Wafer film pasting flaw detection system and method

Abstract

The invention discloses a wafer film defect detection system and method, and relates to the technical field of wafer film defect detection.The wafer film defect detection method comprises the following steps that in the wafer film defect detection process, laser beams are emitted through a laser scanner according to the preset light beam coverage rate, and the surface of a wafer is covered for preliminary scanning; in the laser beam scanning process, the intensity and the consistency of laser reflection signals are analyzed in real time. According to the invention, preliminary scanning is carried out by presetting the light beam coverage rate, the significance of the potential flaw area is evaluated by using the convolutional neural network, accurate classification of the flaw obvious area and the flaw non-obvious area is realized, high-density secondary scanning is carried out by dynamically adjusting the light beam coverage rate for the flaw non-obvious area, and finer surface information is captured. The detection efficiency is improved, the risks of missing detection and false detection are effectively reduced, the accurate detection of tiny flaws is ensured, the quality control of wafer film pasting is remarkably improved, and the production loss is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of wafer film defect detection, and in particular to a wafer film defect detection system and method. Background Art

[0002] Wafer film defect detection refers to the process of quality inspection to identify possible defects or flaws in the process of attaching protective film or functional film to the surface of the wafer. In semiconductor manufacturing, wafers are a key basic material, and the film attaching step is used to prevent the wafer surface from being contaminated or damaged during subsequent processing. However, defects such as bubbles, wrinkles, scratches, particles, etc. may be generated during the film attaching process, which will seriously affect the overall quality of the wafer and the performance of the final product. In order to ensure the surface integrity and film attaching effect of the wafer, defect detection usually uses high-precision optical inspection equipment or machine vision systems to scan and analyze the film on the surface of the wafer through image processing technology to promptly detect any subtle problems generated during the film attaching process. This inspection process is crucial to improving the yield and performance of semiconductor products, ensuring that there will be no greater production losses and product scrapping due to film attaching defects in subsequent manufacturing links.

[0003] Laser scanners are commonly used in the prior art to detect wafer film defects. Laser scanners have extremely high detection accuracy and resolution, and can accurately identify tiny defects such as bubbles, scratches, and uneven film layers. Laser scanning, as a non-contact detection method, avoids physical damage to the wafer surface and is particularly suitable for sensitive semiconductor manufacturing processes. In addition, laser scanners can also provide three-dimensional topographic data, accurately measure the depth and shape of defects, and have automated detection capabilities, thereby improving production efficiency and product quality stability.

[0004] In order to improve detection efficiency and reduce costs, laser scanners in the prior art usually do not set the beam coverage too high, because too high coverage will significantly increase detection time and cost. Wafer film defect detection mainly targets specific types of defects, such as bubbles, scratches, and uneven film layers, which are often concentrated in local areas. Therefore, by reasonably adjusting the beam coverage, it is possible to optimize detection efficiency and reduce unnecessary time consumption while ensuring sufficient detection accuracy. However, lower coverage may not detect all defects, especially when minor defects are difficult to predict, some areas may be ignored, resulting in missed detection. Undetected defects may cause more serious quality problems in subsequent production, or even cause the entire wafer to be scrapped, thereby increasing production costs and reducing product yield and reliability.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a wafer film defect detection system and method, which performs preliminary scanning by presetting the light beam coverage, utilizes convolutional neural network to evaluate the significance of potential defect areas, realizes accurate classification of areas with obvious defects and areas with less obvious defects, and dynamically adjusts the light beam coverage to perform high-density secondary scanning for areas with less obvious defects, captures more subtle surface information, and accurately locates the position and shape of defects in combination with three-dimensional morphological data, which can effectively reduce the risk of missed detection and false detection while improving detection efficiency, ensure accurate detection of tiny defects, significantly improve quality control of wafer film, avoid production losses, and solve the problems in the above-mentioned background technology.

[0007] In order to achieve the above object, the present invention provides the following technical solution: a method for detecting wafer film defects, comprising the following steps:

[0008] During the wafer film defect detection process, the laser scanner emits a laser beam according to the preset beam coverage rate to cover the wafer surface for preliminary scanning;

[0009] During the laser beam scanning process, the intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film, and mark the reflection difference area as a suspected defect area;

[0010] Extract features from the reflection signals of suspected defective areas, and perform in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the pre-trained convolutional neural network is used to evaluate the degree of prominence of potential defects in the wafer film, and the suspected defective areas are further divided into obvious defect areas and inconspicuous defect areas.

[0011] For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects.

[0012] By adjusting the beam coverage, a higher density laser beam is emitted in the marked suspected defect area for a second scan. The high-density laser beam is used to capture more subtle surface information, and combined with the three-dimensional morphology data to obtain the three-dimensional features of the defect, accurately locate and identify the film defects.

[0013] Preferably, the intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film. The specific steps are as follows:

[0014] The laser scanner emits a laser beam according to the preset beam coverage, covers the entire wafer surface for scanning, collects the reflected light signal in real time, and records the reflected light intensity value I(x, y) of each scanning point as a two-dimensional coordinate matrix, where I(x, y) represents the reflected light intensity value at (x, y), where x and y represent the spatial coordinates of the scan, and stores the light intensity information of the entire wafer surface through the two-dimensional coordinate matrix;

[0015] The local consistency calculation is performed on the light intensity value I(x, y) of each point (x, y), specifically: the light intensity of the point is compared with the light intensity of other points in its neighborhood, and the local consistency deviation D(x, y) is calculated. The calculation expression is: In the formula, (x l ,y l ) is the neighborhood point of point (x, y), l represents the index of the neighborhood point, and N is the number of points in the neighborhood;

[0016] After calculating the local consistency deviation D(x, y) of each point, D(x, y) is compared with the preset threshold T. If the local consistency deviation D(x, y) of the point is greater than T, the point is marked as a suspected defect point.

