A multi-dimensional defect intelligent detection system for wafer manufacturing process

Abstract

This invention belongs to the field of wafer defect detection technology and relates to a multi-dimensional intelligent defect detection system for wafer manufacturing processes. It includes: an acquisition module configured to acquire bright-field reflection and dark-field scattering images of the wafer to be inspected, and determine the directional consistency value of the bright-field reflection and dark-field scattering images at a pixel; a first determination module configured to determine a dark-field cleanup value of a pixel using the directional consistency value, the dark-field grayscale value of the same pixel, and the bright-field grayscale value; a second determination module configured to determine a bright-field reconstruction value of a pixel; and a third determination module configured to determine the feature vector magnitude using the dark-field cleanup value and the bright-field reconstruction value, and to determine a defect evaluation index using the dark-field cleanup value, the bright-field reconstruction value, and the feature vector magnitude, thereby determining the defect detection result. This system can balance the cost and accuracy of wafer defect detection.

Description

Technical Field

[0001] This invention belongs to the field of wafer defect detection technology and relates to a multi-dimensional intelligent defect detection system for wafer manufacturing processes. It is used in situations where image processing technology is used to detect defects in wafers, and can balance the cost and accuracy of wafer defect detection. Background Technology

[0002] In the field of semiconductor integrated circuit manufacturing technology, wafers are the core basic carrier for semiconductor chip fabrication. Typically made from electronic-grade single-crystal silicon as the substrate, wafers are processed into circular thin films through multiple processes such as crystal pulling and slicing. They serve as the core substrate and functional carrier for manufacturing various semiconductor products, including integrated circuits, discrete devices, and optoelectronic devices. The entire wafer fabrication process, from substrate preparation to chip manufacturing, involves numerous and complex steps. Detecting potential defects in wafers is a core step in quality control during semiconductor manufacturing and a crucial link in ensuring chip mass production capabilities and product performance.

[0003] In Chinese patent applications with publication numbers CN118967600A, CN117252861A, and CN117423639A, the main methods for obtaining defect detection results are to acquire images of the wafer to be inspected and input these images into a pre-trained detection model, or to directly segment the images of the wafer to be inspected. However, training the model requires traversing a large number of wafer sample images of different types, resulting in high costs, including time. Furthermore, directly segmenting the images of the wafer to be inspected to obtain defect detection results makes it difficult to obtain accurate defect detection results for the wafer.

[0004] The wafer manufacturing process involves numerous steps and complex processes, resulting in extremely high processing costs for a single batch of wafers. Even a tiny defect on the wafer surface can cause the complete failure of the corresponding chip unit, and may also lead to hidden failure risks such as drift in chip electrical performance, excessive leakage current, and decreased anti-interference ability. Therefore, it is necessary to conduct defect detection in the wafer manufacturing process, taking into account both the cost and accuracy of wafer defect detection. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art, solve the technical problem of balancing the cost and accuracy of wafer defect detection, and provide a multi-dimensional intelligent defect detection system for the wafer manufacturing process.

[0006] To achieve the aforementioned objectives, this invention provides a multi-dimensional intelligent defect detection system for wafer manufacturing processes. Its main structure includes: an acquisition module configured to acquire a bright-field reflection image and a dark-field scattering image of the wafer to be inspected, and to determine a first local gradient vector of pixels in the bright-field reflection image and a second local gradient vector of pixels in the dark-field scattering image using an edge detection algorithm; and a first determination module configured to determine the directional consistency value of the bright-field reflection image and the dark-field scattering image at a pixel using the first and second local gradient vectors, and to determine the directional consistency value, the dark-field grayscale value of the same pixel, and the bright-field grayscale value. The first determination module is configured to: determine the dark field cleanup value of a pixel in a dark field scattering image; determine the local abrupt change value by using a second-order spatial operator to extract features from the bright field reflection image and determine the thickness background value by estimating the background of the bright field reflection image; and determine the bright field reconstruction value of the pixel using the local abrupt change value and the thickness background value; and determine the feature vector magnitude by orthogonally adding the dark field cleanup value and the bright field reconstruction value, and determine the defect judgment index using the dark field cleanup value, the bright field reconstruction value, and the feature vector magnitude, and determine the defect detection result using the defect judgment index, thereby enabling defect detection in the wafer manufacturing process.

[0007] The directional consistency value described in this invention is determined by the following method: determining the vector dot product result between the first local gradient vector and the second local gradient vector; taking the magnitude of the first local gradient vector as the first magnitude, taking the magnitude of the second local gradient vector as the second magnitude, determining the magnitude product between the first magnitude and the second magnitude, and taking the ratio between the absolute value of the vector dot product result and the magnitude product as the directional consistency value.

