Wafer quality analysis method based on image recognition

Abstract

The invention discloses a wafer quality analysis method based on image recognition, and particularly relates to the field of semiconductor manufacturing quality detection, which comprises the following steps: synchronously acquiring wafer two-dimensional images under visible light, infrared and ultraviolet spectrums through a multispectral and high-resolution fusion camera combination model; the structured light three-dimensional reconstruction technology is used for obtaining surface three-dimensional shape data, a self-adaptive noise suppression algorithm is applied, noise features are recognized through frequency domain analysis, filtering parameters are dynamically adjusted, targeted noise reduction is achieved, a multi-scale feature pyramid and quantum dot mark recognition fusion model is constructed, defect contours and material heterogeneous areas are extracted in a layered mode, and the defect detection precision is improved. The method comprises the following steps: extracting microstructure characteristic parameters by combining a quantum dot fluorescence labeling technology, mapping microscopic characteristics and three-dimensional shape data to a multi-modal space model, generating three-property parameters of a wafer, establishing a three-dimensional evaluation system, comprehensively judging a quality grade and outputting an analysis report, so that the precision and efficiency of wafer quality detection are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor manufacturing quality detection, and more particularly to a wafer quality analysis method based on image recognition. BACKGROUND

[0002] With the continuous progress of semiconductor manufacturing process, wafer quality detection as a key link to ensure chip performance and reliability, is increasingly valued by the industry, and the small defects of wafer surface and internal structure, such as lattice dislocation, impurity pollution, surface fluctuation, may have a significant impact on subsequent processes and final product performance, therefore, developing high-precision and high-efficiency wafer quality detection method has become an important research direction to improve the level of semiconductor manufacturing.

[0003] The current technical solution mainly relies on the method of combining traditional optical imaging and machine vision to detect wafer defects, which usually uses a single spectrum high-resolution camera to collect wafer surface images, then extracts defect features through image processing algorithm, and assists with simple three-dimensional measurement means to obtain surface topography data, and some systems also introduce standard template comparison or statistical analysis to determine wafer quality grade, and complete wafer quality analysis However, when actually used, it still has some disadvantages, such as image acquisition relying on a single spectral band, which is difficult to capture the microscopic characteristics of wafer internal material heterogeneity and lattice defects, resulting in one-sided detection dimension, fixed parameter denoising algorithm, unable to dynamically adapt to different noise types such as high-frequency noise and periodic interference, easy to cause defect feature loss or noise residue, feature extraction is mainly based on a single scale, lacking precise recognition ability of microscopic structure such as nanoscale holes and impurity distribution, and no correlation model between defect features and wafer physical performance and electrical performance, and the final quality evaluation only stays at the surface defect judgment, which is difficult to reflect the comprehensive performance of the wafer and the matching degree with the process standard, and cannot meet the fine detection needs of high-end semiconductor manufacturing. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a wafer quality analysis method based on image recognition, which solves the problems in the above background art by the following scheme.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a wafer quality analysis method based on image recognition, comprising S1: constructing a multi-dimensional image acquisition model, collecting wafer two-dimensional images under visible light, infrared and ultraviolet spectrum, and simultaneously starting structure light three-dimensional reconstruction technology to obtain wafer surface three-dimensional topography data; S2: constructing a dynamic adjustment model, applying an adaptive noise suppression algorithm to the collected wafer two-dimensional images, identifying noise characteristics through frequency domain analysis, and dynamically adjusting filter parameters to realize targeted noise reduction; S3: Construct a feature pyramid fusion model, which is a fusion model of multi-scale feature pyramid and quantum dot labeling, extract the defect contour and material heterogeneity region of the wafer two-dimensional image through multi-scale feature pyramid hierarchical extraction, and specifically label the defect contour and material heterogeneity region of the wafer two-dimensional image through quantum dot fluorescence labeling technology, and extract the wafer microstructure feature parameters combined with fluorescence intensity analysis; S4: Construct a multi-modal space mapping model, map the wafer microstructure feature parameters and surface three-dimensional topography data to different function models, and generate wafer three-property parameters; S5: Construct a quality evaluation model, establish a three-dimensional evaluation system according to the wafer three-property parameters, determine the quality grade, and output the wafer quality analysis report.

