A surface detection method for preparing magnesium oxide single crystal target

Abstract

The present invention relates to the field of target surface detection, and specifically to a surface detection method for magnesium oxide single crystal target preparation. The method comprises the following steps: obtaining a surface morphology image of each detection sub-area and synchronously collecting component spectrum data of the corresponding detection sub-area; removing continuous background noise from the component spectrum data and extracting characteristic X-ray peak intensities and recording them as component parameters; extracting morphology parameters from the surface morphology image, spatially registering the morphology parameters with the component parameters to obtain a morphology-component fusion feature vector; comparing and analyzing the morphology-component fusion feature vector with a preset contamination particle feature library and a micro-protrusion feature library to obtain similarity probabilities, and obtaining defect type discrimination results for the detection sub-areas based on the similarity probabilities; mapping the discrimination results to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a detection report; and effectively improving the accuracy and comprehensiveness of surface defect detection for magnesium oxide single crystal target materials.

Description

Technical Field

[0001] The present invention relates to the field of target material preparation surface detection, in particular to a method for detecting the surface of a magnesium oxide single crystal target material preparation. Background Art

[0002] During the preparation of magnesium oxide single crystal targets, the target's surface quality plays a crucial role in determining subsequent application performance. Magnesium oxide single crystal targets are widely used in high-precision applications such as semiconductor manufacturing and optical coatings, and their surface defects can directly impact product performance and reliability. Therefore, precise and comprehensive surface testing of magnesium oxide single crystal targets, enabling the timely detection and accurate identification of surface defects, is crucial for ensuring target quality.

[0003] Currently, the surface inspection of magnesium oxide single crystal targets primarily utilizes a scanning electron microscope (SEM) combined with energy-dispersive X-ray spectroscopy (EDS). The SEM is used to obtain topographic images of the target surface, while the EDS is used to collect surface composition data. However, in actual inspection, weak signal defects such as tiny contaminant particles (such as metal particles <10 nm) are seriously missed. Because the signals produced by tiny contaminants in the EDS spectrum are extremely weak and easily masked by noise, their characteristic information is difficult to accurately extract. During defect identification, distinguishing between tiny contaminants and micro-bumps relies heavily on the inspector's experience, making it difficult to ensure consistent test results. This lack of consistent inspection not only increases uncertainty in target quality assessment but can also result in defective targets being introduced into subsequent production processes, impacting the quality and performance of the final product and resulting in financial losses and market risks for the company.

[0004] Therefore, there is an urgent need for a surface inspection method for the preparation of magnesium oxide single crystal targets, which effectively improves the detection capability of tiny contamination particles, reduces the missed detection rate of weak signal defects, improves the consistency and accuracy of detection, and provides reliable guarantee for the quality control of magnesium oxide single crystal targets. Summary of the Invention

[0005] (1) Technical problems to be solved

[0006] The purpose of the present invention is to provide a surface detection method for magnesium oxide single crystal target material preparation, which solves the problem that tiny contamination particles and micro-protrusion defects on the surface of magnesium oxide single crystal target material are difficult to automatically distinguish under weak signal and high noise interference.

[0007] (2) Technical solution

[0008] To achieve the above object, the present invention provides a surface detection method for preparing a magnesium oxide single crystal target, the method comprising:

[0009] S1. Divide the target surface into multiple detection sub-areas and determine the priority scanning order of each detection sub-area; obtain the surface morphology image of each detection sub-area through a scanning electron microscope, and simultaneously collect the composition spectrum data of the corresponding detection sub-area through an energy dispersive X-ray spectrometer.

[0010] S2. After multi-frame superposition processing of the component spectrum data, continuous background noise is removed by nonlinear least squares fitting and characteristic X-ray peak intensity is extracted and recorded as component parameters; morphology parameters are extracted from the surface morphology image, and the morphology parameters are spatially aligned with the component parameters to obtain a morphology-component fusion feature vector; the morphology parameters include micro-area height gradient, local surface curvature radius and edge sharpness parameters.

[0011] S3. Compare and analyze the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtain a defect type discrimination result of the detection sub-area based on the similarity probability.

[0012] S4. Mapping the discrimination results to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report.

[0013] Furthermore, the method of removing continuous background noise and extracting characteristic X-ray peak intensity by nonlinear least squares fitting after performing multi-frame superposition processing on the component spectrum data includes:

[0014] The component spectrum data is subjected to multi-frame superposition processing by the weighted averaging method. The number of superposition frames of the multi-frame superposition processing is dynamically adjusted according to the priority of the current detection sub-area. When the priority of the detection sub-area is higher than the preset priority threshold, the number of superposition frames is gradually increased by an adaptive algorithm until the signal-to-noise ratio improvement rate of the component spectrum data after superposition is greater than the preset improvement rate threshold.

[0015] A composite background model including polynomial basis functions and exponential decay basis functions is constructed, and the component spectrum data after multi-frame superposition processing is globally fitted by nonlinear least squares method to obtain weight coefficients of the polynomial basis functions and the exponential decay basis functions. The weight coefficients are substituted into the composite background model to obtain continuous background noise covering the full energy range; the continuous background noise is removed from the component spectrum data after multi-frame superposition processing to obtain denoised component spectrum data.

[0016] According to the preset target element characteristic energy range, the sliding window method is used to screen candidate characteristic peaks that meet the local maximum condition and have a half-height width within a preset width interval in the denoised component spectrum data, and the peak area integral value is calculated as the characteristic X-ray peak intensity according to the energy channel interval of each candidate characteristic peak, and the left energy boundary and the right energy boundary of each candidate characteristic peak are dynamically calibrated. The dynamic calibration is to determine the peak boundary starting point according to the change trend of the first-order derivative of the counting rate of the energy channel corresponding to the candidate characteristic peak. If the counting rate correlation coefficient of adjacent energy channels is lower than the preset correlation threshold, it is determined to be a pseudo characteristic peak formed by an isolated noise point and is eliminated.

[0017] Furthermore, the method of extracting topography parameters from the surface topography image and spatially registering the topography parameters with the component parameters to obtain a topography-component fusion feature vector includes:

[0018] The surface topography image of each detection sub-area is reconstructed into three-dimensional shape to obtain a height distribution matrix. Based on the height distribution matrix, the height gradient modulus of each pixel point in a preset neighborhood window is calculated as the micro-area height gradient, and the local surface curvature radius is calculated by quadratic surface fitting. The pixel grayscale change rate is extracted along the edge contour of the surface topography image, and the ratio of the maximum grayscale change rate to the grayscale variance of the adjacent area is used as the edge sharpness parameter.

