A full-automatic double-station photoetching machine alignment method and device based on a multi-camera vision system

Abstract

The application relates to a double-station photoetching machine alignment technology field, and discloses a full-automatic double-station photoetching machine alignment method and equipment based on a multi-camera vision system, the method comprising the following steps: S10, jointly calibrating four cameras and calculating a target homography matrix of pixel coordinates corresponding to each camera to a machine coordinate system; S20, acquiring four images through parallel image acquisition of the four cameras, performing normalized cross-correlation positioning and ellipse least square fitting on alignment marks in each image, and obtaining first coordinates of four groups of alignment marks; S30, mapping the first coordinates into second coordinates of the four groups of alignment marks in the machine coordinate system, and solving a first substrate pose deviation vector; S40, controlling a carrier to perform compensation movement and iteratively performing S20-S30 until a residual pose error is smaller than a convergence threshold value, and outputting an alignment ready signal; the application realizes full automation of an alignment process while ensuring alignment precision of the double-station photoetching machine.

Description

Technical Field

[0001] This invention relates to the field of lithography machine alignment technology, and in particular to a fully automated dual-station lithography machine alignment method and equipment based on a multi-camera vision system. Background Technology

[0002] Dual-station lithography machines are core equipment in modern semiconductor and printed circuit board manufacturing, and their alignment accuracy directly determines the overlay quality of multi-layer patterns. In existing technologies, dual-station lithography machines generally employ a single-camera sequential imaging scheme, where a single camera is driven sequentially to scan the alignment marks at the four corners of the substrate, and then the local deviations are serially superimposed to obtain the compensation amount. The inherent drawback of this scheme is that during sequential imaging, the stage continuously drifts due to multiple sources of interference such as foundation vibration and airflow disturbances, resulting in inconsistent imaging times for each mark. This leads to the stitched pose result containing timing drift errors, failing to accurately reflect the overall pose of the substrate.

[0003] Existing technologies calculate the local deviations of the four markers independently and then simply sum them up, failing to incorporate the coordinate constraints of all marker points into a unified global optimization framework. This results in a lack of statistical suppression of single-point positioning noise and a systematic upper limit to pose calculation accuracy. Furthermore, there is a lack of a real-time feedforward compensation mechanism for short-term displacement disturbances of the platform during image acquisition waiting. Disturbances may cause the alignment markers to deviate from the camera's depth of field or field of view, leading to positioning failure and forcing repeated image acquisition, thus reducing alignment efficiency and system stability. Summary of the Invention

[0004] This invention provides a fully automated dual-station lithography machine alignment method and equipment based on a multi-camera vision system. This invention enables the dual-station lithography machine to achieve full automation of the alignment process while ensuring alignment accuracy.

[0005] In a first aspect, the present invention provides a fully automated dual-station lithography machine alignment method based on a multi-camera vision system, the fully automated dual-station lithography machine alignment method based on a multi-camera vision system comprising: S10. Jointly calibrate the four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. S20. By acquiring images in parallel through the four cameras, four images are obtained. Normalized cross-correlation localization and elliptic least squares fitting are performed on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. S30. Based on the target homography matrix, the first coordinate is mapped to the second coordinate of the four sets of alignment marks in the machine coordinate system. The second coordinate and the corresponding design coordinate are substituted into the overdetermined equation system to solve the first substrate pose deviation vector. S40. Control the stage to perform compensation motion according to the first substrate pose deviation vector and iteratively execute S20-S30 until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

[0006] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, S10 includes: S11. Place the standard checkerboard target on the platform plane and perform internal parameter calibration on the four cameras respectively to obtain the focal length parameters, principal point coordinates and radial distortion coefficients of each camera, and verify that the root mean square value of the corner reprojection error of each camera is lower than the reprojection error threshold. S12. The standard checkerboard target is sequentially moved to several positions covering the common calibration area of ​​the four cameras. At each position, the four cameras are controlled to synchronously acquire images through a TTL trigger signal, and the pixel coordinates of the checkerboard corner points in the images of each camera at each position are extracted. S13. Combining the pixel coordinates of the checkerboard corner points with the corresponding machine coordinates read by the stage grating ruler, solve for the initial homography matrix of each camera, and then perform nonlinear optimization on the initial homography matrix to obtain the target homography matrix corresponding to each camera.

[0007] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, S13 includes: S131. Construct coordinate point pairs by combining the pixel coordinates of the corner points of the chessboard grid with the corresponding machine coordinates, and perform linear solving on the coordinate point pairs to obtain the initial homography matrix of each camera. S132. Construct the objective function by the sum of squared residuals between the reprojected coordinates of each coordinate point pair after mapping with the initial homography matrix and the corresponding machine coordinates. Iteratively optimize the objective function based on the Levenberg-Marquardt algorithm to obtain the target homography matrix corresponding to each camera. S133. Substitute the pixel coordinates of the corner points of the chessboard grid into the target homography matrix for reprojection, calculate the root mean square value of the residual between the reprojected coordinates and the corresponding machine coordinates, verify that the root mean square value of the residual is lower than the reprojection error threshold, and confirm that the target homography matrix is ​​valid.

[0008] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, S20 includes: S21. Control the four cameras to acquire images in parallel through the TTL trigger signal to obtain four images. Perform normalized cross-correlation positioning on the alignment marks in each image to obtain the coarse positioning coordinates of each alignment mark. S22. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of contour edge pixels of the region of interest to perform elliptical least squares fitting to obtain the first coordinates of the four sets of alignment marks.

[0009] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, S22 includes: S221. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of contour edge pixels of the region of interest; S222. Substitute the coordinates of each edge pixel in the set of contour edge pixels into elliptic algebra, concatenate the coordinates of each edge pixel row to construct a coefficient matrix, and solve the elliptic coefficients of the coefficient matrix by the least squares method to obtain the target elliptic coefficients of each alignment mark. S223. Substitute the target ellipse coefficient into the ellipse center analytical function to calculate the horizontal and vertical coordinates of the ellipse center. Combine the offset of the region of interest in the original image to restore the global coordinates of the image, and obtain the first coordinates of the four sets of alignment marks.

[0010] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, S30 includes: S31. Substitute the first coordinates of the four sets of alignment marks into the target homography matrix of the corresponding camera to perform homogeneous coordinate transformation, and obtain the second coordinates of the four sets of alignment marks in the machine coordinate system. S32. For each group of second coordinates, the difference between the second coordinates and the corresponding design coordinates is calculated. The linearized rigid body transformation is then expanded into a linear residual equation, and an overdetermined system of equations is constructed. The first substrate pose deviation vector is then obtained by solving the equations.