[0006] Preferably, the multi-dimensional image acquisition model is a combination model of multi-spectral and high-resolution fusion cameras, the spectral range is 380-1100nm, the minimum spectral bandwidth is 10nm, and the combination mode is self-defined channel combination. When acquiring images, different spectral channels are selected for synchronous acquisition according to the preset spectral combination mode.

[0007] Preferably, the structured light three-dimensional reconstruction technology uses 405nm laser to project sinusoidal stripes, the frame rate is 30fps, the single-frame three-dimensional point cloud density is ≥100 points / mm², and the wafer three-dimensional topography data is obtained through the principle of triangulation.

[0008] Preferably, the two-dimensional image includes surface defect contour image and material heterogeneity region image, and the three-dimensional topography data includes surface relief and height data, curvature and slope data, and microstructure three-dimensional model data.

[0009] Preferably, the fusion model of multi-scale feature pyramid and quantum dot labeling recognition includes a feature pyramid unit and a quantum dot labeling recognition unit, and realizes real-time data transmission through a data interaction interface with a transmission delay ≤10ms. Each layer scale of the feature pyramid unit is associated through a Gaussian difference operator, and each layer is provided with an independent feature filter to automatically filter invalid features with a signal-to-noise ratio <3.

[0010] Preferably, the quantum dot fluorescence labeling technology uses CdSe / ZnS core-shell structure quantum dots with a particle size of 5-8nm, an excitation wavelength of 488nm, and an emission wavelength of 655nm. The labeling density on the wafer surface is 10 6 -10 7 / mm² and the labeling specificity is ≥99.5%. The specific binding group modified on the surface of the quantum dot is combined with the atoms at the lattice defects to form a stable covalent bond, thereby improving the fluorescence labeling binding rate.

[0011] Preferably, the wafer microstructure feature parameters include lattice defect density, micro stress distribution parameters, nanoscale hole and gap parameters, and impurity atom distribution parameters.

[0012] Preferably, the multi-modal space mapping model is a multi-dimensional image data and wafer trinity correlation mapping function model, the wafer trinity is physical performance, electrical performance and process standard matching degree, wafer microstructure feature parameters and surface three-dimensional topography data are mapped to different function models to generate wafer trinity parameters, wherein the physical performance parameters include wafer surface flatness, the electrical performance parameters include carrier mobility and resistivity deviation rate, and the process standard matching degree parameters include lithography accuracy and film thickness deviation rate.

[0013] Preferably, the three-dimensional evaluation system in the quality evaluation model is specifically physical performance dimension, electrical performance dimension and process standard matching degree dimension, wafer quality three-dimensional evaluation is carried out according to wafer trinity parameters, and the quality grade is determined, and finally a quality analysis report is output, specifically including a three-dimensional evaluation radar chart and a quality grade.