[0019] The micro-area height gradient, local surface curvature radius and edge sharpness parameters are normalized according to the pixel coordinates of the detection sub-area to obtain a morphological feature vector; according to the scanning paths of the scanning electron microscope and the energy dispersive X-ray spectrometer, a mapping relationship between the image pixel coordinates of the detection sub-area and the energy channel position of the component spectrum data is synchronously established; according to the mapping relationship between the image pixel coordinates of the detection sub-area and the energy channel position of the component spectrum data, the characteristic X-ray peak intensity and the morphological feature vector of the corresponding detection sub-area are spliced ​​according to the dimension to obtain a morphology-component fusion feature vector.

[0020] Furthermore, the method for synchronously establishing a mapping relationship between the image pixel coordinates of the detection sub-area and the energy channel position of the component spectrum data according to the scanning paths of the scanning electron microscope and the energy dispersive X-ray spectrometer includes:

[0021] Based on the correspondence between the scanning step size of the scanning electron microscope and the sampling time interval of the energy dispersive X-ray spectrometer, the synchronization mark of the surface morphology image pixel coordinates and the component spectrum data acquisition time of each detection sub-area is determined; based on the said synchronization mark, the real-time coordinate data of the stage position encoder of the scanning electron microscope is extracted, and combined with the calibration parameters of the energy channel and spatial position of the energy dispersive X-ray spectrometer, a geometric transformation matrix from the image pixel coordinates to the energy channel position is established.

[0022] The geometric transformation matrix is ​​decomposed into translation vector and rotation matrix components, and the parameters of the translation vector and rotation matrix components are optimized by the least squares method to minimize the spatial deviation between the central pixel coordinates of the surface morphology image and the central position of the energy channel of the component spectrum data of the same detection sub-area; the optimized geometric transformation matrix is ​​applied to all detection sub-areas to realize point-by-point mapping between the image pixel coordinates and the energy channel position.

[0023] Furthermore, the method of comparing and analyzing the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtaining a defect type discrimination result of the detection sub-area according to the similarity probability includes:

[0024] After the morphology-component fusion feature vector is standardized, a first kernel space similarity is calculated between it and each sample feature vector in the pollution particle feature library according to kernel function mapping, and a second kernel space similarity is calculated between it and each sample feature vector in the micro-bump feature library.

[0025] The probability distribution of the first kernel spatial similarity in the contamination particle category and the probability distribution of the second kernel spatial similarity in the micro-bump category are estimated respectively through the probability density function, and the first posterior probability that the current morphology-component fusion feature vector belongs to the contamination particle and the second posterior probability that it belongs to the micro-bump are obtained; the probability ratio of the first posterior probability to the second posterior probability is obtained, and when the probability ratio exceeds the preset probability ratio threshold, the defect type is determined to be a contamination particle; otherwise, the defect type is determined to be a micro-bump; the similarity probability includes the first posterior probability and the second posterior probability.

[0026] Furthermore, the method of estimating the probability distribution of the first kernel spatial similarity in the pollution particle category and the probability distribution of the second kernel spatial similarity in the micro-bump category using a probability density function to obtain a first posterior probability that the current shape-component fusion feature vector belongs to the pollution particle and a second posterior probability that the current shape-component fusion feature vector belongs to the micro-bump includes:

[0027] The kernel function bandwidth parameter is calculated according to the eigenvalues ​​of the covariance matrix of all sample feature vectors in the preset pollution particle feature library and micro-protrusion feature library, and the bandwidth parameter is dynamically adjusted according to the distribution sparsity of the sample feature vectors in the pollution particle feature library and micro-protrusion feature library.

[0028] The Gaussian kernel function is used to map the first kernel space similarity and the second kernel space similarity to the high-dimensional space respectively, and the non-parametric kernel density estimation is performed on the kernel space similarity of all samples in the pollution particle feature library and the micro-protrusion feature library to generate the kernel density distribution function of the pollution particle category and the kernel density distribution function of the micro-protrusion category respectively.

[0029] The first kernel space similarity of the current morphology-component fusion feature vector is substituted into the kernel density distribution function of the pollution particle category to calculate its probability density value as the first posterior probability, and the second kernel space similarity is substituted into the kernel density distribution function of the micro-bump category to calculate its probability density value as the second posterior probability.

[0030] Furthermore, the method of dynamically adjusting the bandwidth parameter according to the distribution sparsity of the sample feature vectors in the pollution particle feature library and the micro-protrusion feature library includes:

[0031] According to the spatial distribution of all sample feature vectors in the preset pollution particle feature library and micro-protrusion feature library, the average Mahalanobis distance between samples in each feature library is calculated as a distribution sparsity measurement index; and the bandwidth parameter adjustment coefficient of the pollution particle feature library and the micro-protrusion feature library is generated according to the sparsity ratio of the distribution sparsity measurement index to the preset baseline sparsity.

[0032] When the distribution sparsity measurement index is greater than the preset baseline sparsity, the bandwidth parameter adjustment coefficient is the square root of the sparsity ratio, otherwise it is the square value of the sparsity ratio; the bandwidth parameter adjustment coefficient is multiplied by the initial bandwidth parameter to obtain the dynamically adjusted kernel function bandwidth parameter, and is applied to the construction of kernel density distribution functions for the pollution particle category and micro-protrusion category respectively.

[0033] Furthermore, the method of mapping the discrimination result to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report includes:

[0034] A mapping matrix is ​​established between the center point of the detection sub-area and the three-dimensional coordinate system of the target surface according to the division order of the detection sub-area and the spatial position relationship of the scanning path; the positions of the contaminants and micro-protrusions in the defect type discrimination results are converted into three-dimensional spatial coordinates of the target surface according to the mapping matrix, and the corresponding similarity probability is converted into a confidence score and attached to the three-dimensional coordinates of each defect.

[0035] A spatial clustering analysis is performed on the three-dimensional coordinates of all defects to statistically calculate the defect distribution density, spatial aggregation characteristics and weighted average confidence scores of the contamination particles and micro-protrusions. At the same time, the component proportion weights of different elements in the contamination particle area and the micro-protrusion area are calculated based on the characteristic X-ray peak intensity; the defect distribution density, spatial aggregation characteristics, component proportion weights and the corresponding weighted average confidence scores are integrated into a structured data table, and the structured data table is combined with the three-dimensional coordinate mapping result to generate a test report containing a defect spatial position distribution map, an element composition weight distribution heat map and a confidence layered annotation map; the weighted average confidence score is used to correct the statistical results of the defect distribution density through a weighted average algorithm.

