A wafer-scale panoramic stitching method and system for infrared array cameras

Abstract

This invention belongs to the technical field of semiconductor inspection and relates to a wafer cross-scale panoramic stitching method and system using an infrared array camera. The method includes the following steps: extracting initial stress data of the target wafer based on physical polarization modulation; defining the global optical path spatial phase map as a global physical optical fingerprint; extracting the transient distortion phase gradient variation rate of the overlapping field of view; generating a first-scale depth-of-field perspective base image; deriving the three-dimensional absolute embedding depth of internal bubble defects by inverse physical optical path; fusing the three-dimensional absolute embedding depth to generate high-dimensional cross-scale servo micro-motion commands; and outputting a wafer microstructure identification and analysis atlas. This invention solves the image stitching artifact problem caused by internal physical distortion of the wafer and overcomes the technical deficiency of cross-scale positioning difficulties due to the lack of three-dimensional depth information of defects.

Description

Technical Field

[0001] This invention belongs to the technical field of semiconductor testing and relates to a wafer cross-scale panoramic stitching method and system for infrared array cameras. Background Technology

[0002] Comprehensive quality inspection of silicon wafers during semiconductor manufacturing is a crucial step in ensuring chip yield and reliability. As integrated circuit feature sizes continue to shrink, the requirements for detecting microscopic defects within wafers are becoming increasingly stringent. To achieve high-throughput and full-coverage inspection of large-size wafers, the industry typically employs multi-camera or multi-field-of-view scanning and stitching imaging schemes to ensure both inspection accuracy and efficiency. Among these methods, continuous push-broom imaging using a linear scan camera is a common technique for acquiring high-resolution wafer images.

[0003] For example, Chinese Patent Application No. 202511554828.6 discloses a wafer defect detection method, system, and electronic device. This prior art, in the wafer defect detection process, mainly acquires the imaging unit and stage corresponding to the wafer to be inspected, and determines the macroscopic geometric transformation relationship between them based on the camera calibration parameters corresponding to the imaging unit and the reference point parameters corresponding to the stage. Based on this transformation relationship, the acquired strip images are stitched together to obtain a complete wafer image. Subsequently, a variational autoencoder model is used to obtain the defect region of the wafer to be inspected from brightness contrast and structural features.

[0004] The aforementioned existing technologies, when performing image stitching and defect localization, mainly rely on macroscopic geometric coordinate transformations and surface image features, lacking consideration for the complex physical and optical properties inside the wafer. In actual infrared transmission inspection, residual stress fields are easily generated inside the wafer due to growth and processing. These stress fields can cause birefringence in the detection optical path and induce local physical distortions. Traditional stitching methods based on ideal rigidity or affine transformation models struggle to dynamically correct this image point drift caused by the optical properties of the measured object itself, easily introducing artifacts such as misalignment or overlap in the stitching results. Furthermore, the defect information obtained by this method is mainly limited to a two-dimensional plane. When performing high-magnification microscopic verification, secondary localization based solely on two-dimensional coordinates usually requires a lengthy search and focusing process, making it difficult to achieve rapid and accurate focusing on the three-dimensional spatial position of internal defects.

[0005] Therefore, the technical problem to be solved by this invention is how to solve the image stitching artifact problem caused by physical distortion inside the wafer and overcome the difficulty of cross-scale positioning caused by the lack of three-dimensional depth information of defects. Summary of the Invention

[0006] In a first aspect, the present invention provides a wafer-level panoramic stitching method for infrared array cameras, comprising the following steps: Step S1: Receive the basic transmission signal of the target wafer and extract the initial stress data of the target wafer based on physical polarization modulation. Step S2: Analyze the initial stress data to construct a global optical path spatial phase map, and define the global optical path spatial phase map as a global physical optical fingerprint; Step S3: Read the global physical optical fingerprint and extract the transient distortion phase gradient variation rate of the overlapping field of view. Step S4: Convert the transient distortion phase gradient variation rate into the underlying clock control parameter, perform spatiotemporal coherent charge continuous injection and physical stitching, and generate the first-scale depth-of-field perspective base image. Step S5: Analyze the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image, and deduce the three-dimensional absolute embedment depth of the internal bubble defect by inverse physical optical path. Step S6: Extract the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect body and fuse them with the three-dimensional absolute embedment depth to generate high-dimensional cross-scale servo micro-motion commands. Step S7: Drive the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning according to the high-dimensional cross-scale servo micro-motion command, and output the wafer microstructure identification and analysis atlas.

[0007] A further aspect of this invention involves extracting initial stress data from a target wafer using physical polarization modulation, comprising the following steps: A controlled short-wave infrared light source is driven to emit a monochromatic detection beam, which is then shaped into a linear spot covering the scanning area of ​​the target wafer. A polarizing device is used to perform a polarization spatial modulation operation on the linear spot to generate a controlled polarized beam that penetrates the target wafer, which has slight differences in physical thickness and local growth. The basic transmission signal with birefringence interference effect after penetrating the target wafer is acquired at high frequency by a polarization receiver array, and the initial stress data reflecting the microscopic anisotropy of the internal stress field of the target wafer is extracted.

[0008] A further aspect of this invention defines the global optical path spatial phase map as a global physical optical fingerprint, comprising the following steps: Analysis and comparison of the characteristic values ​​of the interference intensity fluctuation of the orthogonal polarization components of the initial stress data under different crystal space grid regions; Substitute the characteristic value of the interference intensity fluctuation of the orthogonal polarization components into the preset elastic optical axis deflection equation to solve the microscopic transmission optical path difference caused by the distortion of the physical quantity of residual stress field distribution in the target wafer. The microscopic transmission optical path difference of the global span is proportionally mapped onto the planar grid system to establish the global optical path spatial phase map with the geometric center of the target wafer as the reference starting point.

[0009] A further aspect of this invention involves extracting the transient distortion phase gradient variation rate of the overlapping field of view, comprising the following steps: The first and second line array cameras are instructed to perform parallel translation and sweeping operations on the target wafer along a preset mechanical planning trajectory. Global physical optical fingerprints are captured within the physical overlap scanning width range of the first and second line-scan cameras; Perform total differential two-dimensional spatial polarization derivative operation on the global physical optical fingerprint residing within the width range of the physical overlap scan, and output the transient distortion phase gradient variation rate.

[0010] A further aspect of the present invention generates a first-scale depth-of-field perspective baseline image, comprising the following steps: The transient distortion phase gradient variation rate is transmitted to the motion control logic drive core of the second linear scan camera via a preset field programmable gate array bypass protocol. The transient distortion phase gradient variation rate is converted into a local phase lead compensation or a local phase lag parameter by calling the charge-coupled device conduction model. The charge transition slip rate of the second linear array camera is controlled by adjusting the local phase advance compensation amount or the local phase lag parameter, and then fused to generate the first-scale depth perspective base image.

[0011] A further aspect of this invention involves deriving the three-dimensional absolute embedment depth of an internal bubble defect through inverse physical optical path derivation, comprising the following steps: Perform grayscale gradient threshold traversal search on the first-scale depth-of-field perspective base image to delineate the local grayscale collapse boundary contour; Extract the difference in the radius of the focal diffusion circle of the forward transmission probe beam that falls at the centroid of the local grayscale collapse boundary contour. According to the preset three-dimensional point diffusion function degradation and broadening mathematical evolution model, the difference in the radius of the caustic surface diffusion circle is inverted into the Z-axis spatial coordinate compensation vector through a blind deconvolution algorithm, and the three-dimensional absolute embedment depth of the internal bubble defect is calculated.