[0014] The technical effects and advantages of the present application are as follows: The present application realizes multi-dimensional analysis of wafer quality through multi-dimensional image acquisition and multi-modal data fusion, the multi-dimensional image acquisition model covers visible light, infrared and ultraviolet spectrum, and three-dimensional topography data are acquired synchronously, the hierarchical extraction of defect profile and material heterogeneity region and the specific identification of quantum dot markers are combined with the feature pyramid fusion model, macro three-dimensional defects and microstructure features can be captured at the same time, compared with traditional single spectrum detection, the integrity and accuracy of feature extraction are greatly improved, and more abundant basic data is provided for subsequent quality evaluation. The present application effectively guarantees the reliability of the analysis result through dynamic adjustment of the noise reduction mechanism and fine feature processing, the dynamic adjustment model identifies the noise type through frequency domain analysis and adjusts the filter parameters accordingly, the noise is suppressed while the defect details are preserved to the greatest extent, independent filters are arranged in each layer of the feature pyramid to filter invalid features, the quantum dot marker technology is combined with the defect area with high specificity, the influence of interference factors on the extraction of microstructure feature parameters is significantly reduced, and the measurement accuracy of key parameters such as lattice defect density and impurity distribution is improved. The present application realizes comprehensive quantitative determination of wafer quality through the three-dimensional evaluation system of wafer trinity parameters, the multi-modal space mapping model maps micro features and three-dimensional topography data into physical performance, electrical performance and process standard matching degree parameters, the quality evaluation model obtains an objective quality grade through weighted calculation, compared with the traditional grading method based on surface defects only, the actual performance of the wafer and the degree of fit with the process requirements can be more comprehensively reflected, a more scientific basis is provided for quality control in the semiconductor manufacturing process, and the yield and reliability of chip production are improved. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a schematic diagram of the targeted noise reduction structure of the present invention; Figure 3 This is a schematic diagram of the wafer microstructure feature parameter mapping structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] As attached Figure 1 Appendix Figure 2 and attached Figure 3 The wafer quality analysis method based on image recognition shown includes S1: constructing a multi-dimensional image acquisition model, acquiring two-dimensional images of the wafer under visible light, infrared and ultraviolet spectra, and simultaneously starting structured light three-dimensional reconstruction technology to obtain three-dimensional morphology data of the wafer surface.

[0018] It should be specifically noted that the multi-dimensional image acquisition model is a combination of a multispectral and high-resolution fusion camera. The spectral detection range covers a wide band of 380-1100nm. It captures wafer surface defects through the visible light region of 380-760nm, detects internal structural anomalies by penetrating the surface coating through the near-infrared region of 760-1100nm, and identifies fluorescent defects in materials through the ultraviolet band of 200-380nm. The system's minimum spectral bandwidth is 10nm. Fine band division is achieved by subdividing the visible light region into narrow bands of 10nm, acquiring high-resolution imaging components and clear images of each band, presenting micron-level wafer details.

[0019] Regarding channel combination, the model adopts a custom channel combination mechanism, equipped with preset spectral combination modes and a synchronous triggering module. The preset spectral combination modes include three scenario-based schemes: for conventional surface inspection, it enables simultaneous acquisition of three visible light channels (red, green, and blue) and a single near-infrared channel to quickly obtain wafer surface morphology and basic material information; for high-precision defect screening, it enables simultaneous acquisition of a combination of an ultraviolet channel and five narrow-band near-infrared channels to identify subtle material inhomogeneities through multi-band comparison; for wafers with special processes, it supports user-defined channel parameters, such as specifying three discrete bands of 400nm, 800nm, and 1000nm for specific inspection. In the aspect of acquisition control, the high-resolution fusion camera and the structured light three-dimensional reconstruction system realize precise synchronization in time and space. The structured light three-dimensional reconstruction technology adopts a 405 nm laser to project a sinusoidal fringe, and the frame rate is 30 fps, and the single-frame three-dimensional point cloud density is ≥100 points / mm². When the wafer is in the detection position, the system first triggers the structured light projector to project the coded grating pattern to the wafer surface, and synchronously starts the high-resolution fusion camera to capture the deformed grating image. The three-dimensional topography data of the wafer surface is quickly calculated through the principle of triangulation. The three-dimensional topography data includes surface relief and height data, curvature and slope data, and microstructure three-dimensional model data. At the same time, the high-resolution fusion camera synchronously acquires two-dimensional image data of multiple spectral channels in the same field of view range according to the preset spectral combination mode. The two-dimensional image includes a surface defect contour image and a material heterogeneity region image.

[0020] It needs to be further explained that the multi-dimensional image acquisition model synchronously acquires three-dimensional topography data when acquiring two-dimensional images of the wafer. The two-dimensional image reflects the characteristics of material chemical component difference and surface pollution, and the three-dimensional topography data represents the geometric parameters of wafer warpage, flatness, and defect degree.

[0021] S2: A dynamic adjustment model is constructed, and an adaptive noise suppression algorithm is applied to the acquired wafer two-dimensional image. After identifying the noise characteristics through frequency domain analysis, the filter parameters are dynamically adjusted to realize targeted noise reduction.