[0036] Furthermore, the method of correcting the statistical result of defect distribution density by using a weighted average algorithm includes:

[0037] The weight factors of contamination particles and micro-protrusions in the detection sub-area are calculated based on the confidence scores attached to the three-dimensional coordinates of each defect. The weight factor is the ratio of the confidence score of the current defect coordinate to the sum of the confidence scores of all similar defects. The defect distribution density is weighted and summed according to the weight factor to obtain the confidence-weighted mean of the defect distribution density.

[0038] The defect distribution density mean and the preset confidence interval are compared and analyzed with the unweighted defect distribution density to generate a corrected defect distribution density statistical result and embed it into a structured data table of the inspection report.

[0039] Furthermore, the method of converting the contamination particles and micro-protrusion positions in the defect type discrimination result into three-dimensional spatial coordinates of the target surface according to the mapping matrix includes:

[0040] According to the division index number of the detection sub-area, the detection sub-area number marked as contamination particles or micro-protrusions in the judgment result is extracted. According to the correspondence between the detection sub-area number and the three-dimensional coordinates stored in the mapping matrix, the detection sub-area number corresponding to each defect is mapped to the three-dimensional spatial coordinates of the target surface.

[0041] (3) Beneficial effects

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. By combining scanning electron microscopy to obtain surface morphology images and energy dispersive X-ray spectrometers to collect component spectrum data, information is extracted from two dimensions: morphology (micro-area height gradient, curvature radius, edge sharpness) and composition (characteristic X-ray peak intensity). A morphology-composition fusion feature vector is formed. By comparing and analyzing it with the preset contamination particle feature library and micro-bump feature library, the defect type discrimination result of the inspection sub-area can be accurately obtained, effectively improving the accuracy and comprehensiveness of surface defect detection of magnesium oxide single crystal targets, and avoiding missed detection or misjudgment that may be caused by a single detection method.

[0044] 2. In terms of component spectrum data processing, multi-frame overlay processing is used, with the number of overlay frames dynamically adjusted to improve the signal-to-noise ratio based on the priority of the detection sub-region. A composite background model is constructed and nonlinear least squares fitting is used to deduct continuous background noise. At the same time, candidate characteristic peaks are dynamically calibrated and pseudo-characteristic peaks are eliminated. This significantly improves the quality of component spectrum data, removes noise interference, and more accurately extracts characteristic X-ray peak intensities, providing a reliable data foundation for subsequent defect identification and enhancing the reliability of detection results.

[0045] 3. By establishing a mapping relationship between the pixel coordinates of the inspection sub-area image and the energy channel position of the component spectrum data, as well as a mapping matrix between the center point of the inspection sub-area and the three-dimensional coordinate system of the target surface, the defect type identification results are accurately mapped to the three-dimensional coordinates of the target surface, achieving precise marking of the defect location. At the same time, a test report is generated that includes a defect spatial location distribution map, a heat map of the elemental composition weight distribution, and a confidence level layered annotation map. This can intuitively display the test results, facilitate a comprehensive understanding of the distribution, composition, and confidence level of target surface defects, and provide strong support for subsequent process improvement and quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flowchart of a surface detection method for preparing a magnesium oxide single crystal target according to Example 1 of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] Before giving examples, it is necessary to explain the application scenarios of the present invention. The present invention is a surface detection method for the preparation of magnesium oxide single crystal targets. It is applied to the problems that tiny contamination particles (such as metal particles <10nm) and micro-protrusion defects on the surface of magnesium oxide single crystal targets are difficult to distinguish automatically under weak signals and high noise interference, and traditional detection relies on manual experience, resulting in poor consistency of results. It effectively improves the accuracy and comprehensiveness of surface defect detection of magnesium oxide single crystal targets, and provides strong support for the quality control of magnesium oxide single crystal targets.

[0049] Example 1: Figure 1 As shown, this embodiment provides a surface detection method for preparing a magnesium oxide single crystal target, the method comprising:

[0050] S1. Divide the target surface into multiple detection sub-areas and determine the priority scanning order of each detection sub-area; obtain a surface morphology image of each detection sub-area through a scanning electron microscope, and simultaneously collect component spectrum data of the corresponding detection sub-area through an energy dispersive X-ray spectrometer; in the embodiment of this specification, the target surface is divided into multiple detection sub-areas by a uniform division method. The number of detection sub-areas depends on the specific business scenario and is not specifically limited here. In a specific embodiment, the surface of a magnesium oxide single crystal target material with a size of 100mm×100mm is evenly divided into 10×10 detection sub-areas, each sub-area is 10mm×10mm. Each sub-area included in the multiple detection sub-areas is set with a priority scanning order. Specifically, the priority scanning order is determined based on the preliminary analysis results of the target surface roughness. Specifically, the preliminary analysis results of the target surface roughness can be determined based on production experience. The edge area is prone to defects due to concentrated processing stress, so the detection sub-areas in the edge area are given a higher priority; the center area is relatively stable and is given a lower priority. The target surface was inspected using a Zeiss GeminiSEM 500 scanning electron microscope (SEM) equipped with an Oxford X-MaxN 80 energy-dispersive X-ray spectrometer (EDX-R). Within each subregion, the SEM was set at 10 kV, a working distance of 10 mm, and a magnification of 8000× to acquire surface topography images. Simultaneously, the EDXR acquired component spectrum data with a 120-second acquisition time, scanning all subregions sequentially in an S-shaped pattern.

[0051] S2. After multi-frame superposition processing of the component spectrum data, continuous background noise is removed by nonlinear least squares fitting and characteristic X-ray peak intensity is extracted and recorded as component parameters; morphology parameters are extracted from the surface morphology image, and the morphology parameters are spatially aligned with the component parameters to obtain a morphology-component fusion feature vector; the morphology parameters include micro-area height gradient, local surface curvature radius and edge sharpness parameters.

[0052] S3. Compare and analyze the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtain a defect type discrimination result of the detection sub-area based on the similarity probability.

[0053] S4. Mapping the discrimination results to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report.

[0054] The method of removing continuous background noise and extracting characteristic X-ray peak intensity by nonlinear least squares fitting after performing multi-frame superposition processing on the component spectrum data includes:

[0055] The component spectrum data is stacked over multiple frames using a weighted averaging method. The number of stacked frames is dynamically adjusted based on the priority of the current detection subregion. When the priority of a detection subregion is higher than a preset priority threshold, the number of stacked frames is gradually increased using an adaptive algorithm until the signal-to-noise ratio improvement rate of the component spectrum data after stacking exceeds the preset improvement rate threshold. For each detection subregion, eight frames of original component spectrum data are first collected for stacking. Taking the high-priority edge detection subregion as an example, the priority threshold (set to 7 out of 10) determines that this region has a priority of 8, which is higher than the threshold. Therefore, the number of stacked frames is gradually increased to 12 frames. At this point, the signal-to-noise ratio improvement rate reaches 35%, exceeding the preset 30% threshold, completing the multi-frame stacking.