[0008] This invention constructs a standardized directional consistency quantification index by operating on vector dot product and feature modulus. Combined with a preset zero-prevention constant to avoid overflow anomalies in the denominator calculation process, it can stably and objectively reflect the degree of coordination between bright field and dark field in local structural change trends, thereby providing a reliable mathematical basis for accurately distinguishing between coherent scattering of normal patterns and independent scattering of random defects.

[0009] The dark field purification value described in this invention is determined as follows: a first adjustment coefficient, which is linearly negatively correlated with the directional consistency value, is determined using the difference between the upper limit of the directional consistency value and the directional consistency value; a second adjustment coefficient is determined using the difference between the global bright field extremum in the bright field reflection image and the bright field grayscale value of the pixel; the product of the first adjustment coefficient and the second adjustment coefficient is normalized to obtain the target adjustment coefficient of the target pixel in the dark field scattering image; the product of the dark field grayscale value of the target pixel and the target adjustment coefficient is used as the dark field purification value of the target pixel; the target pixel is any pixel in the dark field scattering image.

[0010] The bright-field reconstruction value described in this invention is determined as follows: the second derivative of the bright-field grayscale value of the bright-field reflection image is extracted using the Laplacian operator to determine the local abrupt change value; the bright-field grayscale value within the local calculation window is averaged to determine the local average grayscale; the bright-field grayscale value is spatially integrated using a weighted averaging mechanism to determine the thickness background value; the absolute difference between the bright-field grayscale value and the thickness background value is determined, the ratio of the local abrupt change value to the local average grayscale is used as the abrupt change coefficient, and the product of the abrupt change coefficient and the absolute difference is used as the bright-field reconstruction value.

[0011] This invention utilizes the Laplace operator to extract local abrupt values ​​that characterize high-frequency details, and uses a weighted average mechanism to separate the thickness background values ​​that characterize low-frequency thin-film interference. By redistributing the weights of local features through the product operation of the abrupt coefficient and the absolute difference, it is possible to effectively remove the wafer thickness variation interference in bright-field reflection signals and adaptively enhance the abnormal abrupt features.

[0012] The thickness background value described in this invention is determined as follows: the ratio of the square of the gradient modulus corresponding to the bright-field reflection image to the square of the preset smoothing coefficient is used as the exponential base, and the Gaussian weight is determined using the exponential base; the product of the Gaussian weight and the bright-field gray value is double-integrated within the spatial integration interval to determine the first integral term; the Gaussian weight is double-integrated within the spatial integration interval to determine the second integral term; and the ratio of the first integral term to the second integral term is used as the thickness background value.

[0013] The defect evaluation index described in this invention is determined in the following way: First, the first sum of the squared value of the dark field purification value and the weighted squared value of the bright field reconstruction value is determined; the square root of the first sum is used as the feature vector magnitude; a first product term is obtained by multiplying the dark field purification value and the bright field reconstruction value; a second product term is obtained by multiplying the global dark field extremum and the global bright field extremum; the ratio of the first product term to the second product term is used as the feature ratio; the sum of the feature ratio and a preset positive number is used as the enhancement factor value; the product of the feature vector magnitude and the enhancement factor value is used as the defect evaluation index.

[0014] The bright-field reflection image and dark-field scattering image described in this invention are acquired in the following manner: the original bright-field image and the original dark-field image of the wafer to be inspected are acquired; the original dark-field image is subjected to sub-pixel registration processing using a phase correlation algorithm combined with the original bright-field image to determine the coordinate-aligned image; the original bright-field image is used as the bright-field reflection image, and the coordinate-aligned image is used as the dark-field scattering image.

[0015] This invention employs a phase correlation algorithm to perform sub-pixel-level spatial transformation analysis in the frequency domain, which can effectively overcome the differences in grayscale distribution characteristics between the original bright field image and the original dark field image, achieve spatial coordinate alignment of cross-modal images, and avoid positional deviations caused by mechanical scanning.

[0016] The present invention describes a process for sub-pixel registration of an original dark-field image using a phase correlation algorithm combined with the original bright-field image to determine a coordinate-aligned image. This process includes: performing a discrete Fourier transform on the original bright-field image to obtain first frequency domain data; performing a discrete Fourier transform on the original dark-field image to obtain second frequency domain data; determining the cross-power spectrum matrix of the first and second frequency domain data in the complex frequency domain; performing cross-normalization on the cross-power spectrum matrix to determine a normalized matrix; performing an inverse Fourier transform on the normalized matrix to obtain a spatial pulse signal; searching for the position coordinates corresponding to the pulse extrema in the spatial pulse signal to determine the offset; and using the offset to perform bilinear interpolation resampling and completion on the original dark-field image to obtain a coordinate-aligned image.

