Semiconductor wafer damage evaluation method based on dry stripping stripping machine
Calculate wafer surface morphology data through threshold segmentation algorithm and RNN network model, and solve the problem of damage to wafers by existing loss evaluation methods, realize the rapid and accurate evaluation of deglassing wafer damage, and improve identification and evaluation Accuracy.
Patent Information
- Application Number
- CN202410965168.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-18
AI Technical Summary
The existing wafer damage evaluation methods are mostly destructive testing, which leads to excessive damage to the deglass wafer and makes it difficult to accurately evaluate the degree of damage.
Using threshold segmentation algorithm and RNN network model, by obtaining the surface morphology point cloud data set of deglass degrease, calculating the local desubtracting factor and local desubtracting index, building an overall reduction vector to achieve rapid and accurate evaluation of wafer surface damage .
It improves the identification effect and evaluation accuracy of damage when the die-filled wafer surface is cut, reduces the impact of local plane interference, and achieves faster and more accurate damage assessment.
Smart Images

Figure CN118505706B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of semiconductor debonding, and in particular to a semiconductor wafer damage evaluation method of a debonding stripping machine based on dry debonding. Background Art
[0002] The stripping process is a key step in the semiconductor manufacturing process, and the stripping machine based on dry stripping is a device that can effectively remove photoresist from the surface of semiconductor wafers. This device usually uses high-energy plasma to process photoresist. Compared with traditional wet stripping, dry stripping has the advantages of thorough stripping and fast speed, and is widely used in semiconductor stripping processes.
[0003] During the dry stripping process of semiconductor wafers, the high energy of plasma may cause physical damage to the microstructure of the semiconductor wafer surface. The surface integrity of the semiconductor wafer is crucial to its electrical performance. Therefore, it is necessary to conduct damage evaluation on the semiconductor wafer after stripping.
[0004] At present, the evaluation of wafer damage requires the use of a variety of detection technologies, such as scanning electron microscopes, atomic force microscopes, etc., to observe and measure the microstructure and damage layer on the wafer surface. In addition, the mechanical properties of the wafer can also be evaluated through methods such as fracture tests and bending tests, but fracture tests and bending tests are destructive evaluation methods that cause excessive damage to the wafer. Summary of the invention
[0005] The present application provides a semiconductor wafer damage evaluation method based on dry debonding for a debonding stripping machine, in order to solve the problem of excessive damage to the debonded wafer when the debonding evaluation method is used. The technical solution adopted is as follows:
[0006] One embodiment of the present application provides a semiconductor wafer damage evaluation method based on a stripping machine for dry stripping, the method comprising the following steps:
[0007] (1) Obtaining a surface morphology point cloud dataset of a debonded wafer;
[0008] (2) A threshold segmentation algorithm is used to obtain a single wafer bare die data point set based on the projection result of the surface morphology point cloud data set, and a local subtraction depression factor of each supervoxel is determined based on the segmentation result of each single wafer bare die data point set. The local subtraction depression factor is obtained by:
[0009] (2.1) determining a micro-local smoothness index for each data point in each supervoxel based on the Gaussian curvature and normal vector of each data point in each supervoxel and its neighboring data points;
[0010] (2.2) determining the local plane approximation of each supervoxel based on the similarity of the Gaussian curvature of the data points on the fitted surface of each supervoxel, and calculating the local subtraction concavity factor of each supervoxel based on the local plane approximation and the micro-local smoothness index of all the data points in each supervoxel;
[0011] (3) determining a local subtraction index for each supervoxel based on the distance from the data point in each supervoxel to the nearest neighbor die plane corresponding to each supervoxel and the local subtraction depression factor of each supervoxel;
[0012] (4) The RNN network model is used to obtain the evaluation results of the debonded wafer to be evaluated based on the local reduction index of all supervoxels contained in the surface morphology point cloud dataset;
[0013] The method for calculating the local subtraction sag factor of each supervoxel based on the local plane approximation and the micro-local smoothness index of all data points in each supervoxel is:
[0014] The reciprocal of the sum of the mean of the square of the difference between the micro-local smoothness indexes of any two data points in each supervoxel and 0.1 is taken as the change smoothness index of each supervoxel;
[0015] The product of the mean of the micro-local smoothness index of all data points in each supervoxel and the variation smoothness index of each supervoxel is taken as the numerator;
[0016] The denominator is the sum of the local plane approximation of each supervoxel and 0.1;
[0017] The ratio of the numerator to the denominator is taken as the local subtraction depression factor for each supervoxel.
