Wafer sorting method based on sparse leader dynamic compensation

Abstract

The application discloses a wafer sorting method based on sparse vanguard dynamic compensation, and relates to the technical field of semiconductor wafer sorting. The method determines the row-column index and theoretical physical coordinates of each die as the position basis of each die under ideal assumption conditions through a fast pre-scanning method, selects a die as a vanguard in a global sparse and local encryption manner, determines the high-precision actual physical coordinates and basic offset vector of the vanguard in combination with an image processing technology, dynamically corrects the real-time compensation offset on the basis of the basic offset vector of the vanguard along with the sorting process, and correspondingly calculates the real-time compensation offset of other to-be-compensated dies in the wafer to compensate the theoretical physical coordinates, and then performs sorting. The method guarantees the sorting efficiency, improves the position accuracy of the dies in long-time sorting operation, and then improves the positioning success rate of the dies during sorting and the sorting yield, and has high universality and stability.

Description

Technical Field

[0001] This application relates to the field of semiconductor wafer sorting technology, and in particular to a wafer sorting method based on sparse marker dynamic compensation. Background Technology

[0002] Before semiconductor wafer sorting, a pre-scan step is needed to quickly establish the mapping relationship between the physical coordinates and row and column indices of the dies. Sorting is then performed according to the physical coordinates and row and column indices of the dies. Currently, a commonly used pre-scan method is the logic pre-scan method. This method, after knowing the physical coordinates of the first and adjacent dies, can obtain the row and column indices and physical coordinates of all dies on the wafer through logical reasoning. Sorting is then performed according to the row and column indices and physical coordinates of all dies on the wafer. For 300,000 dies in an 8-inch wafer, the logic pre-scan method can be controlled to within approximately 2 minutes, demonstrating extremely high pre-scan efficiency. While the aforementioned logic pre-scanning method has high pre-scanning efficiency, it is based on the arrangement of the dies in an ideal regular array. However, during the sorting process, the ejector pins repeatedly puncture the blue film on the back of the wafer to lift the dies, causing the blue film to shrink and deform. This phenomenon is called film shrinkage. After the blue film supporting the dies is stretched and expanded, there is generally local non-uniform film shrinkage, which causes the actual position of the dies to often deviate from the ideal regular array. The success rate of locating the dies according to the physical coordinates determined by the aforementioned logic pre-scanning method during sorting is not high, resulting in a low sorting yield.

[0003] To address the blue film shrinkage problem, some pre-scanning methods currently employ marker dies as calibration points for accuracy compensation. Examples include the method in patent publication CN117810124A, entitled "An Algorithm for F-Dieter Dies in a Sorting Machine," and the method in patent publication CN 117878046A, entitled "A Method for Compensating for Film Shrinkage with Marker Dies in a Sorting Machine." These methods utilize marker dies to achieve film shrinkage compensation to some extent, thereby improving pre-scanning accuracy. However, whether relying on pre-fabricated F-die dies for zonal coarse compensation or constructing a bilinear interpolation grid using dense ordinary dies, achieving relatively high compensation accuracy requires image acquisition and positioning of a large number of densely distributed marker dies one by one. This results in pre-scanning time for 8-inch wafers generally exceeding 8 to 10 minutes, leading to extremely low pre-scanning efficiency and directly impacting sorting efficiency. On the other hand, as the sorting process continues and the grains are removed one by one, the stress distribution of the blue film will continue to change. This local non-uniform shrinkage and dynamic accumulation effect of the blue film will cause the compensation accuracy of the existing compensation method to continue to decline, and the actual sorting accuracy and yield are not ideal. Summary of the Invention

[0004] This application addresses the aforementioned problems and technical requirements by proposing a wafer sorting method based on sparse marker dynamic compensation. The technical solution of this application is as follows: A wafer sorting method based on sparse marker dynamic compensation, comprising: Under the ideal assumption that the grains are arranged in an equally spaced array, the wafer is initially pre-scanned to determine the row and column indices of each grain in the wafer and its theoretical physical coordinates in the mechanical coordinate system. Multiple grains are uniformly selected throughout the effective region of the wafer as global benchmarks. Furthermore, multiple grains were selected as localized densification markers in areas of severe film shrinkage on the wafer. Together they form an adaptive sparse vanguard mesh; For each soldier in the adaptive sparse soldier grid Control the local vision camera to move to the vanguard Theoretical physical coordinates Local images were collected and used in conjunction with image processing techniques to identify outstanding individuals. actual physical coordinates and its relationship with theoretical physical coordinates The basic offset vector between ; During the sorting process, the base offset vector of each marker in the adaptive sparse marker grid is dynamically corrected to obtain the real-time compensation offset. Based on the real-time compensation offset of each marker, other grains in the wafer that need compensation are calculated. Real-time compensation offset Utilizing the grains to be compensated Real-time compensation offset Correcting their respective theoretical physical coordinates Physical coordinates after compensation Grain sorting is performed according to the row and column coordinates and the compensated physical coordinates of each grain.

[0005] A further technical solution involves, during the sorting process, processing any marker in the adaptive sparse marker grid. The real-time compensation offset is obtained by dynamically correcting the base offset vector. for:

[0006] in, It is the cumulative sorting time from the start of sorting to the current time. It is related to the cumulative sorting time The corresponding time accumulation coefficient, cumulative sorting time The longer the time, the higher the cumulative coefficient. The larger the value; Therefore, the model soldier This represents the ratio of the number of grains picked up and sorted in a predetermined local area of ​​the region center at the current moment to the total number of grains in that local area. It is related to the grain pickup ratio The corresponding picking influence coefficient, and the proportion of grain picking The larger the value, the greater the influence coefficient. The larger the value, the better.

[0007] A further technical solution involves calculating the real-time compensation offset of other grains in the wafer based on the real-time compensation offset of each benchmark, including: Using each marker in the adaptive sparse marker mesh as a vertex, the effective area of ​​the wafer is divided into multiple adjacent polygonal compensation units; For each die to be compensated located within the polygonal compensation unit, based on the current die to be compensated The theoretical physical coordinates of the markers at each vertex of the polygon compensation unit are obtained. The real-time compensation offset of the markers at each vertex of the current polygon compensation unit is interpolated using the bilinear interpolation algorithm to obtain the real-time compensation offset of the grain to be compensated. For each die to be compensated located in the wafer edge region but outside the polygonal compensation unit, the real-time compensation offset of the die to be compensated is obtained by weighting the real-time compensation offset of multiple benchmarks within a predetermined row and column range around the die to be compensated.

[0008] A further technical solution involves weighting the real-time compensation offsets of multiple benchmarks within a predetermined range around the grain to be compensated to obtain the real-time compensation offset of the grain to be compensated, including: Calculate the grains to be compensated Theoretical physical coordinates and its surrounding predetermined range Any model worker actual physical coordinates Euclidean distance between The real-time compensation offset of each benchmark is weighted by the reciprocal of the calculated Euclidean distance to obtain the real-time compensation offset of the grain to be compensated. :

[0009] in, It is a grain to be compensated Surrounding Pre-determined Row Range Any model worker Real-time compensation offset.