[0056] A composite background model including polynomial basis functions and exponential decay basis functions is constructed, and weight coefficients of the polynomial basis functions and the exponential decay basis functions are obtained by performing global fitting on the component spectrum data after multi-frame superposition processing through nonlinear least squares method, and the weight coefficients are substituted into the composite background model to obtain continuous background noise covering the full energy range; the continuous background noise is removed from the component spectrum data after multi-frame superposition processing to obtain denoised component spectrum data; the constructed composite background model includes cubic polynomial basis functions and two exponentially decaying basis functions , where E represents the energy value. The parameters are solved by nonlinear least squares method. The optimal value of The fitting error with the superimposed component spectrum data is the smallest. is the constant term of the polynomial basis function, which characterizes the baseline shift of the background noise near zero energy; is the linear coefficient of energy, which fits the linear variation trend of background noise with energy; is the quadratic coefficient of energy, fitting the parabolic change of background noise; is the cubic coefficient of energy, which fits the high-order nonlinear changes of background noise. , is the amplitude coefficient of the exponential decay basis function, which is used to control the initial intensity of the background noise; , is the decay rate coefficient of the exponential decay basis function, which determines how quickly the background noise decreases with increasing energy. The composite background model can effectively fit the background noise characteristics of different energy ranges and is particularly suitable for processing the complex background noise of magnesium oxide single crystal targets.

[0057] Based on the preset target element characteristic energy range, a sliding window method is used to screen candidate characteristic peaks in the denoised component spectrum data that meet the local maximum condition and have a half-height width within a preset width range. The peak area integral value is calculated based on the energy channel range of each candidate characteristic peak as the characteristic X-ray peak intensity, and the left and right energy boundaries of each candidate characteristic peak are dynamically calibrated. The dynamic calibration determines the peak boundary starting point based on the trend of the first-order derivative of the count rate of the energy channel corresponding to the candidate characteristic peak. If the count rate correlation coefficient of adjacent energy channels is lower than a preset correlation threshold, it is determined to be a pseudo-characteristic peak formed by an isolated noise point and is eliminated. The count rate is the number of X-ray photons detected by the energy dispersive X-ray spectrometer per unit time in each energy channel. When screening candidate characteristic peaks in the denoised component spectrum data, the sliding window width is set to 50eV, and the screening conditions include: local maximum determination (the count rate of the current point is higher than the count rates of the three energy channel points on the left and right) and the half-height width is within the preset range of 40-200eV. For common metal contamination elements such as Fe, Cu, and Al on the surface of magnesium oxide single crystal targets, energy channel intervals of ±50eV are set near their characteristic energies (Fe Kα 6.4keV, Cu Kα 8.0keV, Al Kα ~1.5keV), and the peak area integral value is calculated as the characteristic X-ray peak intensity. During the dynamic calibration process, the peak boundary starting point is determined by calculating the trend of the first-order derivative change of the count rate of each candidate characteristic peak energy channel. For example, for the Fe Kα characteristic peak, the correlation coefficient of the count rate of adjacent energy channels in the range of 6.35-6.45keV is calculated. If it is lower than the preset correlation threshold of 0.65, it is determined to be a pseudo-characteristic peak and is eliminated, effectively avoiding the interference of false peaks caused by noise fluctuations.

[0058] The method of extracting topography parameters from the surface topography image and spatially registering the topography parameters with the component parameters to obtain a topography-component fusion feature vector includes:

[0059] The surface topography image of each detection sub-area is reconstructed in three dimensions to obtain a height distribution matrix; based on the height distribution matrix, the height gradient modulus of each pixel point is calculated as the micro-area height gradient in a preset neighborhood window, and the local surface curvature radius is calculated by quadratic surface fitting; the pixel grayscale change rate is extracted along the edge contour of the surface topography image, and the ratio of the maximum grayscale change rate to the grayscale variance of the adjacent area is used as the edge sharpness parameter; the SEM image is reconstructed in three dimensions, and a multi-light source stereo imaging algorithm combined with shadow information analysis is used to construct a height distribution matrix with a resolution of 1024×1024 pixels. In a 9×9 pixel neighborhood window, the height gradient vector of the central pixel point relative to the surrounding pixels is calculated, and the modulus of the height gradient vector is taken as the micro-area height gradient value. The quadratic surface z=ax²+by²+cxy+dx+ey+f is fitted to each central pixel point and its neighborhood pixels by the least squares method, and the principal curvature is calculated based on the fitting coefficients a, b, and c. , take the radius of curvature As the local surface curvature radius, along the edge contour of the SEM image, the pixel grayscale change rate (maxgrad) perpendicular to the edge direction is calculated, and the grayscale variance (var) of the 2.5-pixel area on both sides of the edge (5 pixels in total) is calculated, and the edge sharpness parameter s = maxgrad / var is taken.

[0060] The micro-region height gradient, local surface curvature radius, and edge sharpness parameters are normalized according to the pixel coordinates of the detection sub-region to obtain a morphological feature vector. Based on the scanning paths of the scanning electron microscope and energy dispersive X-ray spectrometer, a mapping relationship between the image pixel coordinates of the detection sub-region and the energy channel position of the component spectrum data is synchronously established. Based on this mapping relationship between the image pixel coordinates of the detection sub-region and the energy channel position of the component spectrum data, the characteristic X-ray peak intensity and the morphological feature vector of the corresponding detection sub-region are spliced ​​according to the dimension to obtain a morphological-component fusion feature vector. To achieve spatial registration of morphological parameters and component parameters, the synchronous control interface of the scanning electron microscope and energy dispersive X-ray spectrometer is used to record the acquisition time and corresponding energy channel position of each pixel during the scanning process. Based on the synchronization signal, a mapping relationship between the image pixel coordinates (x, y) and the energy channel position (E, n) of the component spectrum data is established, where E represents the energy value and n represents the channel number.