[0018] Preferably, the method of obtaining a single wafer bare die data point set based on the projection result of the surface morphology point cloud data set by using a threshold segmentation algorithm is:
[0019] A right-handed coordinate system is constructed with the z-axis perpendicular to the plane of the debonded wafer as the debonded wafer coordinate system, and the surface morphology point cloud data set of the debonded wafer to be evaluated is projected into the debonded wafer coordinate system to obtain a data space as the point cloud space;
[0020] The z coordinate values of all data points in the point cloud space are taken as input, the maximum inter-class variance method is used to obtain the coordinate threshold, and the set of all data points in the point cloud space whose z coordinates are greater than the coordinate threshold is taken as the bare chip data point set;
[0021] The coordinates of all data points in the bare die data point set are taken as input, and the set of data points in each cluster obtained by the clustering algorithm is taken as a single wafer bare die data point set.
[0022] Preferably, the method for determining the micro-local smoothness index of each data point in each supervoxel based on the Gaussian curvature and normal vector of each data point in each supervoxel and its adjacent data points is:
[0023] Taking each single wafer bare die data point set as input, each single wafer bare die data point set is segmented into a preset number of supervoxels using a supervoxel-based point cloud segmentation algorithm; taking the coordinates of the data points in each supervoxel as input, a least squares fitting algorithm is used to obtain a fitting surface for each supervoxel;
[0024] The sum of the absolute value of the difference between the Gaussian curvature of each data point in each supervoxel and any adjacent data point of each data point on the fitting surface of each supervoxel and 0.1 is used as the denominator;
[0025] The micro-local smoothness index of each data point in each supervoxel is taken as the average of the ratio of the cosine similarity between the normal vector of each data point in each supervoxel and any adjacent data point of each data point on the fitting surface of each supervoxel and the denominator accumulated over the set of adjacent data points of each data point.
[0026] Preferably, a method for acquiring any adjacent data point of each data point is:
[0027] The Euclidean distance between each data point and the remaining data points in each supervoxel is calculated respectively, and a set consisting of a preset number of remaining data points with the smallest Euclidean distance between each data point is used as a neighboring data point set for each data point, and each element in the neighboring data point set is used as a neighboring data point for each data point.
[0028] Preferably, the method for determining the local plane approximation of each supervoxel based on the similarity of the Gaussian curvature of the data points on the fitted surface of each supervoxel is:
[0029] The accumulation result of the cosine similarity between the normal vectors of any two data points in each supervoxel on the fitting surface of the supervoxel on each supervoxel is taken as a numerator;
[0030] The denominator is the sum of the absolute values of the Gaussian curvatures of all data points in each supervoxel on the fitting surface of the supervoxel and 0.1;
[0031] The ratio of the numerator to the denominator is taken as the local planar proximity of each supervoxel.
[0032] Preferably, the method for determining the local reduction index of each supervoxel based on the distance from the data point in each supervoxel to the nearest neighbor die plane corresponding to each supervoxel and the local reduction sag factor of each supervoxel is:
[0033] The sum of the Euclidean distances between each data point in each supervoxel and all adjacent data points of each data point to the nearest neighboring die plane corresponding to each supervoxel is calculated as a numerator, and the ratio of the numerator to a preset parameter is used as the local damage distance of each data point in each supervoxel;
[0034] The local reduction index of each supervoxel is the product of the mean of the local damage distances of all data points in each supervoxel, the maximum of the local damage distances of all data points in each supervoxel and the local reduction sag factor of each supervoxel.