[0011] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present invention, S32 includes: S321. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value. Substitute the coordinate difference value into the linearized rigid body transformation equation after performing a first-order Taylor expansion on the rotation angle. Expand it into a linear residual equation with translation and rotation angle as unknowns. Concatenate the four groups of linear residual equations row by row to construct an overdetermined equation set. S322. After transposing the coefficient matrix of the overdetermined equation system, multiply it by itself to obtain the normal equation coefficient matrix. After inverting the normal equation coefficient matrix, multiply it by the product of the transposed coefficient matrix and the difference vector to obtain the first substrate pose deviation vector.

[0012] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, S321 includes: S3211. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value of each alignment mark; S3212. Substitute the coordinate difference into the sine and cosine terms of the rotation angle and perform the linearized rigid body transformation equation after the first-order Taylor expansion. Expand each set of coordinate differences into a linear residual equation with the translation and rotation angle as unknowns and the corresponding design coordinate components as coefficients, to obtain four sets of linear residual equations. S3213. The coefficient rows and difference rows of the four sets of linear residual equations are concatenated row by row in the order of the groups to construct an overdetermined system of equations represented by a coefficient matrix and a difference vector.

[0013] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, S40 includes: S41. The translation and rotation angle in the first substrate pose deviation vector are superimposed on the stage target position register, the stage is controlled to perform compensation motion, and S20-S30 are iteratively executed after the stage is in place to obtain the second substrate pose deviation vector. S42. Convert the translation residual and rotation angle residual in the second substrate pose deviation vector into equivalent linear displacements and then calculate the norm to obtain the residual pose error. Compare the residual pose error with the convergence threshold. If the residual pose error is less than the convergence threshold, output the alignment ready signal; otherwise, return the second substrate pose deviation vector to S41 to continue iterative compensation.

[0014] Secondly, the present invention provides a fully automated dual-station lithography machine alignment device based on a multi-camera vision system, the fully automated dual-station lithography machine alignment device based on a multi-camera vision system comprising: The calibration module is used to jointly calibrate four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. The positioning module is used to acquire images in parallel by the four cameras to obtain four images, and to perform normalized cross-correlation positioning and elliptic least squares fitting on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. The pose solving module is used to map the first coordinates to the second coordinates of four sets of alignment marks in the machine coordinate system according to the target homography matrix, and to substitute the second coordinates and the corresponding design coordinates into the overdetermined equation system to solve the first substrate pose deviation vector. The output module is used to control the stage to perform compensation motion and iteratively execute the steps of the positioning module-pose solving module according to the first substrate pose deviation vector until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

[0015] The technical solution provided by this invention establishes a unified machine coordinate system through multi-camera joint calibration, enabling the pixel coordinates of all four cameras to be accurately mapped to the same coordinate reference, thus solving the technical problem of systematic inconsistency errors between the independent coordinate systems of each camera. Based on TTL hardware synchronous triggering, the four cameras acquire images in parallel at the same physical moment, eliminating the temporal pose error introduced by stage disturbance drift in the single-camera sequential image acquisition scheme, ensuring that the four sets of alignment mark images spatially correspond to the true pose of the substrate at the same moment. Through a two-stage positioning process of normalized cross-correlation coarse positioning and elliptic least squares fitting refinement, the positioning accuracy of the alignment mark center is improved to the sub-pixel level. The least squares solution of the coefficients of the elliptic algebraic equation and the analytical calculation of the center are completely analytical, avoiding the uncertainty brought about by iterative convergence. This invention uniformly substitutes the coordinate constraints of the four sets of marks into an overdetermined system of equations constructed based on linearized rigid body transformation, and solves the global least squares solution using the normal equation method. Compared with the local deviation superposition method, this method provides a unique optimal pose estimate in the sense of minimizing the sum of squares of the global residuals, and has a statistical averaging effect on suppressing random errors in single-point positioning. The real-time disturbance feedforward compensation mechanism based on acceleration quadratic integral pre-cancels short-term stage displacement disturbances before each image acquisition trigger, ensuring that each alignment mark remains within the camera's depth of field and field of view during parallel image acquisition, effectively avoiding positioning failures caused by disturbances. This invention enables dual-station lithography machines to achieve full automation of the alignment process while maintaining alignment accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram illustrating the steps of the alignment method for a fully automated dual-station lithography machine based on a multi-camera vision system in an embodiment of the present invention. Figure 2 This is a schematic diagram of the alignment equipment of a fully automated dual-station lithography machine based on a multi-camera vision system in an embodiment of the present invention. Detailed Implementation

[0018] This invention provides a fully automated dual-station lithography machine alignment method and apparatus based on a multi-camera vision system. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 An embodiment of the alignment method for a fully automated dual-station lithography machine based on a multi-camera vision system in this invention includes: S10. Jointly calibrate the four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. S20. By acquiring images in parallel using four cameras, four images are obtained. Normalized cross-correlation localization and elliptic least squares fitting are performed on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. S30. Based on the target homography matrix, map the first coordinate to the second coordinate of the four sets of alignment marks in the machine coordinate system. Substitute the second coordinate and the corresponding design coordinate into the overdetermined equation system to solve the first substrate pose deviation vector. S40. Control the stage to perform compensation motion according to the first substrate pose deviation vector and iteratively execute S20-S30 until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

[0020] Specifically, a unified machine coordinate system is established through multi-camera joint calibration, ensuring that the pixel coordinates of all four cameras can be accurately mapped to the same coordinate reference, thus solving the technical problem of systematic inconsistency errors between the independent coordinate systems of each camera. Based on TTL hardware synchronous triggering, parallel image acquisition by the four cameras at the same physical moment is achieved, eliminating the temporal pose error introduced by stage disturbance drift in the single-camera sequential image acquisition scheme, ensuring that the four sets of alignment mark images spatially correspond to the true pose of the substrate at the same moment. Through a two-stage positioning process of normalized cross-correlation coarse positioning and elliptic least-squares fitting refinement, the positioning accuracy of the alignment mark center is improved to the sub-pixel level. The least-squares solution of the coefficients of the elliptic algebraic equation and the analytical calculation of the center are completely analytical, avoiding the uncertainty brought about by iterative convergence. This invention uniformly substitutes the coordinate constraints of the four sets of marks into an overdetermined system of equations constructed based on linearized rigid body transformation, and solves the global least-squares solution using the normal equation method. Compared with the local deviation superposition method, this method provides a unique optimal pose estimate in the sense of minimizing the sum of squares of the global residuals, and has a statistical averaging effect on suppressing random errors in single-point positioning. The real-time disturbance feedforward compensation mechanism based on acceleration quadratic integral pre-cancels short-term stage displacement disturbances before each image acquisition trigger, ensuring that each alignment mark remains within the camera's depth of field and field of view during parallel image acquisition, effectively avoiding positioning failures caused by disturbances. This invention enables dual-station lithography machines to achieve full automation of the alignment process while maintaining alignment accuracy.