[0017] The first local gradient vector of the present invention is determined by the following method: constructing a horizontal difference operator and a vertical difference operator in two-dimensional space using an edge detection algorithm; performing a convolution operation on the bright field gray value corresponding to the bright field reflection image using the horizontal difference operator to determine the horizontal gradient component; performing a convolution operation on the bright field gray value corresponding to the bright field reflection image using the vertical difference operator to determine the vertical gradient component; and merging the horizontal gradient component and the vertical gradient component into a vector to determine the first local gradient vector.

[0018] The multi-dimensional intelligent defect detection system for wafer manufacturing processes described in this invention determines the defect detection results in the following manner: when the defect evaluation index of a pixel is greater than a preset threshold, the pixel position is marked as a real defect point; when the defect evaluation index of a pixel is less than or equal to the preset threshold, the pixel position is marked as a noise point, and the defect definition process for the pixel position ends; using at least one of the obtained real defect points and noise points, the defect detection results of the wafer to be inspected are output.

[0019] This invention can classify continuously distributed defect evaluation indices by setting a preset threshold, accurately mark areas exceeding the threshold as real defect points, and eliminate the interference of noise points.

[0020] Compared with existing technologies, this invention has at least the following advantages: First, it calculates the local gradient vectors of the bright field and dark field of the wafer to be inspected, and extracts the direction consistency value to quantify the correlation between the two modes in the spatial structure. The correlation is used to purify the dark field image, which can effectively suppress the dark field scattering noise caused by the edge of the normal pattern. The bright field image is reconstructed by combining the second-order spatial operator and background estimation to eliminate the low-frequency interference caused by the change in film thickness. Second, it constructs a defect evaluation index by orthogonally adding the dual-mode purity features. This can accurately separate the normal structure signal from the real defect signal in the background of complex wafers, improve the accuracy of the detection results, and can reduce the cost of wafer defect detection while ensuring the accuracy of wafer defect detection. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structural principle of a multi-dimensional intelligent defect detection system for wafer manufacturing processes, which relates to the present invention.

[0022] Figure 2 This is a schematic diagram comparing the defect evaluation index with the histogram of the dark field scattering image, which is an important part of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings.

[0024] Example 1: This embodiment provides a multi-dimensional intelligent defect detection system 1000 for wafer manufacturing processes. Figure 1 This is a schematic diagram illustrating the structure of a multi-dimensional intelligent defect detection system 1000 for a wafer manufacturing process, according to an exemplary embodiment. Figure 1 As shown, the multi-dimensional defect intelligent detection system 1000 for wafer manufacturing process includes: an acquisition module 1100, a first determination module 1200, a second determination module 1300, and a third determination module 1400.

[0025] The acquisition module 1100 is configured to acquire a bright-field reflection image and a dark-field scattering image of the wafer to be inspected, and to use an edge detection algorithm to determine the first local gradient vector of the pixel in the bright-field reflection image and the second local gradient vector of the pixel in the dark-field scattering image.

[0026] The first determining module 1200 is configured to use a first local gradient vector and a second local gradient vector to determine the directional consistency value of the bright field reflection image and the dark field scattering image at a pixel point, and to use the directional consistency value, the dark field gray value of the same pixel point and the bright field gray value to determine the dark field purification value of the pixel point in the dark field scattering image.

[0027] The second determining module 1300 is configured to use a second-order spatial operator to extract features from the bright-field reflectance image to determine local abrupt change values, and to perform background estimation on the bright-field reflectance image to determine the thickness background value. The bright-field reconstructed value of the pixel is then determined using the local abrupt change value and the thickness background value.

[0028] The third determining module 1400 is configured to orthogonally add the dark field purification value and the bright field reconstruction value to determine the feature vector magnitude, and use the dark field purification value, the bright field reconstruction value and the feature vector magnitude to determine the defect evaluation index, and use the defect evaluation index to determine the defect detection result.

[0029] By integrating multimodal image processing with local feature orthogonal mapping, the system fully considers the different response characteristics of different wafer surface structures to different lighting modes. It uses pixel-level directional consistency to weaken background interference caused by regular patterns and uses second-order spatial domain extraction to filter out film thickness fluctuations. The multidimensional feature space is mapped into a unified defect evaluation index system, which improves the sensitivity to weak defect signals.

[0030] In one embodiment, the bright-field reflection image and the dark-field scattering image are acquired by: acquiring the original bright-field image and the original dark-field image of the wafer to be inspected; using a phase correlation algorithm to perform sub-pixel registration processing on the original dark-field image in combination with the original bright-field image to determine the coordinate-aligned image; using the original bright-field image as the bright-field reflection image and the coordinate-aligned image as the dark-field scattering image.