[0012] A further aspect of the present invention involves extracting the difference in the radius of the caustic dispersion circle of the forward transmission probe beam falling at the centroid of the local grayscale collapse boundary contour, comprising the following steps: By fitting the edge intensity profile of the local grayscale collapse boundary contour with a Gaussian error function, the standard deviation characterizing the edge ambiguity is obtained. The standard deviation is compared with the preset system baseline ambiguity standard deviation to obtain the difference in the radius of the caustic surface blurring circle, thus quantifying the additional optical blurring introduced by the internal bubble defect.

[0013] A further aspect of this invention involves fusing three-dimensional absolute embedment depth to generate high-dimensional cross-scale servo micro-motion commands, including the following steps: Calculate the geometric centroid of the local grayscale collapse boundary profile, map the geometric centroid back to the preset global reference physical base frame according to polynomial registration, and extract the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect body. The absolute two-dimensional hash coordinates of the associated reference surface and the three-dimensional absolute embedding depth are stacked to construct a high-dimensional cross-scale servo micro-motion command. An asymmetric hardware communication handshake feedback link is established to pump high-dimensional cross-scale servo micro-motion commands into the second-scale high-magnification verification scanning module.

[0014] A further aspect of this invention involves outputting a set of microscopic solidity determination and identification analysis images for wafers, including the following steps: The high-dimensional cross-scale servo micro-motion command is analyzed to obtain the horizontal axis drive differential signal, and the second-scale high-magnification verification scanning module is linked to make the cross-shaped anchor of the entrance pupil of the re-examination optical microscopy fixed in the physical space directly above the absolute two-dimensional hash coordinates of the associated reference plane. The depth quantization value stored in the high-dimensional cross-scale servo micro-motion command is extracted and injected into the preset micro-displacement piezoelectric ceramic longitudinal constant force actuator, so that the focusing objective lens unit is locked on the three-dimensional absolutely embedded depth-related astigmatic micro static equilibrium focusing surface. A secondary coaxial stimulated emission coherent light source is ignited within the depth-of-field envelope of the astigmatically neutral microscopic static equilibrium focusing surface to capture and analyze the microscopic slice feature system, and then spliced ​​to output a set of images for the identification and analysis of the microscopic solidity of the wafer.

[0015] Secondly, this invention provides a wafer-level panoramic stitching system for infrared array cameras, comprising the following modules: The initial stress data acquisition module receives the basic transmission signal of the target wafer and extracts the initial stress data of the target wafer based on physical polarization modulation. The global optical fingerprint construction module analyzes the initial stress data to construct a global optical path space phase map, which is defined as a global physical optical fingerprint. The distortion phase gradient calculation module reads the global physical optical fingerprint and extracts the transient distortion phase gradient variation rate of the overlapping field of view. The spatiotemporal coherence stitching control module converts the transient distortion phase gradient variation rate into the underlying clock control parameter, performs continuous spatiotemporal coherence charge injection and physical stitching, and generates the first-scale depth-of-field perspective base image. The 3D defect depth calculation module analyzes the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image and derives the 3D absolute embedment depth of the internal bubble defect body by inverse physical optical path. The cross-scale servo command generation module extracts the absolute two-dimensional hash coordinates of the associated reference surface for locating the internal bubble defect and fuses the three-dimensional absolute embedding depth to generate high-dimensional cross-scale servo micro-motion commands. The high-magnification verification scanning execution module drives the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning based on high-dimensional cross-scale servo micro-motion commands, and outputs a set of wafer microstructure identification and analysis images.

[0016] In summary, the present invention has the following beneficial technical effects: 1. During the scanning process, the system calculates the gradient of the optical fingerprint in the overlapping area in real time and converts this gradient into a dynamic adjustment signal for the charge transfer clock of adjacent camera sensors. This method pre-compensates for image point drift caused by optical distortion at the physical level of charge accumulation, reducing the software stitching difficulties caused by the sparse texture or physical distortion of the object being measured, thus providing technical support for generating a base image with good geometric fidelity.

[0017] 2. After obtaining the two-dimensional basic image, the degree of blurring at the defect edge due to optical effects is further analyzed. Based on the preset point spread function model, this scheme correlates the features of the two-dimensional image with the embedding depth of the defect in the Z-axis direction through inverse solving. Compared with methods that require additional tomographic scanning or switching between multiple imaging modes, this scheme can obtain the three-dimensional coordinate information of the defect in a single macroscopic scanning process, which helps to improve the dimensionality of the data obtained and the detection efficiency without increasing the scanning complexity.

[0018] 3. The 3D coordinates of the defects calculated during the macroscopic scanning stage are encapsulated into high-dimensional servo commands. These commands are transmitted to the high-magnification verification module via a dedicated hardware link, directly driving its XY displacement stage and Z-axis focusing mechanism to perform spatial positioning. This feedforward control method based on prior information reduces the time required for secondary searching and autofocus after scale switching, reduces the impact of factors such as mechanical backlash and thermal drift on positioning accuracy, and helps improve the overall process efficiency and positioning accuracy of cross-scale defect verification. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0020] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0021] Figure 2 A schematic diagram of the framework in the embodiments of this application is disclosed. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0023] See attached document Figure 1 This invention proposes a wafer-level panoramic stitching method for infrared array cameras, comprising the following steps: Step S1: Receive the basic transmission signal of the target wafer and extract the initial stress data of the target wafer based on physical polarization modulation. Step S2: Analyze the initial stress data to construct a global optical path spatial phase map, and define the global optical path spatial phase map as a global physical optical fingerprint; Step S3: Read the global physical optical fingerprint and extract the transient distortion phase gradient variation rate of the overlapping field of view. Step S4: Convert the transient distortion phase gradient variation rate into the underlying clock control parameter, perform spatiotemporal coherent charge continuous injection and physical stitching, and generate the first-scale depth-of-field perspective base image. Step S5: Analyze the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image, and deduce the three-dimensional absolute embedment depth of the internal bubble defect by inverse physical optical path. Step S6: Extract the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect body and fuse them with the three-dimensional absolute embedment depth to generate high-dimensional cross-scale servo micro-motion commands. Step S7: Drive the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning according to the high-dimensional cross-scale servo micro-motion command, and output the wafer microstructure identification and analysis atlas.

[0024] In one embodiment of the present invention, step S1 includes the following steps: A controlled short-wave infrared light source is driven to emit a monochromatic detection beam, which is then shaped into a linear spot covering the scanning range of the target wafer. A polarizing device is used to perform a polarization spatial modulation operation on the linear spot to generate a controlled polarized beam that penetrates the target wafer, which has slight differences in physical thickness and local growth. The basic transmission signal with birefringence interference effect after penetrating the target wafer is acquired at high frequency by a polarization receiver module array, and the initial stress data reflecting the microscopic anisotropy of the internal stress field of the target wafer is extracted.

[0025] Specifically, the execution device of the method in this embodiment includes a controlled short-wave infrared light source, which is a laser light source capable of outputting wavelengths covering the range of 900 nm to 1700 nm. Silicon material exhibits high intrinsic transmittance in the near-infrared band with wavelengths greater than 1100 nm, while 900-1100 nm covers the absorption edge of silicon. This wide spectral range can be used to balance deep penetration and the resonant scattering effect of specific defects. The implementation method can be a supercontinuum light source with a built-in acousto-optic tunable filter, a cascaded collimating lens group, a polarizing device, a two-dimensional precision motion platform, a lower-mounted polarization receiver module array, and a main control computer.

