Wafer edge collapse visual inspection method

Abstract

The present application relates to the technical field of semiconductor wafer defect detection, and particularly relates to a wafer edge collapse visual detection method. The method comprises the following steps: acquiring multi-corner edge image data and corner calibration data of a carrier bonded wafer; performing polar coordinate expansion according to the multi-corner edge image data and the corner calibration data to obtain wafer edge expansion data; performing contour layering according to the wafer edge expansion data to obtain layered edge contour data; performing topological construction on the layered edge contour data to obtain multi-layer edge topological data; performing edge level determination according to the multi-layer edge topological data to obtain edge level data; and performing defect recognition according to the edge level data to obtain collapse detection data. The present application can accurately distinguish between device wafer collapse, carrier damage and glue layer pseudo-notch, and improve the detection accuracy and stability in the multi-layer edge scene.

Description

Technical Field

[0001] This invention relates to the field of semiconductor wafer defect detection technology, and in particular to a visual detection method for wafer edge chipping. Background Technology

[0002] During wafer thinning, temporary bonding, debonding, and subsequent packaging processes, the edge areas are prone to chipping, notches, or cracks due to grinding stress, clamping action, material brittleness, and uneven local stress. For ultra-thin device wafers after thinning, edge chipping not only affects the wafer's appearance and dimensional integrity but also becomes the starting point for cracks to continue propagating into the wafer. Therefore, timely and accurate identification of wafer edge chipping while the carrier is bonded is of great significance for evaluating wafer processing quality and reducing risks in subsequent processes.

[0003] Existing wafer edge visual inspection methods typically acquire wafer edge images using a camera and then employ grayscale thresholding, edge operators, contour fitting, or image classification algorithms to identify edge defects. However, in carrier-bonded wafers, the device wafer is temporarily fixed to a glass or silicon substrate by bonding adhesive. The acquired edge images simultaneously contain the device wafer edge, substrate edge, bonding adhesive boundary, adhesive overflow boundary, and bubble boundary. These boundaries are close to, overlap, or intersect with each other in local areas, making it difficult for a single edge detection result to directly reflect the actual layer level of each contour. This leads to difficulties in accurately distinguishing between genuine wafer chipping, substrate defects, and pseudo-notches in the adhesive layer, easily resulting in false detections, missed detections, and errors in defect layer classification. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a visual detection method for chip edge chipping, thereby resolving at least one of the aforementioned technical issues.

[0005] This application provides a visual detection method for chip edge chipping, including the following steps:

[0006] S1: Acquire multi-corner edge image data and corner calibration data of the carrier bonded wafer; perform polar coordinate expansion based on the multi-corner edge image data and corner calibration data to obtain wafer edge expansion data;

[0007] S2: Perform contour layering based on the wafer edge unfolding data to obtain layered edge contour data;

[0008] S3: Perform topology construction on the layered edge contour data to obtain multi-layer edge topology data;

[0009] S4: Determine the edge level based on the multi-layer edge topology data to obtain the edge level data;

[0010] S5: Defect identification is performed based on edge-level data to obtain edge breakage detection data.

[0011] This application acquires multi-corner edge image data of carrier-bonded wafers and performs polar coordinate expansion by combining corner calibration data. This transforms the wafer edges, which are originally distributed in a ring shape and affected by the shooting angle, into a continuous and alignable angle-radial expansion structure, thereby improving the consistency and comparability of edge information at different corners. By distinguishing device wafer edges, wafer edges, and adhesive layer interference contours through contour layering, and using topology construction to characterize the intralayer continuity and interlayer correlation between each layer contour, the interference of eccentric misalignment, local warping, adhesive overflow delamination, and reflective pseudo-boundaries on the identification of true edge breakage can be reduced. By determining the actual layer to which the abnormal contour belongs through edge level determination, and then identifying defects at different levels, it is possible to effectively distinguish device wafer edge breakage, wafer defects, and adhesive layer pseudo-gaps, avoiding false detections and missed detections caused by multi-layer edge overlap in traditional single contour detection, and improving the accuracy, stability, and layer positioning capability of edge breakage detection for ultra-thin carrier-bonded wafers.

[0012] Preferably, S1 specifically comprises:

[0013] Acquire multi-corner edge image data and corner calibration data of the carrier-bonded wafer; perform coordinate calibration based on the multi-corner edge image data and corner calibration data to obtain edge calibration data; perform polar coordinate transformation based on the edge calibration data to obtain wafer edge unfolding data.

[0014] This application acquires edge image data of the carrier-bonded wafer at multiple rotation angles and performs coordinate calibration by combining the corresponding rotation angle calibration data. This establishes a unified mapping relationship between the pixel position of each frame image and the actual circumferential angle of the wafer, thereby eliminating the influence of stage rotation angle deviation, wafer center position change, and projection differences from different viewing angles on edge position characterization. Through polar coordinate transformation, the originally ring-shaped wafer edges are mapped into a strip-like structure continuously unfolding along the angular direction. This ensures that the device wafer edges, carrier edges, and adhesive layer boundaries have clear radial and circumferential positional relationships in a unified coordinate system, facilitating subsequent contour layering, cross-rotation angle correspondence, and local anomaly comparison. This not only improves the alignment accuracy and data comparability between multi-rotation edge images but also reduces the interference of edge curvature and shooting posture changes on edge chipping feature extraction.

[0015] Preferably, S2 specifically comprises:

[0016] Edge detection is performed on the wafer edge unfolding data to obtain preliminary edge data; noise is removed from the preliminary edge data to obtain edge noise reduction data; contour link construction is performed on the edge noise reduction data to obtain continuous contour data; the continuous contour data is radially sorted to obtain contour sorting data; contour layering is performed on the contour sorting data to obtain layered edge contour data.

[0017] This application utilizes edge detection on wafer edge unfolding data to extract preliminary response positions corresponding to device wafer edges, wafer edges, and adhesive layer boundaries from the angle-radial unfolding structure. Noise removal eliminates false edges formed by particle contamination, local reflections, image texture fluctuations, and isolated pixel responses, reducing interference from invalid edge points on subsequent contour analysis. Contour link construction connects discrete edge points into continuous contours based on circumferential adjacency, radial continuity, and directional consistency, repairing edge breaks caused by local weak responses while preserving the morphological changes of the true gap region. Radial sorting clarifies the internal and external positional relationships of each continuous contour at the same circumferential angle, enabling contour layering. This stablely distinguishes device wafer edges, adhesive layer interference edges, and wafer edges, reducing mis-layering problems caused by close or locally overlapping multiple edges.

[0018] Preferably, S3 specifically comprises:

[0019] Eccentric misalignment analysis is performed on the layered edge contour data to obtain misalignment contour association data; overflow fault analysis is performed on the layered edge contour data and the misalignment contour association data to obtain overflow fault association data; node relationship mapping is performed on the misalignment contour association data and the overflow fault association data to obtain edge node relationship data; topology generation is performed on the edge node relationship data to obtain multi-layer edge topology data.

[0020] This application utilizes eccentric misalignment analysis on layered edge contour data to identify regular contour relationships between device wafers and carriers caused by center offset, local misalignment, and radial spacing variations, avoiding misjudging overall assembly deviations as local edge chipping anomalies. By combining misalignment contour correlation data with adhesive overflow tomography analysis, cross-layer interference relationships formed by adhesive overflow, bubble boundaries, contour bifurcation, and local interruptions can be extracted, thus distinguishing between real wafer edges and pseudo-contours of the adhesive layer. Through node relationship mapping, continuous nodes, abnormal nodes, and corresponding nodes between layers in each contour are uniformly associated, and multi-layer edge topology data with intra-layer continuity, inter-layer correspondence, and anomaly propagation relationships are generated through topology structure. Thus, dispersed contour geometry information can be transformed into structured data with clear connectivity and hierarchical constraints.