[0010] Further technical solutions include: the wafer sorting method based on sparse marker dynamic compensation also includes: During the sorting process, the sorted grains that have already been picked up are dynamically removed from the adaptive sparse grain grid. When a new region of severe shrinkage is detected, multiple grains are selected in the new region of severe shrinkage as new local densification grains. Add to the adaptive sparse vanguard mesh, and determine the newly added locally encrypted vanguards. The actual physical coordinates and the fundamental offset vector between them and the theoretical physical coordinates.

[0011] Further technical solutions include: the wafer sorting method based on sparse marker dynamic compensation also includes: During the sorting process, when the ratio of the number of grains picked up and sorted in a local area within a predetermined range to the total number of grains in the local area reaches a threshold, the local area within the predetermined range is determined to be a newly added area of ​​severe film shrinkage. Furthermore, during the sorting process, the basic offset vector of each marker in the adaptive sparse marker grid is updated synchronously at predetermined sorting intervals. When the updated basic offset vector of a marker reaches a preset offset threshold, the local area within a predetermined range centered on the marker is determined to be a newly added area of ​​severe membrane shrinkage.

[0012] A further technical solution is that, during the sorting process, the basic offset vector of each marker in the adaptive sparse marker grid is synchronously updated at predetermined sorting intervals, including: During the sorting process, at intervals of the first predetermined sorting time... Utilizing any global benchmark in an adaptive sparse benchmark grid Real-time compensation offset Correcting its own theoretical physical coordinates Physical coordinates after compensation Control the local vision camera to move to the current global benchmark. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current global benchmark. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between When a global benchmark exists When the updated base offset vector reaches the preset offset threshold and the number does not exceed the first quantity threshold, the global benchmark is determined. The predetermined range of the region center is a newly added area of ​​severe membrane shrinkage; when the updated base offset vector reaches the global benchmark threshold... When the number exceeds the first threshold, the base offset vector of all soldiers in the adaptive sparse soldier grid is updated again, and the local area within a predetermined range centered on the soldier whose updated base offset vector reaches the preset offset threshold is determined to be the newly added area of ​​severe membrane shrinkage. Furthermore, during the sorting process, every second predetermined sorting time interval... Utilizing arbitrary local densification of vanguards in an adaptive sparse vanguard mesh Real-time compensation offset Correcting their respective theoretical physical coordinates Physical coordinates after compensation Control the local vision camera to move to the current local encryption vanguard. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current local encryption leader. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between When a local encryption vanguard exists When the updated base offset vector reaches the preset offset threshold and the number does not exceed the second quantity threshold, a local encryption benchmark is determined. The predetermined range of the region center is a newly added area of ​​severe membrane shrinkage; when the updated base offset vector reaches the preset offset threshold, the local densification benchmark is established. When the number exceeds the second threshold, the base offset vectors of all soldiers in the adaptive sparse soldier grid are updated again. A predetermined range of local areas centered on soldiers whose updated base offset vectors reach the preset offset threshold is identified as newly added areas of severe membrane shrinkage. .

[0013] A further technical solution involves uniformly selecting multiple grains as global benchmarks within the effective area of ​​the wafer. Furthermore, multiple grains were selected as localized densification markers in areas of severe film shrinkage on the wafer. include: Within the entire effective area of ​​the wafer, according to row and column spacing Multiple grains are selected uniformly as global benchmarks. In regions of severe film shrinkage on the wafer, according to row and column spacing Uniformly supplementing multiple grains as local densification benchmarks The areas of severe wafer shrinkage include regions where the pin operation frequency reaches a frequency threshold as determined by the map and sorting operation path, and / or wafer edge regions.

[0014] A further technical solution involves performing an initial pre-scan on the wafer under the ideal assumption that the grains are arranged in an equally spaced array to determine the row and column indices and theoretical physical coordinates of each grain in the wafer, including: Select at least three non-collinear grains in the wafer as seed grains, and determine the row and column indices and actual physical coordinates of each type of seed grain; The actual physical coordinates at the wafer origin are obtained by fitting the actual physical coordinates of the seed grain. The row and column indices at the wafer origin are obtained by fitting the row and column indices of the seed grains. And under the ideal assumption that the grains are arranged in an equally spaced array, the mechanical coordinates of adjacent grains were calculated. Theoretical spacing in the direction and adjacent grains in the mechanical coordinate system Theoretical spacing in the direction The row and column indices are obtained by fitting. Theoretical physical coordinates in the mechanical coordinate system Initial linear mapping relationship between ; The local vision camera is moved to the actual physical coordinates of the seed grain and a local image is acquired. Image processing techniques are then used to determine the actual physical coordinates of each grain within the local image where the seed grain is located, and the coordinates of adjacent grains are calculated. Actual spacing in the direction and adjacent grains in Actual spacing in the direction ; Utilizing grains in direction and Actual spacing in the direction and By correcting the initial linear mapping relationship, the row and column indices of each grain on the wafer are obtained. With theoretical physical coordinates The global linear mapping relationship between them is ; The local vision camera is controlled to move to different positions on the wafer according to the pre-scanning path and acquire each target surface view in sequence. Based on the relative arrangement structure of the dies in each target surface view, row and column indices are assigned to each die in the target surface view, and the corresponding theoretical physical coordinates are determined by combining the global linear mapping relationship.

[0015] A further technical solution involves assigning row and column indices to each grain within the target view based on the relative arrangement of grains within each target view, and determining the corresponding theoretical physical coordinates by combining this with a global linear mapping relationship. Based on the pre-scan path, the size of a single target view, the size of a single grain, and the overlap rate between adjacent target views, estimate the row and column index range covered by each target view and the row and column index range covered by the target overlap area between adjacent target views. Based on the relative arrangement of the grains within each target view, pre-assign row and column indices to each grain within the range of row and column indices covered by the current target view; Perform deduplication on the pre-allocated row and column indices of the grains within the range of row and column indices covered by the overlapping area between adjacent target views; The theoretical physical coordinates corresponding to the row and column indices of each grain are determined based on the global linear mapping relationship.

[0016] The beneficial technical effects of this application are: This application discloses a wafer sorting method based on dynamic compensation using sparse markers. Under ideal assumptions, this method determines the row and column indices and theoretical physical coordinates of each die using a rapid pre-scan method as the basis for their position. Dies are selected as markers through global sparsity and local density. After determining the high-precision actual physical coordinates and basic offset vector of the markers using image processing technology, the method dynamically corrects the real-time compensation offset based on the markers' basic offset vector during the sorting process. Correspondingly, the real-time compensation offsets of other dies in the wafer to be compensated are calculated to compensate for their theoretical physical coordinates. This method dynamically compensates during the sorting process based on the initial pre-scan results and sorts according to the compensated physical coordinates. The initial pre-scan stage provides the basic position of the dies and ensures pre-scan efficiency. The dynamic compensation operation fully considers the non-uniform shrinkage and time accumulation effect of the blue film to compensate the position of each die. This method improves the position accuracy of each die while ensuring pre-scan efficiency, guaranteeing the positioning success rate and sorting yield in long-term sorting operations. The method exhibits strong versatility and stability.