[0026] During the steps of acquiring the basic transmission signal of the target wafer and extracting the initial stress data inside the silicon material, the main control computer sends control commands to the drive module of the controlled short-wave infrared light source via the digital acquisition card, driving the supercontinuum laser inside to emit broadband infrared light, which is then split by an acousto-optic tunable filter, outputting a center wavelength of... A monochromatic detection beam, with reference to industrial testing standards and the optical properties of silicon materials, at this wavelength The wavelength is set to 1050 nm, which is close to the indirect bandgap energy edge of single-crystal silicon at room temperature. This value can maximize the phase delay signal-to-noise ratio generated by the photoelastic effect. This wavelength can be efficiently received by standard InGaAs detectors while ensuring sufficient penetration depth into silicon wafers.

[0027] Subsequently, the monochromatic probe beam passes sequentially through a cascaded collimating lens group consisting of a Schwarzschild objective and a cylindrical mirror. This lens group reshapes the beam from a point source to a length of [missing information]. Width is The linear light spot, its length The design values ​​ensure complete coverage of the effective area of ​​a single scan of the target wafer.

[0028] Next, a two-dimensional precision motion platform carries the target wafer, which is exemplarily a wafer with a guide electrical type of P-type and a crystal orientation of [missing information]. <100> A single-crystal silicon wafer with a diameter of 300 mm is subjected to a preset constant linear velocity. Move along the first axis, speed Based on the system's data processing bandwidth and the required scanning resolution, it is set to 200 mm / s. This speed ensures that a full-coverage scan of a 300 mm large-size wafer can be completed within tens of seconds, meeting the high-throughput requirements of industrial production lines and allowing linear light spots to be continuously and orthogonally projected onto the target wafer surface.

[0029] Before the light beam enters the target wafer, a polarizing device, such as a GranThompson prism, is placed in its path, and its transmission direction is adjusted to be perpendicular to the principal crystal axis of the target wafer. The included angle is used to perform spatial polarization modulation on the linear light spot, forming a controlled polarized beam with a uniform polarization state. Because its wavelength is located within the infrared high transmission window of silicon material, this controlled polarized beam can deeply penetrate the target wafer substrate with submicron-level physical thickness differences and local lattice strain.

[0030] Below the target wafer, facing the optical path, is a bottom-mounted polarization analyzer and receiver array. Its front end includes a polarizer configured orthogonally to the polarizing device, and its rear end is a one-dimensional InGaAs linear array detector. This polarization analyzer and receiver array... The frequency synchronous acquisition system collects the interference light intensity signal generated by the birefringence effect caused by the internal stress field after penetrating the target wafer. The analog light intensity signal acquired by each pixel is quantized into a digital grayscale value via an analog-to-digital converter. To ensure sampling accuracy along the scanning direction, the Nyquist theorem is applied in conjunction with the scanning speed. Set in conjunction with the pixel size, typically at 20 kHz, this frequency setting matches the linear velocity of 200 mm / s and the subsequently set pixel size to ensure that the Nyquist spatial sampling rate along the push-broom direction is not distorted.

[0031] The main control computer aggregates the temporal grayscale values ​​of all pixels during the scanning process and, based on the photoelastic effect theory, inversely solves them into a spatially correlated phase delay. This solution process follows a specific form of Malus's law, namely, the intensity of transmitted light collected by the lower polarizer receiver array under orthogonal polarizer configuration. Phase delay caused by the difference in principal stress within the target wafer The following relationship exists between them: In order to obtain the measured light intensity Extracting phase delay The following inverse operation formula needs to be performed: in, coordinates The actual digital value of light intensity captured at this pixel; Under the current system lighting and gain conditions, when the phase delay is exactly... When the light intensity is an integer multiple of the value, it is the maximum digital value of light intensity that the detector can receive, which is usually obtained through a pre-calibrated process; This is the phase delay extracted at that coordinate point, and its unit is radians.

[0032] Through the above process, the system constructs initial stress data in the form of a two-dimensional matrix, reflecting the microscopic anisotropy of the stress field inside the target wafer. This initial stress data is a two-dimensional floating-point matrix, and each element of the matrix... This represents the target wafer in physical coordinates. The position is determined by the phase delay corresponding to the optical path difference between two orthogonally polarized light components caused by internal residual stress, with a value ranging from 0 to... In radians, this numerical range is set based on the fact that the phase retardation induced by typical residual stress in semiconductor manufacturing is usually much smaller than half a wavelength, and is limited to... Within radians, the phase entanglement and unpacking ambiguity problem of sinusoidal functions can be eliminated from a physical perspective.

[0033] For example, suppose a two-dimensional precision motion platform carries a target wafer with a diameter of 300 mm for scanning. The system is pre-calibrated to obtain the maximum detectable light intensity. It has 62,500 least significant bits (LSBs). When scanning to a specific physical location on the target wafer, such as a coordinate point... At that time, the transmitted light signal collected by a pixel unit on the polarization receiver array corresponding to that point is converted from analog to digital to obtain the actual digital value of light intensity. It is 15625 LSB.

[0034] At this point, the main control computer calls the inverse operation formula to calculate the phase delay at that point. The calculation process first calculates the light intensity ratio, i.e. Next, we take the square root of the ratio to obtain... Then the arcsine operation is performed, that is... Radius. Finally, multiply the result by 2 to get... Radius. Therefore, in the generated initial stress data matrix, the coordinates of the points... The corresponding matrix element values ​​are denoted as This process is executed synchronously and in parallel across all pixels within the entire scanning area, ultimately forming a complete phase delay map that characterizes the global stress distribution, i.e., the initial stress data.

[0035] In one embodiment of the present invention, step S2 includes the following steps: The orthogonal polarization component interference intensity fluctuation characteristic values ​​of the initial stress data under different crystal space grid regions are analyzed and compared; the orthogonal polarization component interference intensity fluctuation characteristic values ​​are substituted into the preset elastic optical axis deflection equation to solve the microscopic transmission optical path difference caused by the physical quantity of residual stress field distribution distortion of the target wafer; the global span microscopic transmission optical path difference is mapped proportionally to the planar grid system to establish the global optical path space phase map with the geometric center of the target wafer as the reference starting point.

[0036] Specifically, after extracting the initial stress data, the main control computer continues the analysis process to construct a global optical path space phase map with global coordinate space continuity. The main control computer iterates through the two-dimensional matrix of initial stress data, assigning the phase delay represented by each element in the matrix... As a characteristic value reflecting the intensity fluctuation of interference between orthogonal polarization components in different crystal lattice regions, this characteristic value directly quantifies the intensity of the polarization interference effect caused by stress. Subsequently, based on the phase delay, the main control computer calls the preset elastic optical axis deflection equation. The elastic optical axis deflection equation is a macroscopic mathematical description of the photoelastic effect, establishing the physical relationship between stress, refractive index change, optical path difference, and the final phase delay for each coordinate point. phase delay value The solution is performed to transform it into the microscopic transmission optical path difference caused by the distortion of the residual stress field distribution inside the target wafer. .