[0031] The wafer under inspection is usually composed of a silicon substrate and multiple layers of dielectric thin films and conductive metal layers with different refractive indices. During the manufacturing cycle, it needs to be illuminated by a broadband light source or a monochromatic laser source with a wavelength range of 200 nanometers to 800 nanometers. Since the original bright field image mainly receives the specular reflection beam after the light source is vertically illuminated, it is highly sensitive to the flat areas and large-scale film thickness distribution on the wafer surface. On the other hand, the original dark field image relies on an off-axis illumination system to receive the spatial scattering light caused by the undulation of the wafer surface or foreign particles, thus presenting high-frequency detail features. The two have significant energy inversion and morphological differences in imaging mechanism and grayscale distribution.

[0032] When performing stage displacement scanning on a semiconductor metrology instrument, micron-level mechanical vibration and servo positioning errors are unavoidable. If pixel-level algebraic operations are performed directly on the two original images with spatial deviations, it will cause serious feature misalignment, which will lead to the normal wafer structure edge being misjudged as an abnormal defect signal.

[0033] By using a phase correlation algorithm to convert the spatial domain image translation into a phase difference in the frequency domain, and then performing sub-pixel level geometric alignment operations, a spatial mapping relationship between the original bright-field image and the original dark-field image can be established. This ensures that the bright-field reflection image and the dark-field scattering image represent the same microstructure on the wafer surface under the same pixel coordinates, providing a spatially consistent data foundation for subsequent gradient direction consistency calculations and joint modal feature purification.

[0034] In one embodiment, a phase correlation algorithm is used to perform sub-pixel registration processing on the original dark field image in conjunction with the original bright field image to determine the coordinate-aligned image. This includes: performing a discrete Fourier transform on the original bright field image to obtain first frequency domain data, and performing a discrete Fourier transform on the original dark field image to obtain second frequency domain data; determining the cross-power spectrum matrix of the first and second frequency domain data in the complex frequency domain; performing cross-normalization processing on the cross-power spectrum matrix to determine a normalized matrix, and performing an inverse Fourier transform on the normalized matrix to obtain a spatial pulse signal; searching for the position coordinates corresponding to the pulse extrema in the spatial pulse signal to determine the offset; and using the offset to perform bilinear interpolation resampling and completion on the original dark field image to obtain the coordinate-aligned image.

[0035] By calculating the cross-power spectrum matrix, the relative phase difference between corresponding frequency components of two images can be extracted. The cross-normalization operation eliminates the influence of the intensity difference between the original bright field image and the original dark field image, as well as the uneven illumination, by dividing by the amplitude of the frequency domain component. This ensures that the normalized matrix after processing retains only the pure relative displacement phase information.

[0036] After the normalized matrix is ​​remapped back to the spatial domain by inverse Fourier transform, a spatial pulse signal with obvious peaks is formed. The coordinate center point where the pulse extremum is located corresponds to the global offset of the two images in the horizontal and vertical directions.

[0037] After determining the sub-pixel precision offset, such as an offset parameter of 1 or 3 pixels, bilinear interpolation resampling completion technology is applied. For example, the gray values ​​on the translated grid are recalculated based on the weighted average of the four surrounding adjacent pixels. Interpolation resampling can smoothly compensate for the grid misalignment error caused by non-integer coordinate translation while keeping the high-frequency edge information of the image from being over-smoothed, thus obtaining a coordinate-aligned image that is precisely aligned with the original bright-field image, so as to achieve a one-to-one correspondence between the two different optical modes at the spatial feature level.

[0038] In one embodiment, the first local gradient vector is determined by: constructing a horizontal difference operator and a vertical difference operator in two-dimensional space using an edge detection algorithm; performing a convolution operation on the bright field grayscale value corresponding to the bright field reflectance image using the horizontal difference operator to determine the horizontal gradient component; performing a convolution operation on the bright field grayscale value corresponding to the bright field reflectance image using the vertical difference operator to determine the vertical gradient component; and merging the horizontal gradient component and the vertical gradient component into a vector to determine the first local gradient vector.

[0039] In bright-field reflection images of wafer surfaces, the junctions of different dielectric layers or different lithographic patterns usually exhibit significant step changes in bright-field grayscale values. By utilizing the derivative characteristics of this grayscale distribution, the geometric features of local areas of the wafer can be extracted.

[0040] The horizontal and vertical difference operators constructed by the edge detection algorithm are essentially matrices that approximate the first derivative of the neighborhood of the center pixel. The common operator template scale is a matrix structure of 3 rows and 3 columns.

[0041] When the horizontal difference operator slides along the horizontal direction to perform convolution operation on the bright field reflectance image matrix, the intensity of the vertical edge structure is quantified by calculating the difference in gray values ​​between the left and right adjacent pixels. When the vertical difference operator slides along the vertical direction to perform convolution operation, the intensity of the horizontal edge structure is quantified by calculating the difference in gray values ​​between the upper and lower adjacent pixels.