[0035] Preferably, the method for obtaining the nearest neighbor die plane corresponding to each supervoxel is:
[0036] The coordinates of all data points in each single wafer die data point set in the point cloud space are taken as input, and the axis-aligned bounding box algorithm is used to obtain the coordinates of 8 vertices;
[0037] All planes determined by the 8 vertex coordinates and having a perpendicular or parallel relationship to the xoy plane in the point cloud space are used as the die surface planes of each single wafer die data point set;
[0038] For any supervoxel corresponding to each single wafer die data point set, the average of the Euclidean distances from all data points in each supervoxel to each die surface plane of each single wafer die data point set is taken as the supervoxel plane distance between each supervoxel and each die surface plane;
[0039] The die surface plane with the smallest supervoxel plane distance to each supervoxel is taken as the nearest neighbor die plane of each supervoxel.
[0040] Preferably, the method of obtaining the evaluation result of the debonded wafer to be evaluated by using the RNN network model based on the local reduction index of all supervoxels contained in the surface morphology point cloud data set is:
[0041] The cosine similarity between the overall reduction vector of the surface morphology point cloud dataset of the de-bonded wafer to be evaluated and the overall reduction vector of the surface morphology point cloud dataset of the lossless de-bonded wafer is taken as input, and the RNN network model is used to obtain the damage evaluation result of the de-bonded wafer to be evaluated. The damage evaluation result includes two types: meeting the wafer de-bonding standard and not meeting the wafer de-bonding standard.
[0042] Preferably, the method for obtaining the overall reduction vector is:
[0043] The vector composed of the local reduction indices of all supervoxels in each single wafer bare die data point set is used as the local feature vector of each single wafer bare die data point set, and the vector composed of the local feature vectors of all single wafer bare die data point sets contained in the surface morphology point cloud dataset arranged in ascending order of the single wafer bare die data point set numbers is used as the overall reduction vector of the surface morphology point cloud dataset.
[0044] The beneficial effects of the present application are as follows: the present application determines the micro-local smoothness index of the data points in the supervoxel and the supervoxel change smoothness index based on the data distribution characteristics of the damage caused by the reduction phenomenon on the surface of the bare die of the debonded wafer in the supervoxel, and constructs a local reduction depression factor based on the micro-local smoothness index of the data points in each supervoxel and the local plane approximation of the supervoxel. The beneficial effects of the present application are that the geometric characteristics when the reduction phenomenon occurs on the bare die surface are considered, the recognition effect of the damage caused by the reduction phenomenon on the bare die surface of the wafer is improved, and the interference effect of the local plane of the bare die surface is reduced; and the local reduction index is constructed based on the local damage distance of each data point in the supervoxel. The beneficial effects of the present application are that the accuracy of the assessment of the degree of damage caused by the reduction phenomenon on the bare die surface of the wafer is improved; the overall reduction vector is constructed based on the local reduction index, and the recurrent neural network model is used to complete the evaluation of the damage on the surface of the debonded wafer. The present application can evaluate the damage caused by the reduction phenomenon on the bare die surface of the wafer more quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0046] Figure 1 A schematic flow chart of a semiconductor wafer damage evaluation method for a stripping machine based on dry stripping provided in one embodiment of the present application;
[0047] Figure 2 A flowchart of a semiconductor wafer damage evaluation method for a stripping machine based on dry stripping provided in one embodiment of the present application;
[0048] Figure 3 A schematic diagram of clustering data points in a die data point set provided by one embodiment of the present application;
[0049] Figure 4 A schematic diagram of supervoxel segmentation provided by an embodiment of the present application;
[0050] Figure 5A schematic diagram of a supervoxel corresponding fitting surface provided in one embodiment of the present application. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0052] See also Figure 1 , which shows a flow chart of a semiconductor wafer damage evaluation method based on dry degumming provided by an embodiment of the present application, the method comprising the following steps:
[0053] Step S001, obtaining a surface morphology point cloud data set of a debonded wafer.
[0054] During the debonding process, a certain thickness of damage layer may be formed on the wafer, and the thickness of the damage layer is inversely proportional to the fracture strength of the wafer, that is, the thinner the damage layer, the higher the fracture strength of the wafer. Therefore, when evaluating the damage of semiconductor wafers in the debonding stripping machine, this application considers obtaining the damage evaluation result by analyzing the reduction phenomenon and flatness of the bare die on the surface of the wafer after debonding. The implementation process of the whole scheme is as follows: Figure 2 shown.