[0017] A clustering algorithm is applied to process the marked suspected defect points, and the defect points are aggregated according to spatial distribution and density to form a continuous defect area. The continuous defect area is the reflection difference area between the wafer surface and the film.

[0018] Preferably, feature extraction is performed on the reflection signal of the suspected defective area, and the extracted features include the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology. After deep analysis of the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology, a scattering asymmetric reference value and a depth change reference value are generated respectively. The scattering asymmetric reference value is used to quantify the degree of unevenness of signal scattering after laser reflection, and the depth change reference value is used to quantify the height difference of the three-dimensional morphology of the wafer surface. The scattering asymmetric reference value and the depth change reference value obtained after the depth analysis are pre-trained to predict the defect conspicuity coefficient of the convolutional neural network, and the conspicuity of potential defects in the wafer film is evaluated by the defect conspicuity coefficient.

[0019] Preferably, the defect obvious coefficient generated by predicting the potential defects in the suspected defect area through the pre-trained convolutional neural network is compared and analyzed with the pre-set defect obvious coefficient reference threshold, and the suspected defect area is further divided into an obvious defect area and an inconspicuous defect area. The specific division steps are:

[0020] If the defect obvious coefficient is greater than or equal to the defect obvious coefficient reference threshold, the suspected defect area is divided into an obvious defect area; if the defect obvious coefficient is less than the defect obvious coefficient reference threshold, the suspected defect area is divided into an inconspicuous defect area.

[0021] Preferably, the specific steps of generating a scattering asymmetry reference value after performing a deep analysis on the asymmetry of the signal scattering after the laser is reflected on the wafer surface are as follows:

[0022] In the suspected defect area, the laser reflection signal is collected in multiple directions. The collected signals are expressed as reflection intensities in different directions. The laser reflection signal is expressed as a vector field S (x, y, z), where the signal at each point is decomposed into vector components in different directions. The decomposition expression is: Where S x , S y , S z They represent the components of the laser signal in three dimensions, is the corresponding unit direction vector;

[0023] The signal scattering inhomogeneity is evaluated by comparing the vector differences of adjacent signal points, and the signal difference between each point and its neighboring points is defined as ΔS ij , then: ΔS ij =‖S(x i ,y i , z i )-S(x j ,y j , z j )‖, where ΔS ij is the vector difference between the i-th point and the j-th point, ‖·‖ represents the Euclidean distance of the vector; then the signal difference matrix D is constructed, where each element D ij It represents the difference between the i-th point and the j-th point. The construction expression of the signal difference matrix is: D = [D ij ]=[ΔS ij ];

[0024] Based on the signal difference matrix, the scattering asymmetric reference value is calculated. The scattering asymmetric reference value is generated by extracting the high-order change information in the signal difference matrix and defining the scattering asymmetric reference value as the harmonic incremental accumulation of the signal difference matrix. The calculation expression of the scattering asymmetric reference value is:

[0025]

[0026] , where γ is the weight parameter controlling spatial attenuation, r ij represents the Euclidean distance between point i and point j, S asym is the reference value of scattering asymmetry.

[0027] Preferably, the specific steps of generating a depth change reference value after performing depth analysis on the depth change on the three-dimensional topography are as follows:

[0028] For the suspected defect area, obtain the three-dimensional morphological data of the wafer surface in the suspected defect area, and assume that the dot matrix data of the laser scanning is (x u ,y u , z u ), where (x u ,y u ) is the two-dimensional coordinate of the wafer surface, z u The spatial distribution of the wafer surface height is represented by a three-dimensional height function, where the expression of the three-dimensional height function is: Z(x u ,y u )=f(x u ,y u ), where Z(x u ,y u ) The three-dimensional height data of the wafer surface reflects the surface morphology, f(x u ,y u ) is the height distribution function collected from the two-dimensional coordinates of the wafer surface;

[0029] The 3D height gradient and 3D height curvature of the collected 3D topography data are calculated to quantify the severity of the surface change in the suspected defect area. The 3D height gradient is defined as the directional change of the 3D surface height, and the 3D height curvature indicates the degree of curvature of the wafer surface. The calculation expression is:

[0030]

[0031] In the formula, Represents the two-dimensional coordinates (x u ,y u ), Represents the three-dimensional height gradient change of the three-dimensional height function along the x-axis, It represents the three-dimensional height gradient change of the three-dimensional height function along the y-axis, k(x u ,y u ) represents the two-dimensional coordinate (x u ,yu ) at the three-dimensional height curvature;

[0032] By calculating the two-dimensional coordinates (x u ,y u ) and the three-dimensional height curvature κ(x u ,y u ) calculates the two-dimensional coordinates (x u ,y u ) is the depth variation S(x u ,y u ), the expression for calculation is:

[0033] According to the two-dimensional coordinates (x u ,y u ) is the depth variation S(x u ,y u ) calculates the depth variation reference value, which is used to quantify the height difference of the three-dimensional topography of the wafer surface. The calculation expression of the depth variation reference value is: var =∫ A S(x u ,y u )dxdy,Depth var It represents the depth change reference value, which is defined as the integral of the depth change severity in the entire suspected defect area, that is, the comprehensive measure of the surface change of all points in the suspected defect area.