[0037] This solution process essentially restores the phase information of light propagating in a medium to the difference in physical path length. Its core calculation follows the basic physical relationship of light wave propagation, which is specified by the following expression: in, Coordinates within the target wafer The difference in actual geometric path length accumulated between two orthogonally polarized light components due to stress birefringence is called the microscopic transmission path difference, which is measured in nm and typically fluctuates between 0 nm and 500 nm. It is the center wavelength of the monochromatic probe beam; The phase delay is obtained from the initial stress data, and the unit is radians. After completing the calculation for all coordinate points, the system obtains a two-dimensional data map with the same size as the initial stress data, storing all microscopic transmission path differences within the global span.

[0038] The main control computer performs coordinate system mapping operations, determining the geometric center of the target wafer through edge recognition algorithms or mechanical reference points, and establishing a Cartesian plane grid system as the origin of the global reference coordinate system based on this center. The two-dimensional coordinate system has a resolution corresponding to the physical pixel size of the polarization receiver array, such as 14 µm. This size is the typical physical and technological limit of current industrial-grade high-resolution short-wave infrared linear array sensors, which can balance the spatial resolution of photon collection trap capacity and microscopic defects.

[0039] Then, the global microscopic transmission optical path difference data calculated in the previous step is mapped point-by-point proportionally onto the planar grid system according to its corresponding physical location at the time of acquisition, thereby establishing a global optical path spatial phase map with the geometric center of the target wafer as the reference starting point. Due to its inherent physical uniqueness and global continuity, this map is defined as a hardware-layer anchored global physical optical fingerprint, used for subsequent scanning and stitching calibration. This global optical path spatial phase map is ultimately generated as a two-dimensional array, where the index corresponds to the coordinate in the planar grid system, and the value is the microscopic transmission optical path difference at that coordinate. .

[0040] For example, continuing the previous steps, the master computer receives a data set containing coordinate points. Phase delay The initial stress data. The system uses the center wavelength of the monochromatic probe beam set in step S1. nm. To calculate The microscopic transmission optical path difference at the point is calculated by the main control computer using the formula: Through reduction calculations, we obtain... nm.

[0041] Subsequently, the system performs Hough circle transform on the wafer profile image to locate its geometric center, and uses this as the global coordinate origin. (Assuming a point...) The physical location relative to the center is defined as The calculated optical path difference of 175 nm will then be stored in the global optical path spatial phase map and its coordinates. On the corresponding grid nodes. This process is repeated for each point in the initial stress data, and finally the phase delay distribution map of the entire wafer is transformed into a global optical path space phase map with an absolute coordinate reference, expressed in physical units of nm, i.e., a global physical optical fingerprint.

[0042] In one embodiment of the present invention, step S3 includes the following steps: The system instructs the first and second linear array cameras to perform parallel translation and scanning operations on the target wafer along a preset mechanical planning trajectory; it captures the global physical optical fingerprint within the physical overlap scanning width range of the first and second linear array cameras; it performs a total differential two-dimensional spatial polarization derivative operation on the global physical optical fingerprint residing within the physical overlap scanning width range, and outputs the transient distortion phase gradient variation rate.

[0043] Specifically, after generating the global physical optical fingerprint and loading it into memory in the aforementioned steps, the main control computer switches to the first resolution scanning mode, which uses a scanning configuration with a pixel resolution of 7 µm / pixel to perform line tracking and gradient extraction. The main control computer issues commands via the motion control bus to drive the two-dimensional precision motion platform at a constant speed. The target wafer moves along a pre-planned "N"-shaped mechanical trajectory.

[0044] At the same time, the first and second line scan cameras, which are rigidly connected to the same gantry bracket, are activated simultaneously. Both the first and second line scan cameras are 4096-pixel InGaAs line scan cameras, and they are deployed side by side with a center-to-center distance of 21.504 mm between their scan lines. This creates a physical overlap scan width of 7.168 mm, corresponding to 1024 pixels, and performs continuous parallel translational scanning on the target wafer according to the first resolution scale. The first and second linear scan cameras are physically calibrated to ensure their scan lines are strictly parallel, forming a fixed-width physical overlap scan interval at the edge of their field of view. During scanning, the main control computer synchronously extracts and reads the optical path difference data corresponding to the position of each physical pixel within the physical overlap scan width interval from memory based on the real-time physical coordinates fed back by the motion encoder.

[0045] Next, the system performs a total differential two-dimensional spatial polarization derivative operation on the subset of local optical path difference data residing within the physically overlapping scan width acquired at each dynamic scan moment. This operation is an algorithm for real-time calculation of local gradient field intensity, used to quantify the drastic spatial variation of the optical path difference. In engineering practice, it is implemented by convolving a two-dimensional gradient operator, such as the Sobel operator, with the subset of optical path difference data to calculate two spatial partial derivatives along and perpendicular to the scan direction. Then, by vector synthesis of these two partial derivative components and calculation of their magnitude, the system instantly outputs a scalar value as the transient distortion phase gradient variation rate. Its mathematical expression is the magnitude of the gradient vector: In discrete digital image processing, the formula for approximating component calculation using the central difference method is: in, For gradient operators, Represents partial derivatives; and These are the gradient components along the x-axis and y-axis, respectively; This is the optical path difference between the corresponding coordinate points read from the global physical optical fingerprint. In this discrete difference formula, The physical absolute coordinates of the current center pixel are represented by the addition and subtraction of the independent variable 'p' in the numerator expression, which represents the optical path difference between the centers of the adjacent upper or lower physical pixels along the corresponding direction in the grid system. p represents the physical pixel size of the line scan camera. This is an unsigned floating-point number. The larger the value, the more drastic the internal stress changes in the scanned area, and the higher the potential for optical distortion. The system continuously outputs and updates this value to reflect changes in the scan position.

[0046] For example, suppose the physical pixel size of the first line scan camera and the second line scan camera are... All are 7 µm. At a certain scan time, the system captures a portion of the physical overlap scan width with a specific physical image point. This is the global physical optical fingerprint data corresponding to a 3x3 pixel neighborhood centered on the image. The data matrix is ​​in nm units, and the specific values ​​are as follows: Among them, the optical path difference of the center pixel The value is 182 nm. The system uses the central difference method to calculate the gradient components at this point. The gradient components along the x-axis are... The gradient component along the y-axis is: Subsequently, the system calculates the transient distortion phase gradient variation rate. Therefore, when passing over that physical image point At that instant, the transient distortion phase gradient change rate output by the system was 0.833 nm / µm, which will serve as the direct control basis for the continuous injection of spatiotemporal coherent charge in subsequent steps.

[0047] In one embodiment of the present invention, step S4 includes the following steps: The transient distortion phase gradient variation rate is transmitted to the motion control logic drive core of the second linear array camera via a preset field-programmable gate array bypass protocol; the charge coupler conduction model is called to convert the transient distortion phase gradient variation rate into a local phase advance compensation amount or a local phase lag parameter; the local phase advance compensation amount or local phase lag parameter is used to adjust the charge transition slip rate of the second linear array camera, and the images are fused to generate the first-scale depth perspective base image.

[0048] Specifically, during the parallel pushbroom process of the first and second line scan cameras, the field-programmable gate array (FPGA) in the main control computer receives the transient distortion phase gradient change rate calculated in real time by the aforementioned steps. To ensure minimal transmission latency, this transient distortion phase gradient change rate data stream is stripped from the conventional CameraLink camera link or CoaXPress high-speed coaxial vision bus image bus and bypassed via a dedicated FPGA bypass protocol based on low-voltage differential signaling (LVDS). This bypass protocol is a low-latency hardware communication link that does not rely on the operating system or standard bus protocols. Typically, several pairs of LVDS cables are used to directly connect the FPGA control boards of the two cameras, achieving microsecond-level data transmission and directly transmitting it point-to-point to the motion control logic drive core of the second line scan camera.