[0021] Preferably, the eccentric misalignment analysis specifically includes:

[0022] The layered edge contour data is fitted with contour center to obtain layered contour center data; the layered contour center data is offset by calculating the center offset to obtain inter-layer center offset data; the inter-layer center offset data and the layered edge contour data are used to perform radial spacing statistics to obtain radial spacing data; the radial spacing data is used to identify eccentricity changes to obtain eccentricity misalignment feature data; the eccentricity misalignment feature data and the layered edge contour data are used to match misalignment relationships to obtain misalignment contour association data.

[0023] This application establishes geometric benchmarks for the device wafer edge, adhesive layer boundary, and wafer edge by fitting the center of the layered edge contour data, reducing the impact of local gaps, edge interruptions, and imaging noise on overall position determination. Subsequently, center offset calculation quantifies the eccentricity direction and degree of offset between different layer contours, thus initially distinguishing between overall assembly deviations and local edge anomalies during carrier bonding. By combining interlayer center offset data with radial spacing statistics, the actual distance distribution between different layer contours at various circumferential angles can be obtained. Furthermore, by identifying and extracting eccentric components that change regularly with the circumferential angle and locally abrupt misalignment components through eccentricity variation, misalignment relationships can be established by matching the eccentric misalignment characteristics with the original layered contours. This prevents layer mismatches caused by center non-coincidence, local warping, or contour overlap.

[0024] Preferably, the tomographic analysis of the overflow adhesive is as follows:

[0025] Interlayer residuals are extracted based on layered edge contour data and misaligned contour correlation data to obtain interlayer contour residual data; abnormal contours are screened from the interlayer contour residual data to obtain abnormal contour data; contour bifurcation is identified based on the abnormal contour data to obtain exudate bifurcation feature data; contour interruption is determined based on the abnormal contour data and exudate bifurcation feature data to obtain fault contour feature data; fault relationship matching is performed based on the fault contour feature data and misaligned contour correlation data to obtain exudate fault correlation data.

[0026] This application combines layered edge contour data with misaligned contour correlation data for interlayer residual extraction, which eliminates baseline differences between the device wafer and the carrier caused by overall eccentricity, normal misalignment, and slow radial changes. This allows residual contour changes to more effectively reflect local adhesive overflow, bubble boundaries, and tortuosity anomalies. Abnormal contour screening eliminates invalid fluctuations with small amplitudes, short durations, or lack of interlayer correlation, reducing false detections caused by reflective noise and local texture disturbances. Contour bifurcation identification extracts multiple contour structures extending from the adhesive layer to the edge of the device wafer or carrier, and contour interruption determination identifies continuity disruptions caused by adhesive overflow obstruction, adhesive layer tearing, and bubbles. By combining misaligned contour correlation for tortuosity relationship matching, the connection positions and propagation relationships of abnormal bifurcations and interruptions between different edge levels can be determined, forming adhesive overflow tortuosity correlation data with hierarchical affiliation and structural correlation.

[0027] Preferably, the interlayer residual extraction specifically involves:

[0028] Radial spacing is calculated based on the layered edge contour data and the staggered contour correlation data to obtain radial spacing data; eccentric period fitting is performed on the radial spacing data to obtain inter-layer eccentricity variation data; warping trend is extracted from the radial spacing data and inter-layer eccentricity variation data to obtain inter-layer warping variation data; and benchmark residuals are calculated based on the radial spacing data, inter-layer eccentricity variation data, and inter-layer warping variation data to obtain inter-layer contour residual data.

[0029] This application calculates the radial spacing at various angular positions by combining layered edge contour data and misaligned contour correlation data, which can uniformly convert the interlayer distance changes between device wafers, adhesive layers, and carriers into comparable circumferential sequence data. Periodic spacing changes caused by the offset of the device wafer and carrier centers are extracted through eccentric period fitting, thereby eliminating the interference of overall eccentricity on the judgment of local anomalies. Warping trend extraction is performed by combining radial spacing data and interlayer eccentricity change data, which can identify slow non-periodic changes caused by thinning stress, bonding temperature difference, or uneven adhesive layer thickness. Periodic eccentricity components and gradually changing warping components are subtracted by benchmark residual calculation, ensuring that the obtained interlayer contour residual mainly retains abrupt features such as adhesive overflow, delamination, local warping, and edge chipping. This avoids misjudging normal assembly deviations and continuous deformations as defects, improving the sensitivity, stability, and interlayer positioning accuracy of local anomaly contour screening.

[0030] Preferably, the contour bifurcation recognition specifically includes:

[0031] Based on the abnormal contour data, the contour endpoints are extracted to obtain abnormal contour endpoint data; neighborhood direction analysis is performed on the abnormal contour endpoint data to obtain abnormal contour direction data; branch node detection is performed on the abnormal contour direction data to obtain contour bifurcation data; and glue overflow bifurcation determination is performed on the contour bifurcation data and abnormal contour data to obtain glue overflow bifurcation feature data.

[0032] This application extracts the endpoints of local abnormal contour data to accurately locate the start and end points of abnormal contours and areas of continuity disruption, providing a clear spatial reference for subsequent bifurcation structure analysis. Through neighborhood orientation analysis, the extension direction, tangent changes, and local turning trends of each abnormal contour near its endpoints can be obtained, thus distinguishing between normal edge bending and multi-branch contour expansion. Based on abnormal contour orientation data, branch node detection can identify separation, merging, and intersection nodes of the same contour within a local area, extracting characteristic structures formed by adhesive overflow extending along the edge of the device wafer, wafer edge, or interlayer gap. Finally, combining contour bifurcation data with local abnormal contour data for adhesive overflow bifurcation determination eliminates false branches caused by particle noise, reflective edge breaks, and actual edge chipping, enabling the obtained adhesive overflow bifurcation feature data to more accurately reflect the abnormal expansion direction of the adhesive layer and cross-layer connectivity.

[0033] Preferably, S4 specifically comprises:

[0034] Node features are extracted from multi-layer edge topology data to obtain topology node feature data; inter-layer relationships are calculated based on the topology node feature data to obtain edge inter-layer relationship data; common-circle feature verification is performed on the topology node feature data and edge inter-layer relationship data to obtain edge common-circle verification data; layer affiliation matching is performed based on the edge inter-layer relationship data and edge common-circle verification data to obtain edge layer matching data; edge layer determination is performed based on the edge layer matching data to obtain edge layer data.

[0035] This application extracts node features from multi-layer edge topology data, transforming the radial position, local curvature, connection direction, continuity, and adjacent node relationships of each layer's contour nodes into calculable structural features. This preserves the differences between device wafer edges, wafer edges, and adhesive layer boundaries. Inter-layer relationship calculations establish distance correspondences, angular overlaps, and stable connection relationships between nodes at different levels. The co-circle feature is used to verify whether candidate nodes conform to the overall circumferential geometry of the corresponding wafer layer, effectively eliminating false edges formed by adhesive overflow, local reflections, and contour overlap. Based on the co-circle verification results, layer affiliation matching and edge layer determination are performed. Under conditions of eccentric misalignment, local warping, and contour interruption, the device wafer layer, adhesive layer, or wafer layer to which abnormal nodes belong can be accurately determined. This ensures that the obtained edge layer data possesses both geometric consistency and topological relevance, reducing the risk of misjudgment caused by multi-layer edge confusion.

[0036] Preferably, S5 specifically includes:

[0037] Abnormal regions are extracted from edge-level data to obtain layered abnormal region data; defect features are extracted from the layered abnormal region data to obtain layered defect feature data; defect types are determined based on the layered defect feature data to obtain defect determination data; and edge collapse results are generated based on the defect determination data to obtain edge collapse detection data.

[0038] This invention extracts abnormal regions based on edge-level data, enabling the location of contour depressions, interruptions, bifurcations, and localized missing areas in the device wafer layer, adhesive layer, and carrier layer, respectively, thus avoiding the overlapping of abnormalities at different levels. Defect feature extraction yields the notch width, radial intrusion depth, contour curvature abrupt changes, texture continuity, and interlayer correlations for each abnormal region, ensuring that defect characterization considers both geometric morphology and hierarchical structure. Defect type determination based on layered defect features effectively differentiates between device wafer edge chipping, carrier defects, adhesive layer overflow, bubble boundaries, and reflective false notches, reducing false detections caused by adhesive layer interference and multi-layer contour overlap. Based on the defect determination results, edge chipping detection data is generated, including defect type, layer level, circumferential position, and size range, thereby improving the accuracy, traceability, and completeness of result representation in carrier-bonded wafer edge defect detection.