[0042] The independently calculated horizontal and vertical gradient components are merged according to the vector law of the Cartesian coordinate system. The resulting first local gradient vector contains the gradient magnitude and gradient direction angle at the current pixel position. The first local gradient vector fully describes the maximum variation trend of bright-field optical reflection energy in the two-dimensional plane, providing a mathematical representation in the local differential dimension for subsequent comparison of the response mechanisms of bright-field and dark-field to the edge of the same wafer structure.

[0043] In one embodiment, the second local gradient vector is determined as follows: a horizontal difference operator and a vertical difference operator are constructed in two-dimensional space using an edge detection algorithm; the horizontal difference operator is used to perform a convolution operation on the dark field grayscale value corresponding to the dark field scattering image to determine the first dark field component; the vertical difference operator is used to perform a convolution operation on the dark field grayscale value corresponding to the dark field scattering image to determine the second dark field component; and the first dark field component and the second dark field component are vector-merged to determine the second local gradient vector.

[0044] Dark-field scattering images record the photon energy scattered from the wafer surface towards the collecting lens under illumination at a specific angle. Because the regular wiring patterns or contact hole arrays on the wafer have specific periodic spatial structures, these structures also generate diffraction and scattering signals with directional characteristics at their edges.

[0045] The first dark field component, determined by performing convolution operations on the dark field grayscale values ​​using the horizontal difference operator, characterizes the rate of change of scattered light intensity in the horizontal dimension. The second dark field component, determined by performing convolution operations using the vertical difference operator, characterizes the rate of change of scattered light intensity in the vertical dimension.

[0046] The second local gradient vector is formed by orthogonally splicing the first dark field component and the second dark field component. The second local gradient vector represents the direction of the most intense increase in local dark field scattering energy.

[0047] At the edge of the normal pattern on the wafer surface, due to the uniformity of the light field modulation caused by structural undulations, the second local gradient vector often has a fixed geometric relationship with the first local gradient vector of the bright field reflection image in terms of spatial orientation.

[0048] At the locations of randomly generated contaminated particles or irregular etching defects, isotropic scattering characteristics are usually observed, and edge step responses corresponding to bright-field images are often lacking. This makes the second local gradient vector the core feature carrier for identifying abnormal scattering sources.

[0049] In one embodiment, the directional consistency value is determined by: determining the vector dot product between the first local gradient vector and the second local gradient vector; taking the magnitude of the first local gradient vector as the first magnitude, taking the magnitude of the second local gradient vector as the second magnitude, determining the product of the magnitudes of the first and second magnitudes, and taking the ratio between the absolute value of the vector dot product and the product of the magnitudes as the directional consistency value.

[0050] Complex and dense circuit patterns on the wafer surface can lead to high-intensity regular background noise in dark field scattering images, while orientation consistency values ​​can provide a feature correlation metric that is not affected by absolute grayscale light intensity.

[0051] When the structural edges corresponding to two local gradient vectors originate from the same normal pattern, the normal directions pointed to by the first local gradient vector and the second local gradient vector are either highly parallel or collinear in opposite directions.

[0052] The process of determining the directional consistency value involves using the dot product operation between the first local gradient vector and the second local gradient vector to evaluate the degree of overlap of the projections of the two vectors in the spatial direction, and then dividing the dot product result by the scalar multiplication result of the corresponding feature magnitude operation of the two vectors to complete the process of solving the cosine value of the angle between the vectors.

[0053] The directional consistency value can characterize the consistency of the structure at the pixel coordinate position under multimodal observation conditions. The larger the directional consistency value, the more likely the pixel belongs to the high-frequency edge of the fixed pattern of the wafer. The smaller the directional consistency value, the closer it is to 0, the more likely the pixel may have local defects with random shape and lack of spatial coordination in the bright and dark fields.

[0054] In one embodiment, the dark field purification value is determined as follows: a first adjustment coefficient that is linearly negatively correlated with the directional consistency value is determined using the difference between the upper limit of the directional consistency value and the directional consistency value; a second adjustment coefficient is determined using the difference information between the global bright field extremum in the bright field reflection image and the bright field gray value of the pixel; the product of the first adjustment coefficient and the second adjustment coefficient is normalized to obtain the target adjustment coefficient of the target pixel in the dark field scattering image, and the product of the dark field gray value of the target pixel and the target adjustment coefficient is used as the dark field purification value of the target pixel; the target pixel is any pixel in the dark field scattering image.

[0055] Since the directional consistency value is normalized, and its upper limit is 1, an exemplary implementation of the first adjustment coefficient could be... , Position coordinates The orientation consistency value of the pixels; when the orientation consistency value is large, the current pixel position is more likely to be a normal wafer line edge, which can attenuate the dark field gray value at the corresponding position, so that the regular strong scattering background signal can be suppressed.