[0021] In one specific embodiment, S10 includes: S11. Place the standard checkerboard target on the platform plane and perform internal parameter calibration on the four cameras respectively to obtain the focal length parameters, principal point coordinates and radial distortion coefficients of each camera, and verify that the root mean square value of the corner reprojection error of each camera is lower than the reprojection error threshold. S12. The standard checkerboard target is sequentially moved to several positions covering the common calibration area of ​​the four cameras. At each position, the four cameras are controlled to acquire images synchronously through a TTL trigger signal. The pixel coordinates of the checkerboard corner points in the images of each camera at each position are extracted. S13. Combining the pixel coordinates of the checkerboard corner points with the corresponding machine coordinates read by the stage grating ruler, solve for the initial homography matrix of each camera, and then perform nonlinear optimization on the initial homography matrix to obtain the target homography matrix corresponding to each camera.

[0022] Specifically, the calibration process of the four cameras is used as the coordinate reference establishment process for the alignment process of the dual-station lithography machine. After the equipment is installed and debugged, the standard checkerboard target is placed stably on the working plane of the stage, ensuring that the calibration plane of the checkerboard target is consistent with the substrate bearing plane, and that the checkerboard target can cover the effective imaging area of ​​each camera. The grid spacing of the standard checkerboard target can be set to 5mm, and the target manufacturing accuracy is better than 1μm. The grid spacing of 5mm is mainly considered because the field of view of the alignment camera of the lithography machine needs to balance the number of corner points and the spacing between corner points. If the grid is too dense, the corner point detection will be more sensitive to noise and distortion. If the grid is too sparse, it will reduce the effective constraint points in a single image. The manufacturing accuracy of better than 1μm is used to reduce the impact of the target's own error on the camera parameter solution results. During the intrinsic parameter calibration of the four cameras, each camera acquired no fewer than 20 checkerboard images. During acquisition, the translational position and slight tilt of the checkerboard target within the field of view were adjusted to ensure corner point distribution covered the image center, edges, and different distortion regions. This allowed for the determination of the focal length, principal point coordinates, and radial distortion coefficients for each camera. After solving, the calibrated intrinsic parameter model was used to reproject and verify the corner points in each calibration image. The root mean square value of the corner point reprojection error needed to be below 0.3 pixels. Using 0.3 pixels as the reprojection error threshold strikes a relatively stable balance between the sub-pixel corner point extraction capability of industrial cameras and the alignment accuracy requirements of lithography machines. If the reprojection error threshold is exceeded, additional calibration images or images with poor corner point extraction quality need to be removed until all four cameras meet the reprojection error requirement.

[0023] After the intrinsic parameter calibration of the four cameras meets the accuracy requirements, the standard checkerboard target is moved sequentially to several positions that cover the common calibration area of ​​the four cameras. For example, the number of positions can be set to no less than eight, which can form a sufficiently dispersed spatial constraint in the machine coordinate system, so that the planar mapping relationship from pixel coordinates to machine coordinates is not limited to a local area, thereby reducing the extrapolation error of homography at the edge of the field of view. After moving to each calibration position, the stage performs positioning and reads the corresponding position of the checkerboard target in the machine coordinate system through a grating ruler. The resolution of the grating ruler can be 0.1μm, so that the machine coordinate reading error is lower than the allowable error of visual calibration, avoiding the machine coordinate reference from becoming the main source of error in solving the homography matrix. The controller sends a synchronous image acquisition command to the four cameras through the same TTL trigger signal, so that the four cameras simultaneously expose and output four checkerboard images at the same calibration position. Hardware synchronous triggering can avoid inconsistencies in the machine coordinates corresponding to different camera images caused by slight stage drift. When performing corner detection on each image, grayscale equalization and noise suppression are performed on the image, and the pixel coordinates of the checkerboard corners are extracted. The stability of the corner position is improved by sub-pixel refinement. For calibration images with missing corners, blurred edges, or severe local reflections, abnormal corners are removed or re-acquired, so that the data points entering the external parameter joint calibration simultaneously meet the requirements of clear pixel coordinates, accurate machine coordinates, and sufficient coverage.

[0024] After obtaining the pixel coordinates of the checkerboard corner points and the corresponding machine coordinates of the platform grating ruler at each calibration position, a point-to-point relationship between pixel coordinates and machine coordinates is established for each camera as an independent object, and the initial homography matrix for each camera is solved accordingly. The initial homography matrix describes the projection mapping relationship from the image plane of a single camera to a unified machine coordinate plane, enabling the corner positions acquired by the four cameras to be transformed to the same coordinate reference for comparison and calculation. The initial homography matrix can be obtained through linear solution, and the linear solution result serves as the starting point for nonlinear optimization, avoiding convergence instability caused by iterating from unconstrained initial values. In the nonlinear optimization stage, the residual between the reprojected coordinates of the pixel coordinates after mapping by the initial homography matrix and the machine coordinates read by the grating ruler is used as the optimization object. The Levenberg-Marquardt algorithm is used to correct the elements of the homography matrix, gradually reducing the overall residual of each calibration point in the machine coordinate system, and finally obtaining the target homography matrix corresponding to each of the four cameras.

[0025] In one specific embodiment, S13 includes: S131. Construct coordinate point pairs by matching the pixel coordinates of the corner points of the chessboard with the corresponding machine coordinates, and perform linear solving on the coordinate point pairs to obtain the initial homography matrix of each camera. S132. Construct the objective function by the sum of squared residuals between the reprojected coordinates of each coordinate point pair after mapping with the initial homography matrix and the corresponding machine coordinates. Iteratively optimize the objective function based on the Levenberg-Marquardt algorithm to obtain the target homography matrix corresponding to each camera. S133. Substitute the pixel coordinates of the corner points of the chessboard grid into the target homography matrix for reprojection, calculate the root mean square value of the residual between the reprojected coordinates and the corresponding machine coordinates, verify that the root mean square value of the residual is lower than the reprojection error threshold, and confirm that the target homography matrix is ​​valid.

[0026] Specifically, taking a single camera as the processing object, the pixel coordinates of the checkerboard corner points extracted at the same calibration position are paired one by one with the machine coordinates read by the stage grating ruler to form a set of coordinate point pairs. Indicates the first The first camera in Taiwan The x-coordinates of the corner points of the chessboard grid are in pixels. This represents the ordinate of the corresponding pixel, in pixels. This represents the horizontal coordinate of the same corner point in the machine coordinate system, in units of 1. m, This represents the vertical coordinate of the same corner point in the machine coordinate system, in units of 1. m, Indicates the first The homography matrix corresponding to each camera is used to map the pixel plane coordinates to the platform machine coordinate plane. After the coordinate point pairs are constructed, a direct linear transformation method is used to solve the homography matrix for each camera. The projection constraints between multiple corner points are organized into homogeneous linear equations, and the initial homography matrix is ​​obtained through singular value decomposition in the least squares sense or equivalent linear solution. Since the linear solution directly comes from the geometric constraints of all effective corner points, the resulting initial homography matrix can reflect the basic mapping relationship between the camera imaging plane and the platform plane. However, the linear solution does not fully consider the influence of pixel noise, corner sub-pixel extraction errors, and small perturbations in platform readings on the residual distribution. Therefore, nonlinear optimization is required.