[0039] The purpose of this invention is to acquire edge image data of carrier-bonded wafers at multiple rotation angles, and combine this with rotation angle calibration data for polar coordinate unfolding. This transforms the originally ring-shaped distribution of device wafer edges, wafer edges, and adhesive layer boundaries, which are affected by rotational attitude, into a radially unfolded structure continuously arranged along the circumferential angle, thereby providing a unified spatial representation basis for edge information at the same physical location under different rotation angles. By continuously connecting, radially sorting, and hierarchically dividing multiple similar contours in the unfolded image through contour layering, it is possible to initially distinguish between device wafer contours, wafer contours, and adhesive layer interference contours, reducing the impact of local reflections, adhesive overflow, and contour overlap on edge recognition. By characterizing the intralayer continuity, interlayer correspondence, and abnormal connection relationships between contour nodes in each layer through topology construction, complex edge changes caused by eccentric misalignment, local warping, adhesive layer bifurcation, and contour interruption can be transformed into structured relational information. By determining the edge hierarchy, combined with topological node features, interlayer distance, and cocircular consistency, the actual device wafer layer, adhesive layer, or wafer layer to which the abnormal contour belongs can be determined. Defect identification based on edge-level data can accurately distinguish between real chipping on device wafers, wafer defects, and pseudo-notches in the adhesive layer, and generate chipping detection data with defect level, circumferential position, and morphological information, thereby improving the accuracy, stability, and traceability of chipping detection on carrier-bonded ultrathin wafers. Attached Figure Description

[0040] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0041] Figure 1 A flowchart illustrating the steps of a visual detection method for chip edge chipping according to an embodiment is shown.

[0042] Figure 2 A flowchart illustrating the steps of a polar coordinate expansion method according to an embodiment is shown;

[0043] Figure 3 A flowchart illustrating the steps of a contour layering method according to one embodiment is shown.

[0044] Figure 4 A flowchart illustrating the steps of a topology construction method according to an embodiment is shown;

[0045] Figure 5 A flowchart illustrating the steps of an edge level determination method according to an embodiment is shown.

[0046] Figure 6 A flowchart illustrating the steps of a defect identification method according to an embodiment is shown. Detailed Implementation

[0047] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0048] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0049] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] A device wafer with a diameter of 200 mm and a reduced thickness of 50 μm was selected and temporarily bonded to a glass substrate with a thickness of 700 μm using bonding adhesive. The imaging calibration ratio was 2 μm / pixel. The carrier-bonded wafer was placed on a rotating stage for imaging. The angle interval is 180 frames, and the resolution is The edge image of pixels, and according to angular resolution Polar coordinate unfolding with a radial resolution of 1 pixel yields wafer edge unfolding data covering the entire circumference. Subsequently, edge detection thresholds are applied. Gray level / pixel, radial continuity threshold Pixel, Tangent Direction Threshold and angular coverage threshold Edge detection, noise removal, contour link construction, radial sorting, and contour layering are performed to extract the contours of the device wafer layer, adhesive layer, and carrier layer, respectively.

[0051] After center fitting, the center offset between the device wafer layer and the carrier layer is 8.6 pixels, and the offset direction is... Average inter-story spacing The value is 24.7 pixels, and the eccentricity variation amplitude is... It is 8.4 pixels. Single-sided pixel count. After extracting the warping trend using the cyclic sliding median, to The interval yielded continuous interlayer residuals, with a maximum residual of 10.8 pixels, exceeding the interlayer residual threshold. The system constructs a multi-layered edge topology map containing 1264 edge nodes and 2786 connecting edges, based on the pixel count and more than 5 consecutive angle points, and combines this with a co-circular residual threshold. The pixel completed edge-level determination. Detection results indicate that the anomaly in this region is located in the device wafer layer, with a maximum radial penetration depth. It has 11 pixels, or 22μm, a angular width of approximately 1.05 mm, and a missing outline area. With a pixel size of 2140 square pixels, or 8560 μm², it is determined to be a chip chip chip; at the same time, Anomalies with cross-layer bifurcation but continuous main contours of the device wafer layer and carrier layer were identified as pseudo-gap overflows of adhesive layer and were not included in the edge breakage results.

[0052] 120 carrier-bonded wafers with manually verified results were selected for inspection, including 146 device wafer edge chips, 82 adhesive layer pseudo-notches, and 39 carrier defects. This method correctly detected 142 device wafer edge chips, with a detection rate of 97.3%, and only 4 non-device wafer anomalies were misclassified as edge chips, with a misclassification rate of 3.3%. The defect level determination accuracy reached 96.6%. When using a single-contour inspection method without contour layering and topology determination, the device wafer edge chip detection rate was 89.0%, the non-device wafer anomaly misclassification rate was 19.0%, and the layer determination accuracy was 79.0%. Therefore, this method improves the edge chip detection rate by 8.3 percentage points, reduces the misclassification rate by 15.7 percentage points, and improves the defect level determination accuracy by 17.6 percentage points, effectively suppressing the interference of eccentric misalignment, local warping, and adhesive overflow pseudo-contours on the edge chip detection results.

[0053] Please see Figures 1 to 6 This application provides a visual detection method for chip edge chipping, comprising the following steps:

[0054] S1: Acquire multi-corner edge image data and corner calibration data of the carrier bonded wafer; perform polar coordinate expansion based on the multi-corner edge image data and corner calibration data to obtain wafer edge expansion data;

[0055] Specifically, the carrier-bonded wafer is placed on a rotating stage, and the rotating stage is controlled to rotate at preset angle intervals. (e.g., 2°) Rotate sequentially, and at the k-th turn... Image of the wafer edge is acquired. ,in, , Simultaneously record the rotation angles corresponding to each image. This generates multi-corner edge image data and corner calibration data. Circle fitting is performed on the outer edge pixels of the slide in each image, minimizing... Determine a unified reference center. and reference radius , The x-coordinate of the outer edge pixels of the slide. The vertical coordinate of the outer edge pixels of the slide. For pixel order, the upper limit is 1. Based on the rotation angle corresponding to the k-th frame image, perform reverse rotation correction on the image coordinates: ,in, This is the corner compensation matrix. For lateral correction coordinates, This is the vertical correction coordinate. The correction coordinates are then converted to polar coordinates: , ,in, This represents the radial distance of an edge pixel relative to the center of the reference circle. This application relates to the circumferential angle, and in the context of trigonometric functions, arc length, and integral operations, the circumferential angle is used. All values ​​are converted to radians; angle thresholds expressed in degrees are converted to radians before being used in calculations. This is done according to the preset angle resolution. (e.g., 0.1°) and radial resolution (e.g., 1 pixel), map each frame's pixels to an angle-radial grid; when multiple frames of data are mapped to the same grid position, select the pixel with the larger edge gradient magnitude as the valid pixel, and then follow... By splicing the data in an incremental order, we obtain the wafer edge unfolding data, which covers the entire circumference of the wafer, with the horizontal axis representing the circumferential angle and the vertical axis representing the radial distance.

[0056] S2: Perform contour layering based on the wafer edge unfolding data to obtain layered edge contour data;

[0057] Specifically, the wafer edge unfolding data is represented as ,in, Let r be the circumferential angle and r be the radial distance. Calculate the angular gradient at each grid location. and radial gradient And calculate the gradient magnitude. Will satisfy The grid positions are determined as preliminary edge points, where, To obtain preliminary edge data, a preset edge detection threshold, such as 30, is set based on the gradient magnitude distribution of defect-free samples. This threshold is used to collect all preliminary edge points.

[0058] Perform connected component partitioning on the initial edge data, and calculate the number of pixels in the q-th connected component. Angular span (The coverage area of ​​the q-th connected component in the circumferential angular direction) and the average gradient Will satisfy , or Isolated connected components are removed, and those with intervals not exceeding a preset compensation distance are excluded. Connect the adjacent edge points to fill in the gaps. =8, =0.5, to obtain edge noise reduction data.