[0022] It needs to be specifically explained that the targeted noise reduction is to perform two-dimensional fast Fourier transform (2D-FFT) on the wafer two-dimensional image and perform frequency spectrum shift. The noise characteristics are identified by analyzing the amplitude spectrum, the dominant frequency band of the noise is determined by calculating the power spectral density, the interference intensity is judged by quantifying the noise energy proportion, the types of Gaussian noise and periodic noise are distinguished according to the frequency spectrum mode, the filter parameters are dynamically adjusted based on the identification results, the standard deviation of the adaptive Gaussian filter is enhanced for high-frequency noise, the center frequency and bandwidth of the notch filter are calibrated for periodic interference, and the wavelet threshold and decomposition layer number are optimized for mixed noise. After the filtering process is completed in the frequency domain, the inverse Fourier transform is used to convert back to the spatial domain. At the same time, a real-time feedback mechanism is introduced, the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) are used as indexes, the filter strength is increased by 10% at the noise residual position, and the filter strength is weakened by 10% at the detail loss position, so as to realize targeted noise suppression and accurate preservation of wafer defect details.

[0023] It needs to be further explained that the standard for identifying noise characteristics by analyzing the amplitude spectrum is that high-energy distribution appears in the high-frequency region (the Euclidean distance from any point in the region to the center is greater than one-fourth of the distance from the edge of the entire region), it is determined that it is high-frequency noise, a specific frequency point appears an isolated high-energy peak, it is determined that it is periodic noise, and the energy distribution has no obvious rules and covers multiple frequency bands, it is determined that it is mixed noise.

[0024] The standard for determining the noise dominant frequency band by calculating the power spectral density is to square the amplitude spectrum to obtain the power spectral density (PSD), the formula is , sliding window along the frequency axis, calculating the mean value of PSD in each window, when the window mean value is more than 3 times the signal frequency band mean value, mark the frequency range corresponding to the window as the noise dominant frequency band.

[0025] The standard for judging the interference intensity by quantifying the noise energy ratio is to calculate the ratio of the total energy in the noise dominant frequency band to the total energy of the full frequency band, and the noise energy ratio is obtained. The interference intensity is determined by comparing the ratio with the set threshold. The specific analysis method includes: Extracting the frequency coordinate , the total energy of the full frequency band ; Analyzing the total energy in the noise dominant frequency band , where is the region corresponding to the noise dominant frequency band in the frequency domain, represents the power spectral density at the coordinate in the frequency domain; Analyzing the noise energy ratio , when r < 10%, it is low interference, when 10% ≤ r ≤ 30%, it is medium interference, and when r > 30%, it is high interference.

[0026] The standard for distinguishing Gaussian noise, periodic noise and mixed noise type combined with the spectrum mode is that the Gaussian noise: PSD is continuously and uniformly distributed in the high frequency band, there is no obvious peak, and the noise energy ratio gradually decreases with the increase of frequency, the periodic noise: PSD appears a sharp peak at a specific frequency, the peak energy is more than 5 times the energy of the surrounding area, and the mixed noise: both high-frequency uniform distribution region and discrete peak value exist, and the energy ratio of the two types of regions is more than 5%.

[0027] It needs to be further explained that the standard for dynamically adjusting the filtering parameters based on the identification result is to increase the standard deviation of the adaptive Gaussian filter for high-frequency noise, when , when , when , the center frequency and bandwidth of the notch filter for periodic interference are calibrated, the calibrated center frequency is the peak frequency coordinate of the periodic noise, and the bandwidth , low interference medium interference high interference optimizing wavelet threshold and decomposition layer number for mixed noise, for decomposition layer number low interference medium interference high interference for wavelet threshold the specific analysis mode includes: analyzing noise standard deviation ; analyzing wavelet threshold .