[0027] In nonlinear optimization, the sum of squared residuals between the reprojected coordinates of the checkerboard corner pixel coordinates after mapping with the initial homography matrix and the corresponding machine coordinates is used as the optimization objective. The objective function can be expressed as:

[0028] in Indicates the first The sum of squares of the reprojection residuals of the cameras; the smaller the value, the smaller the mapping error from pixel coordinates to machine coordinates. This represents the homogeneous coordinate normalization operation, which divides the first two dimensions of the 3D homogeneous result by the third dimension to obtain the 2D coordinates on the machine coordinate plane. The Levenberg-Marquardt algorithm uses the initial homography matrix as the starting point for iteration. In each iteration, it adjusts the homography matrix elements based on the residual changes, reducing the overall reprojection bias of all coordinate point pairs. It stops when the residual decrease is less than a set stopping condition or when the maximum number of iterations is reached. The stopping condition can be that the rate of change of the sum of squares of two consecutive residuals is less than 10. -6 The maximum number of iterations can be set to 100. The rate of change threshold can prevent the optimization from ending prematurely when the residuals are still decreasing significantly. The maximum number of iterations is used to limit long periods of non-convergence caused by abnormal data. All checkerboard corner pixel coordinates are resubstituted into the target homography matrix. The root mean square value of the residual between the reprojected coordinates and the raster ruler machine coordinates is calculated and compared with the reprojection error threshold. The corner reprojection error threshold during the intrinsic parameter calibration stage can be set to 0.3 pixels. The machine coordinate error threshold is used during the target homography matrix verification stage. The machine coordinate error threshold is calculated based on the camera calibration ratio, lens magnification, and the Jacobian relationship of local homography mapping. When the root mean square value of the residuals is lower than the reprojection error threshold, the target homography matrix is ​​confirmed as valid and written to the system parameter file. When the root mean square value of the residuals does not meet the requirements, abnormal corners, corners in edge distortion areas, or low-resolution images are removed, and linear and nonlinear optimizations are re-executed to ensure that the final retained target homography matrix has stable global mapping accuracy.

[0029] In one specific embodiment, S20 includes: S21. Control four cameras to acquire images in parallel through TTL trigger signals to obtain four images. Perform normalized cross-correlation positioning on the alignment marks in each image to obtain the coarse positioning coordinates of each alignment mark. S22. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of contour edge pixels of the region of interest to perform elliptical least squares fitting to obtain the first coordinates of the four sets of alignment marks.

[0030] Specifically, the stage positioning confirmation signal is used as the start condition for the parallel image acquisition process. After the grating ruler confirms that the stage has entered the alignment position and the position lock meets the image acquisition conditions, the controller sends rising edge pulses to the external trigger input ports of the four cameras through the same hardware TTL trigger bus. The TTL trigger pulse width can be set to 10μs, which can meet the recognition requirements of the industrial camera's external trigger port for a stable high-level holding time, and will not occupy an excessively long control cycle. The difference in the exposure start time of the four cameras is determined by the trigger bus propagation delay, which can be controlled within 100ns under the conditions of the same control board wiring and equal-length wiring. The exposure time is adjusted during the equipment calibration stage based on the illumination brightness of the alignment mark and the camera gain, so that the average gray level of the alignment mark area falls within the range of 45% to 65% of the sensor's full-scale gray level value, which corresponds to approximately 115 to 166 gray levels for an 8-bit grayscale image. This range can keep the gray level gradient of the mark edge in a relatively stable linear response region, reducing edge widening caused by overexposure and insufficient edge contrast caused by underexposure.

[0031] After acquiring four images, the positioning module performs coarse positioning of the alignment marks on each image. The original grayscale image is processed by Gaussian filtering, with the filter kernel size set to 5×5 pixels and the standard deviation set to 1.2 pixels to suppress camera thermal noise and shot noise, while avoiding significant smoothing shifts to the mark edges. The standard template cropped from the clear alignment mark image during the calibration stage is used as the matching template. The template size can be 1.2 times the nominal diameter of the alignment mark multiplied by 1.2, so that the template covers the complete mark outline while retaining a limited background area to enhance matching discrimination. The normalized cross-correlation response value can be calculated by the following formula:

[0032] in, Indicates the first Road image at translation position Normalized cross-correlation response value at the location; Indicates the first Road smoothing image with template coordinates The corresponding grayscale value; Represents coordinates in the standard template grayscale value; This represents the average grayscale value of the current candidate matching region; This represents the average grayscale value of the standard template. Normalized cross-correlation processing can reduce the impact of overall brightness variations on the matching results, making it suitable for situations where there are slight fluctuations in illumination intensity during lithography machine alignment. After calculating the response map, the positioning module searches for the peak value of the response map and converts the peak position into coarse positioning coordinates of the alignment mark center. The response peak value needs to be no less than 0.85. If the response peak value is lower than this threshold, it can be determined that the alignment mark is off-field, severely contaminated, or partially occluded, triggering a resampling image or an anomaly warning.

[0033] After obtaining four sets of coarse localization coordinates, a region of interest (ROI) is extracted centered on each set of coordinates. The ROI side length is approximately 1.5 times the nominal diameter of the alignment marker, covering the marker edges while retaining a small amount of background buffer. This limits the edge detection range and reduces interference from unmarked textures on the fitting results. Local grayscale normalization and edge enhancement are performed within the ROI, followed by Canny edge detection to extract the contour edge pixel set. The Canny high threshold is set to the 95th percentile of the gradient magnitude of the ROI, and the low threshold is set to 0.5 times the high threshold. This threshold combination preserves the main contour edges and suppresses random texture edges. After obtaining the contour edge pixel set, the localization module performs elliptic least-squares fitting on the edge pixel set and, combined with the offset of the ROI in the original image, restores the fitted center coordinates to the entire image coordinate system, forming the first coordinates of the four alignment markers.

[0034] S21. Perform normalized cross-correlation localization on the alignment marks in each image to obtain the coarse localization coordinates of each alignment mark. This includes: performing Gaussian filtering preprocessing on each image; using the pre-stored alignment mark template as the core, calculating the normalized cross-correlation response value pixel by pixel on the filtered image; constructing a normalized cross-correlation response map of the same size as the image; performing global peak search on the normalized cross-correlation response map; extracting the position with the largest response value as the coarse localization coordinates of the mark center; comparing the maximum response value with the normalized cross-correlation threshold; if the maximum response value is lower than the normalized cross-correlation threshold, determining that the current alignment mark localization is invalid and triggering a resampling image; outputting the coarse localization coordinates that have passed the threshold verification along with the corresponding local peak neighborhood of the normalized cross-correlation response map for use in S22 to extract the region of interest, thus obtaining the coarse localization coordinates of each alignment mark.