[0059] Contour links are constructed from the edge noise reduction data in ascending order of circumferential angle. For angles of... edge point At the next angular position Searching for the content that satisfies and Candidate points, among which, For radial continuity thresholds, such as 3. The direction of the local tangent at the edge. Set a threshold for the tangent direction, such as 15°; select the candidate point with the smallest radial distance difference. Connection. If the number of consecutive missing angles exceeds... If the current link is terminated, multiple continuous contours are formed, resulting in continuous contour data.

[0060] For each angular position The contour numbers are sorted in ascending order according to the radial coordinates corresponding to each continuous contour to obtain the contour sequence set. This leads to the formation of contour sorting data. The radial median value of each continuous contour is then calculated. Angular coverage and the spacing between adjacent contours The radial position should be located closer to the inside, and the angular coverage should not be lower than the preset coverage threshold. The contour is divided into the device wafer edge layer; the contour that is radially outward and continuously covers the circumference is divided into the wafer edge layer; the contour located between the two, with an angular coverage of not less than 0.85 and a continuous angular span of not less than 5°, and the number of angle points with radial variation of adjacent angle positions exceeding 3 pixels, accounting for not less than 20% of the total number of effective angle points of the contour, is divided into the adhesive layer edge layer, thus obtaining layered edge contour data containing layer identifier, contour point coordinates, angular range and radial position.

[0061] S3: Perform topology construction on the layered edge contour data to obtain multi-layer edge topology data;

[0062] Specifically, the layered edge contour data is represented as a contour set. ,in, , and These represent the contour point sequences of the device wafer layer, adhesive layer, and carrier layer, respectively. Circle fitting is performed on the contours of the device wafer layer and carrier layer to obtain the circle centers. , and corresponding radius , And calculate the interlayer center offset. and offset direction At each circumferential angle Below, the radial spacing between the wafer outline and the carrier outline of the computing device. ,according to , and The continuous change relationship is identified by the periodic spacing change caused by the overall eccentricity, and the corresponding relationship is established for the contour points that satisfy the continuous angle and consistent radial change trend to obtain the staggered contour association data.

[0063] A baseline for normal inter-layer spacing is constructed based on the correlation data of staggered layer contours. Calculate the residual between the actual spacing and the reference spacing. Will satisfy And the number of consecutive angles is not less than The contour segment was identified as an abnormal segment, among which, This is the interlayer residual threshold, such as 3. The minimum number of consecutive angles, such as 5. The contour endpoints, tangent directions, and number of branches within the abnormal section are detected; when the same node connects to no fewer than three contour links, it is marked as an overflow fork node; when the angle interval between adjacent contour endpoints is greater than... When the angle is 0.5° and the radial extension direction is opposite, it is marked as a fault node, forming overflow fault correlation data.

[0064] Set each contour point or continuous contour segment as a topology node. Node attributes include their level, angle range, average radial position, local curvature, and continuity. For nodes within the same level with adjacent angles and a radial difference not exceeding [a certain threshold], [the following parameters are considered]. For nodes with a value of 3, establish intra-layer edges; for nodes with staggered correspondences between different layers, establish inter-layer corresponding edges; for nodes with overflowing, branching, or contour interruption relationships, establish abnormal association edges to obtain edge node relationship data. All topological nodes and their connecting edges are combined to form a multi-layer edge topology graph G=(V,E), where V is the set of edge nodes, and E is the set of intra-layer edges, inter-layer corresponding edges, and abnormal association edges, thus obtaining multi-layer edge topology data.

[0065] S4: Determine the edge level based on the multi-layer edge topology data to obtain the edge level data;

[0066] Specifically, the multi-layer edge topology data is represented as a topology graph G=(V,E), where, Let E be the set of edge nodes, including intra-layer connections, inter-layer corresponding edges, and abnormal association edges. For each edge node... Extract its angle range Mean radial position Local curvature Node connectivity Continuous length of the profile and radial variation This yields the topological node feature data.

[0067] Based on the topological node feature data, inter-layer relationships are calculated for nodes with overlapping angle intervals. For nodes... and Calculate the angular overlap length and radial spacing , For nodes Mean radial position For nodes Average radial position. When and At that time, the two were identified as corresponding nodes between candidate layers; the order of candidate layers for device wafer, adhesive layer, and carrier layer was established according to radial position from small to large, and the edge layer relationship data was obtained. The minimum angular overlap threshold, such as 0.5 degrees. This represents the maximum interlayer matching distance, such as 20 pixels.

[0068] By performing circle fitting on the node coordinates within the same candidate layer, the candidate circle center can be obtained. and radius Compute nodes cocircular residuals relative to the candidate circle: .when And the radius difference between this node and its adjacent nodes does not exceed When this happens, mark it as a valid cocircular node. =2; otherwise, mark it as a non-circular abnormal node and obtain edge circularity verification data.

[0069] Layer assignment matching is performed based on edge inter-layer relationship data and edge co-circle verification data. Layers with radially inner positions, stable co-circle residuals, and continuous intra-layer lengths not less than [amount missing] are assigned. Node matching at 0.5° is used as the edge node of the device wafer. The stable co-circular residual means that at least 85% of the contour points within the node angle range satisfy the co-circular residual. Each node has a radial position of at least 3 pixels, and the radial difference between adjacent contour points does not exceed 3 pixels. "Inner radial position" means that among the candidate node sets with overlapping angle intervals, the average radial position of the node is the minimum, and the difference between it and the second smallest radial position node is not less than a radial continuity threshold of 3 pixels. Nodes with outer radial positions and continuous circumferential coverage are matched as carrier edge nodes. "Continuous circumferential coverage" refers to the contour angular coverage rate. Furthermore, the number of angle points in any consecutive missing segment does not exceed 3. The radial position being "outer" means that in the set of candidate nodes with overlapping angle intervals, the average radial position of the node is at its maximum value, and the difference between it and the node with the second largest radial position is not less than the radial continuity threshold of 3 pixels. Nodes located between these two, with abnormally associated edges, or with insufficient co-circular stability are matched as adhesive layer edge nodes. If a single node simultaneously satisfies multiple hierarchical conditions, the hierarchical continuity is maintained based on its intra-layer connection relationship with its preceding and following adjacent nodes to obtain edge hierarchical matching data. Insufficient co-circular stability means that more than 15% of the contour points within the node's angle interval satisfy the co-circular residual. The edge layer data is obtained by summing the final layer identifier, angle range, radial position, co-circular state, and abnormal connection state of each node, or the radial difference between adjacent contour points exceeding 3 pixels.

[0070] S5: Defect identification is performed based on edge-level data to obtain edge breakage detection data.

[0071] Specifically, edge-level data is represented as a set of nodes. Each node contains a hierarchy identifier. Angle range Radial position Concyclic residuals Continuous and abnormal connection states. Reference edge curves are constructed for the device wafer layer, adhesive layer, and carrier layer, respectively. The reference edge curve is obtained by circular extension of the normal co-circular nodes on both sides of the abnormal node. That is, normal co-circular nodes within a preset range (e.g., 1°) outside the start and end angles of the abnormal node are extracted, and circle fitting is performed on the normal co-circular nodes to obtain the fitting circle center of the corresponding level. and the fitted radius In the abnormal angle range Inside, according to A corresponding circular arc is generated, ensuring that the generated arc maintains positional and tangential continuity with the normal contours on both sides of the abnormal region at its endpoints. This yields the reference edge curve used to characterize the defect-free theoretical edge. The radial deviation of each node relative to its respective level's reference edge is calculated. , will satisfy And the range of continuous angles is not less than The nodes are merged into anomaly regions, resulting in hierarchical anomaly region data.

[0072] Extract the angular width of the q-th anomaly region Maximum radial penetration depth Area of ​​missing outline Curvature mutation amount Continuous interruption length And abnormal connection types, to obtain hierarchical defect feature data, among which, The fitted radius is the value of the corresponding level.