[0017] This method directly selects ordinary wafers as benchmarks, eliminating the need for pre-defined special-structure benchmarks, pre-fabricated benchmarks with special appearances during wafer manufacturing, and pre-loading of pre-made benchmark map documents. It is adaptable to all conventional wafers and all sorting scenarios, offering wider versatility and applicability. Furthermore, this method employs a global sparse and local dense approach to benchmark selection, achieving high compensation accuracy with a smaller number of benchmarks while reducing the time spent on visual benchmark positioning, thus improving both sorting efficiency and yield.

[0018] This method considers the local non-uniform shrinkage and dynamic accumulation effect of the blue film during the sorting process to update the basic offset vector of the benchmark to obtain the real-time compensation offset. Then, based on the location of the die to be compensated, it provides two different strategies to use the real-time compensation offset of the benchmark to calculate the real-time compensation offset of other dies to be compensated, ensuring that there is a good dynamic compensation effect for dies to be compensated located in the center and edge regions of the wafer. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the wafer sorting method based on sparse marker dynamic compensation in Example 1.

[0020] Figure 2 This is a schematic diagram of an adaptive sparse vanguard mesh constructed in Example 1.

[0021] Figure 3 It is according to the method of Example 3, with Figure 2 A schematic diagram showing how each marker in the adaptive sparse marker mesh divides the effective area of ​​the wafer into multiple adjacent polygonal compensation units, with each marker as a vertex.

[0022] Figure 4 yes Figure 3 A magnified view of the area within the dashed line.

[0023] Figure 5 This is the global benchmark in Example 4. and local encryption pioneer The flowchart illustrates the process of updating the base offset vector and adding locally encrypted vanguards in the adaptive sparse vanguard mesh.

[0024] Figure 6 This is a schematic diagram of the initial pre-scan of the wafer in Example 5.

[0025] Figure 7 This is a schematic diagram showing the overlap of two adjacent target surface views during a moving pre-scan of a wafer in an example. Detailed Implementation

[0026] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0027] Example 1 discloses a wafer sorting method based on sparse marker dynamic compensation. Please refer to [link / reference]. Figure 1 The flowchart shown illustrates that this wafer sorting method includes the following steps: Step 110 involves performing an initial pre-scan on the wafer under the ideal assumption of an equally spaced array of grains to determine the row and column indices of each grain and its theoretical physical coordinates in the mechanical coordinate system. The theoretical physical coordinates determined in this step do not consider the effects of actual film shrinkage. This step aims to quickly pre-scan and determine the reference positions of all grains on the entire wafer, and can be achieved using existing logic pre-scan methods or other high-speed pre-scan methods.

[0028] In addition, to ensure the effectiveness of the initial pre-scan, the wafer's deflection angle has been corrected before the initial pre-scan, ensuring that, within the allowable error range, the wafer's row direction is aligned with the machine coordinate system. Direction, column direction along the machine coordinate system Orientation. This step can be accomplished by identifying flat edges or notches on the wafer; specific methods can be found in existing literature.

[0029] Step 120: Uniformly select multiple grains within the effective area of ​​the wafer as global benchmarks. Furthermore, multiple grains were selected as localized densification markers in areas of severe film shrinkage on the wafer. Together, they form an adaptive sparse vanguard grid.

[0030] Therefore, this application divides all dies in the wafer into two categories: benchmark dies and dies to be compensated. Benchmark dies are further divided into global benchmark dies. and local encryption pioneer Together they form an adaptive sparse vanguard mesh. The vanguards in the adaptive sparse vanguard mesh referred to in this application thereafter include all global baseline vanguards. and local encryption pioneer .

[0031] This application directly selects ordinary dies on the wafer as markers, eliminating the need to prefabricate markers with special appearances during the wafer manufacturing process or to load prefabricated marker maps. Therefore, it has high versatility and can be adapted to conventional wafers and sorting scenarios.

[0032] In addition, this application selects global benchmark pacesetters. As a global coarse-scale benchmark, and selecting localized refinement benchmarks in areas of severe membrane shrinkage. Increasing the density of markers allows for fine-tuning in localized areas, and the marker positions better reflect the actual membrane shrinkage. This enables two-level compensation, which in turn allows for achieving higher compensation accuracy with a smaller number of markers.

[0033] The areas of severe wafer shrinkage include regions where the pin operation frequency reaches a threshold, as determined by the map and sorting path, and / or wafer edge regions. The specific range of the wafer edge regions can be customized. The Bin value of each die can be determined from the map, and the current sorting logic prioritizes sorting dies with the same Bin value. The sorting path for dies with the same Bin value is also pre-planned before sorting; the sorting path is the pin operation path. Therefore, the pin operation frequency of each region can be evaluated based on the positions within the pin operation path. For example, the ratio of the number of pin operation positions within a certain range to the total number of dies covered by that range can be used as the pin operation frequency. This allows for the selection of high-frequency operation regions. The specific region division range, calculation method, and frequency threshold can all be customized. For example, in a range containing 50×50=2500 grains, there are 2000 grains with the same Bin value. According to the planned sorting operation path, the grains with the same Bin value in this area will be sorted continuously. Therefore, the number of pin operation positions in this range is 2000, and the pin operation frequency is 2000 / 2500=80%. When the frequency threshold is set to 70%, this area is the area of ​​severe film shrinkage.

[0034] In one embodiment, the row and column spacing is applied globally across the effective area of ​​the wafer. Multiple grains are selected uniformly as global benchmarks. In regions of severe film shrinkage on the wafer, according to row and column spacing Uniformly supplementing multiple grains as local densification benchmarks Within the entire effective region of the wafer, the larger row and column spacing value is used. Uniformly select global benchmark soldiers This method achieves full-wafer coverage with the fewest number of benchmarks, while the blue film deformation in the region of severe film shrinkage on the wafer is 3 to 5 times that in the central region. In the region of severe film shrinkage, the density is increased to the row and column spacing. Afterwards, local densification markers can be added as little as possible while meeting the sorting accuracy requirements.

[0035] The row and column spacing here Selected based on practical experience; for example, in one instance, row and column spacing. Taking the distance between 100 grain rows and columns as an example, when the grain size is 0.2 mm and the kerf width between grains is 0.02 mm, the adjacent global benchmark is... The physical spacing between them is approximately 22 mm. When the spacing between markers in the region of severe film shrinkage is 22 mm, the compensation error for other grains will exceed 10 μm. Therefore, in the region of severe film shrinkage, additional locally densified markers are selected. This reduces the spacing between markers in the area of ​​severe film shrinkage to 11mm, thereby lowering the compensation error for other grains to within 5μm, meeting the sorting accuracy requirements.

[0036] In one instance, such as Figure 2 As shown, the dies are arranged in an array structure on the wafer, with dicing channels between adjacent dies. The actual dies do not cover the entire wafer area, but are distributed only within the central effective region. In this example, a global benchmark is selected within the global effective region of the wafer at 100 row and column intervals. A total of 25 global benchmark soldiers were selected. Then, at every 50 rows and columns, 12 local encryption markers are selected in the edge region of the wafer. In this example, the wafer edge region includes the area of ​​the two outermost rows and columns of grains. Additionally, within the high-frequency operation area of ​​the ejector pins shown in the dashed box, three local densification markers are selected at intervals of 50 rows and columns. Therefore, a total of 25 global benchmark soldiers were selected. and 15 locally encrypted benchmarks The resulting adaptive sparse scavenger mesh contains only 40 scavengers, a relatively limited number, and still exhibits sparse characteristics. It should be noted that... Figure 2 For ease of illustration, only a schematic diagram of the grain arrangement is shown, and does not represent the actual distance between the grains.