[0056] An exemplary implementation of the second adjustment coefficient could be... ; This represents the global bright-field extremum in the bright-field reflectance image. Position coordinates The bright field grayscale value of the pixel in the bright field reflective image; the global bright field extremum is equal to the maximum grayscale value of all pixels in the bright field reflective image; since actual foreign particles often exhibit low bright field grayscale values ​​but high brightness characteristics in dark field scattering, a second adjustment coefficient based on the difference between the bright field grayscale value and the global bright field extremum is introduced, which plays an adaptive modulation role based on the bright field brightness background. When the local bright field grayscale value is larger, that is, the background reflectivity is stronger, it can suppress false anomalies that are not actually defects.

[0057] By dynamically scaling the directional consistency and relative brightness of the bright field, the output dark field purification value retains and amplifies the actual defect signal that only has random scattering characteristics and is relatively dim in the bright field, thereby improving the screening ability for small particles.

[0058] In one embodiment, the bright-field reconstructed value is determined as follows: the second derivative of the bright-field grayscale value of the bright-field reflectance image is extracted using the Laplacian operator to determine the local abrupt change value; the bright-field grayscale value within the local calculation window is averaged to determine the local average grayscale; the bright-field grayscale value is spatially integrated using a weighted averaging mechanism to determine the thickness background value; the absolute difference between the bright-field grayscale value and the thickness background value is determined, the ratio of the local abrupt change value to the local average grayscale is used as the abrupt change coefficient, and the product of the abrupt change coefficient and the absolute difference is used as the bright-field reconstructed value.

[0059] After the chemical mechanical polishing process of a wafer, microscopic thin film thickness variations may remain on the wafer surface. These thickness fluctuations can cause slowly changing interference bands and grayscale shifts in bright-field reflection images, potentially masking microscopic abrupt defects caused by scratches or pores.

[0060] As an isotropic second-order differential operator, the Laplacian operator is more sensitive to isolated points and thin lines in images. By using the Laplacian operator to extract local abrupt changes through second-order derivative extraction, it can keenly capture the curvature maxima on the grayscale profile, thereby highlighting the high-frequency signals that actually represent structural damage.

[0061] To correct for uneven local illumination, averaging can be performed on local calculation windows, such as a 5x5 pixel region, to determine the local average gray level and establish a local energy reference. The background estimation uses a weighted averaging mechanism combined with large-scale spatial integration to extract the thickness background value containing only slow low-frequency components. By calculating the absolute difference between the original bright field gray value and the thickness background value, the intermediate residual signal that eliminates the influence of macroscopic film thickness fluctuations is obtained.

[0062] The constructed mutation coefficients achieve feature standardization based on local contrast by dividing the local mutation value by the local average gray level. By multiplying the mutation coefficients with the absolute difference, the amplification capability of the second derivative for high-frequency details is combined with the resistance of the first residual to background drift. This allows the bright-field reconstruction value to retain and enhance the actual bright-field local structural mutation information while reducing low-frequency thin-film interference noise.

[0063] In one embodiment, the background thickness value is determined as follows: the ratio of the square of the gradient modulus corresponding to the bright-field reflectance image to the square of the preset smoothing coefficient is used as the exponential base, and the Gaussian weight is determined using the exponential base; a first integral term is determined by performing a double integral operation on the product of the Gaussian weight and the bright-field gray value within the spatial integration interval; a second integral term is determined by performing a double integral operation on the Gaussian weight within the spatial integration interval; and the ratio of the first integral term to the second integral term is used as the background thickness value.

[0064] The ratio of the square of the gradient magnitude corresponding to the bright field reflectance image to the square of the preset smoothing coefficient is used as the exponential base and substituted into the natural exponential function to determine the Gaussian weight. The purpose is to dynamically adjust the fusion weight between pixels based on the intensity of the local gradient.

[0065] When the local gradient magnitude is small, i.e. in a region of gradual film thickness variation, the exponential base approaches a small value, resulting in a higher Gaussian weight. This allows for spatial smoothing of depth within the region. However, when the local gradient magnitude is much larger than the preset smoothing coefficient, it indicates a strong structural edge position. In this case, the Gaussian weight can rapidly decay to a minimum value, thus avoiding pixel fusion between different edges.

[0066] Within a given spatial integration interval, such as a circular region with a radius of 15 pixels, a double integration operation is performed on the product of Gaussian weights and bright-field gray values ​​to determine the first integral term representing the weighted energy sum. A double integration operation is then performed on the Gaussian weights to determine the second integral term used for normalization. The ratio of the first integral term to the second integral term eliminates the influence of high-frequency structural edges and reflects the changes in low-frequency thin-film optical interference. This provides a parameter basis for accurately separating background interference and minor defect residuals from the bright-field signal.