[0073] Defect type is determined based on its layer and defect characteristics: when the abnormal area is located in the device wafer layer, and , Simultaneously, when the actual contour is missing from the reference edge towards the wafer interior, it is determined to be a wafer chipping issue. The missing actual contour from the reference edge towards the wafer interior refers to the actual contour's radial position being smaller than the corresponding reference edge's radial position, and the radial deviation... 1 pixel, and continuously reach 1 pixel in the angular direction When an abnormal area is located in the slide layer and causes a disruption of the slide's outer edge continuity, it is determined to be a slide defect. The disruption of the slide's outer edge continuity refers to an angular interval between adjacent effective contour endpoints of the slide layer exceeding a certain value. Furthermore, the middle region lacks any valid contour points whose radial difference from the two endpoints does not exceed 3 pixels; when the abnormal region is located in the adhesive layer, or has cross-layer bifurcated connections, but the device wafer layer and carrier layer contours remain continuous, it is determined to be adhesive overflow or adhesive layer false gaps; for regions that only exhibit increased co-circular residuals without forming contour defects, they are determined to be reflective or local deformation interference. The increased co-circular residuals refer to more than 15% of the contour points in the abnormal region satisfying the co-circular residuals. A radial deviation of more than 3 pixels, or a radial difference between adjacent contour points, constitutes a contour loss. This means that the actual contour within the abnormal region is continuously located inside the reference edge, with a radial deviation of [missing information]. Each pixel corresponds to a missing area of ​​the outline. When satisfied 1 pixel and When the area is squared, the chip edge chipping standard of the device wafer is determined; defect judgment data is obtained. Based on the defect judgment data, the target defect's level, start and end angles, maximum intrusion depth, angular width, missing area, and defect type are recorded, and adjacent defects with an angle interval of less than [missing information] are grouped together. Defects of the same type with a radius of 0.3° were merged to obtain edge breakage detection data.

[0074] Preferably, S1 specifically comprises:

[0075] S11: Acquire multi-corner edge image data and corner calibration data of the carrier bonded wafer;

[0076] Specifically, the carrier-bonded wafer is placed on a rotating stage at a preset angular interval. Rotation, at the kth turning angle Edge images are collected at the location. ,in , This yields multi-corner edge image data and corner calibration data.

[0077] S12: Perform coordinate calibration based on multi-corner edge image data and corner calibration data to obtain edge calibration data;

[0078] Specifically, a circle is fitted to the outer edge of each frame of the slide to determine the reference circle center. And rotate in the opposite direction according to the corresponding angle: , For lateral correction coordinates, For longitudinal correction coordinates, This is the corner compensation matrix. The horizontal image coordinates of the original edge pixels. The vertical image coordinates of the original edge pixels are used to obtain the edge calibration data.

[0079] S13: Perform polar coordinate conversion based on edge calibration data to obtain wafer edge unfolding data.

[0080] Specifically, polar coordinate transformation is performed on the edge calibration data: , , Let be the radial distance of the edge pixel in the k-th frame relative to the center of the reference circle. Let be the circumferential angle of the edge pixel in the k-th frame relative to the center of the reference circle. Let be the lateral correction coordinates of the edge pixels in the k-th frame. The vertical correction coordinates of the edge pixels in the k-th frame are given; the radial pixels are arranged in ascending order of angle, and the overlapping areas of adjacent corners are stitched together to obtain the wafer edge unfolding data.

[0081] Preferably, S2 specifically comprises:

[0082] S21: Perform edge detection on the wafer edge unfolding data to obtain preliminary edge data;

[0083] Specifically, the wafer edge unfolding data is represented as Calculate the angular gradient radial gradient and gradient magnitude ,Will The grid positions are determined as edge points, and preliminary edge data is obtained.

[0084] S22: Noise removal is performed based on the preliminary edge data to obtain edge noise reduction data;

[0085] Specifically, the preliminary edge data is divided into connected components, and the number of pixels is reduced. Angular span or average gradient Connected components, and fill in gaps not exceeding [a certain value]. Edge points are used to obtain edge noise reduction data.

[0086] S23: Contour link construction is performed based on edge noise reduction data to obtain continuous contour data;

[0087] Specifically, edge points are searched in ascending order of circumferential angle; when the radial difference between adjacent points does not exceed a certain threshold... The difference in tangent direction does not exceed Establish a link when the number of consecutive missing angles exceeds [a certain threshold]. When the value is 3, the link is terminated, and continuous contour data is obtained.

[0088] S24: Perform radial sorting on the continuous contour data to obtain contour sorting data;

[0089] Specifically, at every angle Below, the radial coordinates of the continuous contours are arranged from smallest to largest to form a set of contour numbers. This yields the outline sorting data.

[0090] S25: Perform contour layering based on contour sorting data to obtain layered edge contour data.

[0091] Specifically, the radial median value of each contour is statistically analyzed. and angular coverage ; will be closer to the inside and The outline is defined as the device wafer layer, the outermost and continuous outline is defined as the carrier layer, and the outline located between the device wafer layer and the carrier layer, with angular coverage <0.85, continuous angular span not less than 0.5°, and radial variation of more than 3 pixels, accounting for not less than 20% of the angle points, is defined as the adhesive layer edge layer, thus obtaining the layered edge outline data.

[0092] Preferably, S3 specifically comprises:

[0093] S31: Perform eccentric layering analysis on the layered edge contour data to obtain the layered contour association data;

[0094] Specifically, circle fitting is performed on the device wafer layer contour and the carrier layer contour in the layered edge contour data to obtain the corresponding circle center and radius; the interlayer circle center offset is calculated. and offset direction and at each circumferential angle Calculate the radial spacing below .according to , and The continuous change of angles is matched with the radial spacing change of adjacent angle positions and the change of less than 3 and the continuous maintenance of no less than 5 angle points to obtain the cross-layer contour association data.

[0095] S32: Perform overflow tomography analysis based on the layered edge contour data and the misaligned contour correlation data to obtain overflow tomography correlation data;

[0096] Specifically, interlayer reference spacing is established based on the cross-layer profile correlation data. Based on the correlation data of the staggered contours, the effective circumferential angles are extracted. Calculate the interlayer radial spacing by matching device wafer profile points and wafer profile points. , For the first The radial position of the slide layer profile at the angle. For the first The radial position of the device wafer layer profile at the angle; eccentric periodic fitting of the radial spacing sequence yields... Then adjust the eccentric spacing. Using single-sided point count Cyclic sliding median processing yields the warping curve. , For the first The inter-layer periodicity variation at the angle caused by overall eccentricity The average interlayer spacing between the device wafer layer and the carrier layer. This represents the eccentricity variation component in the horizontal direction. This represents the eccentric variation component in the vertical direction. For the first The de-eccentric spacing after subtracting the eccentric periodic component at the angle. The inter-story reference spacing is established by superimposing the eccentric period variation with the gradual warping trend: For angular positions lacking valid corresponding contour points, continuous interpolation is performed using the reference spacing between their preceding and following normal angular positions to complete the representation. Residuals are then calculated based on the layered edge contour data. Will satisfy And the number of consecutive angles is not less than The contour segments were identified as anomalous segments; the contour bifurcation, endpoint interruption and cross-layer connection relationships in the anomalous segments were marked to obtain the overflow fault correlation data.

[0097] S33: Map the node relationships based on the cross-sectional contour association data and the overflow fault association data to obtain the edge node relationship data;

[0098] Specifically, continuous contour segments in the cross-sectional contour association data and the overflow fracture association data are set as edge nodes; for those segments that are angularly adjacent and whose radial difference does not exceed a certain value within the same layer... For nodes with a value of 3, establish intra-layer connections; for nodes with misaligned layer correspondences, establish inter-layer correspondences; for nodes with glue overflow forks or contour interruptions, establish abnormal associations to obtain edge node relationship data. The presence of glue overflow forks refers to a branch node connecting at least 3 contour links, where the difference in tangent direction between any two contour links is greater than [value missing]. At least one branch extends into the adhesive layer region between the device wafer layer and the carrier layer, and the continuous angle range of this branch is not less than [missing information]. Meanwhile, the main contours of the device wafer layer and the carrier layer remain continuous within the corresponding angle range; the contour interruption refers to the angle interval between adjacent endpoints in the same contour link being greater than the interruption angle threshold. Furthermore, there are no valid edge points between the two endpoints that have a radial difference of no more than 3 pixels from the adjacent contour points.