[0061] The method for synchronously establishing a mapping relationship between the image pixel coordinates of the detection sub-region and the energy channel position of the component spectrum data according to the scanning paths of the scanning electron microscope and the energy dispersive X-ray spectrometer includes:

[0062] According to the corresponding relationship between the scanning step of the scanning electron microscope and the sampling time interval of the energy dispersive X-ray spectrometer, the synchronization mark of the surface morphology image pixel coordinates and the component spectrum data acquisition time of each detection sub-area is determined; according to the synchronization mark, the real-time coordinate data of the stage position encoder of the scanning electron microscope is extracted, and the calibration parameters of the energy channel and spatial position of the energy dispersive X-ray spectrometer are combined to establish a geometric transformation matrix from the image pixel coordinates to the energy channel position; according to the corresponding relationship between the scanning step of the scanning electron microscope (set to 10nm) and the sampling time interval of the energy dispersive X-ray spectrometer (set to 10μs), the synchronization mark is determined. For example, when the scanning electron microscope completes a pixel scan, a synchronization pulse signal is triggered, and the energy dispersive X-ray spectrometer begins to collect the component spectrum data of the current position after receiving the synchronization pulse signal. By extracting the real-time coordinate data (x, y, z) of the stage position encoder of the scanning electron microscope and combining the calibration parameters of the energy channel and spatial position of the energy dispersive X-ray spectrometer, a geometric transformation matrix T is established. The geometric transformation matrix can be expressed as ,in is the translation vector [tx,ty,tz], and R is the rotation matrix containing the rotation angle parameters [θx,θy,θz].

[0063] The geometric transformation matrix is ​​decomposed into translation vector and rotation matrix components. The parameters of the translation vector and rotation matrix components are optimized by the least squares method to minimize the spatial deviation between the center pixel coordinates of the surface topography image and the center position of the energy channel of the component spectrum data in the same detection sub-region. The optimized geometric transformation matrix is ​​applied to all detection sub-regions to achieve point-by-point mapping between the image pixel coordinates and the energy channel position. To optimize the geometric transformation matrix parameters, the least squares method is used to select 10 feature points, and their pixel coordinates in the surface topography image and the energy channel position in the component spectrum data are recorded respectively. The error function is constructed. , where pi is the image pixel coordinate and qi is the corresponding energy channel position. The error function is minimized by iteratively optimizing the parameters using the gradient descent method, resulting in the optimal geometric transformation matrix and achieving accurate mapping between pixel coordinates and energy channel positions.

[0064] The method of comparing and analyzing the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtaining a defect type discrimination result of the detection sub-region according to the similarity probability includes:

[0065] After normalizing the morphology-component fusion feature vector, calculate the first kernel space similarity with each sample feature vector in the pollution particle feature library according to the kernel function mapping, and calculate the second kernel space similarity with each sample feature vector in the micro-bump feature library; normalize the morphology-component fusion feature vector and convert the eigenvalues ​​of each dimension to the [0,1] interval. Use the radial basis function kernel , parameter γ is set to 0.1, and the kernel space similarity between the current feature vector and each sample in the feature library is calculated. The first kernel space similarity between the current feature vector and each sample in the pollution particle feature library (containing 100 typical samples) is calculated by kernel function mapping. , and the second kernel space similarity with each sample in the micro-protrusion feature library (also containing 100 typical samples) .

[0066] The probability distribution of the first kernel space similarity in the pollution particle category and the probability distribution of the second kernel space similarity in the micro-protrusion category are estimated respectively through the probability density function, and the first posterior probability that the current morphology-component fusion feature vector belongs to the pollution particle and the second posterior probability that it belongs to the micro-protrusion are obtained; the probability ratio of the first posterior probability to the second posterior probability is obtained, and when the probability ratio exceeds the preset probability ratio threshold, the defect type is determined to be a pollution particle, otherwise the defect type is determined to be a micro-protrusion; the similarity probability includes the first posterior probability and the second posterior probability. In the pollution particle category and Probability distribution of micro-protrusion categories. For example, for a suspected defect detected in the edge area of ​​the target, the average similarity between the calculated morphology-composition fusion feature vector and the first kernel space of the contamination particle feature library is 0.85, and the average similarity between the calculated morphology-composition fusion feature vector and the second kernel space of the micro-protrusion feature library is 0.35. The posterior probability that the feature vector belongs to a contamination particle is obtained by kernel density estimation. =0.78, the posterior probability of being a micro-bump =0.22. Calculate the probability ratio / =3.55, which exceeds the preset probability ratio threshold of 2.5, so the defect is determined to be a contamination particle.

[0067] The method of estimating the probability distribution of the first kernel space similarity in the pollution particle category and the probability distribution of the second kernel space similarity in the micro-bump category by using a probability density function to obtain a first posterior probability that the current shape-component fusion feature vector belongs to the pollution particle and a second posterior probability that the current shape-component fusion feature vector belongs to the micro-bump includes:

[0068] The kernel function bandwidth parameter is calculated based on the eigenvalues ​​of the covariance matrix of all sample feature vectors in the preset pollution particle feature library and micro-protrusion feature library. The bandwidth parameter is dynamically adjusted according to the distribution sparsity of the sample feature vectors in the pollution particle feature library and micro-protrusion feature library. The kernel function bandwidth parameter is calculated based on the eigenvalues ​​of the covariance matrix of all sample feature vectors in the pollution particle feature library and micro-protrusion feature library. For the pollution particle feature library, the average eigenvalue of the covariance matrix is ​​0.15, and the initial bandwidth parameter is =0.5×0.15^(0.5)=0.194; for the micro-convex feature library, the average eigenvalue of the covariance matrix is ​​0.18, and the initial bandwidth parameter =0.5×0.18^(0.5)=0.212.

[0069] The first and second kernel space similarities are mapped to high-dimensional space respectively using Gaussian kernel functions, and non-parametric kernel density estimation is performed on the kernel space similarities of all samples in the pollution particle feature library and the micro-bump feature library to generate kernel density distribution functions of pollution particle categories and micro-bump categories respectively;

[0070] Substitute the first kernel space similarity of the current shape-component fusion feature vector into the kernel density distribution function of the pollution particle category to calculate its probability density value as the first posterior probability, and substitute the second kernel space similarity into the kernel density distribution function of the micro-bump category to calculate its probability density value as the second posterior probability. Using Gaussian kernel function , perform non-parametric kernel density estimation on the kernel space similarity of all samples in the pollution particle feature library and the micro-protrusion feature library, and generate the kernel density distribution function of the pollution particle category respectively and the kernel density distribution function of the micro-bump category For the current shape-component fusion feature vector, calculate the first kernel space similarity =0.85, substitute Get the probability density value =0.78 as the first posterior probability; calculate the second kernel space similarity =0.35, substitute Get the probability density value =0.22 as the second posterior probability.