[0034] Preferably, for areas where defects are not obvious, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. The specific steps are as follows:

[0035] The relative measure of the defect obvious coefficient in the defect inconspicuous area relative to the reference threshold of the defect obvious coefficient is calculated. The calculation expression is:

[0036]

[0037] , where τ 显著 Represents a relative significance measure, used to measure the current defect significance coefficient Defect sign Defect relative to the defect coefficient reference threshold 阈值 proportion;

[0038] According to the relative significance measure τ 显著 Calculate the coverage adjustment factor to determine the amount of beam coverage that needs to be increased. The calculation expression of the coverage adjustment factor is: Where ΔC is the coverage adjustment factor, which is used to dynamically expand the beam coverage, Cp is the preset beam coverage, the coverage during the initial scan, and ω is the adjustment factor, which controls the sensitivity and nonlinearity of the coverage adjustment factor;

[0039] Apply the calculated coverage adjustment factor ΔC to the preset beam coverage C p , to obtain the new beam coverage, which is used to perform a higher-density secondary scan of the suspected defect area. The calculation expression of the new beam coverage is: In the formula, C new is the coverage of the new beam, which is applied to the secondary high-density scanning.

[0040] A wafer film defect detection system includes a preliminary scanning module, a signal analysis and marking module, a feature extraction and classification module, a dynamic expansion scanning module, and a fine scanning and positioning module;

[0041] A preliminary scanning module, during the wafer film defect detection process, emits a laser beam according to a preset beam coverage rate through a laser scanner to cover the wafer surface for preliminary scanning;

[0042] The signal analysis and marking module analyzes the intensity and consistency of the laser reflection signal in real time during the laser beam scanning process to identify the reflection difference area between the wafer surface and the film, and marks the reflection difference area as a suspected defect area;

[0043] The feature extraction and classification module extracts features from the reflection signals of suspected defective areas and performs in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the module uses a pre-trained convolutional neural network to evaluate the degree of prominence of potential defects in the wafer film and further divides the suspected defective areas into areas with obvious defects and areas with inconspicuous defects.

[0044] Dynamically expand the scanning module. For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects.

[0045] The fine scanning and positioning module uses the adjusted beam coverage to emit a higher-density laser beam in the marked suspected defect area for a second scan. It uses a high-density laser beam to capture more subtle surface information and combines it with three-dimensional morphology data to obtain the three-dimensional features of the defect, accurately locate and identify film defects.

[0046] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0047] The present invention performs a preliminary scan by presetting the beam coverage, quickly identifies potential defect areas, and uses a convolutional neural network to evaluate the significance of potential defect areas, thereby achieving accurate classification of areas with obvious defects and areas with inconspicuous defects. For areas with inconspicuous defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network, and a high-density secondary scan is performed again, which can capture more subtle surface information and accurately locate the specific position and shape of the defects in combination with three-dimensional morphological data. It can effectively reduce the risk of missed detection and false detection while maintaining detection efficiency, ensure accurate detection of minor defects, significantly improve the quality control level of wafer film, and avoid losses and quality problems caused by missed detection in subsequent production. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0049] Figure 1 The present invention is a method flow chart of a wafer film defect detection method.

[0050] Figure 2 It is a module schematic diagram of a wafer film defect detection system of the present invention. DETAILED DESCRIPTION

[0051] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0052] The present invention provides Figure 1 A wafer film defect detection method shown includes the following steps:

[0053] During the wafer film defect detection process, the laser scanner emits a laser beam according to the preset beam coverage rate to cover the wafer surface for preliminary scanning;

[0054] During the wafer film defect detection process, a laser scanner emits a laser beam according to a preset beam coverage rate to perform a preliminary scan of the wafer surface in order to quickly identify potential abnormal areas. Compared with the existing technology, the preset beam coverage rate is slightly lower, mainly used to improve detection efficiency and reduce costs. This lower coverage rate can initially detect abnormal phenomena such as reflection differences in wafer film, but cannot accurately identify defects. In this way, areas where defects may exist can be quickly locked without the need for high-density scanning of the entire wafer surface, thereby ensuring detection speed while avoiding unnecessary waste of resources. The preset beam coverage rate is not limited to a fixed value or setting, but is flexibly adjusted according to actual process requirements and specific product characteristics to meet the detection requirements in different scenarios. By setting the coverage rate reasonably, reference data can be provided for subsequent more accurate detection steps without sacrificing the efficiency of preliminary detection.

[0055] During the laser beam scanning process, the intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film, and mark the reflection difference area as a suspected defect area;

[0056] The intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film. The specific steps are as follows:

[0057] The laser scanner emits a laser beam according to the preset beam coverage rate, covers the entire wafer surface for scanning, collects the reflected light signal in real time, and records the reflected light intensity value I(x, y) of each scanning point as a two-dimensional coordinate matrix, where I(x, y) represents the reflected light intensity value at (x, y), where x and y represent the spatial coordinates of the scan, and the light intensity information of the entire wafer surface is stored through the two-dimensional coordinate matrix;

[0058] In order to identify the reflection difference area between the wafer surface and the film, the light intensity value I(x, y) of each point (x, y) is calculated for local consistency. Specifically, the light intensity of the point is compared with the light intensity of other points in its neighborhood, and the local consistency deviation D(x, y) is calculated. The calculation expression is: In the formula, (x l ,y l ) is the neighborhood point of point (x, y), l represents the index of the neighborhood point, and N is the number of points in the neighborhood;

[0059] A higher D(x, y) value indicates that there is a significant difference in reflection between the point and the surrounding area, which may be a defect area.

[0060] After calculating the local consistency deviation D(x, y) of each point, D(x, y) is compared with the preset threshold T. If the local consistency deviation D(x, y) of the point is greater than T, the point is marked as a suspected defect point.

[0061] The threshold T can be adjusted according to the specific production process and material characteristics, and is not specifically limited here to ensure that the detection sensitivity is appropriate.