[0099] S34: Generate a topology structure based on the edge node relationship data to obtain multi-layer edge topology data.

[0100] Specifically, all edge nodes are combined into a node set V, and the intra-layer connections, inter-layer correspondences, and abnormal associations are combined into an edge set E. Node level identifiers are set according to the device wafer layer, adhesive layer, and carrier layer respectively, and a topology graph G=(V,E) is generated to obtain multi-layer edge topology data.

[0101] Preferably, the eccentric misalignment analysis specifically includes:

[0102] By fitting the center of the contour circle to the layered edge contour data, the center of the contour circle is obtained.

[0103] Specifically, circle fitting is performed on the device wafer layer contour points and the carrier layer contour points in the layered edge contour data. For the contour points of the l-th layer... By minimizing The center of the contour circle of this layer is obtained. and the fitted radius ; center of the device wafer layer and the center of the slide layer The data are then summarized to obtain the center data of the layered contour circles.

[0104] The center offset of the layered contour circles is calculated to obtain the interlayer center offset data;

[0105] Specifically, the center offset between the device wafer layer and the carrier layer is calculated based on the layer contour center data. And calculate the offset direction. ,Will , The corresponding layer identifiers constitute the inter-layer center offset data.

[0106] Radial spacing data is obtained by statistically analyzing the interlayer center offset data and the layer edge contour data;

[0107] Specifically, the angular correspondence between the device wafer profile and the carrier profile is determined based on the interlayer center offset data; at each circumferential angle... At each location, the radial position of the device wafer layer is read. and radial position of the substrate layer Calculate the radial spacing and in accordance with The radial spacing sequence is formed by ascending order, thus obtaining the radial spacing data.

[0108] Eccentricity variation is identified from radial spacing data to obtain eccentric misalignment feature data;

[0109] Specifically, the radial spacing data is fitted with periodic variations: ,in, Let a be the average inter-story spacing, and b be the eccentricity components. Calculate the eccentricity variation amplitude. and direction of change and the actual radial spacing with Angular segments with continuous and consistent changes are marked as eccentric segments, and segments with local abrupt changes are marked as candidate segments for misalignment, thus obtaining eccentric misalignment feature data. The local abrupt change segment refers to a segment where the deviation between the actual radial spacing and the eccentric periodic change curve satisfies... Each pixel, and the misalignment deviation direction of adjacent angular positions is the same, and the deviation change satisfies the following conditions: Pixels, and this state is maintained continuously for an angle segment of no less than 5 angle points; for Establish a normal eccentric correspondence based on the angle and position.

[0110] Based on the eccentric misalignment feature data and the layered edge contour data, misalignment relationship matching is performed to obtain misalignment contour association data.

[0111] Specifically, based on the eccentric layering feature data, search for contour points in the layered edge contour data that have the same angular position or whose angular difference does not exceed a preset angular interval; when the actual radial spacing of the corresponding contour points is... When the difference is no more than 3 pixels and the radial change direction of adjacent angle positions is consistent, the misalignment correspondence between the device wafer contour points and the carrier wafer contour points is established, and the corresponding angle, radial spacing, eccentricity amplitude and misalignment state are recorded to obtain misalignment contour association data.

[0112] Preferably, the tomographic analysis of the overflow adhesive is as follows:

[0113] Interlayer residuals are extracted based on the layered edge contour data and the staggered contour correlation data to obtain interlayer contour residual data;

[0114] Specifically, the angles of each circle are read based on the correlation data of the staggered contour. Corresponding radial position of the device wafer layer and radial position of the slide layer Calculate the radial spacing Using the aforementioned eccentric periodic curve Characterizes the overall eccentricity component, and for By performing continuous trend fitting, the warping variation curve is obtained. ; Calculate interlayer residuals The interlayer profile residual data were obtained.

[0115] Abnormal contour data are obtained by filtering out abnormal contours from the interlayer contour residual data.

[0116] Specifically, the selection is based on an increasing order of circumferential angles. The residual position, and the angles are continuous with a number of continuous angles not less than The locations are merged into abnormal segments; consecutive segments with a number less than 1 are removed. The isolated residual points are identified, and the contour coordinates, residual amplitudes, and hierarchical levels corresponding to each abnormal segment are retained to obtain abnormal contour data.

[0117] Based on the abnormal contour data, contour bifurcation identification is performed to obtain the overflow bifurcation feature data.

[0118] Specifically, the contour endpoints and intersections in the abnormal contour data are extracted, and the tangent directions of adjacent contour segments at each point are calculated. A condition is defined as follows: when the same intersection point connects to at least three contour links, and the difference in tangent direction between any two links is greater than the aforementioned tangent direction threshold. When the angle is 15°, the intersection point is marked as a bifurcation node; based on the angular position, number of branches, radial extension direction and cross-layer status of the bifurcation node, the overflow bifurcation characteristic data are obtained.

[0119] Based on the abnormal contour data and the overflow bifurcation feature data, the contour interruption is determined to obtain the fault contour feature data.

[0120] Specifically, based on abnormal contour data, adjacent endpoints in the same contour link are searched, and when the angular interval between the two endpoints is greater than the interruption angle threshold... =0.5°, and the intermediate region does not satisfy the radial continuity threshold. When there are 3 valid contour points, the corresponding contour is determined to be interrupted; combined with the overflow bifurcation feature data, the interruption position is marked as adjacent to the bifurcation node, and the fault contour feature data including the interruption angle, endpoint position, interruption length and adjacent bifurcation state is obtained.

[0121] Fault relationship matching was performed based on fault contour feature data and fault-related contour data to obtain overflow fault correlation data.

[0122] Specifically, the device wafer layer, adhesive layer, or carrier layer corresponding to each fault endpoint is determined based on the fault profile correlation data; when the angle range of the fault profile feature overlaps with that of the fault profile correlation profile, and the radial distance between the endpoints does not exceed... During the process, the correlation between the fault endpoints and the corresponding layer contours is established. Abnormal relationships between the connecting device wafer layer and the adhesive layer, the adhesive layer and the substrate layer, or the simultaneous connection of the three layers are recorded separately, and their angle range, bifurcation nodes, interruption endpoints, and layer connection status are summarized to obtain the adhesive overflow fault correlation data.

[0123] Preferably, the interlayer residual extraction specifically involves:

[0124] Radial spacing data is obtained by calculating the radial spacing based on the layered edge contour data and the staggered contour correlation data;

[0125] Specifically, each effective circumferential angle is read based on the staggered contour correlation data. The corresponding device wafer layer contour points and wafer layer contour points are located, and their radial positions are obtained from the layer edge contour data. and Calculate the radial spacing: Angular positions without valid misalignment relationships are marked as invalid positions and will not participate in subsequent fitting; according to... The effective angles, radial spacing, and corresponding contour levels are summarized in ascending order to obtain the radial spacing data.

[0126] Eccentric periodic fitting was performed on the radial spacing data to obtain the interlayer eccentricity variation data;

[0127] Specifically, an eccentric periodic fitting is performed on the radial spacing data. Using the first-order circumferential periodic change caused by the offset between the device wafer and the carrier's center as the fitting benchmark, the following is constructed: And by minimizing Seeking a and b, among which, This is a set of angles with valid staggered layer correspondences. Let be the average inter-story spacing, and a and b be the eccentricity components in two orthogonal directions. The eccentricity variation curve... eccentricity amplitude and eccentric direction Encapsulation yields data on interlayer eccentricity changes.

[0128] Warping trend was extracted from radial spacing data and inter-layer eccentricity variation data to obtain inter-layer warping variation data.