[0071] The method for dynamically adjusting the bandwidth parameter according to the distribution sparsity of the sample feature vectors in the pollution particle feature library and the micro-protrusion feature library includes:

[0072] According to the spatial distribution of all sample feature vectors in the preset pollution particle feature library and micro-protrusion feature library, the average Mahalanobis distance between samples in each feature library is calculated as a distribution sparsity measurement index; the bandwidth parameter adjustment coefficient of each pollution particle feature library and micro-protrusion feature library is generated according to the sparsity ratio of the distribution sparsity measurement index to the preset reference sparsity; for each pair of sample feature vectors xi and xj in the pollution particle feature library, its Mahalanobis distance is calculated. , where Σ is the sample covariance matrix. The average value of the Mahalanobis distance of all sample pairs is used to obtain the distribution sparsity metric of the pollution particle feature library. =2.8; Similarly, calculate the distribution sparsity metric of the micro-protrusion feature library =2.3.

[0073] When the distribution sparsity metric is greater than the preset baseline sparsity, the bandwidth parameter adjustment coefficient is the square root of the sparsity ratio, otherwise it is the square value of the sparsity ratio; the bandwidth parameter adjustment coefficient is multiplied by the initial bandwidth parameter to obtain the dynamically adjusted kernel function bandwidth parameter, and is applied to the kernel density distribution function construction of the pollution particle category and the micro-protrusion category respectively; the baseline sparsity is obtained by statistically calculating the logarithm of the covariance matrix determinant of the sample feature vectors in the preset pollution particle feature library and the micro-protrusion feature library. Baseline sparsity It is determined by the logarithm of the covariance matrix determinant of the sample eigenvector in the feature library, specifically: , where n is the dimension of the feature vector. Calculate the sparsity ratio , When the sparsity ratio is greater than 1, the bandwidth parameter adjustment coefficient is the square root of the sparsity ratio; when the sparsity ratio is less than 1, the bandwidth parameter adjustment coefficient is the square value of the sparsity ratio. Therefore, the bandwidth parameter adjustment coefficient of the pollution particle feature library is =1.058; bandwidth parameter adjustment coefficient of micro-protrusion feature library =0.846. Finally, the kernel function bandwidth parameter after dynamic adjustment is , .

[0074] The method of mapping the discrimination result to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report includes:

[0075] Based on the order in which the detection subregions are divided and the spatial relationship of the scanning path, a mapping matrix is ​​established between the center points of the detection subregions and the three-dimensional coordinate system of the target surface. Based on this mapping matrix, the locations of the contaminant particles and micro-bumps in the defect type discrimination results are converted into three-dimensional spatial coordinates on the target surface, and the corresponding similarity probabilities are converted into confidence scores and attached to each defect's three-dimensional coordinates. A mapping matrix M is established between the center points of the detection subregions and the three-dimensional coordinate system of the target surface. For a 100mm×100mm target surface, it is divided into 10×10 detection subregions, with the center point of each subregion corresponding to a three-dimensional coordinate (X, Y, Z). For example, the center point of the detection subregion at position (3, 4) is mapped to the three-dimensional coordinates (35mm, 45mm, 0mm) on the target surface. Based on the discrimination results, the defect type (contaminant particle or micro-bump) is associated with the detection subregion number in which it is located and converted into three-dimensional spatial coordinates on the target surface using the mapping matrix M. For example, the defect identified as a contamination particle is located in the detection sub-area (3, 4). According to the mapping matrix, its three-dimensional coordinates are (35mm, 45mm, 0mm), and the corresponding similarity probability (such as 0.78) is converted into a confidence score of 78.

[0076] A spatial cluster analysis is performed on all three-dimensional coordinates of defects to calculate the weighted average of the defect distribution density, spatial aggregation characteristics, and confidence scores of the contamination particles and micro-protrusions. At the same time, the component weights of different elements in the contamination particle area and the micro-protrusion area are calculated based on the characteristic X-ray peak intensity. The defect distribution density, spatial aggregation characteristics, component weights, and the corresponding weighted average of the confidence scores are integrated into a structured data table. The structured data table is combined with the three-dimensional coordinate mapping results to generate a test report containing a defect spatial position distribution map, a heat map of the element composition weight distribution, and a confidence layered annotation map. The weighted average of the confidence scores is used to correct the statistical results of the defect distribution density. A spatial cluster analysis is performed on all three-dimensional coordinates of defects using the DBSCAN algorithm with parameters set to eps=5mm (cluster radius) and minPts=3 (minimum number of points) to calculate the defect distribution density of contamination particles and micro-protrusions. For example, in the edge area of ​​the target (0-10mm range), the density of contamination particles is 3.6 / cm², and the density of micro-protrusions is 1.2 / cm²; in the center area of ​​the target, the density of contamination particles is 0.8 / cm², and the density of micro-protrusions is 0.5 / cm². At the same time, the composition weights of different elements in the contamination particle area and the micro-protrusion area are calculated based on the intensity of the characteristic X-ray peak. For example, in the contamination particle area, Fe accounts for 15.3%, Cu accounts for 8.7%, and Al accounts for 5.2%; in the micro-protrusion area, the main components are Mg (70.5%) and O (29.2%), and almost no metal elements are contained. The defect distribution density, spatial aggregation characteristics, component proportion weights, and the weighted average of the corresponding confidence scores are integrated into a structured table, and three types of visualization charts are generated: a defect spatial location distribution map, which uses a scatter plot method, with different colors representing different types of defects; an element composition weight distribution heat map, which uses a color gradient to represent the distribution concentration of different elements; and a confidence layered annotation map, which divides the confidence into three levels: high, medium, and low, and is marked with red, yellow, and green respectively.

[0077] The method for correcting the statistical result of defect distribution density by using a weighted average algorithm includes:

[0078] The weight factors for contaminants and micro-bumps within the detection sub-region are calculated based on the confidence scores assigned to each defect's three-dimensional coordinates. The weight factor is the ratio of the confidence score for the current defect coordinate to the sum of the confidence scores for all similar defects. The defect distribution density is weighted and summed according to the weight factors to obtain the confidence-weighted mean defect distribution density. For example, for a contaminant defect with a confidence score of 78, if the sum of the confidence scores for all contaminants within the region is 300, the weight factor for this defect is w = 78 / 300 = 0.26. The defect distribution density is weighted and summed according to the weight factors. Raw statistics show a contaminant particle density of 3.6 particles / cm² at the target edge. After factoring in confidence weighting, the weighted average density is 3.2 particles / cm².

[0079] The defect distribution density mean and the preset confidence interval are compared and analyzed with the unweighted defect distribution density to generate a revised defect distribution density statistical result and embed it into the structured data table of the test report. The preset confidence interval is [2.4, 4.0] pieces / cm², with a confidence level of 95%. The unweighted 3.6 pieces / cm² is within the confidence interval but deviates towards the upper limit, indicating that the defect density in the high-confidence area is slightly lower than the average level. Finally, in the test report, the revised contamination particle distribution density is expressed as 3.2±0.8 pieces / cm². This weighted average method can more accurately reflect the actual situation of the defect distribution and reduce the interference of low-confidence judgment results on statistical analysis.