[0049] The second linear array camera's drive core contains a pre-defined charge-coupled device (CCD) conduction model. This model is a pre-calibrated conversion relationship based on optical design and sensor physical characteristics, expressed as a linear conversion function or a multidimensional lookup table. It maps the optical spatial distortion gradient to the time-domain compensation required for charge transfer. When a transient distortion phase gradient change rate is received... Subsequently, the model immediately converts its numerical equivalent into an adjustment amount for the charge transfer clock frequency specifically used to intervene in the time delay integral (TDI) external trigger action cycle. The transformation relationship, in the simplified linear model, is represented by the frequency adjustment amount. With transient distortion phase gradient variation rate Proportional: in, This is the conversion coefficient of the conduction model obtained from system calibration. According to the calculation results, if this adjustment is positive, it is defined as the local phase advance compensation, which means that the clock frequency needs to be increased to accelerate charge transfer and catch up with the image point that arrives "earlier" due to refraction; if it is negative, it is defined as the local phase lag parameter, which means that the clock frequency needs to be reduced to slow down charge transfer and wait for the image point that arrives "delayed" due to refraction.

[0050] Subsequently, a local phase lead compensation or local phase lag parameter is dynamically applied to the clock frequency synthesizer of the second linear array camera, causing the actual charge transfer clock frequency of the second linear array camera to change. By its nominal frequency This is combined with the adjustment amount: in, The unit is Hz; It is the nominal clock frequency that matches the mechanical scanning speed; It is the scanning speed of a two-dimensional precision motion platform; This refers to the physical pixel size of the sensor. By adjusting the aforementioned clock frequency, the transition and sliding rate of the charge packets within the TDI sensor in the vertical transfer register is controlled pixel by pixel in real time.

[0051] The purpose of this approach is to ensure that the step vector of charge accumulation and movement on the sensor is strictly synchronized with the displacement of the physical image point on the sensor plane caused by stress refraction within the target wafer on the time axis, both in velocity and direction. Through this underlying physical-level synchronization, signals acquired by the two cameras within the physically overlapping scan width can be seamlessly fused at the charge level, ultimately generating a seamless, color-aberration-free, continuous infrared transmission first-scale depth-of-field perspective image covering the entire target wafer with a resolution of 7 µm / pixel.

[0052] For example, continuing from the previous step, the system obtains physical image points. Transient distortion phase gradient variation rate at The system parameters in this embodiment are: scanning speed. pixel size Calculate the nominal charge transfer clock frequency. .

[0053] Before leaving the factory, the system calibrates the conversion coefficients of the charge-coupled device (CCD) conduction model by scanning a standard etalon with a known gradient change. At this moment, the FPGA of the second linear array camera receives a gradient value of 0.833 nm / µm and immediately calculates the frequency adjustment: Since this value is positive, it is determined to be the local phase lead compensation amount.

[0054] Ultimately, the dynamically adjusted clock frequency applied to the TDI sensor is: The adjusted frequency signal is sent to the clock drive circuit, so that in The charge transfer rate of the second linear array camera matches the image point drift caused by the internal stress of the wafer, thereby achieving physical seamless stitching with the data from the first linear array camera, together forming part of the basic image of the first-scale depth-of-field perspective.

[0055] In one embodiment of the present invention, step S5 includes the following steps: A grayscale gradient threshold traversal search is performed on the first-scale depth-of-field perspective base image to delineate the local grayscale collapse boundary contour; the difference in the radius of the caustic diffusion circle of the forward transmission probe beam falling at the centroid of the local grayscale collapse boundary contour is extracted; according to the preset three-dimensional point spread function degradation broadening mathematical evolution model, the difference in the radius of the caustic diffusion circle is inverted into the Z-axis spatial coordinate compensation vector through a blind deconvolution algorithm, and the three-dimensional absolute embedment depth of the internal bubble defect is calculated.

[0056] Specifically, upon receiving the first-scale depth-of-field perspective baseline image generated in the previous step, the image analysis engine within the main control computer begins scanning it to analyze the radiometric optical degradation pathological features present in the image. The engine performs a grayscale gradient threshold traversal search on all array pixels contained within the first-scale depth-of-field perspective baseline image. This process iteratively sets thresholds from the low end to the high end of the grayscale histogram, identifying all grayscale values ​​below a preset collapse threshold. pixels, collapse threshold It is based on the gray-level average of defect-free regions in the first-scale depth-of-field perspective baseline image. and standard deviation Dynamically set, usually taken This ensures high-confidence identification of anomalous dark areas. These pixels correspond to attenuation bands formed by total internal reflection and severe scattering of infrared light due to internal bubbles or void structures.

[0057] The system uses a connected component labeling algorithm to cluster these low grayscale pixels, and applies a contour tracing algorithm to the outer boundary of each cluster to close the loop and define the local grayscale collapse boundary contour of each potential defect. This contour is an ordered set of two-dimensional pixel coordinates used to describe the projection shape of the defect on the XY plane.

[0058] For each defined local grayscale collapse boundary contour, the system calculates its geometric centroid position. Then, using this centroid position as the center, an image sub-block containing the defect and its edge region is extracted. Within this sub-block, the system analyzes the difference in the radius of the caustic blurring circle caused by refraction and diffraction effects induced by lattice discontinuities as the forward transmission probe beam passes through the defect boundary. This difference is extracted by fitting a Gaussian error function to the intensity profile of the defect edge, obtaining the standard deviation characterizing the edge ambiguity. And compare it with the standard deviation of the system baseline ambiguity measured from the adjacent defect-free area. The difference in the radius of the dispersion circle, obtained through comparison, is quantified in engineering as the difference in standard deviations obtained from Gaussian fitting. This reflects the additional degree of optical blurring introduced by the defect.

[0059] The system invokes a pre-loaded mathematical evolution model of the degradation and broadening of the three-dimensional point spread function within a polycrystalline silicon dielectric. This model is a database or function model pre-established through optical simulation software or experimental measurements on calibration pieces with defects of known depth. This describes the functional relationship between the radius of the dispersion circle and the embedment depth of the defect. By using the difference in the radius of the dispersion circle extracted in the preceding steps as input, and executing a blind deconvolution algorithm based on this model, such as the Richardson-Lucy algorithm, the system inversely retrieves a unique Z-axis spatial coordinate compensation vector. The value of this vector is the three-dimensional absolute embedment depth of the internal bubble defect, i.e., the vertical distance from the upper surface of the target wafer to the geometric center of the identified defect. The core of this inverse inversion lies in solving an inverse optimization problem, namely, finding the depth value. This makes the point spread function model at this depth... Compared to an ideal image without fuzziness or defects The result obtained after convolution is compared with the actual observed defect image sub-blocks. The optimization process, which minimizes the difference between the two, is expressed by the following equation: in, It is the three-dimensional absolute embedment depth to be solved; This indicates the search for parameters that minimize the objective function. ; It is an image sub-block extracted from the first-scale depth-of-field perspective base image, surrounding the centroid of the defect; It is an idealized, unblurred defect image model, such as a binary mask with the same shape as the boundary contour of the local grayscale collapse; It is a three-dimensional point spread function degenerate broadening mathematical evolution model in depth The specific function form at the location; Represents a two-dimensional convolution operation; This represents the sum of squared L2 norms of the differences between all pixels in the image.