[0037] In another embodiment, to ensure the effectiveness of the benchmarks in the adaptive sparse benchmark mesh, when a grain is selected as a candidate benchmark, the local vision camera is moved to the theoretical physical coordinates of the candidate benchmark and a local image is acquired. The grain image of the candidate benchmark is captured and matched with a pre-made standard grain image template to detect whether the candidate benchmark has quality defects. This step is performed according to existing general methods and will not be elaborated here. When the candidate benchmark has no quality defects, it is added to the adaptive sparse benchmark mesh as the final selected benchmark. When the candidate benchmark has quality defects, other grains without quality defects near the candidate benchmark are selected as the final selected benchmarks and added to the adaptive sparse benchmark mesh.

[0038] Step 130, for each soldier in the adaptive sparse soldier grid Control the local vision camera to move to the vanguard Theoretical physical coordinates Local images were collected and used in conjunction with image processing techniques to identify outstanding individuals. actual physical coordinates and its relationship with theoretical physical coordinates The basic offset vector between .

[0039] The local vision camera in this step is an inherent device in the sorting machine. While its field of view is limited, it boasts high resolution and a large depth of field, enabling high-definition imaging of each grain within its range. Typically, a local vision camera has a field of view of 10mm × 10mm and a resolution of 1024 × 1024. Taking a single grain size of 0.2mm and a kerf width of 0.02mm as an example, the field of view includes 45 × 45 = 2025 grains. Although the field of view of the local vision camera is limited, as the example above shows, a single field of view still contains a considerable number of grains. This step involves acquiring theoretical physical coordinates... When capturing local images, it is not necessary to acquire images of the entire target surface within the field of view. In one embodiment, a local image of a 3x3=9 grain area is acquired. Then, image processing techniques can be used to determine the target. actual physical coordinates This part is accomplished using existing typical methods, including matching local images with pre-made standard grain image templates to obtain the matched benchmarks. The center pixel coordinates are then determined, and based on the pre-calibration results of the local vision camera and mechanical axes, the center pixel coordinates are converted into actual physical coordinates in the mechanical coordinate system. The matching accuracy reaches the sub-pixel level, enabling the acquisition of outstanding individuals. High-precision actual physical coordinates .

[0040] Step 140: During the sorting process, the base offset vector of each marker in the adaptive sparse marker grid is dynamically corrected to obtain the real-time compensation offset. Based on the real-time compensation offset of each marker, the other grains in the wafer that need compensation are calculated. Real-time compensation offset Using the grains to be compensated Real-time compensation offset Correcting their respective theoretical physical coordinates Physical coordinates after compensation Grain sorting is performed according to the row and column coordinates and the compensated physical coordinates of each grain.

[0041] The core of this application is to use the positional offset of the marker to calculate the positional offset of all the grains in the whole wafer, so as to dynamically correct and obtain the compensated physical coordinates of each grain. However, under this framework, each marker does not always use the basic offset vector determined in the pre-scanning stage before sorting. Instead, it takes into account the non-uniform shrinkage of the blue film and the time accumulation effect, and continuously updates the real-time compensation offset on the basic offset vector as the sorting operation progresses. This achieves dynamic compensation as the sorting operation progresses, improves the compensation accuracy, and thus improves the accuracy of the compensated physical coordinates of each grain.

[0042] In summary, the method provided in this embodiment first determines the row and column indices and theoretical physical coordinates of each grain under ideal assumptions using a rapid pre-scan method. The initial pre-scan serves as the positional basis for each grain, and a benchmark is selected through global sparsity and local densification. After determining the high-precision actual physical coordinates and basic offset vector of the benchmark using image processing technology, the real-time compensation offset is obtained by dynamically correcting the benchmark's basic offset vector during the sorting process. Correspondingly, the real-time compensation offset of other grains to be compensated in the wafer is calculated to compensate for the theoretical physical coordinates. This method integrates the initial pre-scan with dynamic compensation during the sorting process. The initial pre-scan stage provides the basic position of the grains and ensures pre-scan efficiency, while the dynamic compensation operation fully considers the non-uniform shrinkage and time accumulation effect of the blue film to compensate the position of each grain. This method improves positional accuracy while ensuring pre-scan efficiency, guaranteeing the stability of coordinate accuracy during long-term sorting operations.

[0043] In Example 2, step 140 above requires dynamic correction of the basic offset vector of each marker in the adaptive sparse marker mesh to obtain real-time compensation offset as the sorting process progresses. As the sorting process continues and grains are removed one by one, the blue film stress distribution will continuously change. This local non-uniform shrinkage and dynamic accumulation effect of the blue film will cause the position of each grain to shift. Therefore, in this example, considering the influence of the local non-uniform shrinkage and dynamic accumulation effect of the blue film on the grain position, the position of any marker in the adaptive sparse marker mesh is calculated according to the following formula. The real-time compensation offset is obtained by dynamically correcting the base offset vector. :

[0044] in, It is the cumulative sorting time from the start of sorting to the current time. It is related to the cumulative sorting time The corresponding time accumulation coefficient, cumulative sorting time The longer the time, the higher the cumulative coefficient. The larger the value, the more likely the mapping relationship between the two is predetermined based on empirical values ​​or pre-defined conditions. Therefore, the model soldier This represents the ratio of the number of grains picked up and sorted in a predetermined local area of ​​the region center at the current moment to the total number of grains in that local area. It is related to the grain pickup ratio The corresponding picking influence coefficient, and the proportion of grain picking The larger the value, the greater the influence coefficient. The larger the value, the more likely the mapping relationship between the two is predetermined based on experience or pre-defined conditions.

[0045] Based on the physical law that the adhesion of the blue film decreases over time and the deformation rate increases, the cumulative sorting time is... With time cumulative factor The mapping relationship is expressed as a piecewise function. In one embodiment, the cumulative sorting time... The numerical unit is minutes (min), when the cumulative sorting time Time cumulative coefficient When the cumulative sorting time Time cumulative coefficient .

[0046] To align with the physical law that the blue film deformation accelerates nonlinearly during the grain sorting and picking process, the picking influence coefficient is... Grain Pickup Ratio The mapping relationship adopts a piecewise function form. In one embodiment, when Time corresponding ,when Time corresponding ,when Time corresponding .

[0047] For example, in one instance, the base offset vector of a certain benchmark is... As the sorting process continues, the cumulative sorting time at the current moment is determined. The percentage of grains picked up within a predetermined local area centered on the model worker. At that time, the mapping relationship is determined according to the above embodiments. , The real-time compensation offset of the model soldier at the current moment is obtained by dynamically correcting the error. .