[0080] The method of converting the contamination particles and micro-protrusion positions in the defect type discrimination result into three-dimensional spatial coordinates of the target surface according to the mapping matrix includes:

[0081] The detection subregion numbers marked as contaminant particles or micro-bumps in the discrimination results are extracted based on the detection subregion index numbers. Based on the correspondence between detection subregion numbers and 3D coordinates stored in a mapping matrix, the detection subregion number corresponding to each defect is mapped to the 3D coordinates of the target surface. A mapping matrix is ​​established to store the correspondence between detection subregion numbers and 3D coordinates. For a standard 100mm×100mm magnesium oxide single crystal target, the surface is typically evenly divided into 10×10 detection subregions, each measuring 10mm×10mm. The mapping matrix M is a two-dimensional array, where M[i][j] represents the 3D coordinates (X, Y, Z) corresponding to detection subregion (i, j), where i and j represent the row and column indices of the detection subregion, respectively, and range from 1 to 10. The coordinate calculation formula is X = (i-0.5) × 10mm, Y = (j-0.5) × 10mm, and Z = 0mm (representing the target surface plane). For example, the three-dimensional coordinates corresponding to the detection sub-area (1,1) (i.e., the first area in the lower left corner) are (5mm, 5mm, 0mm); the three-dimensional coordinates corresponding to the detection sub-area (10,10) (i.e., the last area in the upper right corner) are (95mm, 95mm, 0mm).

[0082] Finally, it should be noted that although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A surface detection method for preparing a magnesium oxide single crystal target, characterized in that: The method comprises: The target surface is divided into multiple detection sub-areas, and the priority scanning order of each detection sub-area is determined; the surface topography image of each detection sub-area is obtained by scanning electron microscopy, and the composition spectrum data of the corresponding detection sub-area is collected simultaneously by energy dispersive X-ray spectrometry; After multi-frame superposition processing of the component spectrum data, continuous background noise is removed by nonlinear least squares fitting and characteristic X-ray peak intensities are extracted and recorded as component parameters; topography parameters are extracted from the surface topography image, and the topography parameters are spatially aligned with the component parameters to obtain a topography-component fusion feature vector; the topography parameters include micro-area height gradient, local surface curvature radius, and edge sharpness parameters; Comparing and analyzing the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtaining a defect type discrimination result of the detection sub-area based on the similarity probability; Mapping the discrimination results to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report; The method of removing continuous background noise and extracting characteristic X-ray peak intensity by nonlinear least squares fitting after performing multi-frame superposition processing on the component spectrum data includes: The component spectrum data is subjected to multi-frame superposition processing by a weighted average method, the number of superposition frames of the multi-frame superposition processing is dynamically adjusted according to the priority of the current detection sub-region, and when the priority of the detection sub-region is higher than a preset priority threshold, the number of superposition frames is gradually increased by an adaptive algorithm until the signal-to-noise ratio improvement rate of the component spectrum data after superposition is greater than the preset improvement rate threshold; A composite background model including a polynomial basis function and an exponential decay basis function is constructed, and weight coefficients of the polynomial basis function and the exponential decay basis function are obtained by performing a global fit on the component spectrum data after multi-frame superposition processing using a nonlinear least squares method. The weight coefficients are substituted into the composite background model to obtain a continuous background noise covering the entire energy range; and the continuous background noise is removed from the component spectrum data after multi-frame superposition processing to obtain denoised component spectrum data. According to the preset target element characteristic energy range, the sliding window method is used to screen candidate characteristic peaks that meet the local maximum condition and have a half-height width within a preset width interval in the denoised component spectrum data, and the peak area integral value is calculated as the characteristic X-ray peak intensity according to the energy channel interval of each candidate characteristic peak, and the left energy boundary and the right energy boundary of each candidate characteristic peak are dynamically calibrated. The dynamic calibration is to determine the peak boundary starting point according to the change trend of the first-order derivative of the counting rate of the energy channel corresponding to the candidate characteristic peak. If the counting rate correlation coefficient of adjacent energy channels is lower than the preset correlation threshold, it is determined to be a pseudo characteristic peak formed by an isolated noise point and is eliminated.

2. The surface detection method for preparing a magnesium oxide single crystal target according to claim 1, characterized in that: The method of extracting topography parameters from the surface topography image and spatially registering the topography parameters with the component parameters to obtain a topography-component fusion feature vector includes: A three-dimensional topography reconstruction is performed on the surface topography image of each detection sub-region to obtain a height distribution matrix. Based on the height distribution matrix, the height gradient modulus of each pixel point in a preset neighborhood window is calculated as the micro-area height gradient, and the local surface curvature radius is calculated by quadratic surface fitting. The pixel grayscale change rate is extracted along the edge contour of the surface topography image, and the ratio of the maximum grayscale change rate to the grayscale variance of the adjacent area is used as the edge sharpness parameter. The micro-area height gradient, local surface curvature radius and edge sharpness parameters are normalized according to the pixel coordinates of the detection sub-area to obtain a morphological feature vector; according to the scanning paths of the scanning electron microscope and the energy dispersive X-ray spectrometer, a mapping relationship between the image pixel coordinates of the detection sub-area and the energy channel position of the component spectrum data is synchronously established; according to the mapping relationship between the image pixel coordinates of the detection sub-area and the energy channel position of the component spectrum data, the characteristic X-ray peak intensity and the morphological feature vector of the corresponding detection sub-area are spliced ​​according to the dimension to obtain a morphology-component fusion feature vector.

3. The surface detection method for preparing a magnesium oxide single crystal target according to claim 2, characterized in that: The method for synchronously establishing a mapping relationship between the image pixel coordinates of the detection sub-region and the energy channel position of the component spectrum data according to the scanning paths of the scanning electron microscope and the energy dispersive X-ray spectrometer includes: Determining synchronization marks for the pixel coordinates of the surface topography image and the component spectrum data acquisition moment of each detection sub-region based on the corresponding relationship between the scanning step size of the scanning electron microscope and the sampling time interval of the energy dispersive X-ray spectrometer; extracting real-time coordinate data of the scanning electron microscope stage position encoder based on the synchronization marks, and establishing a geometric transformation matrix from the image pixel coordinates to the energy channel position in combination with the calibration parameters of the energy dispersive X-ray spectrometer energy channel and spatial position; The geometric transformation matrix is ​​decomposed into translation vector and rotation matrix components, and the parameters of the translation vector and rotation matrix components are optimized by the least squares method to minimize the spatial deviation between the central pixel coordinates of the surface morphology image and the central position of the energy channel of the component spectrum data of the same detection sub-area; the optimized geometric transformation matrix is ​​applied to all detection sub-areas to realize point-by-point mapping between the image pixel coordinates and the energy channel position.