[0060] For example, suppose the main control computer analyzes the first-scale depth-of-field perspective baseline image. Statistically, the average grayscale value of the defect-free region... 52,000 LSB, standard deviation The collapse threshold is set at 400 LSB. LSB. The system retrieved pixels in the region whose grayscale values ​​were generally below the collapse threshold and successfully delineated the local grayscale collapse boundary contour. After calculating its centroid, the system analyzed the intensity profile of the defect edge and fitted the standard deviation of its edge blur. pixels. Meanwhile, the baseline ambiguity of the system, measured from a neighboring uniform background region, is [value missing]. pixels.

[0061] Therefore, the calculated difference in the radius of the dispersion circle is pixels. Considering a resolution of 7 µm / pixel, this difference translates to a physical size of... µm. The system calls a pre-loaded mathematical evolution model, which simplifies to a linear relationship. Among them, the calibration coefficient Solving by reverse engineering ,get Ultimately, the system outputs that the three-dimensional absolute embedment depth of the internal bubble defect is 700 µm.

[0062] In one embodiment of the present invention, step S6 includes the following steps: The geometric centroid of the local grayscale collapse boundary contour is calculated, and the geometric centroid is mapped back to the preset global reference physical base frame according to polynomial registration. The absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect are extracted. The absolute two-dimensional hash coordinates of the associated reference surface and the three-dimensional absolute embedment depth are stacked to form a high-dimensional cross-scale servo micro-motion command. An asymmetric hardware communication handshake feedback link is built to pump the high-dimensional cross-scale servo micro-motion command into the second-scale high-magnification verification scanning module.

[0063] Specifically, after the main control computer completes the three-dimensional absolute embedment depth calculation of a single internal bubble defect, it immediately initiates the process of extracting the coordinates of the associated reference surface and generating cross-scale instructions. First, the system reverse-engineers the local grayscale collapse boundary contour determined in step S5, which is composed of a series of pixel coordinate points. By calculating the arithmetic mean of the x and y components of these coordinate points, the system calculates the coordinates of the geometric centroid of the contour in the first-scale depth-of-field perspective base image coordinate system. The calculation formula is as follows: in, It is the first one that constitutes the boundary contour of the local grayscale collapse. The coordinates of each pixel. This represents the total number of pixels in the contour. The system calls a polynomial registration model pre-calibrated using a standard calibration plate, performs a nonlinear transformation on the pixel coordinates of the geometric centroid, and re-maps them precisely back to the source device, i.e., the global reference physical framework of the two-dimensional precision motion platform. This allows the extraction of the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect. This coordinate system consists of a high-precision coordinate pair containing two double-precision floating-point numbers, used to uniquely specify the target location point. The underlying second-order polynomial registration model can be expressed as: in, , The absolute two-dimensional hash coordinates of the associated reference surface; , The coordinates of the geometric centroid of the boundary contour of the local grayscale collapse; and The polynomial coefficients are predetermined. This transformation maps the unit from pixels to millimeters (mm). Subsequently, the main control computer performs a stacking construction operation of heterogeneous multidimensional information tensor data packets, that is, it integrates the absolute two-dimensional hash coordinates of the associated reference surface extracted from the planar topology and represented in floating-point form, containing two components, X and Y, with the three-dimensional absolute embedment depth Z component obtained from spatial perspective deduction and represented in integer form, and adds metadata such as instruction header, defect unique identifier, and checksum, to jointly construct a high-dimensional cross-scale servo micro-motion instruction with a fixed byte length and carrying prior information of the cross-precision target spatial positioning origin. .

[0064] The instruction is constructed as a structured vector: in, It is the instruction header identifier; It is a unique identifier for defects; It is the three-dimensional absolute embedment depth; It is the checksum of the data packet; This represents vector transpose. The micro-motion instruction is a 32-byte binary data packet, which, as an information set, satisfies the instruction communication structure characteristics. Finally, the main control computer, through its FPGA coprocessor, grants access to the local area bus and sends a wake-up request signal to the second-scale high-magnification verification scanning module, which is in a dormant standby configuration. This scanning module is a precision optical detection unit integrating a high numerical aperture microscope objective, a high-resolution area array camera, and a piezoelectric ceramic Z-axis micro-motion platform.

[0065] Upon receiving the readiness confirmation signal from the module, a dedicated asymmetric hardware communication handshake feedback link, based on the LVDS physical layer and a custom serial protocol, is established with a data transmission rate of up to 1.25 Gbps. The FPGA coprocessor immediately uses this link, in memory access mode, to force the assembled high-dimensional cross-scale servo micro-motion instructions to be pre-pumped into the onboard cache of the second-scale high-magnification verification scanning module as a high-speed serial data stream.

[0066] For example, continuing from the previous step, the system has determined the three-dimensional absolute embedment depth of an internal bubble defect. µm. The system further obtained the local grayscale collapse boundary contour of the defect, which is composed of a set of pixels. For the sake of simplification, it is assumed to be a rectangle with four vertices at coordinates (1020, 550), (1024, 550), (1024, 554), and (1020, 554).

[0067] First, calculate its geometric centroid: pixels; pixels. The system calls a simplified affine transformation, such as a first-order polynomial, as the registration model, with coefficients as follows: mm / pixel mm, mm / pixel mm, with all other coefficients set to zero. Substitute the centroid coordinates for calculation: mm; mm.

[0068] The absolute two-dimensional hash coordinates of the associated reference surface are thus obtained as (57.154 mm, 123.864 mm). The system then packages this information into a high-dimensional, multi-scale servo micro-motion command. Assuming the command header is 0xABCD, the defect identifier ID is 0x0001, and the checksum is calculated to be 0xFAFBFCFDFEFF, the data packet is assembled as [0xABCD, 0x0001, 57.154, 123.864, 700, 0xFAFBFCFDFEFF]. This data packet is rapidly pumped into the FPGA of the second-scale high-magnification verification scanning module, which is already awake and awaiting commands, via an asymmetric hardware communication handshake feedback link, preparing data for its next step of precise positioning and focusing.

[0069] In one embodiment of the present invention, step S7 includes the following steps: The high-dimensional cross-scale servo micro-motion command is analyzed to obtain the horizontal axis drive differential signal. This signal is then linked to the second-scale high-magnification verification scanning module to anchor the crosshair of the entrance pupil of the re-examination optical microscope directly above the physical space of the absolute two-dimensional hash coordinates of the associated reference plane. The depth quantization value stored in the high-dimensional cross-scale servo micro-motion command is extracted and injected into the preset micro-displacement piezoelectric ceramic longitudinal constant force actuator, which locks the focusing objective lens unit onto the three-dimensional absolutely embedded depth-correlated astigmatic microstatic equilibrium focusing surface. Within the depth envelope of the astigmatic microstatic equilibrium focusing surface, a secondary coaxial stimulated emission detection coherent light source is ignited to capture and analyze the microscopic slice feature system, and then stitched together to output a wafer microscopic solidity identification and analysis atlas.