[0062] Apply a clustering algorithm (such as K-Means or DBSCAN) to process the marked suspected defect points, aggregate the defect points according to spatial distribution and density, and form a continuous defect area. The continuous defect area is the reflection difference area between the wafer surface and the film;

[0063] During the inspection process, all detected suspected defect points are grouped according to their distribution positions in space and the density of surrounding points through clustering algorithms. When some defect points are close in space and have similar distribution characteristics, these points will be classified into the same group to form an overall defect area. Through this process, large-area defects or continuously distributed defect areas can be more clearly identified, which is convenient for subsequent analysis and further processing. The aggregation step helps to integrate discrete points into a meaningful area, reflecting a more obvious defect pattern.

[0064] These differences may be caused by defects such as bubbles, scratches, uneven film layers, etc. When the reflected signal of the laser beam deviates significantly from the preset normal range, the suspected defect area will be automatically marked as a suspected defect area. This step of optimization can improve the sensitivity of difference detection through high-precision sensors and signal analysis algorithms, ensuring that subtle reflection changes can be detected, especially for the early identification of minor defects.

[0065] Extract features from the reflection signals of suspected defective areas, and perform in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the pre-trained convolutional neural network is used to evaluate the degree of prominence of potential defects in the wafer film, and the suspected defective areas are further divided into obvious defect areas and inconspicuous defect areas.

[0066] Feature extraction is performed on the reflection signal of the suspected defective area. The extracted features include the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology. After in-depth analysis of the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology, a scattering asymmetry reference value and a depth change reference value are generated respectively. The scattering asymmetry reference value is used to quantify the degree of unevenness of signal scattering after laser reflection. The larger the value, the more significant the wafer surface unevenness, and there may be obvious defects. The depth change reference value is used to quantify the height difference of the three-dimensional morphology of the wafer surface. The larger the value, the greater the degree of surface unevenness, indicating that there may be serious defects. The scattering asymmetry reference value and the depth change reference value obtained after the in-depth analysis are pre-trained with the convolutional neural network to predict the defect obvious coefficient, and the defect obvious coefficient is used to evaluate the degree of obviousness of potential defects in wafer film.

[0067] The pre-trained convolutional neural network is a deep learning model specifically designed to process data with a grid structure. The convolutional neural network extracts the features of the data through a series of convolutional layers, pooling layers, and fully connected layers, and can automatically identify and learn feature representations at different levels. In wafer film defect detection, the convolutional layer is responsible for extracting local features of the original data (such as scattering asymmetric reference values ​​and depth variation reference values), while the pooling layer is used to reduce feature dimensions, retain important information, and prevent overfitting. This structure makes the convolutional neural network both efficient and powerful in feature learning when processing high-dimensional data, and is particularly suitable for image and signal data.

[0068] Pre-trained convolutional neural networks are usually trained on large-scale datasets to learn effective feature representations. When applied to specific tasks (such as wafer film defect detection), the pre-learned weights and features can be directly used for new tasks through transfer learning. This method not only saves training time, but also improves the accuracy and generalization ability of the model. In this application, the trained convolutional neural network is able to combine the scattering asymmetry reference value and the depth variation reference value to generate a defect visibility coefficient. The defect visibility coefficient is used to quantify and evaluate the severity of potential defects, thereby helping to identify and repair defects in a timely manner in the production process and improve the yield and reliability of the product.

[0069] When the signal scattering asymmetry after the laser is reflected on the wafer surface is high, it usually indicates that there are obvious defects in the suspected defective area. Scattering asymmetry reflects the difference in reflection intensity of the laser signal in different directions, which is usually caused by the unevenness of the wafer surface or film, such as bubbles, scratches, uneven film layers, etc. These defects will cause the optical path to deviate when the laser is reflected, which increases the difference in scattering intensity in different directions, thus showing a higher asymmetry. On the contrary, if the scattering asymmetry is low, it means that the laser is not significantly disturbed when reflected on the surface, the surface is relatively smooth and uniform, so the defects in the suspected area are not obvious or do not exist. By analyzing the scattering asymmetry, the severity of the suspected defective area can be quickly assessed. Higher asymmetry means that the reflected signal is disturbed by the irregular structure of the surface, which can be used as a reliable indicator to mark the defective area.

[0070] The specific steps for generating a scattering asymmetry reference value after in-depth analysis of the asymmetry of the signal scattering after the laser is reflected on the wafer surface are as follows:

[0071] In the suspected defect area, the laser reflection signal is collected in multiple directions. The collected signal can be expressed as the reflection intensity in different directions. The laser reflection signal is expressed as a vector field S (x, y, z), where the signal of each point is decomposed into vector components in different directions. The decomposition expression is: Where S x , S y , S z They represent the components of the laser signal in three dimensions, is the corresponding unit direction vector;

[0072] The purpose of this step is to extract the reflection intensity components of each point in different directions, laying the foundation for subsequent inhomogeneity analysis.

[0073] The signal scattering inhomogeneity is evaluated by comparing the vector differences of adjacent signal points, and the signal difference between each point and its neighboring points is defined as ΔS ij , then: ΔS ij =‖S(x i ,y i , z i )-S(x j ,y j , z j )‖, where ΔS ij is the vector difference between the i-th point and the j-th point, ‖·‖ represents the Euclidean distance of the vector; then the signal difference matrix D is constructed, where each element D ij It represents the difference between the i-th point and the j-th point. The construction expression of the signal difference matrix is: D = [D ij ]=[ΔSij ], the larger the element in the difference matrix, the greater the signal difference between the point and its neighboring points, reflecting the inhomogeneity of the scattering signal;

[0074] This step quantifies the spatial variation pattern of the signal into a difference matrix that captures the variation in signal scattering in different directions.