[0035] In one specific embodiment, S22 includes: S221. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of pixels at the contour edge of the region of interest; S222. Substitute the coordinates of each edge pixel in the contour edge pixel set into elliptic algebra, concatenate the coordinates of each edge pixel row to construct a coefficient matrix, and solve the elliptic coefficients of the coefficient matrix by the least squares method to obtain the target elliptic coefficients of each alignment mark. S223. Substitute the target ellipse coefficients into the ellipse center analytical function to calculate the x-coordinate and y-coordinate of the ellipse center. Combine this with the offset of the region of interest in the original image to restore the global coordinates of the image, and obtain the first coordinates of the four sets of alignment marks.

[0036] Specifically, the coarse positioning coordinates are used as the local search center. A region of interest (ROI) is established around the alignment mark in each image path, ensuring that edge extraction only affects the neighborhood of the mark, reducing interference from line textures, dust highlights, and uneven background grayscale on ellipse fitting. After the ROI is extracted, grayscale smoothing and gradient calculation are performed on the local image, and a dual-threshold edge detection method is used to extract the mark contour. Only pixels connected to the coarse positioning center, with contour lengths satisfying the mark size constraints, and continuous edge gradient directions are retained in the edge detection results, thus forming a set of contour edge pixels. Discrete noise edges caused by local reflections, dirt, or gaps can be removed based on the distance distribution from the edge points to the coarse positioning center, ensuring that edge pixels entering ellipse fitting correspond as closely as possible to the same mark outer contour, avoiding unmarked edges pulling the fitting center off-center.

[0037] Substituting the coordinates of each edge pixel in the set of contour edge pixels into the elliptic algebraic relation, the elliptic algebraic relation can be expressed as:

[0038] in, This represents the x-coordinate of the edge pixel in the coordinate system of the region of interest. This represents the ordinate of the edge pixel in the coordinate system of the region of interest. , , , , , Let represent the elliptic algebraic coefficients to be solved. For each edge pixel, a row of coefficient constraints is constructed. The constraints corresponding to all edge pixels are concatenated row by row to form an overdetermined system of linear equations. The elliptic algebraic coefficients are then solved using the least squares method. After the solution is completed, based on... The fitting results are verified to satisfy the elliptic constraint. If the constraint is not satisfied, it indicates that there are many abnormal edges mixed in with the edge pixel set. It is necessary to first remove the edge points that deviate from the main contour, and then reconstruct the coefficient matrix and solve it. After verification by the elliptic constraint, the obtained elliptic algebraic coefficients are used as the target elliptic coefficients in the central analytical calculation.

[0039] Substituting the target ellipse coefficients into the ellipse center analytical relation, we obtain the x-coordinate and y-coordinate of the ellipse center in the coordinate system of the region of interest. The ellipse center analytical relation can be expressed as:

[0040]

[0041] in, This represents the x-coordinate of the center of the ellipse in the coordinate system of the region of interest. This represents the ordinate of the ellipse center in the region of interest coordinate system. Since ellipse fitting is performed within the region of interest, the center resolution result is still a local coordinate. Therefore, it is necessary to combine the horizontal and vertical offsets of the upper left corner of the region of interest relative to the original image to reconstruct the coordinates, obtaining the alignment mark center coordinates in the global coordinate system of the original image. After the above processing is performed on the four images, each of the four cameras outputs a set of alignment mark center coordinates, which serve as the first coordinates of the four sets of alignment marks.

[0042] S222: Substitute the coordinates of each edge pixel in the contour edge pixel set into the elliptic algebraic equation, concatenate the coordinates of each edge pixel row to construct a coefficient matrix, and solve the elliptic coefficients of the coefficient matrix using the least squares method to obtain the target elliptic coefficients of each alignment mark. This also includes: calculating a discriminant for the elliptic coefficients obtained by the least squares method, comparing the discriminant with zero, and if the discriminant does not satisfy the elliptic constraint conditions, determining that there are abnormal edge pixels in the current contour edge pixel set, performing abnormal point removal on the contour edge pixel set, reconstructing the coefficient matrix, and solving the elliptic coefficients again; calculating the ratio of the major axis to the minor axis of the ellipse for the elliptic coefficients that pass the elliptic constraint verification, comparing the ratio of the major and minor axes with a preset roundness threshold, and if the ratio of the major and minor axes exceeds the roundness threshold range, determining that the current alignment mark image has contour distortion due to defocus or dirt, triggering the re-sampling process; outputting the elliptic coefficients that simultaneously satisfy the elliptic constraint conditions and the roundness threshold as the target elliptic coefficients of each alignment mark for use in S223 for ellipse center coordinate analysis.

[0043] In one specific embodiment, S30 includes: S31. Substitute the first coordinates of the four sets of alignment marks into the target homography matrix of the corresponding camera to perform homogeneous coordinate transformation, and obtain the second coordinates of the four sets of alignment marks in the machine coordinate system. S32. For each set of second coordinates, subtract the corresponding design coordinates, expand the linearized rigid body transformation into a linear residual equation, construct an overdetermined system of equations, and solve to obtain the first substrate pose deviation vector.

[0044] Specifically, the first coordinates of the four sets of alignment marks output by the four cameras are used as input for unified pose calculation, and processed according to the fixed correspondence between camera number and alignment mark number to avoid mismatch between camera field of view and mark position. For the first... The first coordinate output by the camera ,in Indicates the first The x-coordinate of the pixel aligned to the center of the marker in the image coordinate system of the camera. Indicates the first The ordinate of the pixel aligned to the center of the marker in the camera image coordinate system is obtained by normalized cross-correlation coarse localization and elliptic least-squares fitting. During coordinate mapping, the controller reads the first pixel coordinate from the system parameter file. Target homography matrix corresponding to each camera Then, after expanding the first coordinate to homogeneous coordinate form, matrix multiplication is performed:

[0045] in, , , These represent the three intermediate components after homogeneous coordinate transformation. This is used to perform normalization. Then, the first two dimensions of the homogeneous coordinates are divided by the third dimension to obtain the second coordinate in the machine coordinate system:

[0046] in, Indicates the first The measured lateral coordinates of the alignment mark in the machine coordinate system, in units of... m, Indicates the first The longitudinal measured coordinates of the alignment mark in the machine coordinate system, in units of... m. Through homogeneous coordinate transformation, the marker center positions of the four cameras, which were originally in different image coordinate systems, were transformed to the same platform machine coordinate system.