[0028] S3: Construct a feature pyramid fusion model, the model is a fusion model of multi-scale feature pyramid and quantum dot labeling, the defect contour and material heterogeneity area of the wafer two-dimensional image are extracted by multi-scale feature pyramid hierarchical extraction, the defect contour and material heterogeneity area of the wafer two-dimensional image are specifically labeled by quantum dot fluorescence labeling technology, and the wafer microstructure characteristic parameters are extracted by combining fluorescence intensity analysis.

[0029] It should be specifically pointed out that the fusion model of multi-scale feature pyramid and quantum dot labeling recognition includes a feature pyramid unit and a quantum dot labeling recognition unit, real-time data transmission is realized through a data interaction interface and the transmission delay is ≤10ms, wherein the scale of each layer of the feature pyramid unit is realized by a Gaussian difference operator to realize hierarchical correlation and the feature mapping error of adjacent hierarchical layers is ≤2%, an independent feature filter is arranged in each layer to automatically filter invalid features with a signal-to-noise ratio < 3.

[0030] The quantum dot fluorescence labeling technology adopts CdSe / ZnS core-shell structure quantum dots, the particle size of the quantum dots is 5-8nm, the excitation wavelength is 488nm, the emission wavelength is 655nm, the labeling density on the wafer surface is 10 6 -10 7 / mm² and the labeling specificity is ≥99.5%, the specific binding group for wafer lattice defects on the surface of the quantum dots forms a stable covalent bond with the atoms at the lattice defects to improve the fluorescence labeling binding rate, and the wafer microstructure characteristic parameters include lattice defect density, microstress distribution parameters, nanoscale hole and gap parameters, and impurity atom distribution parameters.

[0031] Further, the 5-layer multi-scale feature pyramid is constructed to extract the defect profile and material heterogeneous region of the two-dimensional image from the scale level of 1 um x 1 um to 100 um x 100 um, the lattice defects on the wafer surface are specifically labeled by the quantum dot fluorescence labeling technology, the number of labeled points in the two-dimensional image is counted to convert the defect number per unit area, which is recorded as the lattice defect density, the wavelength shift of the quantum dot fluorescence emission at different positions on the wafer is measured, and the stress distribution parameter is converted and obtained, which is recorded as the micro stress distribution parameter, the quantum dots are filled in the nano holes and gaps, and the geometric parameters of the holes and gaps are measured by high-resolution microscopy combined with fluorescence intensity, which is recorded as the nano hole and gap parameter, and the distribution of impurity atoms in the wafer microstructure is obtained by detecting the spatial distribution of the quantum dot fluorescence intensity of the labeled impurity atoms, which is recorded as the impurity atom distribution parameter.

[0032] Further, the method for converting the wavelength shift of the fluorescence emission to obtain the stress distribution parameter is as follows: according to the quantitative relationship between the offset of the wavelength of the quantum dot fluorescence emission and the stress sensitivity coefficient, the stress value and spatial distribution of the corresponding region of the wafer are calculated, and the specific analysis method includes: Measuring the reference fluorescence emission wavelength of the quantum dot in the stress-free state Actual emission wavelength after stress Obtaining the wavelength offset Calling the stress sensitivity coefficient of the quantum dot According to the ratio of the wavelength offset to the stress sensitivity coefficient , the local stress is calculated, and the above measurement and calculation are repeated for different positions on the wafer, and the stress distribution parameter is generated by combining the spatial coordinate information.

[0033] S4: Constructing a multi-modal spatial mapping model to map the wafer microstructure feature parameters and surface three-dimensional topography data to different function models to generate wafer three-property parameters.

[0034] Specifically, the multi-modal spatial mapping model is a correlation mapping function model of multi-dimensional image data and wafer three properties, and the wafer three properties are physical performance, electrical performance and process standard matching degree. The wafer microstructure feature parameters and surface three-dimensional topography data are mapped to different function models to generate wafer three-property parameters, wherein the physical performance parameters include wafer surface flatness, the electrical performance parameters include carrier mobility, resistivity deviation rate, and the process standard matching degree parameters include lithography accuracy and film thickness deviation rate.