[0075] Based on the signal difference matrix, the scattering asymmetric reference value is calculated. The scattering asymmetric reference value is generated by extracting the high-order change information in the signal difference matrix. The scattering asymmetric reference value is defined as the harmonic incremental accumulation of the signal difference matrix. The calculation expression of the scattering asymmetric reference value is:

[0076]

[0077] , where γ is the weight parameter controlling spatial attenuation, r ij represents the Euclidean distance between point i and point j, S asym is the scattering asymmetry reference value;

[0078] The weight parameter that controls spatial attenuation refers to an adjustment factor used to adjust the degree of mutual influence between adjacent signal points when analyzing signals. It reduces the influence of distant signal points on the overall signal analysis by assigning smaller weights to distant signal points, thereby highlighting the relationship between close-range signals. The setting of this parameter needs to be adjusted according to the specific application scenario: when paying attention to local signal changes, the weight parameter γ should be set larger to enhance the influence of close-range signals; when it is necessary to observe signal trends in a larger range, γ should be set smaller so that distant signals also have a certain influence. In general, the value of γ is tuned through experiments or based on data features to obtain the best signal attenuation effect.

[0079] The formula harmonizes and weights the difference values ​​in the signal difference matrix to highlight the asymmetric characteristics of the reflected signal in different directions. The larger the scattering asymmetry reference value, the higher the degree of signal non-uniformity, which in turn reflects the degree of obvious defects on the wafer surface.

[0080] It can be seen from the scattering asymmetry reference value that the larger the performance value of the scattering asymmetry reference value generated after the in-depth analysis of the asymmetry of the signal scattering after the laser is reflected on the wafer surface, the stronger the unevenness of the signal scattering after the laser is reflected on the wafer surface, which usually means that there is a more obvious defect in the suspected defect area. A larger scattering asymmetry reference value reflects that the intensity difference of the reflected signal in multiple directions is significant, indicating that there may be large irregularities or defects on the surface; conversely, when the scattering asymmetry reference value is smaller, it means that the signal scattering is more uniform, the surface is smoother or uniform, and the possibility of defects is smaller or less obvious. Therefore, the larger the scattering asymmetry reference value, the more obvious the defect, and the smaller the scattering asymmetry reference value, the less obvious the defect.

[0081] The depth variation in the three-dimensional morphology can be used as an important feature to judge suspected defective areas. When the depth variation of the three-dimensional morphology is large, it usually means that there are significant unevenness or irregular structures on the surface, which may be caused by serious defects, such as bubbles, scratches, or separation of film layers. These obvious depth differences often indicate that the defects in the suspected defective area not only affect the surface flatness, but may also have a negative impact on subsequent processing or the function of the wafer. Therefore, the greater the depth variation, the more severe the defect. When the depth variation of the three-dimensional morphology is small or close to zero, it means that the surface of the suspected defective area is relatively flat, the defects may be relatively small or insignificant, and are unlikely to have a significant impact on the overall quality.

[0082] The specific steps of generating a depth change reference value after depth analysis of the depth change on the three-dimensional topography are as follows:

[0083] For the suspected defect area, obtain the three-dimensional morphological data of the wafer surface in the suspected defect area, and assume that the dot matrix data of the laser scanning is (x u ,y u , z u ), where (x u ,y u ) is the two-dimensional coordinate of the wafer surface, z u The spatial distribution of the wafer surface height is represented by a three-dimensional height function, where the expression of the three-dimensional height function is: Z(x u ,y u )=f(x u ,y u ), where Z(x u ,y u ) The three-dimensional height data of the wafer surface reflects the surface morphology, f(x u ,y u ) is the height distribution function collected from the two-dimensional coordinates of the wafer surface;

[0084] Height data can be collected directly by laser scanners, representing surface height variations.

[0085] The 3D height gradient and 3D height curvature of the collected 3D topography data are calculated to quantify the severity of the surface changes in the suspected defect area. By calculating the 3D height gradient and 3D height curvature, the height difference and the change of the surface slope can be more accurately reflected, so as to identify the area with large depth changes. The 3D height gradient is defined as the directional change of the 3D surface height, and the 3D height curvature indicates the curvature of the wafer surface. The calculation expression is:

[0086]

[0087] In the formula, Represents the two-dimensional coordinates (x u ,y u ), Represents the three-dimensional height gradient change of the three-dimensional height function along the x-axis, represents the three-dimensional height gradient change of the three-dimensional height function along the y-axis, κ(x u ,y u ) represents the two-dimensional coordinate (x u ,y u ) at the three-dimensional height curvature;

[0088] By calculating the two-dimensional coordinates (x u ,y u ) and the three-dimensional height curvature κ(x u ,y u ) calculates the two-dimensional coordinates (x u ,y u ) is the depth variation S(x u ,y u ), the expression for calculation is:

[0089] The depth variation of the wafer surface S(x u ,y u ) through the three-dimensional height gradient and the three-dimensional height curvature κ(x u ,y u ) is used to comprehensively measure the severity of the three-dimensional height change on the wafer surface. The larger the three-dimensional height gradient, the faster the surface changes; the larger the three-dimensional height curvature, the more severe the surface curvature. The combination of the two can quantify complex surface changes.

[0090] According to the two-dimensional coordinates (xu ,y u ) is the depth variation S(x u ,y u ) calculates the depth variation reference value, which is used to quantify the height difference of the three-dimensional topography of the wafer surface. The calculation expression of the depth variation reference value is: var =∫ A S(x u ,y u )dxdy,Depth var It represents the depth change reference value. The depth change reference value is defined as the integral of the depth change severity in the entire suspected defect area, that is, the comprehensive measure of the surface change of all points in the suspected defect area. The larger the depth change reference value, the more severe the surface change of the suspected defect area is and the more obvious the defect is.