[0047] After obtaining four sets of second coordinates, the controller calls the four sets of alignment mark design coordinates pre-stored in the layout file or process parameter file, and establishes the relationship between the second coordinates and the corresponding design coordinates according to the same mark number. The difference between each set of second coordinates and the corresponding design coordinates reflects the local deviation of the current substrate relative to the theoretical loading position. However, the difference of a single mark is easily affected by edge extraction noise, local contamination, or target manufacturing deviation. Therefore, the four sets of coordinate constraints are simultaneously incorporated into the linearized rigid body transformation relationship. The overall pose deviation of the substrate is described as lateral translation, longitudinal translation, and in-plane rotation angle. Under the condition that mechanical coarse alignment has been completed, the residual rotation angle is in a small angle range. The sine and cosine terms can be linearized according to the first-order Taylor expansion, so that each set of mark points forms two linear residual relationships with respect to translation and rotation angle. The four mark points form a total of eight linear residual relationships, while the unknowns are only lateral translation, longitudinal translation, and rotation angle. Therefore, the resulting equation system belongs to the overdetermined equation system, which can use the spatial distribution of the four corner marks to suppress random errors by averaging.

[0048] When constructing the overdetermined equation system, the coefficient row consists of the lateral and longitudinal components of the corresponding design coordinates, and the difference vector consists of the difference between the second coordinate and the design coordinates. The matrix solution result is the first substrate pose deviation vector. Since the four alignment marks are arranged at the four corners of the substrate and their spatial positions are not collinear, the coefficient matrix has a good rank condition, which can stably solve for the lateral translation, longitudinal translation, and rotation angle. If the residual of a certain mark point deviates significantly from other mark points, it can be eliminated or weighted before solving by combining positional confidence or residual consistency, to prevent a single abnormal mark from dominating the overall pose result. After obtaining the first substrate pose deviation vector, the controller uses the first substrate pose deviation vector as the input of the stage compensation motion, so that the stage can be corrected according to the overall pose estimation in the sense of global least squares.

[0049] In one specific embodiment, S32 includes: S321. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value. Substitute the coordinate difference value into the linearized rigid body transformation equation after performing a first-order Taylor expansion on the rotation angle. Expand it into a linear residual equation with translation and rotation angle as unknowns. Concatenate the four groups of linear residual equations row by row to construct an overdetermined system of equations. S322. After transposing the coefficient matrix of the overdetermined equation system, multiply it by itself to obtain the normal equation coefficient matrix. After inverting the normal equation coefficient matrix, multiply it by the product of the transpose of the coefficient matrix and the difference vector to obtain the first substrate pose deviation vector.

[0050] Specifically, using four sets of second coordinates as the basis for actual measurements, each set of second coordinates is registered with the corresponding design coordinates in the layout file according to the same marker number, and the difference between the horizontal and vertical coordinates is calculated separately. The coordinate differences of the four alignment markers are then substituted into the linearized rigid body transformation relationship. For the first... One alignment mark, and This represents the measured horizontal and vertical coordinates in the machine coordinate system, in units of... m, and This represents the corresponding horizontal and vertical coordinates of the design, in units of... m, Indicates the lateral translation of the substrate, in units of m, This indicates the longitudinal translation amount of the substrate, in units of... m, This represents the in-plane rotation angle of the substrate, in rad. During linearization, the sine term in small-angle rotations is approximated as the rotation angle, and the cosine term is approximated as 1, resulting in two linear residual equations for each alignment mark:

[0051]

[0052] In the two relationships above, the left side represents the difference between the measured coordinates and the design coordinates, in units of... m; the right-side translation amount itself is m, the rotation angle can be considered a dimensionless quantity, and the product of the rotation angle and the design coordinates is still m. m, therefore the horizontal and vertical residuals are consistent in dimension. After expanding the four alignment marks sequentially, a total of eight linear residual relationships are obtained. The coefficients of each linear residual relationship are concatenated row by row to form a coefficient matrix. The coordinate differences are then concatenated in the same order to form a difference vector. Simultaneously, the lateral translation, longitudinal translation, and rotation angle to be determined are used to form the first substrate pose deviation vector. This leads to the overdetermined system of equations:

[0053] in, This represents a coefficient matrix consisting of four sets of design coordinates, with a size of 8×3. This represents the pose deviation vector of the first substrate; This represents the difference vector consisting of four sets of differences between measured coordinates and design coordinates. Since the four alignment marks are distributed at the four corners of the substrate, and the spatial positions of the mark points are not collinear, the coefficient matrix can provide effective constraints on the translation and rotation components, avoiding instability caused by calculating the overall pose solely from local points.

[0054] The overdetermined system of equations is solved using normal equations. The coefficient matrix... Transpose the matrix and multiply it by the coefficient matrix itself to obtain the coefficient matrix of the normal equation. Then transpose the coefficient matrix and multiply it by the difference vector on the left. The right-hand side of the normal equation is obtained. Then, the coefficient matrix of the normal equation is inverted and the pose deviation vector of the first substrate is calculated. The solution relationship is:

[0055] obtained The solution includes lateral translation, longitudinal translation, and in-plane rotation vectors. The result is a global pose estimate obtained by minimizing the sum of squared residuals of the four marker points. Compared with the method of directly adding the four local deviations, solving the normal equation can simultaneously consider the constraints of all marker points on translation and rotation, so that the sub-pixel localization error of individual marker points, edge deviation caused by local contamination, and slight image noise are suppressed in the overall solution.

[0056] Between S322 and S41, there is also a step of identifying and correcting the systematic residuals caused by the non-rigid deformation of the substrate: substituting the first substrate pose deviation vector back into the overdetermined equations, calculating the point-by-point residuals between the theoretical coordinates and the second coordinates of each alignment mark under rigid body transformation, obtaining four sets of mark point residual vectors; calculating the difference between the mean and maximum residual values ​​of the four sets of mark point residual vectors, comparing the difference with the non-rigid deformation judgment threshold, if the difference exceeds the non-rigid deformation judgment threshold, it is determined that the current substrate has non-rigid local deformation; otherwise, the first substrate pose deviation vector is directly output to S41; for the judgment... For substrates with non-rigid local deformation, the four sets of residual vectors of marked points and their corresponding design coordinates are substituted into the affine transformation model. Scale scaling components and shear components are added to the rigid transformation parameters to construct an extended linear equation system. The affine correction parameters are then solved using the normal equation method to obtain an extended substrate pose deviation vector containing rigid and non-rigid correction components. The translation and rotation components in the extended substrate pose deviation vector are output to the S41 control stage to perform compensation motion. Simultaneously, the non-rigid correction components are recorded in the deformation compensation parameter file for the exposure station to perform corresponding pattern corrections on local areas during pattern projection.