[0035] Further, the specific analysis method of the wafer surface flatness includes: in the surface relief and height data, the height value of each sampling point on the wafer surface is ​For the number of sampling points, For the average height, the standard deviation of the surface fluctuation is analyzed , further analyze the flatness of the wafer surface , wherein is a proportionality coefficient determined by the wafer material properties and detection standards.

[0036] The specific analysis method of wafer carrier mobility includes: extracting wafer lattice defect density, impurity atom distribution concentration, respectively denoted as , analyzing the wafer carrier mobility , wherein is the mobility of an ideal defect-free, impurity-free crystal, and a, b are proportionality coefficients, the specific values of which are determined by wafer material properties and detection standards.

[0037] The specific analysis method of wafer resistivity deviation rate includes: extracting the maximum stress in the wafer micro stress distribution parameter, denoted as , extracting the volume fraction of nanoscale pores and gaps, denoted as , analyzing the wafer resistivity deviation rate , wherein are the target threshold values of wafer micro stress maximum and nanoscale pore and gap parameter volume fraction, respectively, determined by wafer material properties and detection standards.

[0038] The specific analysis method of wafer lithography precision includes: measuring the wafer curvature radius and slope data , calculating the defocus amount and projection offset, both of which reflect the wafer lithography precision, the defocus amount , wherein is the diameter of the lithography field of view, and the projection offset , wherein is the distance from the lithography lens to the wafer surface.

[0039] The specific analysis method of wafer thin film thickness deviation rate includes: obtaining the actual thickness of each point of the thin film through the microstructure three-dimensional model data in the wafer three-dimensional topography, and the ratio of the difference between the actual thickness and the target thickness to the target thickness is the thickness deviation rate.

[0040] S5: Construct a quality evaluation model, establish a three-dimensional evaluation system according to the wafer three-property parameters, determine the quality grade, and output the wafer quality analysis report.

[0041] It should be specifically pointed out that the three-dimensional evaluation system in the quality evaluation model is specifically the physical performance dimension, the electrical performance dimension and the process standard matching degree dimension, the wafer quality three-dimensional evaluation is carried out according to the wafer three-property parameters, the quality grade is determined, and finally the quality analysis report is output, which specifically includes three-dimensional evaluation radar chart and quality grade.

[0042] It needs to be further explained that in the construction of the three-dimensional evaluation system, the wafer physical performance score adopts a 100-point deduction system, the surface flatness is set as a reference value of total thickness variation (TTV) ≤5um, and the actual value is deducted by 5 points for every 10% exceeding the reference value, 0 points for exceeding 100% and above, and the total physical performance score is composed.

[0043] The electrical performance carrier mobility takes the target lower limit (determined by wafer material characteristics and detection standards) as the reference, deducts 10 points for every 5% below the target lower limit, and 0 points for 50% below, the resistivity deviation rate takes ±5% as the reference, deducts 5 points for every 1% exceeding, and 0 points for reaching ±25% and above, and the two parameters each account for 50 points, and together constitute the total score of electrical performance.

[0044] The process standard matching degree takes ±2nm as the reference, deducts 5 points for every 0.5nm exceeding, and 0 points for deviation ≥±8nm, the thin film thickness deviation rate takes ±3% as the reference, deducts 5 points for every 1% exceeding, and 0 points for reaching ±10% and above, and together constitute the total score of process standard matching degree.

[0045] The quality level is divided into excellent, good, qualified and unqualified four levels, the three-dimensional evaluation system comprehensive score is calculated by weighting the three-dimensional scores according to the weight of physical performance 30%, electrical performance 40% and process standard matching degree 30%, the full score is 100 points, the comprehensive score of 90 points and above is excellent, 80-89 points is good, 60-79 points is qualified, and less than 60 points is unqualified.

[0046] The three-dimensional evaluation radar chart takes three dimensions as the coordinate axis, marks and connects the comprehensive scores of each dimension, and directly displays the performance of the wafer in each aspect.