[0048] In Example 3, after obtaining the real-time compensation offset of each benchmark in step 140 via dynamic correction, it is necessary to calculate the other dies in the wafer that need to be compensated based on the real-time compensation offset of each benchmark. Real-time compensation offset This embodiment targets the grains to be compensated at different locations. Two different computation strategies are provided: First, using each marker in the adaptive sparse marker mesh as a vertex, the effective area of ​​the wafer is divided into multiple adjacent polygonal compensation units. The specific division method can refer to existing commonly used methods; a common approach is to divide the area into multiple quadrilateral compensation units using every four neighboring markers as vertices. For example, based on... Figure 2 An example is shown in the schematic diagram of the polygonal compensation unit obtained by dividing the effective area of ​​the wafer with each marker as a vertex. Figure 3 As shown, the red lines represent the boundaries of the polygonal compensation units, formed by... Figure 3It can be seen that, due to the higher density of markers in the region of severe membrane shrinkage, the size of the polygonal compensation units obtained in the region of severe membrane shrinkage is smaller.

[0049] Most of the grains will be divided into polygonal compensation units, and each grain to be compensated within a polygonal compensation unit will be... Based on the current grains to be compensated The theoretical physical coordinates of the markers at each vertex of the polygon compensation unit are used, and the real-time compensation offset of the markers at each vertex of the polygon compensation unit is interpolated using a bilinear interpolation algorithm to obtain the grain to be compensated. The real-time compensation offset is calculated, and the specific interpolation algorithm formula is not shown here.

[0050] For example in Figure 3 In the example, a magnified view of the dashed box area is shown below. Figure 4 As shown, the grains to be compensated Located in the model , , , The polygon compensation unit is located inside the vertex, therefore based on the current grain to be compensated. and the benchmark at each vertex of the polygon compensation unit , , , The theoretical physical coordinates are used to determine the benchmarks using a bilinear interpolation algorithm. , , , The real-time compensation offset is interpolated to obtain the grain to be compensated. Real-time compensation offset.

[0051] Depend on Figure 3 and Figure 4 It can be intuitively seen that, because the benchmarks are sparsely selected, some of the dies to be compensated at the wafer edge inevitably do not lie within any arbitrary polygonal compensation unit. For example... Figure 4 In the middle, the grains to be compensated It is not located within any polygon compensation unit. In this case, the grain to be compensated cannot be obtained using the above interpolation method. The real-time compensation offset. Then, for each grain to be compensated outside the polygonal compensation unit... For the grain to be compensated Surrounding Pre-determined Row Range The real-time compensation offset of multiple benchmarks within the cell is weighted to obtain the grain to be compensated. Real-time compensation offset, the grain to be compensated Surrounding Pre-determined Row Range Custom divisions are possible. These include: Calculate the grains to be compensated Theoretical physical coordinates and its surrounding predetermined range Any model worker actual physical coordinates Euclidean distance between The real-time compensation offset of each benchmark is weighted by the reciprocal of the calculated Euclidean distance to obtain the grain to be compensated. Real-time compensation offset :

[0052] in, It is a grain to be compensated Surrounding Pre-determined Row Range Any model worker Real-time compensation offset.

[0053] For example in Figure 4 In the example, for the grain to be compensated that is not located within any polygon compensation unit... First, determine the range of the predetermined rows and columns around it. Includes model workers and model soldier Calculate the grains to be compensated respectively Theoretical physical coordinates and benchmarks and model soldier The Euclidean distance between the actual physical coordinates, and then the standardization of the standardization according to the above formula. and model soldier The real-time compensation offset is weighted to obtain the grain to be compensated. Real-time compensation offset.

[0054] In Example 4, since the selected benchmarks in this application are ordinary dies on the wafer, the benchmarks themselves will be picked up and sorted as the sorting process progresses. Therefore, during the sorting operation, it is also necessary to dynamically remove the benchmarks that have already been picked up and sorted from the adaptive sparse benchmark mesh. However, continuously removing benchmarks will make position compensation difficult. On the other hand, the operation of selecting areas of severe film shrinkage based on the Map and sorting operation path when initially selecting benchmarks also relies on experience and is not absolutely accurate and reasonable. As the sorting process progresses, the actual film shrinkage of the blue film is constantly changing. Therefore, in order to adapt to the actual film shrinkage changes and to replenish the number of benchmarks to ensure the compensation effect, new benchmarks will be dynamically added to the adaptive sparse benchmark mesh during the sorting operation. The principle of adding is: During the sorting process, the system dynamically monitors for newly identified areas of severe film shrinkage. When such areas are detected, multiple grains are selected within these areas as additional localized reinforcement markers. Add to the adaptive sparse vanguard mesh, and determine the newly added locally encrypted vanguards. The actual physical coordinates and their fundamental offset vectors relative to the theoretical physical coordinates, local encryption benchmark The selection strategy is similar to step 120, determining the local encryption vanguard. The method for determining the actual physical coordinates and the base offset vector is similar to step 130. It mainly includes two aspects: During the sorting process, when the ratio of sorted grains within a predetermined local area to the total number of grains in that local area reaches a threshold, the predetermined local area where the grain sorting ratio reaches the threshold is determined to be a newly added area of ​​severe film shrinkage. The detection frequency can be set to check for the occurrence of predetermined local areas where the grain sorting ratio reaches the threshold.

[0055] Furthermore, during the sorting process, the base offset vector of each marker in the adaptive sparse marker grid is updated synchronously at predetermined sorting intervals. When the updated base offset vector of a marker reaches a preset offset threshold, a predetermined local area centered on the marker whose updated base offset vector reaches the preset offset threshold is determined to be a newly added region of severe film shrinkage. The preset offset threshold here is an empirical value, usually taken as half the grain width.

[0056] Considering the overall benchmark and local encryption pioneer Different compensation effects are achieved; in another embodiment, different update intervals are used for the global benchmark. and local encryption pioneer The base offset vector is updated, including the following steps, please refer to [reference needed]. Figure 5 Flowchart: (1) For the global benchmark in the adaptive sparse benchmark grid

[0057] During the sorting process, at each interval of the first predetermined sorting time... Utilizing any global benchmark in an adaptive sparse benchmark grid Real-time compensation offset Correcting its own theoretical physical coordinates Obtain the compensated physical coordinates at the current moment Then, control the local vision camera to move to the current global benchmark. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current global benchmark. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between The specific method is similar to step 130 above, and will not be repeated here.

[0058] Simultaneously update global benchmarks After the base offset vector, when a global benchmark exists... When the updated base offset vector reaches a preset offset threshold, it is further determined whether the number reaches a first quantity threshold, which is a custom value. When the number does not exceed the first quantity threshold, a global benchmark soldier whose updated base offset vector reaches the preset offset threshold is identified. The predetermined area within the regional center is a newly added region of intense membrane contraction, and new local densification markers are added. .

[0059] When the updated base offset vector reaches the preset offset threshold, the global benchmark... When the number exceeds a first threshold, a global rescan of the benchmarks is first triggered to update the base offset vectors of all benchmarks in the adaptive sparse benchmark grid, avoiding large-scale sorting errors caused by overall positional shifts. Then, the global baseline benchmark is considered simultaneously. and local encryption pioneer The positional offset change determines the local area centered on the marker whose updated base offset vector reaches the preset offset threshold. This area is a newly added region of severe membrane contraction.