[0057] In one specific embodiment, S321 includes: S3211. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value of each alignment mark; S3212. Substitute the coordinate difference into the sine and cosine terms of the rotation angle and perform the linearized rigid body transformation equation after the first-order Taylor expansion. Expand each set of coordinate differences into a linear residual equation with the translation and rotation angle as unknowns and the corresponding design coordinate components as coefficients, to obtain four sets of linear residual equations. S3213. Concatenate the coefficient rows and difference rows of the four sets of linear residual equations row by row in the order of the sets to construct an overdetermined system of equations represented by a coefficient matrix and a difference vector.

[0058] Specifically, after the four sets of second coordinates have been unified into the machine coordinate system, a one-to-one correspondence between the measured coordinates and the design coordinates is established according to the alignment mark numbers, and the coordinate differences with the same number are used as the observations for pose solving. The controller reads the... The second coordinate of the alignment mark and corresponding design coordinates The difference between the horizontal and vertical components is obtained by subtracting the horizontal coordinates. Difference from vertical coordinate .in, and All with m is the unit, representing the first... The local deviation of each alignment mark relative to the design position. After the difference calculation is completed for the four alignment marks in sequence, four sets of coordinate difference values ​​are obtained. The coordinate difference values ​​are not directly used as the platform compensation command, but as the right-hand observation of the linearized rigid body transformation equation, so that the translational deviation and rotational deviation can be solved together in a unified system of equations.

[0059] The overall deviation of the substrate relative to the design position is decomposed into lateral translation, longitudinal translation, and in-plane rotation angle. Since mechanical coarse alignment has already limited the angular deviation to a small range, the sine term of the rotation angle can be approximated as the rotation angle using a first-order Taylor expansion, and the cosine term can be approximated as 1. This transforms the rigid body transformation relationship, which originally contained trigonometric functions, into a linear relationship. For each alignment mark, the lateral coordinate difference is determined by the lateral translation and the equivalent rotational displacement caused by the designed longitudinal coordinate, while the longitudinal coordinate difference is determined by the longitudinal translation and the equivalent rotational displacement caused by the designed lateral coordinate. Therefore, each set of coordinate differences can be expanded into two linear residual relationships. The coefficient behavior of the lateral relationship... The coefficient behavior of vertical relationships The two coefficient rows correspond to three unknowns: horizontal translation, vertical translation, and in-plane rotation angle. After processing, the first alignment mark provides two constraints, the second alignment mark provides two more constraints, and so on, until the four alignment marks form eight linear constraints, transforming the pose deviation solution from local coordinate difference judgment to parameter estimation under global constraints.

[0060] Following the alignment mark numbers from 1 to 4, write the horizontal and vertical coefficient rows corresponding to each alignment mark into the coefficient matrix row by row. The horizontal and vertical coordinate differences under the same order are written into the difference vector. This leads to an overdetermined system of equations. The splicing order must remain strictly consistent, that is, the first... The horizontal coefficient row of each alignment mark corresponds to , No. The vertical coefficient row of each alignment mark corresponds to To avoid misalignment between the coefficient matrix rows and the difference vector elements, four alignment marks are placed at the four corners of the substrate. Eight linear constraints act on three unknowns, resulting in an overdetermined coefficient matrix that can simultaneously constrain lateral translation, longitudinal translation, and rotation errors. After constructing the coefficient matrix and difference vector, the overdetermined equations can proceed to the normal equation solution process, yielding the first substrate pose deviation vector used for stage compensation.

[0061] In one specific embodiment, S40 includes: S41. The translation and rotation angle in the first substrate pose deviation vector are superimposed on the stage target position register, the stage is controlled to perform compensation motion, and S20-S30 are iteratively executed after the stage is in place to obtain the second substrate pose deviation vector. S42. Convert the translation residual and rotation angle residual in the second substrate pose deviation vector into equivalent linear displacements and then calculate the norm to obtain the residual pose error. Compare the residual pose error with the convergence threshold. If the residual pose error is less than the convergence threshold, output the alignment ready signal; otherwise, return the second substrate pose deviation vector to S41 to continue iterative compensation.

[0062] Specifically, the first substrate pose deviation vector is used as the input for stage closed-loop compensation, allowing the visual calculation results to directly affect the target position register of the XYθ stage. Upon receiving the first substrate pose deviation vector, the controller determines whether the stage is currently in a locked state where compensation is permitted. If the grating ruler reading is stable and the driver is not in an alarm state, the target position of the stage is corrected according to the negative feedback principle. The first substrate pose deviation vector includes lateral translation, longitudinal translation, and in-plane rotation angle. During compensation, these are written to the stage target position register in the opposite direction to the deviation, gradually bringing the measured substrate pose closer to the designed pose. For lateral and longitudinal translation residuals, the controller writes the corresponding compensation amount to the position target register of the linear motor or lead screw drive shaft; for in-plane rotation residuals, the controller writes the rotation compensation amount to the angle target register of the θ-axis turntable or differential drive mechanism. The stage driver executes the compensation motion according to the updated target register and confirms that the stage has entered a stable state after the motion is completed through grating ruler feedback, servo positioning flags, and speed below a stop threshold, avoiding the introduction of dynamic errors due to re-acquiring images before the stage is fully stable. After the first compensation is completed, the controller triggers the four-camera synchronous image acquisition and alignment mark positioning in S20 again, and continues to execute the target homography matrix mapping and global pose solution in S30, thereby obtaining the second substrate pose deviation vector. The second substrate pose deviation vector reflects the residual error that still exists after the first compensation, and can be used to determine whether the compensation has met the overlay accuracy requirements.

[0063] The controller uniformly converts the lateral translation residual, longitudinal translation residual, and rotation angle residual to a linear displacement scale for evaluation, ensuring a unified dimension for convergence judgment. The rotation angle residual, expressed in angle form, cannot be directly added to the μm-level translation residual; therefore, it needs to be equivalently converted using the maximum diagonal spacing among the four alignment marks, allowing the displacement effect of in-plane rotation on the substrate edge region to be included in the same evaluation metric. The residual pose error can be expressed as:

[0064] in, This represents the residual pose error, in μm. This represents the lateral translation residual in the pose deviation vector of the second substrate, in μm. This represents the longitudinal translational residual in the pose deviation vector of the second substrate, in μm. This represents the in-plane rotation angle residual in the pose deviation vector of the second substrate, in rad. This represents the maximum diagonal spacing among the four alignment marks, in μm. Since rad can be treated as dimensionless, The unit remains μm, and the three components of the residual pose error are consistent in dimension.