[0129] Specifically, the de-eccentricity spacing is calculated based on the radial spacing data and the inter-layer eccentricity variation data: According to the preset warping trend window, in the circumferential direction... Perform sliding median extraction: ,in, The number of angle points on one side of the warp trend window is set using a first-to-last loop, and is greater than the maximum number of consecutive angle points expected for a single overflow or edge breakage anomaly, to preserve the slowly changing warp trend and suppress local abrupt changes. The corresponding angle positions are then... The data on interlayer warping changes were then compiled.

[0130] Based on the radial spacing data, inter-layer eccentricity variation data, and inter-layer warping variation data, the baseline residual is calculated to obtain the inter-layer profile residual data.

[0131] Specifically, the baseline residuals are calculated based on radial spacing data, inter-layer eccentricity variation data, and inter-layer warping variation data: .in, For angle The interlayer profile residuals after deducting the overall eccentric periodic component and the gradually varying warping component. , The corresponding device wafer contour points, wafer contour points, and layer identifiers are associated and packaged to obtain interlayer contour residual data, which serves as a basis for subsequent steps. and the number of consecutive angles is not less than Input is provided to filter out abnormal contours.

[0132] Preferably, the contour bifurcation recognition specifically includes:

[0133] Extract the endpoints of the abnormal contours from the abnormal contour data to obtain the abnormal contour endpoint data;

[0134] Specifically, for each abnormal contour link in the abnormal contour data, node connectivity is statistically analyzed. Nodes connected to only one adjacent contour point are identified as contour endpoints, while nodes connected to two adjacent contour points are identified as ordinary contour nodes. For contours with local breaks, if the angular interval between the two end nodes is greater than the interruption angle threshold... If the angle is 0.5°, then each endpoint is retained as an independent endpoint. Record the position of each endpoint. The level to which it belongs, the corresponding contour number, and the continuous sequence along the contour inwards. By identifying neighboring points, we obtain the endpoint data of the abnormal contour.

[0135] Neighborhood orientation analysis is performed on the endpoint data of abnormal contours to obtain abnormal contour orientation data;

[0136] Specifically, endpoints are read based on the abnormal contour endpoint data. and its neighboring points on the same contour link Calculate the endpoint directional vector using the most distant effective neighboring point: ,in Let be the unit direction vector of the i-th anomalous contour endpoint, used to characterize the contour originating from the endpoint. Pointing to the Neighboring contour points The direction of extension, For the endpoints of the contour, The first one selected inward along the same contour link Neighboring points, It is the Euclidean norm. The modulus is 1, and the tangent direction angles of adjacent contour segments are calculated simultaneously: ,in, Let be the tangent direction angle at the i-th contour endpoint. The radial position difference between the endpoint and its neighboring points. The difference in circumferential angle between the endpoint and its neighboring points. It is a two-parameter arctangent function. This represents the actual arc distance corresponding to the circumferential angle difference. Using radians, and recording the radial extension direction of the endpoints, the corresponding layer, and the continuity status of the contour, abnormal contour direction data is obtained.

[0137] Branch node detection is performed based on abnormal contour direction data to obtain contour bifurcation data;

[0138] Specifically, based on the abnormal contour direction data, when the angle difference does not exceed a preset angle interval and the radial distance does not exceed the aforementioned radial continuity threshold... Candidate adjacency relationships are established between endpoints of 3 pixels. If at least three abnormal contour segments are connected at the same candidate position, and the difference in tangent direction between any two contour segments satisfies... =15°, then this position is determined as a branch node; record the branch node position, number of branches, direction of each branch and hierarchical connection relationship to obtain the outline fork data.

[0139] Based on the contour bifurcation data and abnormal contour data, the overflow bifurcation is determined, and the overflow bifurcation feature data is obtained.

[0140] Specifically, based on the profile bifurcation data, the complete profile segments corresponding to each branch node are traced back from the abnormal profile data. When a branch node is located between the device wafer layer and the carrier layer, and at least one branch extends along the radial region of the adhesive layer, and the main profiles of the device wafer layer and the carrier layer remain continuous within this angular range, it is determined to be an adhesive overflow bifurcation. When the branch is only located in a single wafer layer and the profile is missing towards the wafer interior, it is not determined to be an adhesive overflow bifurcation. The angular position, number of branches, extension direction, cross-layer status, and associated profile number of the adhesive overflow bifurcation node are summarized to obtain the adhesive overflow bifurcation feature data.

[0141] Preferably, S4 specifically comprises:

[0142] S41: Extract node features from multi-layer edge topology data to obtain topology node feature data;

[0143] Specifically, multi-layer edge topology data is represented as a topology graph. ,in, E is a set of edge nodes, including intra-layer connections, inter-layer corresponding edges, and abnormal association edges. For each edge node... Read the coordinates of its contour points and extract the angle range. Mean radial position Local curvature Node connectivity Continuous length of the profile and radial variation The topology node feature data is obtained by associating the intra-layer connection state and abnormal association state of the corresponding node.

[0144] S42: Calculate the inter-layer relationships based on the topological node feature data to obtain the edge inter-layer relationship data;

[0145] Specifically, based on the topological node feature data, matching is performed on nodes in different layers that overlap in angle intervals. For nodes... and Calculate the angular overlap length: and radial spacing .when and At that time, the two were identified as candidate corresponding nodes between layers, and the inner and outer layer order, angular overlap range, and topological connection type between the nodes were recorded in ascending order of average radial position to obtain the edge layer relationship data. Among them, The minimum angular overlap threshold, such as 0.5°, This represents the maximum interlayer matching distance, such as 20.

[0146] S43: Perform co-circle feature verification on the topology node feature data and the inter-edge layer relationship data to obtain edge co-circle verification data;

[0147] Specifically, based on the inter-layer relationship data, nodes with the same radial sequence and continuously connected by intra-layer connecting edges are classified into the same candidate edge layer. Then, a circle is fitted based on the node contour coordinates in the topological node feature data to obtain the center of the (h)th candidate edge layer. and radius Compute nodes Concyclic residuals: ,in, These are the representative coordinates of the node contour points. =2, and the average radial position difference between the node and the nodes in the adjacent layer does not exceed the aforementioned radial continuity threshold. When the value is 3, the node is marked as a valid cocircular node; otherwise, it is marked as a non-cocircular abnormal node, and the edge cocircular verification data is obtained.

[0148] S44: Perform layer-level matching based on the inter-layer relationship data and edge co-circle verification data to obtain edge layer-level matching data;

[0149] Specifically, the radial layer sequence of each node is determined based on the edge inter-layer relationship data, and layer assignment matching is performed in conjunction with edge co-circle verification data. Nodes with radially inward positions, valid co-circles, and continuous contour lengths are selected. Nodes with a radius of 0.5° are matched as candidate nodes for the device wafer edge; nodes with radially outward positions, effective cocircular connections, and continuous intralayer connectivity are matched as candidate nodes for the wafer edge; nodes located between candidate nodes for the device wafer edge and candidate nodes for the wafer edge, or nodes with adhesive overflow bifurcation, abnormal delamination correlation, and insufficient cocircular stability are matched as candidate nodes for the adhesive layer edge, thus obtaining edge-level matching data.

[0150] S45: Determine the edge level based on the edge level matching data to obtain the edge level data.

[0151] Specifically, the continuity of edge-level matching data is verified. If a node matches only one candidate level, that candidate level is determined as the node's final level. If a node matches multiple candidate levels, the level that is consistent with the adjacent nodes and has a continuous radial position is selected based on the principle of maintaining the level of nodes within the adjacent layers. For nodes marked as non-circular anomalous nodes but connected to normal nodes of a certain level through an intra-layer connecting edge, that level is retained and an anomalous status indicator is added. Finally, the level indicator, angle range, average radial position, co-circular status, and anomalous association status of each node are summarized to obtain the edge-level data.