[0091] The depth variation reference value is used to quantify the height difference in the three-dimensional morphology. The size of the depth variation reference value directly reflects the obviousness of the defect. Specifically, the depth variation reference value is generated by analyzing the height change of the wafer surface. When the depth variation reference value is large, it indicates that the surface height of the suspected defect area changes dramatically, there are large unevenness or irregularities, and the defect is obvious. On the contrary, when the depth variation reference value is small, it means that the height difference of the suspected defect area is small, the surface is relatively flat, and the defect is not obvious. Therefore, the size of the depth variation reference value can be effectively evaluated.

[0092] The convolutional neural network is not specifically limited here, and can realize the scattering asymmetric reference value S asym And depth change reference value Depth var Perform comprehensive analysis to generate defect apparent coefficient Defect sign In order to realize the technical solution of the present invention, the present invention provides a specific implementation method; Defect obvious coefficient Defect sign The generated calculation formula is: Defect sign =α*S asym +β*Depth var , where α and β are the scattering asymmetric reference values ​​S asym And depth change reference value Depth var The preset proportional coefficient, and α and β are both greater than 0.

[0093] It can be seen from the defect visibility coefficient that the larger the expression value of the scattering asymmetry reference value generated after the in-depth analysis of the asymmetry of the signal scattering after the laser is reflected on the wafer surface, the larger the expression value of the depth change reference value generated after the in-depth analysis of the depth change on the three-dimensional morphology. That is, the larger the expression value of the defect visibility coefficient generated by predicting the potential defects in the suspected defect area through the pre-trained convolutional neural network, the more obvious the defects in the suspected defect area are, and vice versa.

[0094] The defect visibility coefficient generated by predicting the potential defects in the suspected defect area through the pre-trained convolutional neural network is compared and analyzed with the pre-set defect visibility coefficient reference threshold, and the suspected defect area is further divided into an obvious defect area and an inconspicuous defect area. The specific division steps are as follows:

[0095] If the defect obvious coefficient is greater than or equal to the defect obvious coefficient reference threshold, the suspected defect area is divided into an obvious defect area; if the defect obvious coefficient is less than the defect obvious coefficient reference threshold, the suspected defect area is divided into an inconspicuous defect area.

[0096] For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects.

[0097] For areas where defects are not obvious, the beam coverage is dynamically expanded based on the prediction results of the convolutional neural network to increase the laser scanning density. The specific steps are as follows:

[0098] The relative measure of the defect obvious coefficient in the defect inconspicuous area relative to the reference threshold of the defect obvious coefficient is calculated. The calculation expression is:

[0099]

[0100] , where τ 显著 Represents a relative significance measure, used to measure the current defect significance coefficient Defect sign Defect relative to the defect coefficient reference threshold 阈值 proportion;

[0101] Defect sign , predicted by a convolutional neural network, and used to quantify the visibility of defects.

[0102] According to the relative significance measure τ 显著Calculate the coverage adjustment factor to determine the amount of beam coverage that needs to be increased. This adjustment factor is based on a nonlinear function to adjust the relative significance metric τ 显著 When approaching the reference threshold of defect noticeability coefficient, more significant adjustment is made. The calculation expression of coverage adjustment factor is: Where ΔC is the coverage adjustment factor, which is used to dynamically expand the beam coverage, C p is the preset beam coverage, the coverage during the initial scan, and ω is the adjustment factor, which controls the sensitivity and nonlinearity of the coverage adjustment factor. It is usually determined by experiments, such as ω = 2;

[0103] Apply the calculated coverage adjustment factor ΔC to the preset beam coverage C p , to obtain the new beam coverage, which is used to perform a higher-density secondary scan of the suspected defect area. The calculation expression of the new beam coverage is: In the formula, C new is the coverage of the new beam, which is applied to the secondary high-density scanning.

[0104] Through the adjusted beam coverage, a higher-density laser beam is emitted in the marked suspected defect area for secondary scanning. The high-density laser beam is used to capture more subtle surface information, and the three-dimensional features of the defect are obtained in combination with the three-dimensional morphology data, so as to accurately locate and identify the film defects.

[0105] Compared with the initial scan, the secondary scan uses a higher-density laser beam, which can capture the tiny information on the wafer surface more carefully, especially those subtle defects that cannot be accurately detected in the initial scan. The high-density laser beam can not only record the optical reflection characteristics of the surface, but also combine the three-dimensional morphology data, that is, the strength and change of the laser beam reflection can be used to obtain the subtle height, convexity, depth and other spatial information of the wafer surface. Through this process, the system can present the appearance, location and severity of the defects in three dimensions. This detection method that combines optical and geometric features helps to more accurately locate the specific location of the defects and identify the specific types of defects, such as bubbles, scratches or uneven film layers. Finally, through the secondary scan of the high-density laser beam, the system can further verify the results of the initial detection and make more accurate defect judgments to ensure that effective treatment measures can be taken for these defects in subsequent processes.

[0106] The present invention performs a preliminary scan with a preset beam coverage, quickly identifies potential defect areas, and uses a convolutional neural network to evaluate the significance of potential defect areas, thereby achieving accurate classification of areas with obvious defects and areas with less obvious defects. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network, and a high-density secondary scan is performed again, which can capture more subtle surface information and accurately locate the specific position and shape of the defects in combination with the three-dimensional morphology data. It can effectively reduce the risks of missed detection and false detection while maintaining detection efficiency, ensure the accurate detection of tiny defects, significantly improve the quality control level of wafer film pasting, and avoid losses and quality problems caused by missed detection in subsequent production.