[0060] (2) For locally denser vanguards in an adaptive sparse vanguard mesh

[0061] During the sorting process, at intervals of the second predetermined sorting time... Utilizing arbitrary local densification of vanguards in an adaptive sparse vanguard mesh Real-time compensation offset Correcting their respective theoretical physical coordinates Obtain the compensated physical coordinates at the current moment Control the local vision camera to move to the current local encryption vanguard. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current local encryption leader. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between The specific method is similar to step 130 above, and will not be repeated here.

[0062] The above-mentioned second predetermined sorting time That is, the global benchmark. The base offset vector is updated at a relatively low frequency, while the local encryption benchmark is updated. The base offset vector is updated at a higher frequency to improve local membrane shrinkage adaptation. First predetermined sorting time. The value range is usually 10 minutes to 15 minutes, corresponding to the second predetermined sorting time. The value range is usually 5 minutes to 10 minutes.

[0063] Similarly, in the synchronous update of local encryption benchmarks After the base offset vector, when there is a local encryption vanguard When the updated base offset vector reaches a preset offset threshold, it is further determined whether the quantity reaches a second quantity threshold, which is customizable. When the quantity does not exceed the second quantity threshold, local encryption vanguards that have reached the preset offset threshold with the updated base offset vector are identified. The predetermined area within the regional center is a newly added region of intense membrane contraction, and new local densification markers are added. When the number reaches the second threshold, a global rescan of the benchmarks is triggered, updating the base offset vectors of all benchmarks in the adaptive sparse benchmark grid to avoid large-scale sorting errors caused by overall positional shifts. Then, the global baseline benchmark is also considered. and local encryption pioneer The positional offset change determines the local area centered on the marker whose updated base offset vector reaches the preset offset threshold. This area is a newly added region of severe membrane contraction.

[0064] It should be noted that in this embodiment, the dynamic detection of whether new areas of severe film shrinkage have appeared, and the dynamic update operation of the adaptive sparse benchmark mesh are carried out simultaneously with the sorting operation. When performing sorting operation on the dies on the wafer, other areas can be detected simultaneously to see whether new areas of severe film shrinkage have appeared, so it will not add too much extra time.

[0065] Example 5: Step 110 above requires an initial pre-scan of the wafer to determine the row and column indices of each die and its theoretical physical coordinates in the mechanical coordinate system. This step can be completed using existing logic pre-scan methods. However, to further improve the efficiency of the initial pre-scan and thus the sorting efficiency, this example provides another initial pre-scan method, including the following steps. Please refer to... Figure 6 The flowchart shown is as follows: Step 510: Select at least three non-collinear grains in the wafer as seed grains, and determine the row and column indices and actual physical coordinates of each seed grain. To ensure a good fit, the row and column spacing between any two seed grains should be greater than 1 / 3 of the total number of rows and columns to avoid excessive fitting error caused by seed grains being too close together.

[0066] One method for selecting seed dies is manual interactive selection, where the sorting machine receives user instructions through its human-machine interface to determine the selected seed die. Another method is automatic visual recognition selection, which uses visual algorithms to automatically identify the wafer center die, the left edge die of the wafer's flat edge, and the upper right edge die as seed dies.

[0067] For example, in one instance, the first seed die selected is the center die of the wafer, with a row and column index of (500, 500) and actual physical coordinates of (100.00mm, 100.00mm). The second seed die selected is the left edge die of the wafer's flat edge, with a row and column index of (100, 100) and actual physical coordinates of (80.00mm, 80.00mm). The third seed die selected is the upper right edge die of the wafer, with a row and column index of (549, 549) and actual physical coordinates of (180.78mm, 180.78mm).

[0068] Step 520: Obtain the actual physical coordinates of the wafer origin by fitting the actual physical coordinates of the seed grain. The row and column indices of the wafer origin are obtained by fitting the row and column indices of the seed grains. Furthermore, under the ideal assumption that the grains are arranged in an equally spaced array, the positions of adjacent grains in the mechanical coordinate system can be calculated. Theoretical spacing in the direction and adjacent grains in the mechanical coordinate system Theoretical spacing in the direction The row and column indices can then be obtained by fitting the data. Theoretical physical coordinates in the mechanical coordinate system The initial linear mapping relationship between them:

[0069] The initial linear mapping relationship can be directly used as the final global linear mapping relationship. However, in order to improve the accuracy of the global linear mapping relationship, this embodiment also adds a verification and correction step 530.

[0070] Step 530: Control the local vision camera to move to the actual physical coordinates of the seed grain and acquire a local image. Combine image processing techniques to determine the actual physical coordinates of each grain within the local image where the seed grain is located, and calculate the coordinates of adjacent grains. Actual spacing in the direction and adjacent grains in Actual spacing in the direction The local image of the seed grain typically covers a 3x3 area around the seed grain. The above results can be obtained by calculating the actual spacing between different adjacent grains and averaging the values.

[0071] In practice, the actual spacing can be obtained based on a local image of any one of the seed grains. and Alternatively, the actual spacing can be calculated from the local images of the multiple seed grains and then averaged.

[0072] Step 540, utilizing grains in direction and Actual spacing in the direction and By correcting the initial linear mapping relationship, the row and column indices of each grain on the wafer are obtained. With theoretical physical coordinates Global linear mapping relationship between them:

[0073] Step 550: Control the local vision camera to move to different positions on the wafer according to the pre-scanning path and acquire each target surface view in sequence. Assign row and column indices to each grain in the target surface view according to the relative arrangement structure of the grains in each target surface view and determine the corresponding theoretical physical coordinates by combining the global linear mapping relationship.

[0074] Similar to existing logical pre-scanning methods, this embodiment, limited by the field of view of the local vision camera, typically requires moving the pre-scanning path. Based on the pre-scanning path, the size of a single target view, the size of a single die, and the overlap rate between adjacent target views, the row and column index range covered by each target view and the row and column index range covered by the overlapping area between adjacent target views can be estimated. Regardless of whether it's logical pre-scanning or the pre-scanning method of this embodiment, to avoid missing dies, adjacent target views need to overlap. The overlap rate between adjacent target views is predetermined. Based on the size of a single target view and the size of a single die, the total number of rows and columns covered by a single target view can be determined. Combining this with the overlap rate, the total number of rows and columns of the overlapping area can be determined. Following the pre-scanning path, the row and column index range covered by each target view and the row and column index range covered by the overlapping area can be estimated sequentially.

[0075] For example, if the size of a single target view is 10mm × 10mm, the size of a single die is 0.2mm, the kerf width between adjacent dies is 0.02mm, and a single target view covers 45 rows and 45 columns, then when pre-scanning along a zigzag path from the starting position along the row direction, if the row and column indices of the starting position are determined to be (1, 20), the estimated row index range covered by the first target view can be determined to be 1~45, and the column index range to be 20~64. If the overlap rate of adjacent target views is predetermined to be 10%, the estimated row index range covered by the second target view in the row direction can be determined to be 41~85, and the column index range to be 20~64. Furthermore, the row index range covered by the overlapping area of ​​the first and second target views can be determined to be 41~45, and the column index range to be 20~64.