[0070] Specifically, upon receiving the high-dimensional cross-scale servo micro-motion command pumped in by step S6, the airborne control system of the second-scale high-magnification verification scanning module immediately begins parsing the command data packet. The control system extracts the absolute two-dimensional hash coordinates of the associated reference surface as the target position from the command and inputs them as setpoints to the closed-loop motion controller of the base plane load screw. This controller compares the target coordinates with the current platform position fed back by the grating ruler in real time, generating a horizontal axis drive differential signal (this signal is the pulse frequency or analog voltage signal output by the servo controller; its magnitude and direction depend on the position difference, and its core relies on the PID proportional-integral-derivative closed-loop control algorithm. Taking the X-axis as an example, the generated differential signal...). The formula for calculation is: in, The current time variable; The time variable is the integral time variable; error ; The target coordinates; Real-time location; These are the proportional, integral, and derivative gain coefficients, respectively. The control system uses these signals to drive the servo motors on the X and Y axes at high speed.

[0071] The motor drives the high-precision ball screw to force the base of the second-scale high-magnification verification scanning module until the center line of the crosshair of the entrance pupil of the optical microscope is precisely aligned and anchored in physical space with the absolute two-dimensional hash coordinates of the associated reference surface, and the position error is less than the preset stability threshold.

[0072] Simultaneously, the control system intercepts and releases the depth quantization value in micrometers, i.e., the three-dimensional absolute embedment depth, stored in the high-dimensional cross-scale servo micro-motion command package. This depth value is directly used as the preset bias parameter for the absolute zero point of the motion origin, converted into a control voltage via a digital-to-analog converter, bypassing all global automatic search and focusing processes, and injected into the micro-displacement piezoelectric ceramic longitudinal constant force actuator controlling the focusing objective unit. The target displacement of this micro-displacement actuator is the three-dimensional absolute embedment depth. That is, the three-dimensional absolute embedment depth and the injected control voltage The following linear relationship exists between them: in, The injected control voltage; This is the voltage-displacement conversion coefficient of the piezoelectric ceramic actuator. The actuator responds to the voltage signal, driving the focusing objective unit to complete a single-step levitation transition, and utilizes its internal displacement sensor feedback to achieve silent hard locking, ensuring its focus is precisely stabilized on the astigmatic-free microscopic static equilibrium focusing plane uniquely associated with the three-dimensional absolute embedment depth. This focusing plane refers to the focal plane where the image is clearest and has the least aberration when the objective lens focus falls precisely on the Z-plane where the geometric center of the defect body is located.

[0073] After locking the XYZ axes, the system briefly ignites a secondary coaxial stimulated emission probe coherent light source, such as a 980 nm confocal laser, within the depth-of-field envelope of the controlled, astigmatic-free, micro-static equilibrium focusing surface. The choice of 980 nm is to achieve a smaller Rayleigh diffraction-limited spot while maintaining silicon transmittance, thereby improving lateral resolution. Using an ultra-shallow microstepping method, for example with a step size of 0.2 µm (smaller than the depth of focus of a high numerical aperture objective), satisfying the axial oversampling theorem for 3D reconstruction, the system scans within a 20 µm range above and below the embedment depth, driving the piezoelectric ceramic actuator to perform a series of fine axial scans. This 20 µm range covers and encompasses the theoretical maximum solution error of the front-end Z-axis blind zone convolution algorithm, ensuring that the physical entity of the defect falls within the scanning envelope. Simultaneously, it captures a series of high-frequency analytical micro-slice features without axial dispersive halos at extremely microscopic limits.

[0074] Finally, the system performs array stacking and three-dimensional reconstruction of all captured microscopic slice feature images, and splices them to output a wafer microscopic substance identification and analysis atlas that spans macroscopic push-broom and microscopic tomography and achieves causal verification. Essentially, it is a set of data containing a large number of high-resolution two-dimensional images, which is used to achieve the final defect nature determination through visual reconstruction.

[0075] For example, the controller of the second-scale high-magnification verification scanning module receives a high-dimensional cross-scale servo micro-motion command generated by S6, the content of which is [0xABCD, 0x0001, 57.154, 123.864, 700, 0xFAFBFCFDFEFF]. The controller immediately resolves the target's physical coordinates as follows: mm, mm. Assuming the current platform position is (57.100 mm, 123.800 mm), the PID controller generates the maximum drive signal based on initial errors of 0.054 mm and 0.064 mm, driving the XY-axis stage to move at high speed. Positioning is complete when the position reading fed back by the grating ruler enters and stabilizes within the tolerance window of (57.154 ± 0.0001 mm, 123.864 ± 0.0001 mm).

[0076] At the same time, the controller parses the depth value µm. The micro-displacement piezoelectric ceramic longitudinal constant force actuator of this module has been calibrated, and its voltage-displacement conversion coefficient... µm / V. Therefore, the controller calculates the required injection voltage as follows: V. This voltage is applied to the piezoelectric ceramic, driving the objective lens from its initial position to a depth plane of 700 µm and locking it in.

[0077] After triaxial positioning, the system triggers a 980 nm confocal laser and scans from 680 µm to 720 µm in 0.2 µm steps, acquiring a 1280x1024 pixel microscopic slice image at each step. After scanning, the system packages the acquired 201 images to generate a wafer microstructure identification and analysis atlas, which is then sent to the defect analysis system for final confirmation, thus completing the full closed loop.

[0078] See appendix Figure 2 The present invention also proposes a wafer-scale panoramic stitching system for infrared array cameras, comprising the following modules: The initial stress data acquisition module receives the basic transmission signal of the target wafer and extracts the initial stress data of the target wafer based on physical polarization modulation. The global optical fingerprint construction module analyzes the initial stress data to construct a global optical path space phase map, which is defined as a global physical optical fingerprint. The distortion phase gradient calculation module reads the global physical optical fingerprint and extracts the transient distortion phase gradient variation rate of the overlapping field of view. The spatiotemporal coherence stitching control module converts the transient distortion phase gradient variation rate into the underlying clock control parameter, performs continuous spatiotemporal coherence charge injection and physical stitching, and generates the first-scale depth-of-field perspective base image. The 3D defect depth calculation module analyzes the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image and derives the 3D absolute embedment depth of the internal bubble defect body by inverse physical optical path. The cross-scale servo command generation module extracts the absolute two-dimensional hash coordinates of the associated reference surface for locating the internal bubble defect and fuses the three-dimensional absolute embedding depth to generate high-dimensional cross-scale servo micro-motion commands. The high-magnification verification scanning execution module drives the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning based on high-dimensional cross-scale servo micro-motion commands, and outputs a set of wafer microstructure identification and analysis images.

[0079] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A wafer cross-scale panoramic stitching method for an infrared array camera, characterized in that, Includes the following steps: Step S1: Receive the basic transmission signal of the target wafer and extract the initial stress data of the target wafer based on physical polarization modulation. Step S2: Analyze the initial stress data to construct a global optical path spatial phase map, and define the global optical path spatial phase map as a global physical optical fingerprint; Step S3: Read the global physical optical fingerprint and extract the transient distortion phase gradient variation rate of the overlapping field of view. Step S4: Convert the transient distortion phase gradient variation rate into the underlying clock control parameter, perform spatiotemporal coherent charge continuous injection and physical stitching, and generate the first-scale depth-of-field perspective base image. Step S5: Analyze the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image, and deduce the three-dimensional absolute embedment depth of the internal bubble defect by inverse physical optical path. Step S6: Extract the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect body and fuse them with the three-dimensional absolute embedment depth to generate high-dimensional cross-scale servo micro-motion commands. Step S7: Drive the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning according to the high-dimensional cross-scale servo micro-motion command, and output the wafer microstructure identification and analysis atlas.