[0107] The present invention provides Figure 2 A wafer film defect detection system shown includes a preliminary scanning module, a signal analysis and marking module, a feature extraction and classification module, a dynamic expansion scanning module, and a fine scanning and positioning module;

[0108] A preliminary scanning module, during the wafer film defect detection process, emits a laser beam according to a preset beam coverage rate through a laser scanner to cover the wafer surface for preliminary scanning;

[0109] The signal analysis and marking module analyzes the intensity and consistency of the laser reflection signal in real time during the laser beam scanning process to identify the reflection difference area between the wafer surface and the film, and marks the reflection difference area as a suspected defect area;

[0110] The feature extraction and classification module extracts features from the reflection signals of suspected defective areas and performs in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the module uses a pre-trained convolutional neural network to evaluate the degree of prominence of potential defects in the wafer film and further divides the suspected defective areas into areas with obvious defects and areas with inconspicuous defects.

[0111] Dynamically expand the scanning module. For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects.

[0112] The fine scanning and positioning module uses the adjusted beam coverage to emit a higher-density laser beam in the marked suspected defect area for secondary scanning. The high-density laser beam is used to capture more subtle surface information, and combined with the three-dimensional morphology data to obtain the three-dimensional features of the defect, accurately locate and identify the film defects;

[0113] A wafer film defect detection method provided in an embodiment of the present invention is implemented by the above-mentioned wafer film defect detection system. The specific method and process of a wafer film defect detection system are detailed in the embodiment of the above-mentioned wafer film defect detection method, which will not be repeated here.

[0114] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0115] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0116] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A wafer film defect detection method, characterized in that: The following steps are involved: During the wafer film defect detection process, the laser scanner emits a laser beam according to the preset beam coverage rate to cover the wafer surface for preliminary scanning; During the laser beam scanning process, the intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film, and mark the reflection difference area as a suspected defect area; Extract features from the reflection signals of suspected defective areas, and perform in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the pre-trained convolutional neural network is used to evaluate the degree of prominence of potential defects in the wafer film, and the suspected defective areas are further divided into obvious defect areas and inconspicuous defect areas. For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects. By adjusting the beam coverage, a higher density laser beam is emitted in the marked suspected defect area for a second scan. The high-density laser beam is used to capture more subtle surface information, and combined with the three-dimensional morphology data to obtain the three-dimensional features of the defect, accurately locate and identify the film defects.

2. A wafer film defect detection method according to claim 1, characterized in that: The intensity and consistency of the laser reflection signal are analyzed in real time to identify the reflection difference area between the wafer surface and the film. The specific steps are as follows: The laser scanner emits a laser beam according to the preset beam coverage, covers the entire wafer surface for scanning, collects the reflected light signal in real time, and records the reflected light intensity value I(x, y) of each scanning point as a two-dimensional coordinate matrix, where I(x, y) represents the reflected light intensity value at (x, y), where x and y represent the spatial coordinates of the scan, and stores the light intensity information of the entire wafer surface through the two-dimensional coordinate matrix; The local consistency calculation is performed on the light intensity value I(x, y) of each point (x, y), specifically: the light intensity of the point is compared with the light intensity of other points in its neighborhood, and the local consistency deviation D(x, y) is calculated. The calculation expression is: In the formula, (x l ,y l ) is the neighborhood point of point (x, y), l represents the index of the neighborhood point, and N is the number of points in the neighborhood; After calculating the local consistency deviation D(x, y) of each point, D(x, y) is compared with the preset threshold T. If the local consistency deviation D(x, y) of the point is greater than T, the point is marked as a suspected defect point. A clustering algorithm is applied to process the marked suspected defect points, and the defect points are aggregated according to spatial distribution and density to form a continuous defect area. The continuous defect area is the reflection difference area between the wafer surface and the film.

3. The method for detecting wafer film defects according to claim 1, characterized in that: Feature extraction is performed on the reflection signal of the suspected defective area, and the extracted features include the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology. After in-depth analysis of the asymmetry of signal scattering after the laser is reflected on the wafer surface and the depth change in the three-dimensional morphology, a scattering asymmetric reference value and a depth change reference value are generated respectively. The scattering asymmetric reference value is used to quantify the degree of unevenness of signal scattering after laser reflection, and the depth change reference value is used to quantify the height difference of the three-dimensional morphology of the wafer surface. The scattering asymmetric reference value and the depth change reference value obtained after the depth analysis are pre-trained with the convolutional neural network to predict the defect conspicuity coefficient, and the conspicuity of potential defects in the wafer film is evaluated by the defect conspicuity coefficient.

4. The method for detecting wafer film defects according to claim 3, characterized in that: The defect visibility coefficient generated by predicting the potential defects in the suspected defect area through the pre-trained convolutional neural network is compared and analyzed with the pre-set defect visibility coefficient reference threshold, and the suspected defect area is further divided into an obvious defect area and an inconspicuous defect area. The specific division steps are as follows: If the defect obvious coefficient is greater than or equal to the defect obvious coefficient reference threshold, the suspected defect area is divided into an obvious defect area; if the defect obvious coefficient is less than the defect obvious coefficient reference threshold, the suspected defect area is divided into an inconspicuous defect area.

5. The method for detecting wafer film defects according to claim 3, characterized in that: The specific steps for generating a scattering asymmetry reference value after in-depth analysis of the asymmetry of the signal scattering after the laser is reflected on the wafer surface are as follows: In the suspected defect area, the laser reflection signal is collected in multiple directions. The collected signals are expressed as reflection intensities in different directions. The laser reflection signal is expressed as a vector field S (x, y, z), where the signal at each point is decomposed into vector components in different directions. The decomposition expression is: Where S x , S y , S z They represent the components of the laser signal in three dimensions, is the corresponding unit direction vector; The signal scattering inhomogeneity is evaluated by comparing the vector differences of adjacent signal points, and the signal difference between each point and its neighboring points is defined as ΔS ij , then: ΔS ij =‖S(x i ,y i , z i )-S(x j ,y j , z j )‖, where ΔS ij is the vector difference between the i-th point and the j-th point, ‖·‖ represents the Euclidean distance of the vector; then the signal difference matrix D is constructed, where each element D ij It represents the difference between the i-th point and the j-th point. The construction expression of the signal difference matrix is: D = [D ij ]=[ΔS ij ]; Based on the signal difference matrix, the scattering asymmetric reference value is calculated. The scattering asymmetric reference value is generated by extracting the high-order change information in the signal difference matrix and defining the scattering asymmetric reference value as the harmonic incremental accumulation of the signal difference matrix. The calculation expression of the scattering asymmetric reference value is: Among them, γ is the weight parameter that controls spatial attenuation, r ij represents the Euclidean distance between point i and point j, S asym is the reference value of scattering asymmetry.