[0076] After obtaining the target surface view, the grains within the target surface view are matched with a preset grain standard template to determine the relative arrangement of the grains within the target surface view. Then, row and column indices are pre-assigned to each grain within the row and column index range covered by the current target surface view.

[0077] For example, in the above example, after determining that the row index range of the first target view is 1~45 and the column index range is 20~64, the row and column indices of the first row and first column of the target view can be pre-assigned as (1,20), the row and column indices of the first row and second column of the target view can be pre-assigned as (1,21), the row and column indices of the first row and third column of the target view can be pre-assigned as (1,22), ... the row and column indices of the first row and 45th column of the target view can be pre-assigned as (1,64), and so on for the other rows. Since the row and column index pre-allocation operation described above is performed independently within each target view, and adjacent target views overlap, overlapping granules within the overlapping target view region will be assigned row and column indices simultaneously in multiple target views. For example, in one instance, such as Figure 7 As shown, the first target image is indicated by the green dashed box, and the second target image is indicated by the red dashed box. The shaded areas represent overlapping grains located within the overlapping region of the target surfaces. When processing the first target image, row and column indices were pre-assigned to each overlapping grain based on the row index range and relative grain arrangement of the first target image. Similarly, when processing the second target image, row and column indices were pre-assigned to each overlapping grain based on the row index range and relative grain arrangement of the second target image. This results in multiple row and column indices being pre-assigned to overlapping grains within the overlapping region of the target surfaces, necessitating a deduplication operation.

[0078] Traditional methods require searching for duplicate row and column indices globally and then removing duplicates. To reduce the time spent on deduplication, this embodiment, after pre-estimating the range of row and column indices covered by the overlapping area between adjacent target views, performs deduplication on overlapping dies whose pre-allocated row and column indices are within the range of row and column indices covered by the overlapping area between adjacent target views. In other words, the row and column indices of each overlapping die are compared with those of other overlapping dies, and the results with consistent row and column indices are merged, retaining only the unique row and column indices.

[0079] This allows us to obtain the row and column indices of each grain, and finally, based on the global linear mapping relationship, we can determine the theoretical physical coordinates corresponding to the row and column indices of each grain within the current target view.

[0080] In a comparative example, a sorting test was performed on an 8-inch wafer containing 300,000 0.2mm x 0.2mm dies. When sorting according to the methods of the above embodiments of this application, it takes about 2 seconds to complete the deflection angle correction of the wafer, and about 76 seconds to perform the initial pre-scan of the wafer by constructing a global linear mapping relationship. Figure 2 The adaptive sparse marker mesh containing 40 markers, and the time taken to visually determine the base offset vector for each marker, is 35 s. The entire pre-scanning process before sorting takes 113 s. At the start of sorting, the global average coordinate error of the compensated physical coordinates of each grain calculated based on the base offset vector of each marker is ±4.2 μm. According to the method of this application, the real-time compensated offset of each grain is dynamically updated during the sorting process, and the compensated physical coordinates are calculated. After 120 minutes of continuous sorting, the global average coordinate error of the compensated physical coordinates of each grain is ±11.6 μm, and the sorting yield is 99.98% after 2 hours of continuous sorting.

[0081] Following the method disclosed in CN 117878046 A, a comparison was conducted under the same conditions. When using a fixed, equidistant, dense selection of markers, a total of 300 markers were required for the entire wafer, far exceeding the number of markers in this application. This resulted in a total time of 512 seconds for the comparative method to complete wafer angle correction, pre-scanning, and determine the positional offset of each marker before sorting. At the start of sorting, the global average coordinate error of the compensated physical coordinates of each die calculated based on 300 markers was ±14.7 μm. After 120 minutes of continuous sorting, the global average coordinate error of the compensated physical coordinates of each die increased to ±83.5 μm, and the sorting yield was 99.21% after 2 hours of continuous sorting. It is evident that using more markers and a longer overall time, the sorting accuracy and sorting yield are inferior to the method in this application. When the number of markers is reduced, the global average coordinate error will further increase, and the sorting yield will further decrease.

[0082] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A wafer sorting method based on sparse marker dynamic compensation, characterized in that, The wafer sorting method based on sparse vanguard dynamic compensation includes: Under the ideal assumption that the grains are arranged in an equally spaced array, the wafer is initially pre-scanned to determine the row and column indices of each grain in the wafer and its theoretical physical coordinates in the mechanical coordinate system. Multiple grains are uniformly selected throughout the effective region of the wafer as global benchmarks. Furthermore, multiple grains were selected as localized densification markers in areas of severe film shrinkage on the wafer. Together they form an adaptive sparse vanguard mesh; For each soldier in the adaptive sparse soldier grid Control the local vision camera to move to the vanguard Theoretical physical coordinates Local images were collected and used in conjunction with image processing techniques to identify outstanding individuals. actual physical coordinates and its relationship with theoretical physical coordinates The basic offset vector between ; During the sorting process, the base offset vector of each marker in the adaptive sparse marker grid is dynamically corrected to obtain the real-time compensation offset. Based on the real-time compensation offset of each marker, other grains in the wafer that need compensation are calculated. Real-time compensation offset Utilizing the grains to be compensated Real-time compensation offset Correcting their respective theoretical physical coordinates Physical coordinates after compensation Grain sorting is performed according to the row and column coordinates and the compensated physical coordinates of each grain.

2. The wafer sorting method based on sparse marker dynamic compensation according to claim 1, characterized in that, During the sorting process, any vanguard in the adaptive sparse vanguard grid is sorted. The real-time compensation offset is obtained by dynamically correcting the base offset vector. for: in, It is the cumulative sorting time from the start of sorting to the current time. It is related to the cumulative sorting time The corresponding time accumulation coefficient, cumulative sorting time The longer the time, the higher the cumulative coefficient. The larger the value; Therefore, the model soldier This represents the ratio of the number of grains picked up and sorted in a predetermined local area of ​​the region center at the current moment to the total number of grains in that local area. It is related to the grain pickup ratio The corresponding picking influence coefficient, and the proportion of grain picking The larger the value, the greater the influence coefficient. The larger the value, the better.

3. The wafer sorting method based on sparse marker dynamic compensation according to claim 1, characterized in that, The calculation of the real-time compensation offset of other grains in the wafer based on the real-time compensation offset of each benchmark includes: Using each marker in the adaptive sparse marker mesh as a vertex, the effective area of ​​the wafer is divided into multiple adjacent polygonal compensation units; For each die to be compensated located within the polygonal compensation unit, based on the current die to be compensated The theoretical physical coordinates of the markers at each vertex of the polygon compensation unit are used to interpolate the real-time compensation offset of the markers at each vertex of the current polygon compensation unit using a bilinear interpolation algorithm to obtain the real-time compensation offset of the die to be compensated. For each die to be compensated located in the wafer edge region but outside the polygonal compensation unit, the real-time compensation offset of the die to be compensated is obtained by weighting the real-time compensation offset of multiple benchmarks within a predetermined row and column range around the die to be compensated.