[0067] In one embodiment, the defect assessment index is determined as follows: a first sum of the squared value of the dark field cleansing value and the weighted squared value of the bright field reconstruction value is determined; the square root of the first sum is used as the feature vector magnitude; a first product term is obtained by multiplying the dark field cleansing value and the bright field reconstruction value; a second product term is obtained by multiplying the global dark field extremum and the global bright field extremum; the ratio of the first product term to the second product term is used as the feature ratio; the sum of the feature ratio and a preset positive number is used as the enhancement factor value; and the product of the feature vector magnitude and the enhancement factor value is used as the defect assessment index.

[0068] The first and second product terms are constructed to obtain the feature ratio. The logic of cross-multiplication is to quantify the confidence level of the synchronous abnormal changes in the dark field and the bright field. Since real defects such as deep cavities usually show obvious abrupt changes in both bright field reflection and dark field scattering, the feature ratio can output a larger nonlinear gain for such high-risk defects.

[0069] By adding the feature ratio to a preset positive number, such as 1, to construct a non-zero guaranteed enhancement factor value, and then multiplying the enhancement factor value with the feature vector magnitude, an amplified response can be given to small or hidden defects.

[0070] Figure 2 This is a schematic diagram comparing the defect evaluation index with the histogram of the dark field scattering image, such as... Figure 2 As shown, in the range of low signal strength, such as level 0 to 100, the original dark field scattering signal contains a large number of dense pixels. These are usually coherent scattering background noise caused by the normal pattern structure on the wafer surface. The distribution of the corresponding defect evaluation index is significantly concentrated in the lower intensity range, and a smoother distribution is formed in the middle section where the original noise is stronger, effectively stripping away the normal structure signal.

[0071] In the high-intensity range above level 150, the original dark field signal exhibits a discrete pulse-like distribution and significant overlap near level 200. In this embodiment, the orthogonal fusion of dark field purification value and bright field reconstruction value makes the final generated defect evaluation index more discernible while maintaining high sensitivity, which helps to accurately remove the real defect signal.

[0072] In one embodiment, the defect detection result is determined as follows: if the defect evaluation index of a pixel is greater than a preset threshold, the pixel position is marked as a real defect point; if the defect evaluation index of a pixel is less than or equal to the preset threshold, the pixel position is marked as a noise point, and the defect definition process for the pixel position ends; using at least one of the obtained real defect points and noise points, the defect detection result of the wafer to be inspected is output.

[0073] After establishing the defect evaluation index distribution matrix, the defect evaluation index distribution matrix can be discretized into a wafer defect distribution map using a hard decision threshold. The preset threshold can be determined in advance by the yield control standards of specific process nodes on the production line and historical scanning data. When processing the wafer map, the defect evaluation index of all pixels can be traversed one by one.

[0074] If the defect evaluation index is greater than the preset threshold, the current coordinates can be recorded in the internal coordinate data structure, and the pixel position can be marked as a real defect point with specific geometric coordinates and confidence level. If the value is less than or equal to the preset threshold, it means that the pixel position belongs to the normal wafer background or is a safety fluctuation caused by the processing tolerance. The pixel position will be marked as a suppressed noise point, and the logical judgment action for the pixel position will end to save computing resources.

[0075] By statistically analyzing the coordinate distribution across the entire wafer, at least one format can be used, including the obtained location distribution cloud map of actual defect points and the spatial density statistics of noise points, to output a defect detection result report for the wafer to be inspected. The defect detection result report can be directly sent to subsequent yield management software to trigger remedial actions such as wafer rework guidance or lithography machine process parameter calibration.

Claims

1. A multi-dimensional intelligent defect detection system for wafer manufacturing processes, characterized in that, include: The acquisition module is configured to acquire the bright-field reflection image and the dark-field scattering image of the wafer to be inspected, and use an edge detection algorithm to determine the first local gradient vector of the pixel in the bright-field reflection image and the second local gradient vector of the pixel in the dark-field scattering image. The first determining module is configured to use a first local gradient vector and a second local gradient vector to determine the directional consistency value of the bright field reflective image and the dark field scattering image at a pixel point, and to use the directional consistency value, the dark field gray value of the same pixel point and the bright field gray value to determine the dark field purification value of the pixel point in the dark field scattering image. The second determining module is configured to use a second-order spatial operator to extract features from the bright-field reflectance image to determine local abrupt change values, and to perform background estimation on the bright-field reflectance image to determine the thickness background value. The bright-field reconstructed value of the pixel is then determined using the local abrupt change value and the thickness background value. The third determination module is configured to orthogonally add the dark field purification value and the bright field reconstruction value to determine the feature vector magnitude, and use the dark field purification value, the bright field reconstruction value and the feature vector magnitude to determine the defect evaluation index, and use the defect evaluation index to determine the defect detection result.

2. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The directional consistency value is determined in the following way: Determine the vector dot product between the first local gradient vector and the second local gradient vector; The magnitude of the first local gradient vector is taken as the first magnitude, and the magnitude of the second local gradient vector is taken as the second magnitude. The product of the magnitudes of the first and second magnitudes is determined, and the ratio between the absolute value of the vector dot product and the product of the magnitudes is taken as the direction consistency value.

3. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The dark field cleanliness value is determined in the following way: The first adjustment coefficient, which is linearly negatively correlated with the direction consistency value, is determined by using the difference between the upper limit of the direction consistency value and the direction consistency value; the second adjustment coefficient is determined by using the difference between the global bright field extremum in the bright field reflectance image and the bright field gray value of the pixel. The product of the first adjustment coefficient and the second adjustment coefficient is normalized to obtain the target adjustment coefficient of the target pixel in the dark field scattering image. The product of the dark field gray value of the target pixel and the target adjustment coefficient is used as the dark field purification value of the target pixel. The target pixel is any pixel in the dark field scattering image.

4. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The bright-field reconstruction value is determined in the following way: The Laplacian operator is used to extract the second derivative of the bright field grayscale values ​​of the bright field reflectance image to determine the local abrupt change value; the bright field grayscale values ​​within the local calculation window are averaged to determine the local average grayscale. The thickness background value is determined by spatial integration of the bright field gray value using a weighted average mechanism; the absolute difference between the bright field gray value and the thickness background value is determined, the ratio of the local mutation value to the local average gray value is used as the mutation coefficient, and the product of the mutation coefficient and the absolute difference is used as the bright field reconstruction value.

5. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 4, characterized in that, The background thickness value is determined in the following way: The Gaussian weights are determined by using the ratio of the square of the gradient magnitude corresponding to the bright-field reflectance image to the square of the preset smoothing coefficient as the exponential base. The first integral term is determined by performing a double integration operation on the product of the Gaussian weight and the bright field gray value within the spatial integration interval. The second integral term is determined by performing a double integral operation on the Gaussian weight within the spatial integral interval; the ratio of the first integral term to the second integral term is used as the thickness background value.

6. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The defect assessment index is determined in the following way: Determine the first sum of the weighted squares of the dark field purification value and the bright field reconstruction value, and use the square root of the first sum as the modulus of the feature vector. Multiply the dark field purification value by the bright field reconstruction value to obtain the first product term, multiply the global dark field extremum by the global bright field extremum to obtain the second product term, use the ratio of the first product term to the second product term as the feature ratio, and use the sum of the feature ratio and a preset positive number as the enhancement factor value. The product of the feature vector magnitude and the enhancement factor value is used as the defect evaluation index.

7. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, Bright-field reflectance images and dark-field scattering images were acquired using the following methods: Acquire the original bright-field image and the original dark-field image of the wafer to be inspected; The phase correlation algorithm is used to perform sub-pixel registration processing on the original dark field image in combination with the original bright field image to determine the coordinate-aligned image. The original bright-field image is used as the bright-field reflection image, and the coordinate-aligned image is used as the dark-field scattering image.

8. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 7, characterized in that, The original dark-field image is subpixel registered using a phase correlation algorithm combined with the original bright-field image to determine the coordinate-aligned image, including: The first frequency domain data is obtained by performing a discrete Fourier transform on the original bright-field image, and the second frequency domain data is obtained by performing a discrete Fourier transform on the original dark-field image; the cross-power spectrum matrix of the first frequency domain data and the second frequency domain data in the complex frequency domain is determined. The cross-power spectrum matrix is ​​cross-normalized to determine the normalized matrix. The normalized matrix is ​​then subjected to inverse Fourier transform to obtain the spatial pulse signal. The position coordinates corresponding to the pulse extrema in the spatial pulse signal are searched to determine the offset. The offset is then used to perform bilinear interpolation resampling to complete the original dark field image and obtain a coordinate-aligned image.

9. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The first local gradient vector is determined in the following way: A horizontal and vertical differential operators in two-dimensional space are constructed using an edge detection algorithm. The horizontal differential operator is used to perform convolution operation on the bright field gray value corresponding to the bright field reflectance image to determine the horizontal gradient component. The vertical differential operator is used to perform convolution operation on the bright field gray value corresponding to the bright field reflectance image to determine the vertical gradient component. The horizontal and vertical gradient components are then merged to determine the first local gradient vector.

10. The multi-dimensional intelligent defect detection system for wafer manufacturing process according to claim 1, characterized in that, The defect detection results are determined in the following ways: If the defect evaluation index of a pixel is greater than a preset threshold, the pixel location is marked as a real defect point; if the defect evaluation index of a pixel is less than or equal to the preset threshold, the pixel location is marked as a noise point, and the defect definition process for the pixel location ends. Using at least one of the obtained real defect points and noise points, the defect detection results of the wafer to be inspected are output.

Images