[0055] This application takes the debonding treatment of wafers by a plasma debonding stripping machine as an example, wherein the specification of the wafer is taken as an empirical value of 1 inch, and the wafer is debonded by using a debonding stripping machine, and the resulting wafer is recorded as a debonded wafer. In this application, the debonding parameters of the debonding stripping machine, the heating plate temperature, RF power, gas flow rate, and reaction pressure are set to 250°C, 900W, 3000SCCM, and 1torr, respectively. The specifications of the debonded wafers and the debonding parameters of the debonding stripping machine can be set by the implementer to appropriate values.
[0056] Secondly, the surface morphology data of the debonded wafer to be evaluated is collected every 5 seconds using a three-dimensional topography instrument until the debonding process of the debonded wafer is completed, and the surface morphology point cloud data set of the debonded wafer at each sampling time is obtained. The surface morphology point cloud data set collected at the end of the debonding process is recorded as .
[0057] At this point, the surface morphology point cloud dataset of the debonded wafer is obtained, which is used for the subsequent construction of the point cloud space.
[0058] Step S002 , using a threshold segmentation algorithm to obtain a single wafer bare die data point set based on the projection result of the surface morphology point cloud data set, and determining a local subtraction depression factor of each supervoxel based on the segmentation result of each single wafer bare die data point set.
[0059] Before the wafer is debonded, the circuit pattern on the wafer is usually divided into independent small squares, namely bare chips, using precise cutting technology. In order to more accurately evaluate the damage, the surface morphology point cloud data set of the debonded wafer to be evaluated is considered to be clustered.
[0060] Specifically, a right-handed coordinate system is constructed with the direction perpendicular to the plane of the debonded wafer as the z-axis as the debonded wafer coordinate system. All data points in the debonded wafer are projected to the debonded wafer coordinate system, and the data space composed of all the projected data points is recorded as the point cloud space. Since the die on the surface of the wafer has a higher height than the substrate of the wafer, the data points corresponding to the die on the surface of the debonded wafer in the point cloud space should have a larger z coordinate value. Therefore, the z coordinate values of all data points in the point cloud space are used as input, and the coordinate threshold of the z coordinate value is obtained by the maximum inter-class variance method. The set of data points in the point cloud space with z coordinates greater than the coordinate threshold is taken as the set of die data points. The maximum inter-class variance method is a well-known technology, and the specific process will not be repeated.
[0061] Furthermore, since there is a certain spacing distance between the dies on the surface of the debonded wafer, the die data point set will form multiple data point clustering areas in the point cloud space, and the number of clustering areas should be equal to the number of dies, that is, each data point clustering area corresponds to one die. Therefore, the coordinates of all data points in the die data point set are used as input, and the DBSCAN (Density-based spatial clustering of applications with noise) clustering algorithm is used to divide the data points in the die data point set into K clusters, and the clusters are numbered 1-K. In this application, the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm are set to 2 and 5 respectively, and the set of data points in each cluster is taken as a single wafer die data point set, such as Figure 3 As shown, the DBSCAN clustering algorithm is a well-known technology, and the specific process will not be repeated here.
[0062] In another embodiment, the implementer may also use the coordinates of all data points in the bare die data point set as input, and use the AP (Affinity propagation) clustering algorithm to divide the data points in the bare die data point set into K clusters. In this application, the Euclidean distance between two data points is used as the metric distance when the AP clustering algorithm is clustered, and the maximum number of iterations is set to 200. The AP clustering algorithm is a well-known technology, and the specific process will not be repeated here.
[0063] Secondly, since photoresist is usually coated on the surface of the wafer, there is photoresist on the top of the die on the surface of the wafer, but not on the side walls of the die. As a result, the top of the die is usually not damaged during the debonding process, while the side walls of the die may be cut under the bombardment and erosion of plasma, causing the side walls of the die to become concave. The more serious the damage to the die, the greater the degree of concavity of the side walls of the die.