6. The method for detecting wafer film defects according to claim 3, characterized in that: The specific steps of generating a depth change reference value after depth analysis of the depth change on the three-dimensional topography are as follows: For the suspected defect area, obtain the three-dimensional morphological data of the wafer surface in the suspected defect area, and assume that the dot matrix data of the laser scanning is (x u ,y u , z u ), where (x u ,y u ) is the two-dimensional coordinate of the wafer surface, z u The spatial distribution of the wafer surface height is represented by a three-dimensional height function, where the expression of the three-dimensional height function is: Z(x u ,y u )=f(x u ,y u ), where Z(x u ,y u ) The three-dimensional height data of the wafer surface reflects the surface morphology, f(x u ,y u ) is the height distribution function collected from the two-dimensional coordinates of the wafer surface; The 3D height gradient and 3D height curvature of the collected 3D topography data are calculated to quantify the severity of the surface change in the suspected defect area. The 3D height gradient is defined as the directional change of the 3D surface height, and the 3D height curvature indicates the degree of curvature of the wafer surface. The calculation expression is: In the formula, Represents the two-dimensional coordinates (x u ,y u ), Represents the three-dimensional height gradient change of the three-dimensional height function along the x-axis, represents the three-dimensional height gradient change of the three-dimensional height function along the y-axis, κ(x u ,y u ) represents the two-dimensional coordinate (x u ,y u ) at the three-dimensional height curvature; By calculating the two-dimensional coordinates (x u ,y u ) and the three-dimensional height curvature κ(x u ,y u ) Calculate the two-dimensional coordinates (x u ,y u ) is the depth variation S(x u ,y u ), the expression for calculation is: According to the two-dimensional coordinates (x u ,y u ) is the depth variation S(x u ,y u ) calculates the depth variation reference value, which is used to quantify the height difference of the three-dimensional topography of the wafer surface. The calculation expression of the depth variation reference value is: var =∫ A S(x u ,y u )dxdy,Depth var It represents the depth change reference value, which is defined as the integral of the depth change severity in the entire suspected defect area, that is, the comprehensive measure of the surface change of all points in the suspected defect area.

7. The method for detecting wafer film defects according to claim 4, characterized in that: For areas where defects are not obvious, the beam coverage is dynamically expanded based on the prediction results of the convolutional neural network to increase the laser scanning density. The specific steps are as follows: The relative measure of the defect obvious coefficient in the defect inconspicuous area relative to the reference threshold of the defect obvious coefficient is calculated. The calculation expression is: In the formula, τ 显著 Represents a relative significance measure, used to measure the current defect significance coefficient Defect sign Defect relative to the defect coefficient reference threshold 阈值 proportion; According to the relative significance measure τ 显著 Calculate the coverage adjustment factor to determine the amount of beam coverage that needs to be increased. The calculation expression of the coverage adjustment factor is: Where ΔC is the coverage adjustment factor, which is used to dynamically expand the beam coverage, C p is the preset beam coverage, the coverage during the initial scan, and ω is the adjustment factor, which controls the sensitivity and nonlinearity of the coverage adjustment factor; Apply the calculated coverage adjustment factor ΔC to the preset beam coverage C p , to obtain the new beam coverage, which is used to perform a higher-density secondary scan of the suspected defect area. The calculation expression of the new beam coverage is: In the formula, C new is the coverage of the new beam, which is applied to the secondary high-density scanning.

8. A wafer film defect detection system, used to implement the wafer film defect detection method according to any one of claims 1 to 7, characterized in that: It includes preliminary scanning module, signal analysis and marking module, feature extraction and classification module, dynamic expansion scanning module and fine scanning and positioning module; A preliminary scanning module, during the wafer film defect detection process, emits a laser beam according to a preset beam coverage rate through a laser scanner to cover the wafer surface for preliminary scanning; The signal analysis and marking module analyzes the intensity and consistency of the laser reflection signal in real time during the laser beam scanning process to identify the reflection difference area between the wafer surface and the film, and marks the reflection difference area as a suspected defect area; The feature extraction and classification module extracts features from the reflection signals of suspected defective areas and performs in-depth analysis on the extracted feature data. Based on the features after in-depth analysis, the module uses a pre-trained convolutional neural network to evaluate the degree of prominence of potential defects in the wafer film and further divides the suspected defective areas into areas with obvious defects and areas with inconspicuous defects. Dynamically expand the scanning module. For areas with obvious defects, the wafer lamination is directly judged as unqualified. For areas with less obvious defects, the beam coverage is dynamically expanded according to the prediction results of the convolutional neural network to increase the laser scanning density. At this time, the laser scanning range no longer covers the entire wafer, but is concentrated in the area with less obvious defects. The fine scanning and positioning module uses the adjusted beam coverage to emit a higher-density laser beam in the marked suspected defect area for a second scan. It uses a high-density laser beam to capture more subtle surface information and combines it with three-dimensional morphology data to obtain the three-dimensional features of the defect, accurately locate and identify film defects.

Images