comprises the steps of measuring feature similarity among defects by using a weighted Euclidean distance, dynamically adjusting each feature weight according to the defect type, optimizing a weighted chi-square test on discrete arc particle clusters, introducing particle size weights, and screening a best fit model by calculating normalized fit goodness of different fit models, wherein the normalized fit goodness formula is as follows: In the middle of For normalizing the fitting goodness, the value range is [0,1], and the closer to 1 is the higher the matching degree of the fitting model and the actual defect data; the total number of the wafer defect data points participating in fitting, namely the number of effective defect points in a single grid; is the first The actual observation values of the defect points comprise fitting target variables such as coordinate values, density characteristic values and the like of the defects; is the first Model predictive values of the defect points are calculated by circular arc, elliptical arc or parabolic arc models; For the average of the actual observations of all defect points, i.e ; The deviation degree of the model predicted value and the actual data is reflected for the sum of squares of the residual errors; the degree of dispersion of the actual data itself is reflected as the sum of the total squares.
10. The method for detecting the circular arc-shaped defects based on the statistical test and the multi-feature fusion according to claim 1, wherein in the step S5, the defect mode report includes defect classification statistics, a key parameter list and process association suggestions, the classification statistics defines the number, the duty ratio and the distribution area of each type of defects, the key parameter list marks the curvature range, the length range and the best fit model of each type of defects, and the process association suggestions output the targeted process investigation direction.

Description

Crystal arc-shaped defect detection method based on statistical inspection and multi-feature fusion Technical Field The invention relates to the technical field of semiconductor manufacturing and wafer defect detection, in particular to a wafer arc defect detection method based on statistical inspection and multi-feature fusion. Background In the semiconductor manufacturing industry, wafers are used as a core base component, and surface defects of the wafers directly determine the chip yield and the product reliability. Arc defects are typical and complex defect types in the wafer production process, cover various forms such as arc scratches, particle clusters, annular belts, segmented arcs and the like, have large parameter differences such as curvature, length, direction angle and the like, often present cross-grid irregular distribution characteristics, and bring great challenges to single-wafer defect detection. The precise identification of the various complex arc defects is a key link for guaranteeing the production quality of wafers and optimizing the process parameters, and has important practical significance for improving the manufacturing efficiency of semiconductors. In the prior art, related wafer defect detection schemes have been developed, but special detection for various complex arc defects still has obvious defects. For example, patent CN119540198a discloses a defect detection method and device for a wafer, in which a comprehensive defect distribution map is generated by stacking a plurality of pieces of wafer defect data, a COP defect distribution pattern is identified by using a spatial clustering algorithm, and a risk score is calculated by combining with the defect identification model to determine COP defects. However, the core focuses on COP defect detection at the batch level, special detection logic is not designed for complex arc defects with multiple types and parameters on a single crystal circle, and a specific extraction and differentiation detection strategy for the core characteristics of the arc defects is lacking, so that the problems of confusion of arc defect classification and poor suitability cannot be solved. Another patent CN113095438a discloses a wafer defect classification method, which realizes defect type identification by secondary classification based on characteristic parameters such as defect size, signal intensity value and the like and a sample image. However, the selected defect characteristic parameters do not cover key attributes (such as curvature value, direction angle change rate, defect continuity and the like) of the arc defects, the classification model is not optimized for morphological characteristics of the arc defects, different types of arc defects are difficult to distinguish, and meanwhile, an effective elimination mechanism for noise interference is lacked, so that false detection is easy to occur. Also, patent CN118096767B discloses a method for detecting defective cutting of wafer, which combines multiple classification models and machine vision techniques, and realizes the grade determination of defective cutting by contour extraction and rectangular distance calculation. However, the method is only aimed at a specific defect type of poor cutting, does not relate to special detection of various arc defects such as arc scratches, particle clusters, annular belts and the like, does not construct a multi-dimensional inspection system and algorithm optimization mechanism which are suitable for different arc defects, and cannot meet the detection requirements of various arc defects. In summary, the prior art has the following defects that firstly, a special detection framework aiming at single wafer diversity complex arc defects is lacked, the existing scheme focuses on specific defect types or batch level detection, multi-form and multi-parameter characteristic adaptation of the arc defects is insufficient, secondly, defect characteristics are selected incompletely, core differentiation characteristics of the arc defects are not covered, classification accuracy is low, a differentiation multi-dimensional detection strategy is not designed, a fixed detection mode is difficult to match detection requirements of the arc defects of different types, fourthly, an interference elimination mechanism is imperfect, noise and non-arc defect interference are prone to occur, the false detection rate is high, algorithm suitability is insufficient, optimization means aiming at characteristic weight adjustment, multi-mode fitting and the like of the arc defects are lacked, detection accuracy is limited, and thirdly, an output result is mainly simple defect judgment, a complete report comprising process association suggestions is not provided, and the supporting effect on production process optimization is weak. Therefore, a method for detecting complex arc defects of wafer diversity with high adaptability and high accuracy is needed to solve the above prob