4. The wafer sorting method based on sparse marker dynamic compensation according to claim 3, characterized in that, The real-time compensation offset of the grain to be compensated is obtained by weighting the real-time compensation offsets of multiple benchmarks within a predetermined range around the grain to be compensated, including: Calculate the grains to be compensated Theoretical physical coordinates and its surrounding predetermined range Any model worker actual physical coordinates Euclidean distance between The real-time compensation offset of the target grain is obtained by weighting the real-time compensation offset of each target grain with the reciprocal of the calculated Euclidean distance as the weight. : in, It is a grain to be compensated Surrounding Pre-determined Row Range Any model worker Real-time compensation offset.

5. The wafer sorting method based on sparse marker dynamic compensation according to claim 1, characterized in that, The wafer sorting method based on sparse vanguard dynamic compensation also includes: During the sorting process, the sorted grains that have already been picked up are dynamically removed from the adaptive sparse grain grid. When a new region of severe shrinkage is detected, multiple grains are selected in the new region of severe shrinkage as new local densification grains. Add to the adaptive sparse vanguard mesh, and determine the newly added locally encrypted vanguards. The actual physical coordinates and the fundamental offset vector between them and the theoretical physical coordinates.

6. The wafer sorting method based on sparse marker dynamic compensation according to claim 5, characterized in that, The wafer sorting method based on sparse vanguard dynamic compensation also includes: During the sorting process, when the ratio of the number of grains picked up and sorted in a local area within a predetermined range to the total number of grains in the local area reaches a threshold, it is determined that the local area within the predetermined range is a newly added area of ​​severe film shrinkage. Furthermore, during the sorting process, the basic offset vector of each marker in the adaptive sparse marker grid is updated synchronously at predetermined sorting intervals. When the updated basic offset vector of a marker reaches a preset offset threshold, it is determined that the local area within a predetermined range centered on the marker is a newly added area of ​​severe membrane shrinkage.

7. The wafer sorting method based on sparse marker dynamic compensation according to claim 6, characterized in that, During the sorting process, the basic offset vector of each soldier in the adaptive sparse soldier grid is updated synchronously at predetermined sorting intervals, including: During the sorting process, at intervals of the first predetermined sorting time... Utilizing any global benchmark in an adaptive sparse benchmark grid Real-time compensation offset Correcting its own theoretical physical coordinates Physical coordinates after compensation Control the local vision camera to move to the current global benchmark. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current global benchmark. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between When a global benchmark exists When the updated base offset vector reaches a preset offset threshold and the number does not exceed a first quantity threshold, the global benchmark is determined. The predetermined range of the region center is a newly added area of ​​severe membrane shrinkage; when the updated base offset vector reaches the global benchmark threshold... When the number exceeds the first threshold, the base offset vector of all soldiers in the adaptive sparse soldier grid is updated again, and the local area within a predetermined range centered on the soldier whose updated base offset vector reaches the preset offset threshold is determined to be the newly added area of ​​severe membrane shrinkage. And, during the sorting process, at each second predetermined sorting interval... Utilizing arbitrary local densification of vanguards in an adaptive sparse vanguard mesh Real-time compensation offset Correcting their respective theoretical physical coordinates Physical coordinates after compensation Control the local vision camera to move to the current local encryption vanguard. Compensated physical coordinates Local images were collected and combined with image processing techniques to redetermine the current local encryption leader. actual physical coordinates And update its coordinates with theoretical physical coordinates. The basic offset vector between When a local encryption vanguard exists When the updated base offset vector reaches a preset offset threshold and the number does not exceed a second quantity threshold, the local encryption vanguard is determined to be the target. The predetermined range of the region center is a newly added area of ​​severe membrane shrinkage; when the updated base offset vector reaches the preset offset threshold, the local densification benchmark is established. When the number exceeds the second threshold, the base offset vectors of all soldiers in the adaptive sparse soldier grid are updated again. A predetermined range of local areas centered on soldiers whose updated base offset vectors reach the preset offset threshold is identified as newly added areas of severe membrane shrinkage. .

8. The wafer sorting method based on sparse marker dynamic compensation according to claim 1, characterized in that, Multiple grains are uniformly selected throughout the effective region of the wafer as global benchmarks. Furthermore, multiple grains were selected as localized densification markers in areas of severe film shrinkage on the wafer. include: Within the entire effective area of ​​the wafer, according to row and column spacing Multiple grains are selected uniformly as global benchmarks. In regions of severe film shrinkage on the wafer, according to row and column spacing Uniformly supplementing multiple grains as local densification benchmarks The areas of severe wafer shrinkage include regions where the pin operation frequency reaches a frequency threshold as determined by the map and sorting operation path, and / or wafer edge regions.

9. The wafer sorting method based on sparse marker dynamic compensation according to claim 1, characterized in that, Under the ideal assumption that the grains are arranged in an equally spaced array, an initial pre-scan of the wafer is performed to determine the row and column indices and theoretical physical coordinates of each grain in the wafer, including: Select at least three non-collinear grains in the wafer as seed grains, and determine the row and column indices and actual physical coordinates of each type of seed grain; The actual physical coordinates at the wafer origin are obtained by fitting the actual physical coordinates of the seed grain. The row and column indices at the wafer origin are obtained by fitting the row and column indices of the seed grains. And under the ideal assumption that the grains are arranged in an equally spaced array, the mechanical coordinates of adjacent grains were calculated. Theoretical spacing in the direction and adjacent grains in the mechanical coordinate system Theoretical spacing in the direction The row and column indices are obtained by fitting. Theoretical physical coordinates in the mechanical coordinate system Initial linear mapping relationship between ; The local vision camera is moved to the actual physical coordinates of the seed grain and a local image is acquired. Image processing techniques are then used to determine the actual physical coordinates of each grain within the local image where the seed grain is located, and the coordinates of adjacent grains are calculated. Actual spacing in the direction and adjacent grains in Actual spacing in the direction ; Utilizing grains in direction and Actual spacing in the direction and By correcting the initial linear mapping relationship, the row and column indices of each grain on the wafer are obtained. With theoretical physical coordinates The global linear mapping relationship between them is ; The local vision camera is controlled to move to different positions on the wafer according to the pre-scanning path and acquire each target surface view in sequence. Based on the relative arrangement structure of the dies in each target surface view, row and column indices are assigned to each die in the target surface view, and the corresponding theoretical physical coordinates are determined by combining the global linear mapping relationship.

10. The wafer sorting method based on sparse marker dynamic compensation according to claim 9, characterized in that, Based on the relative arrangement of grains within each target view, row and column indices are assigned to each grain within the target view, and the corresponding theoretical physical coordinates are determined in conjunction with the global linear mapping relationship, including: Based on the pre-scan path, the size of a single target view, the size of a single grain, and the overlap rate between adjacent target views, estimate the row and column index range covered by each target view and the row and column index range covered by the target overlap area between adjacent target views. Based on the relative arrangement of the grains within each target view, pre-assign row and column indices to each grain within the range of row and column indices covered by the current target view; Perform deduplication on the pre-allocated row and column indices of the grains within the range of row and column indices covered by the overlapping area between adjacent target views; The theoretical physical coordinates corresponding to the row and column indices of each grain are determined based on the global linear mapping relationship.

Images