4. The surface detection method for preparing a magnesium oxide single crystal target according to claim 1, characterized in that: The method of comparing and analyzing the morphology-component fusion feature vector with a preset pollution particle feature library and a micro-protrusion feature library to obtain similarity probability, and obtaining a defect type discrimination result of the detection sub-region according to the similarity probability includes: After normalizing the morphology-component fusion feature vector, a first kernel space similarity is calculated between the feature vector and each sample feature vector in the pollution particle feature library according to kernel function mapping, and a second kernel space similarity is calculated between the feature vector and each sample feature vector in the micro-bump feature library; The probability distribution of the first kernel spatial similarity in the contamination particle category and the probability distribution of the second kernel spatial similarity in the micro-bump category are estimated respectively through the probability density function, and the first posterior probability that the current morphology-component fusion feature vector belongs to the contamination particle and the second posterior probability that it belongs to the micro-bump are obtained; the probability ratio of the first posterior probability to the second posterior probability is obtained, and when the probability ratio exceeds the preset probability ratio threshold, the defect type is determined to be a contamination particle; otherwise, the defect type is determined to be a micro-bump; the similarity probability includes the first posterior probability and the second posterior probability.

5. The surface detection method for preparing a magnesium oxide single crystal target according to claim 4, characterized in that: The method of estimating the probability distribution of the first kernel space similarity in the pollution particle category and the probability distribution of the second kernel space similarity in the micro-bump category by using a probability density function to obtain a first posterior probability that the current shape-component fusion feature vector belongs to the pollution particle and a second posterior probability that the current shape-component fusion feature vector belongs to the micro-bump includes: The kernel function bandwidth parameter is calculated based on the eigenvalues ​​of the covariance matrix of all sample feature vectors in the preset pollution particle feature library and micro-bump feature library, and the bandwidth parameter is dynamically adjusted based on the distribution sparsity of the sample feature vectors in the pollution particle feature library and micro-bump feature library; The first and second kernel space similarities are mapped to high-dimensional space respectively using Gaussian kernel functions, and non-parametric kernel density estimation is performed on the kernel space similarities of all samples in the pollution particle feature library and the micro-bump feature library to generate kernel density distribution functions of pollution particle categories and micro-bump categories respectively; The first kernel space similarity of the current morphology-component fusion feature vector is substituted into the kernel density distribution function of the pollution particle category to calculate its probability density value as the first posterior probability, and the second kernel space similarity is substituted into the kernel density distribution function of the micro-bump category to calculate its probability density value as the second posterior probability.

6. The surface detection method for preparing a magnesium oxide single crystal target according to claim 5, characterized in that: The method for dynamically adjusting the bandwidth parameter according to the distribution sparsity of the sample feature vectors in the pollution particle feature library and the micro-protrusion feature library includes: Based on the spatial distribution of all sample feature vectors in the preset pollution particle feature library and micro-protrusion feature library, the average Mahalanobis distance between samples in each feature library is calculated as a distribution sparsity measurement index; and bandwidth parameter adjustment coefficients for the pollution particle feature library and the micro-protrusion feature library are generated based on the sparsity ratio of the distribution sparsity measurement index to a preset reference sparsity. When the distribution sparsity measurement index is greater than the preset baseline sparsity, the bandwidth parameter adjustment coefficient is the square root of the sparsity ratio, otherwise it is the square value of the sparsity ratio; the bandwidth parameter adjustment coefficient is multiplied by the initial bandwidth parameter to obtain the dynamically adjusted kernel function bandwidth parameter, and is applied to the construction of kernel density distribution functions for the pollution particle category and micro-protrusion category respectively.

7. The surface detection method for preparing a magnesium oxide single crystal target according to claim 1, characterized in that: The method of mapping the discrimination result to the three-dimensional coordinates of the target surface, marking the positions and confidence levels of the contamination particles and micro-protrusions, and generating a test report includes: A mapping matrix is ​​established between the center point of the detection sub-area and the three-dimensional coordinate system of the target surface based on the division order of the detection sub-areas and the spatial position relationship of the scanning path; the positions of the contaminant particles and micro-protrusions in the defect type discrimination results are converted into three-dimensional spatial coordinates on the target surface based on the mapping matrix, and the corresponding similarity probability is converted into a confidence score and attached to the three-dimensional coordinate of each defect; A spatial clustering analysis is performed on the three-dimensional coordinates of all defects to statistically calculate the defect distribution density, spatial aggregation characteristics and weighted average confidence scores of the contamination particles and micro-protrusions. At the same time, the component proportion weights of different elements in the contamination particle area and the micro-protrusion area are calculated based on the characteristic X-ray peak intensity; the defect distribution density, spatial aggregation characteristics, component proportion weights and the corresponding weighted average confidence scores are integrated into a structured data table, and the structured data table is combined with the three-dimensional coordinate mapping result to generate a test report containing a defect spatial position distribution map, an element composition weight distribution heat map and a confidence layered annotation map; the weighted average confidence score is used to correct the statistical results of the defect distribution density through a weighted average algorithm.

8. The surface detection method for preparing a magnesium oxide single crystal target according to claim 7, characterized in that: The method for correcting the statistical result of defect distribution density by using a weighted average algorithm includes: Calculating weight factors for contaminants and micro-protrusions within the detection sub-region based on the confidence scores attached to the three-dimensional coordinates of each defect, where the weight factor is the ratio of the confidence score of the current defect coordinate to the sum of the confidence scores of all similar defects; performing a weighted summation of the defect distribution density based on the weight factor to obtain a confidence-weighted defect distribution density mean; The defect distribution density mean and the preset confidence interval are compared and analyzed with the unweighted defect distribution density to generate a corrected defect distribution density statistical result and embed it into a structured data table of the inspection report.

9. The surface detection method for preparing a magnesium oxide single crystal target according to claim 7, characterized in that: The method of converting the contamination particles and micro-protrusion positions in the defect type discrimination result into three-dimensional spatial coordinates of the target surface according to the mapping matrix includes: According to the division index number of the detection sub-area, the detection sub-area number marked as contamination particles or micro-protrusions in the judgment result is extracted. According to the correspondence between the detection sub-area number and the three-dimensional coordinates stored in the mapping matrix, the detection sub-area number corresponding to each defect is mapped to the three-dimensional spatial coordinates of the target surface.

Images