[0152] Preferably, S5 specifically includes:

[0153] S51: Extract abnormal regions based on edge-level data to obtain hierarchical abnormal region data;

[0154] Specifically, the edge nodes in the edge-level data are arranged according to the device wafer layer, adhesive layer, and carrier layer, respectively. For the edge of the l-th layer, the normal co-circular nodes on both sides of the node to be detected are extracted, and the fitting circle center of the corresponding layer is obtained based on the normal co-circular nodes. and the fitted radius Transform the abnormal region and the normal nodes on both sides to fit a circle centered at their respective hierarchical levels. A local polar coordinate system with the origin as the coordinate system, in which the following is applied: By performing circular arc extension, the reference edge curve is obtained. Compute nodes Radial deviation relative to the reference edge of the corresponding level: ,in, and These represent the circumferential angle and radial position of the node, respectively. This will satisfy... =3. The angle direction is continuous and the cumulative angle range is not less than 1. Nodes with an angle of 0.5° are merged into the same abnormal region, and their level, start and end angles, and associated nodes are recorded to obtain hierarchical abnormal region data.

[0155] S52: Extract defect features from the hierarchical abnormal region data to obtain hierarchical defect feature data;

[0156] Specifically, the angular width of the q-th hierarchical anomaly region is extracted: Maximum radial penetration depth: And the area of ​​missing contours: , The termination angle for the abnormal region. The starting angle of the abnormal region. For the radial deviation of the node, This is the fitted radius for the corresponding layer. Simultaneously, the maximum difference between the actual curvature of the anomaly region and the baseline edge curvature is calculated to obtain the curvature abrupt change. The reference edge curvature is determined based on the reference circle obtained by fitting the normal co-circular nodes on both sides of the abnormal region. Let the fitting radius of the corresponding level be... The curvature of the reference edge is: ,in, The reference edge curvature of the l-th layer, The fitting radius for the reference edge of the l-th layer is used; the continuous arc lengths without effective contour nodes are counted to obtain the continuous interruption lengths. The corresponding overflow branching, faulting, and cross-layer connection states are read to obtain delamination defect characteristic data. Among them, The fitted radius is the value of the corresponding level.

[0157] S53: Defect type determination is performed based on the layered defect feature data to obtain defect determination data;

[0158] Specifically, the defect type is determined based on the level and layered defect characteristics. When the abnormal region is located in the device wafer layer, it meets the following criteria: =5、 =20, and when the actual contour is missing from the reference edge into the wafer, it is judged as a chipped edge of the device wafer; when the abnormal area is located in the carrier layer and is accompanied by an interruption of the main contour of the carrier, it is judged as a carrier defect; when the abnormal area is located in the adhesive layer, or has a cross-layer bifurcation connection but the main contours of the device wafer layer and the carrier layer remain continuous, it is judged as adhesive overflow or adhesive layer false gap; when it only shows an increase in the co-circular residual without producing an actual contour loss, it is judged as reflection or local deformation interference, and defect judgment data is obtained.

[0159] S54: Generate edge breakage results based on defect judgment data to obtain edge breakage detection data.

[0160] Specifically, target areas identified as chipped edges on device wafers are selected from the defect assessment data, and their start and end angles are recorded. Angular width Maximum radial penetration depth Area of ​​missing outline and curvature change amount For defects belonging to the same level and with the same defect type, and with adjacent angle intervals less than [a certain value], The target area with an angle of 0.3° is merged, and the defect range is updated with the minimum starting angle, maximum ending angle and complete missing contour after merging, to obtain edge breakage detection data including edge breakage location, size, level and associated abnormal state.

[0161] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0162] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A visual detection method for chip edge chipping, characterized in that, Includes the following steps: S1: Acquire multi-corner edge image data and corner calibration data of the carrier bonded wafer; perform polar coordinate expansion based on the multi-corner edge image data and corner calibration data to obtain wafer edge expansion data; S2: Perform contour layering based on the wafer edge unfolding data to obtain layered edge contour data; S3: Perform topology construction on the layered edge contour data to obtain multi-layer edge topology data; S4: Determine the edge level based on the multi-layer edge topology data to obtain the edge level data; S5: Defect identification is performed based on edge-level data to obtain edge breakage detection data; S2 specifically involves: performing edge detection on the wafer edge unfolding data to obtain preliminary edge data; performing noise removal on the preliminary edge data to obtain edge noise reduction data; constructing contour links on the edge noise reduction data to obtain continuous contour data; performing radial sorting on the continuous contour data to obtain contour sorting data; and performing contour layering on the contour sorting data to obtain layered edge contour data. S3 specifically involves: performing eccentric misalignment analysis on the layered edge contour data to obtain misalignment contour association data; performing overflow fracture analysis based on the layered edge contour data and the misalignment contour association data to obtain overflow fracture association data; and performing node relationship mapping based on the misalignment contour association data and the overflow fracture association data to obtain edge node relationship data. The topology structure is generated based on the relationship data of edge nodes to obtain multi-layer edge topology data; The specific steps of eccentric misalignment analysis are as follows: Fitting the center of the contour to the layered edge contour data to obtain layered contour center data; calculating the center offset of the layered contour center data to obtain inter-layer center offset data; statistically analyzing the radial spacing based on the inter-layer center offset data and the layered edge contour data to obtain radial spacing data; identifying eccentricity changes in the radial spacing data to obtain eccentric misalignment feature data; and matching the misalignment relationship based on the eccentric misalignment feature data and the layered edge contour data to obtain misalignment contour association data. The specific steps of the exudate tomographic analysis are as follows: Interlayer residuals are extracted based on layered edge contour data and misaligned contour correlation data to obtain interlayer contour residual data; abnormal contours are filtered from the interlayer contour residual data to obtain abnormal contour data; contour bifurcation is identified based on the abnormal contour data to obtain exudate bifurcation feature data; contour interruption is determined based on the abnormal contour data and exudate bifurcation feature data to obtain fault contour feature data; and fault relationship matching is performed based on the fault contour feature data and misaligned contour correlation data to obtain exudate fault correlation data. S4 specifically involves extracting node features from multi-layer edge topology data to obtain topology node feature data. Inter-layer relationships are calculated based on topological node feature data to obtain edge inter-layer relationship data; Perform co-circle feature verification on topological node feature data and edge inter-layer relationship data to obtain edge co-circle verification data; perform layer affiliation matching based on edge inter-layer relationship data and edge co-circle verification data to obtain edge layer matching data; Edge level determination is performed based on edge level matching data to obtain edge level data.

2. The method for visually detecting chipping at the edge of a wafer according to claim 1, characterized in that, S1 specifically refers to: Acquire multi-corner edge image data and corner calibration data of the carrier-bonded wafer; perform coordinate calibration based on the multi-corner edge image data and corner calibration data to obtain edge calibration data; perform polar coordinate transformation based on the edge calibration data to obtain wafer edge unfolding data.

3. The method for visually detecting chipping at the edge of a wafer according to claim 1, characterized in that, Interlayer residual extraction specifically involves: Radial spacing is calculated based on the layered edge contour data and the staggered contour correlation data to obtain radial spacing data; eccentric period fitting is performed on the radial spacing data to obtain interlayer eccentricity variation data; warping trend is extracted from the radial spacing data and interlayer eccentricity variation data to obtain interlayer warping variation data. Based on the radial spacing data, inter-layer eccentricity variation data, and inter-layer warping variation data, the baseline residual is calculated to obtain the inter-layer profile residual data.

4. The method for visually detecting chipping at the edge of a wafer according to claim 1, characterized in that, Contour bifurcation recognition specifically involves: The abnormal contour endpoints are extracted from the abnormal contour data to obtain abnormal contour endpoint data; neighborhood direction analysis is performed on the abnormal contour endpoint data to obtain abnormal contour direction data; branch node detection is performed on the abnormal contour direction data to obtain contour bifurcation data; and glue overflow bifurcation determination is performed on the contour bifurcation data and abnormal contour data to obtain glue overflow bifurcation feature data.

5. The method for visually detecting chipping at the edge of a wafer according to claim 1, characterized in that, S5 specifically refers to: Anomaly regions are extracted from edge-level data to obtain hierarchical anomaly region data; defect features are extracted from the hierarchical anomaly region data to obtain hierarchical defect feature data. Defect type determination is performed based on layered defect feature data to obtain defect determination data; Based on the defect judgment data, edge breakage results are generated to obtain edge breakage detection data.

Images