2. The wafer cross-scale panoramic stitching method of an infrared array camera according to claim 1, wherein, The initial stress data of the target wafer is extracted using physical polarization modulation, including the following steps: A controlled short-wave infrared light source is driven to emit a monochromatic detection beam, which is then shaped into a linear spot covering the scanning range of the target wafer. By using a polarizing device to perform a polarization spatial modulation operation on a linear light spot, a controlled polarized beam is generated to penetrate the target wafer with slight differences in physical thickness and local growth. The basic transmission signal with birefringence interference effect after penetrating the target wafer is acquired at high frequency by a polarization receiver module array, and the initial stress data reflecting the microscopic anisotropy of the stress field inside the target wafer is extracted.

3. The wafer cross-scale panoramic stitching method of an infrared array camera according to claim 1, wherein, Defining the global optical path spatial phase map as a global physical optical fingerprint includes the following steps: Analysis and comparison of the characteristic values ​​of the interference intensity fluctuation of the orthogonal polarization components of the initial stress data under different crystal space grid regions; Substitute the characteristic value of the interference intensity fluctuation of the orthogonal polarization components into the preset elastic optical axis deflection equation to solve the microscopic transmission optical path difference caused by the distortion of the physical quantity of residual stress field distribution in the target wafer. The microscopic transmission optical path difference of the global span is proportionally mapped onto the planar grid system to establish the global optical path spatial phase map with the geometric center of the target wafer as the reference starting point.

4. The wafer-scale panoramic stitching method for an infrared array camera according to claim 1, characterized in that, Extracting the transient distortion phase gradient variation rate of the overlapping field of view beyond the boundary includes the following steps: The first and second line array cameras are instructed to perform parallel translation and sweeping operations on the target wafer along a preset mechanical planning trajectory. Global physical optical fingerprints are captured within the physical overlap scanning width range of the first and second line-scan cameras; Perform total differential two-dimensional spatial polarization derivative operation on the global physical optical fingerprint residing within the width range of the physical overlap scan, and output the transient distortion phase gradient variation rate.

5. The wafer-scale panoramic stitching method for an infrared array camera according to claim 1, characterized in that, Generating a first-scale depth-of-field perspective base image includes the following steps: The transient distortion phase gradient variation rate is transmitted to the motion control logic drive core of the second linear scan camera via a preset field programmable gate array bypass protocol. The transient distortion phase gradient variation rate is converted into a local phase lead compensation or a local phase lag parameter by calling the charge-coupled device conduction model. The charge transition slip rate of the second linear array camera is controlled by adjusting the local phase advance compensation amount or the local phase lag parameter, and then fused to generate the first-scale depth perspective base image.

6. The wafer-scale panoramic stitching method for an infrared array camera according to claim 1, characterized in that, The three-dimensional absolute embedment depth of the internal bubble defect is derived by inverse physics optical path, including the following steps: Perform grayscale gradient threshold traversal search on the first-scale depth-of-field perspective base image to delineate the local grayscale collapse boundary contour; Extract the difference in the radius of the focal diffusion circle of the forward transmission probe beam that falls at the centroid of the local grayscale collapse boundary contour. According to the preset three-dimensional point diffusion function degradation and broadening mathematical evolution model, the difference in the radius of the caustic surface diffusion circle is inverted into the Z-axis spatial coordinate compensation vector through a blind deconvolution algorithm, and the three-dimensional absolute embedment depth of the internal bubble defect is calculated.

7. A wafer-scale panoramic stitching method for an infrared array camera according to claim 6, characterized in that, Extracting the difference in the radius of the caustic circle of confusion of the forward transmission probe beam falling at the centroid of the local grayscale collapse boundary profile includes the following steps: By fitting the edge intensity profile of the local grayscale collapse boundary contour with a Gaussian error function, the standard deviation characterizing the edge ambiguity is obtained. The standard deviation is compared with the preset system baseline ambiguity standard deviation to obtain the difference in the radius of the caustic surface blurring circle, thus quantifying the additional optical blurring introduced by the internal bubble defect.

8. The wafer-scale panoramic stitching method for an infrared array camera according to claim 1, characterized in that, The process of generating high-dimensional, cross-scale servo micro-motion commands by integrating three-dimensional absolute embedment depth includes the following steps: Calculate the geometric centroid of the local grayscale collapse boundary profile, map the geometric centroid back to the preset global reference physical base frame according to polynomial registration, and extract the absolute two-dimensional hash coordinates of the associated reference surface of the internal bubble defect body. The absolute two-dimensional hash coordinates of the associated reference surface and the three-dimensional absolute embedding depth are stacked to construct a high-dimensional cross-scale servo micro-motion command. An asymmetric hardware communication handshake feedback link is established to pump high-dimensional cross-scale servo micro-motion commands into the second-scale high-magnification verification scanning module.

9. A wafer-scale panoramic stitching method for an infrared array camera according to claim 1, characterized in that, Output a set of microstructure identification and analysis images for wafers, including the following steps: The high-dimensional cross-scale servo micro-motion command is analyzed to obtain the horizontal axis drive differential signal, and the second-scale high-magnification verification scanning module is linked to make the cross-shaped anchor of the entrance pupil of the re-examination optical microscopy fixed in the physical space directly above the absolute two-dimensional hash coordinates of the associated reference plane. The depth quantization value stored in the high-dimensional cross-scale servo micro-motion command is extracted and injected into the preset micro-displacement piezoelectric ceramic longitudinal constant force actuator, so that the focusing objective lens unit is locked on the three-dimensional absolutely embedded depth-related astigmatic micro static equilibrium focusing surface. A secondary coaxial stimulated emission coherent light source is ignited within the depth-of-field envelope of the astigmatically neutral microscopic static equilibrium focusing surface to capture and analyze the microscopic slice feature system, and then spliced ​​to output a set of images for the identification and analysis of the microscopic solidity of the wafer.

10. A wafer-scale panoramic stitching system for an infrared array camera, characterized in that, Includes the following modules: The initial stress data acquisition module receives the basic transmission signal of the target wafer and extracts the initial stress data of the target wafer based on physical polarization modulation. The global optical fingerprint construction module analyzes the initial stress data to construct a global optical path space phase map, which is defined as a global physical optical fingerprint. The distortion phase gradient calculation module reads the global physical optical fingerprint and extracts the transient distortion phase gradient variation rate of the overlapping field of view. The spatiotemporal coherence stitching control module converts the transient distortion phase gradient variation rate into the underlying clock control parameter, performs continuous spatiotemporal coherence charge injection and physical stitching, and generates the first-scale depth-of-field perspective base image. The 3D defect depth calculation module analyzes the radiometric optical degradation pathological characteristics of the first-scale depth-of-field perspective base image and derives the 3D absolute embedment depth of the internal bubble defect body by inverse physical optical path. The cross-scale servo command generation module extracts the absolute two-dimensional hash coordinates of the associated reference surface for locating the internal bubble defect and fuses the three-dimensional absolute embedding depth to generate high-dimensional cross-scale servo micro-motion commands. The high-magnification verification scanning execution module drives the preset second-scale high-magnification verification scanning module to complete static focusing and in-situ tomographic scanning based on high-dimensional cross-scale servo micro-motion commands, and outputs a set of wafer microstructure identification and analysis images.

Images