[0064] Furthermore, for any single wafer bare die data point set, the kth single wafer bare die data point set J k For example, J k All data points in J are taken as input and the supervoxel-based point cloud segmentation algorithm is used to segment J k All data points in are divided into M supervoxels, such as Figure 4 As shown, the set of data points in each supervoxel corresponds to a bare chip surface space area in the cloud space. In this application, the size of M takes an empirical value of 20. The point cloud segmentation algorithm based on supervoxels is a well-known technology, and the specific process will not be repeated here.
[0065] Secondly, for J k For any supervoxel in the ath supervoxel, take the ath supervoxel as an example, take the coordinates of the data points in the ath supervoxel as input, use the moving least squares fitting algorithm to perform surface fitting on the data points in the ath supervoxel, and get the fitting surface of the ath supervoxel, as shown in Figure 5 As shown, the Gaussian curvature and normal vector of each data point on the fitting surface of the a-th supervoxel are calculated respectively, wherein the moving least squares fitting algorithm, Gaussian curvature and normal vector are all well-known technologies, and the specific process is not repeated here. For each data point in the a-th supervoxel, taking the i-th data point as an example, the Euclidean distance between the i-th data point and each of the remaining data points is calculated respectively, and the N data points with the smallest Euclidean distance from the i-th data point are all used as the adjacent data points of the i-th data point, and the set composed of the N adjacent data points is used as the adjacent data point set of the i-th data point. In this application, the size of N takes the empirical value of 10.
[0066] Based on the above analysis, a local subtraction depression factor is constructed here to characterize the possibility of depression caused by reduction phenomenon in the spatial area of the die surface corresponding to each supervoxel. Calculate the local subtraction depression factor of the ath supervoxel:
[0067] ;
[0068] ;
[0069] ;
[0070] In the formula, l a,i is the microlocal smoothness index of the ith data point in the ath supervoxel, N is the number of neighboring data points of the ith data point, n is the nth neighboring data point of the ith data point, and g a,i , μ a,i are the normal vector and Gaussian curvature of the ith data point, g a,i (n), μ a,i (n) are the normal vector and Gaussian curvature of the nth adjacent data point, CS( ) is the cosine similarity function, CS(g a,i ,g a,i (n)) is g a,i With g a,i (n) cosine similarity, β is a parameter adjustment factor used to prevent the denominator from being 0, and the empirical value of β is 0.1;
[0071] t a is the local planar approximation of the ath supervoxel, m a is the number of data points in the ath supervoxel, g a,j is the normal vector of the jth data point in the ath supervoxel, CS(g a,i ,g a,j ) is g a,i With g a,j The cosine similarity between a is the sum of the absolute values of the Gaussian curvatures of all data points in the a-th supervoxel;
[0072] L a is the local subtractive sag factor of the a-th supervoxel, is the mean of the microlocal smoothness indices of all data points in the ath supervoxel, pl a is the variation smoothness index of the a-th supervoxel, and the magnitude of the variation smoothness index is equal to the reciprocal of the sum of the mean of the square of the difference between the micro-local smoothness indexes of any two data points in the a-th supervoxel accumulated on the a-th supervoxel and 0.1.
[0073] The greater the possibility that the die surface in the die surface space region corresponding to the a-th supervoxel is a plane, the more similar the normal vectors of the data points in the a-th supervoxel are on the fitting surface of the a-th supervoxel, and the closer the intrinsic curvature of the i-th data point and its adjacent data points on the fitting surface, the greater the probability that the die surface in the a-th supervoxel is a plane. a,i ,g a,i The larger the value of (n)), the closer the Gaussian curvature of the data point on the fitting surface is to 0, a,i The larger the value of t is, the smaller the chip surface is. When the chip surface is cut but not damaged, a relatively regular shallow bowl-shaped depression will be formed on the chip surface. The smoother the chip surface is in the corresponding chip surface space area, the more similar the normal vectors between the data points are, and t a The larger the value of ; therefore, the more likely the die surface in the die surface space area corresponding to the a-th supervoxel is to have a concave area caused by reduction, the greater the difference in the normal vectors of the data points on the fitting surface of the a-th supervoxel, the greater the change in Gaussian curvature, and the greater the probability of a large degree of geometric change on the fitting surface, L a The larger the value of .