[0047] Secondly: the drawings of the disclosed embodiments only involve the structures involved in the disclosed embodiments, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other; Finally: the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A wafer quality analysis method based on image recognition, characterized in that, Includes the following steps: S1: Construct a multi-dimensional image acquisition model to acquire two-dimensional images of the wafer under visible light, infrared and ultraviolet spectra, and simultaneously start structured light three-dimensional reconstruction technology to obtain three-dimensional morphology data of the wafer surface; S2: Construct a dynamic adjustment model, apply an adaptive noise suppression algorithm to the acquired two-dimensional wafer images, identify noise features through frequency domain analysis, and dynamically adjust the filtering parameters to achieve targeted noise reduction; S3: Construct a feature pyramid fusion model, which is a fusion model of multi-scale feature pyramid and quantum dot label recognition. The defect contours and material heterogeneous regions of the wafer's two-dimensional image are extracted layer by layer through the multi-scale feature pyramid. The defect contours and material heterogeneous regions of the two-dimensional image on the wafer surface are specifically labeled by quantum dot fluorescence labeling technology. The microstructure feature parameters of the wafer are extracted by combining fluorescence intensity analysis. S4: Construct a multimodal spatial mapping model to map the wafer microstructure feature parameters and surface three-dimensional morphology data to different function models to generate wafer three-dimensional parameters; S5: Construct a quality assessment model, establish a three-dimensional assessment system based on the three wafer properties parameters, determine the quality level, and output a wafer quality analysis report.

2. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The multi-dimensional image acquisition model is a combination model of multispectral and high-resolution fusion camera, with a spectral range of 380-1100nm and a minimum spectral bandwidth of 10nm. The combination method is a custom channel combination. When acquiring images, different spectral channels are selected for synchronous acquisition according to the preset spectral combination mode.

3. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The structured light 3D reconstruction technology uses a 405nm laser to project sinusoidal stripes at a frame rate of 30fps, with a single-frame 3D point cloud density of ≥100 points / mm², and obtains wafer 3D topography data through the triangulation principle.

4. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The two-dimensional images include surface defect contour images and material heterogeneous region images, and the three-dimensional morphology data includes surface undulation and height data, curvature and slope data, and microstructure three-dimensional model data.

5. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The fusion model of multi-scale feature pyramid and quantum dot tag recognition includes feature pyramid units and quantum dot tag recognition units. It achieves real-time data transmission with a transmission latency of ≤10ms through a data interaction interface. The feature pyramid unit achieves hierarchical association between its various levels through the difference of Gaussian operator, and each level is equipped with an independent feature filter to automatically filter out invalid features with a signal-to-noise ratio of <3.

6. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The quantum dot fluorescent labeling technology uses CdSe / ZnS core-shell structured quantum dots with a particle size of 5-8 nm, an excitation wavelength of 488 nm, an emission wavelength of 655 nm, and a labeling density of 10 on the wafer surface. 6 -10 7 With a labeling specificity of ≥99.5%, the specific binding groups modified on the quantum dot surface bind to atoms at lattice defects to form stable covalent bonds, thereby improving the fluorescent labeling binding rate.

7. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The microstructure characteristics of the wafer include lattice defect density, microstress distribution parameters, nanoscale pore and gap parameters, and impurity atom distribution parameters.

8. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The multimodal spatial mapping model is a correlation mapping function model between multi-dimensional image data and wafer properties, namely physical properties, electrical properties, and process standard matching degree. The wafer microstructure feature parameters and surface three-dimensional morphology data are mapped to different function models to generate wafer property parameters. Among them, physical property parameters include wafer surface flatness, electrical property parameters include carrier mobility and resistivity deviation rate, and process standard matching degree parameters include photolithography accuracy and thin film thickness deviation rate.

9. The wafer quality analysis method based on image recognition according to claim 1, characterized in that: The three-dimensional evaluation system in the quality assessment model specifically includes physical performance, electrical performance, and process standard matching dimensions. Based on the three wafer properties parameters, the quality of the wafer is evaluated in three dimensions and the quality level is determined. Finally, a quality analysis report is output, which includes a three-dimensional evaluation radar chart and the quality level.

Images