[0074] At this point, the local reduction sag factor of each supervoxel is obtained, which is used to subsequently calculate the local reduction index of each supervoxel.
[0075] Step S003 : determining a local reduction index of each supervoxel based on the distance from the data point in each supervoxel to the nearest neighbor die plane corresponding to each supervoxel and the local reduction sag factor of each supervoxel.
[0076] Furthermore, since the top of the wafer die is usually not damaged during the debonding process, and the bottom of the die is connected to the wafer substrate, it is generally not damaged. Therefore, the kth single wafer die data point set J k The coordinates of the data points in the point cloud space are used as input, and the axis-aligned bounding box algorithm is used to output the coordinates of 8 vertices. The axis-aligned bounding box algorithm is a well-known technology, and the specific process is not repeated here. Secondly, the coordinates of the 8 vertices are connected in pairs to determine multiple planes, and all planes that meet the perpendicular or parallel relationship with the xoy plane in the point cloud space are selected from the multiple planes as J k of the die surface plane.
[0077] Secondly, for J k For any corresponding supervoxel, taking the ath supervoxel as an example, calculate the distance from each data point on the ath supervoxel to J k The Euclidean distance of each die surface plane is calculated by moving all data points in the a-th supervoxel to J kThe average of the Euclidean distances of each die surface plane is taken as the supervoxel plane distance between the a-th supervoxel and each of the die surface planes, and the die surface plane with the smallest supervoxel plane distance to the a-th supervoxel is taken as the nearest neighbor die plane G of the a-th supervoxel. a .
[0078] Based on the above analysis, a local reduction index is constructed here to characterize the reduction degree when the reduction phenomenon occurs on the die surface in the die surface space area corresponding to each supervoxel. Calculate the local reduction index of the ath supervoxel:
[0079] ;
[0080] ;
[0081] In the formula, u a,i is the local damage distance of the ith data point in the ath supervoxel, ΔD(ai) is the distance from the ith data point in the ath supervoxel to the nearest neighbor die plane G a The sum of the Euclidean distances between them, N is the number of all adjacent data points of the i-th data point;
[0082] V a is the local reduction index of the ath supervoxel, is the mean of the local damage distances of all data points in the ath supervoxel, D max,a is the maximum value of the local damage distance of all data points in the ath supervoxel, L a is the local subtractive depression factor of the a-th supervoxel.
[0083] The higher the degree of damage to the die surface in the die surface space region corresponding to the a-th supervoxel, the farther the distance between the data point in the a-th supervoxel and the nearest neighboring die plane of the a-th supervoxel, u a,i The larger the value of L a The larger the value of , the greater the probability that the die surface corresponding to the ath supervoxel is damaged by the reduction phenomenon; that is, V a The larger the value of , the greater the degree of damage to the die surface caused by the reduction phenomenon in the die surface spatial region corresponding to the a-th supervoxel.
[0084] At this point, the local reduction index of each supervoxel in each single wafer die data point set is obtained, which is used to subsequently construct the overall reduction vector of the surface morphology point cloud data set.
[0085] Step S004: using the RNN network model to obtain the evaluation result of the debonded wafer to be evaluated based on the local reduction index of all supervoxels contained in the surface morphology point cloud data set.
[0086] According to the above steps, the local reduction index of each supervoxel in all single-wafer bare die data point sets is calculated respectively, and the vector composed of the local reduction indexes of all supervoxels in each single-wafer bare die data point set is used as the local feature vector of each single-wafer bare die data point set, and the vector composed of the local feature vectors of all single-wafer bare die data point sets contained in the surface morphology point cloud dataset of the debonded wafer to be evaluated arranged in ascending order of the single-wafer bare die data point set numbers is used as the overall reduction vector of the surface morphology point cloud dataset of the debonded wafer to be evaluated.
[0087] Furthermore, wafers that have been debonded and have the same specifications as the debonded wafer and no damage on the wafer surface are screened out as lossless debonded wafers, and surface morphology data of the lossless debonded wafers are collected using a three-dimensional profilometer to obtain a surface morphology point cloud data set of the lossless debonded wafer collected at the end of debonding. The above-mentioned process of obtaining the overall reduction vector of the surface morphology point cloud data set is repeated to obtain the overall reduction vector of the surface morphology point cloud data set of the lossless debonded wafer.
[0088] Secondly, the cosine similarity between the overall reduction vector of the surface morphology point cloud data set of the debonded wafer to be evaluated and the overall reduction vector of the surface morphology point cloud data set of the lossless debonded wafer is calculated, and the data labels of the cosine similarity of the wafers that meet the debonding standard and those that do not meet the debonding standard are set to 1 and 0, respectively. Secondly, the cosine similarity between the overall reduction vector of the surface morphology point cloud data set of the debonded wafer to be evaluated and the overall reduction vector of the surface morphology point cloud data set of the lossless debonded wafer is used as input, and the RNN (Recurrent Neural Network) network model is used to obtain the damage evaluation result of the debonded wafer to be evaluated. The training of the neural network is a well-known technology, and the specific process will not be repeated here. The damage evaluation result includes two types: meeting the wafer debonding standard and not meeting the wafer debonding standard.
[0089] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.
Claims
1. A semiconductor wafer damage evaluation method based on a dry debonding stripping machine, characterized in that: The method comprises the following steps: (1) Obtaining a surface morphology point cloud dataset of a debonded wafer; (2) A threshold segmentation algorithm is used to obtain a single wafer bare die data point set based on the projection result of the surface morphology point cloud data set, and a local subtraction depression factor of each supervoxel is determined based on the segmentation result of each single wafer bare die data point set. The local subtraction depression factor is obtained by: (2.1) determining a micro-local smoothness index for each data point in each supervoxel based on the Gaussian curvature and normal vector of each data point in each supervoxel and its neighboring data points; (2.2) determining the local plane approximation of each supervoxel based on the similarity of the Gaussian curvature of the data points on the fitted surface of each supervoxel, and calculating the local subtraction concavity factor of each supervoxel based on the local plane approximation and the micro-local smoothness index of all the data points in each supervoxel; (3) determining a local subtraction index for each supervoxel based on the distance from the data point in each supervoxel to the nearest neighbor die plane corresponding to each supervoxel and the local subtraction depression factor of each supervoxel; (4) Using the RNN network model to obtain the evaluation results of the debonded wafer to be evaluated based on the local reduction index of all supervoxels contained in the surface morphology point cloud dataset; The method for calculating the local subtraction sag factor of each supervoxel based on the local plane approximation and the micro-local smoothness index of all data points in each supervoxel is: The reciprocal of the sum of the mean of the square of the difference between the micro-local smoothness indexes of any two data points in each supervoxel and 0.1 is taken as the change smoothness index of each supervoxel; The product of the mean of the micro-local smoothness index of all data points in each supervoxel and the variation smoothness index of each supervoxel is taken as the numerator; The denominator is the sum of the local plane approximation of each supervoxel and 0.1; The ratio of the numerator to the denominator is taken as the local subtraction depression factor for each supervoxel.
2. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 1 is characterized in that: The method of obtaining a single wafer bare die data point set based on the projection result of the surface morphology point cloud data set by using a threshold segmentation algorithm is: A right-handed coordinate system is constructed with the z-axis perpendicular to the plane of the debonded wafer as the debonded wafer coordinate system, and the surface morphology point cloud data set of the debonded wafer to be evaluated is projected into the debonded wafer coordinate system to obtain a data space as the point cloud space; The z coordinate values of all data points in the point cloud space are taken as input, the maximum inter-class variance method is used to obtain the coordinate threshold, and the set of all data points in the point cloud space whose z coordinates are greater than the coordinate threshold is taken as the bare chip data point set; The coordinates of all data points in the bare die data point set are taken as input, and the set of data points in each cluster obtained by the clustering algorithm is taken as a single wafer bare die data point set.
3. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 1, characterized in that: The method for determining the micro-local smoothness index of each data point in each supervoxel based on the Gaussian curvature and normal vector of each data point in each supervoxel and its adjacent data points is: Taking each single wafer bare die data point set as input, segmenting each single wafer bare die data point set into a preset number of supervoxels using a supervoxel-based point cloud segmentation algorithm; The coordinates of the data points in each supervoxel are taken as input, and a least square fitting algorithm is used to obtain the fitting surface of each supervoxel; The sum of the absolute value of the difference between the Gaussian curvature of each data point in each supervoxel and any adjacent data point of each data point on the fitting surface of each supervoxel and 0.1 is used as the denominator; The micro-local smoothness index of each data point in each supervoxel is taken as the average of the ratio of the cosine similarity between the normal vector of each data point in each supervoxel and any adjacent data point of each data point on the fitting surface of each supervoxel and the denominator accumulated over the set of adjacent data points of each data point.
4. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 3 is characterized in that: The method for obtaining any adjacent data point of each data point is: The Euclidean distance between each data point and the remaining data points in each supervoxel is calculated respectively, and a set consisting of a preset number of remaining data points with the smallest Euclidean distance between each data point is used as a neighboring data point set for each data point, and each element in the neighboring data point set is used as a neighboring data point for each data point.
5. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 1, characterized in that: The method for determining the local plane approximation of each supervoxel based on the similarity of the Gaussian curvature of the data points on the fitting surface of each supervoxel is: The accumulation result of the cosine similarity between the normal vectors of any two data points in each supervoxel on the fitting surface of the supervoxel on each supervoxel is taken as a numerator; The denominator is the sum of the absolute values of the Gaussian curvatures of all data points in each supervoxel on the fitting surface of the supervoxel and 0.1; The ratio of the numerator to the denominator is taken as the local planar proximity of each supervoxel.
6. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 1, characterized in that: The method for determining the local reduction index of each supervoxel based on the distance from the data point in each supervoxel to the nearest neighbor die plane corresponding to each supervoxel and the local reduction depression factor of each supervoxel is: The sum of the Euclidean distances between each data point in each supervoxel and all adjacent data points of each data point to the nearest neighboring die plane corresponding to each supervoxel is calculated as a numerator, and the ratio of the numerator to a preset parameter is used as the local damage distance of each data point in each supervoxel; The local reduction index of each supervoxel is the product of the mean of the local damage distances of all data points in each supervoxel, the maximum of the local damage distances of all data points in each supervoxel and the local reduction sag factor of each supervoxel.
7. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 6, characterized in that: The method for obtaining the nearest neighbor die plane corresponding to each supervoxel is: The coordinates of all data points in each single wafer die data point set in the point cloud space are taken as input, and the axis-aligned bounding box algorithm is used to obtain the coordinates of 8 vertices; All planes determined by the 8 vertex coordinates and having a perpendicular or parallel relationship with the xoy plane in the point cloud space are used as the die surface planes of each single wafer die data point set; For any supervoxel corresponding to each single wafer die data point set, the average of the Euclidean distances from all data points in each supervoxel to each die surface plane of each single wafer die data point set is taken as the supervoxel plane distance between each supervoxel and each die surface plane; The die surface plane with the smallest supervoxel plane distance to each supervoxel is taken as the nearest neighbor die plane of each supervoxel.
8. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 1, characterized in that: The method for obtaining the evaluation result of the debonded wafer to be evaluated by using the RNN network model based on the local reduction index of all supervoxels contained in the surface morphology point cloud data set is: The cosine similarity between the overall reduction vector of the surface morphology point cloud dataset of the de-bonded wafer to be evaluated and the overall reduction vector of the surface morphology point cloud dataset of the lossless de-bonded wafer is taken as input, and the RNN network model is used to obtain the damage evaluation result of the de-bonded wafer to be evaluated. The damage evaluation result includes two types: meeting the wafer de-bonding standard and not meeting the wafer de-bonding standard.
9. The semiconductor wafer damage evaluation method of a stripping machine based on dry stripping according to claim 8, characterized in that: The method for obtaining the overall reduction vector is: The vector composed of the local reduction indices of all supervoxels in each single wafer bare die data point set is used as the local feature vector of each single wafer bare die data point set, and the vector composed of the local feature vectors of all single wafer bare die data point sets contained in the surface morphology point cloud dataset arranged in ascending order of the single wafer bare die data point set numbers is used as the overall reduction vector of the surface morphology point cloud dataset.