[0065] The convergence threshold can be adjusted based on the overlay budget, alignment mark arrangement size, and stage repeatability. In exposure or encapsulation lithography scenarios for substrates with a minimum feature linewidth of approximately 2.5 μm or more, the convergence threshold can be set to 0.5 μm. Under different process conditions with varying alignment accuracy requirements, the convergence threshold is readjusted based on the corresponding overlay budget. The controller compares the residual pose error with the convergence threshold. If the residual pose error is less than the convergence threshold, it indicates that the combined deviation of the current substrate's translational and rotational errors has met the alignment requirements. The controller outputs an alignment ready signal and allows the dual-station state machine to transition the substrate into the exposure preparation or exposure execution process. If the residual pose error does not meet the convergence requirements, the controller returns the second substrate pose deviation vector as a new compensation input to S41, updates the stage target position register again, and executes the compensation motion. To avoid infinite loops caused by abnormal substrates or mark damage, the maximum number of iterations can be set to 5, covering the multi-round convergence requirements when the mechanical coarse alignment residual is large, while also limiting the alignment process time. If the convergence condition is not met even after reaching the maximum number of iterations, the system will no longer continue to perform compensation iterations. Instead, it will output an out-of-tolerance alarm and mark the current substrate as needing to be re-inspected, thereby preventing substrates with obvious loading deviations or abnormal markings from entering the exposure station.

[0066] S41. Before superimposing the translation and rotation angle in the first substrate pose deviation vector onto the stage target position register, the process also includes a step of real-time feedforward compensation for short-term displacement disturbances of the stage: After the stage is confirmed to be in position, disturbance feedforward compensation is initiated. The average zero-bias value collected during the power-on stationary phase is subtracted from the acceleration sample values ​​of each axis output by the stage's built-in accelerometer at a fixed sampling period to obtain the zero-bias corrected acceleration sequence. The zero-bias corrected acceleration sequence is then subjected to stepwise cumulative summation to obtain the velocity sequence, and finally, the velocity sequence is subjected to stepwise cumulative summation to obtain... The system reads the short-term disturbance displacement sequence of the platform during the last control cycle before the TTL trigger signal is issued, obtains the disturbance feedforward compensation amount, adds the disturbance feedforward compensation amount to the platform target position register, and controls the platform to perform disturbance compensation fine-tuning. After completing the disturbance compensation fine-tuning, the system sends a TTL trigger signal to the four cameras to perform parallel image acquisition, and resets the velocity sequence and disturbance displacement sequence to zero and re-integrates each time the platform is confirmed to be in position, ensuring that the disturbance feedforward compensation amount only reflects the short-term platform disturbance drift during the current image acquisition waiting period.

[0067] The alignment method for a fully automated dual-station lithography machine based on a multi-camera vision system, as described above in embodiments of the present invention, is described below. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the fully automated dual-station lithography alignment equipment based on a multi-camera vision system in this invention includes: The calibration module 201 is used to jointly calibrate four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. The positioning module 202 is used to acquire images in parallel by four cameras to obtain four images, and to perform normalized cross-correlation positioning and elliptic least squares fitting on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. The pose solving module 203 is used to map the first coordinates to the second coordinates of four sets of alignment marks in the machine coordinate system according to the target homography matrix, and substitute the second coordinates and the corresponding design coordinates into the overdetermined equation system to solve the pose deviation vector of the first substrate. The output module 204 is used to control the stage to perform compensation motion and iteratively execute the steps of the positioning module-pose solving module according to the first substrate pose deviation vector until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

[0068] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0069] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The above-described 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.

Claims

1. A fully automated dual-station lithography machine alignment method based on a multi-camera vision system, characterized in that, include: S10. Jointly calibrate the four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. S20. By acquiring images in parallel through the four cameras, four images are obtained. Normalized cross-correlation localization and elliptic least squares fitting are performed on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. S30. Based on the target homography matrix, the first coordinate is mapped to the second coordinate of the four sets of alignment marks in the machine coordinate system. The second coordinate and the corresponding design coordinate are substituted into the overdetermined equation system to solve the first substrate pose deviation vector. S40. Control the stage to perform compensation motion according to the first substrate pose deviation vector and iteratively execute S20-S30 until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

2. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 1, characterized in that, S10 includes: S11. Place the standard checkerboard target on the platform plane and perform internal parameter calibration on the four cameras respectively to obtain the focal length parameters, principal point coordinates and radial distortion coefficients of each camera, and verify that the root mean square value of the corner reprojection error of each camera is lower than the reprojection error threshold. S12. The standard checkerboard target is sequentially moved to several positions covering the common calibration area of ​​the four cameras. At each position, the four cameras are controlled to synchronously acquire images through a TTL trigger signal, and the pixel coordinates of the checkerboard corner points in the images of each camera at each position are extracted. S13. Combining the pixel coordinates of the checkerboard corner points with the corresponding machine coordinates read by the stage grating ruler, solve for the initial homography matrix of each camera, and then perform nonlinear optimization on the initial homography matrix to obtain the target homography matrix corresponding to each camera.

3. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 2, characterized in that, S13 includes: S131. Construct coordinate point pairs by combining the pixel coordinates of the corner points of the chessboard grid with the corresponding machine coordinates, and perform linear solving on the coordinate point pairs to obtain the initial homography matrix of each camera. S132. Construct the objective function by the sum of squared residuals between the reprojected coordinates of each coordinate point pair after mapping with the initial homography matrix and the corresponding machine coordinates. Iteratively optimize the objective function based on the Levenberg-Marquardt algorithm to obtain the target homography matrix corresponding to each camera. S133. Substitute the pixel coordinates of the corner points of the chessboard grid into the target homography matrix for reprojection, calculate the root mean square value of the residual between the reprojected coordinates and the corresponding machine coordinates, verify that the root mean square value of the residual is lower than the reprojection error threshold, and confirm that the target homography matrix is ​​valid.

4. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 1, characterized in that, S20 includes: S21. Control the four cameras to acquire images in parallel through the TTL trigger signal to obtain four images. Perform normalized cross-correlation positioning on the alignment marks in each image to obtain the coarse positioning coordinates of each alignment mark. S22. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of contour edge pixels of the region of interest to perform elliptical least squares fitting to obtain the first coordinates of the four sets of alignment marks.

5. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 4, characterized in that, S22 includes: S221. Using the coarse positioning coordinates as the center, extract the region of interest corresponding to each alignment mark, and extract the set of contour edge pixels of the region of interest; S222. Substitute the coordinates of each edge pixel in the set of contour edge pixels into elliptic algebra, concatenate the coordinates of each edge pixel row to construct a coefficient matrix, and solve the elliptic coefficients of the coefficient matrix by the least squares method to obtain the target elliptic coefficients of each alignment mark. S223. Substitute the target ellipse coefficient into the ellipse center analytical function to calculate the horizontal and vertical coordinates of the ellipse center. Combine the offset of the region of interest in the original image to restore the global coordinates of the image, and obtain the first coordinates of the four sets of alignment marks.

6. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 1, characterized in that, S30 includes: S31. Substitute the first coordinates of the four sets of alignment marks into the target homography matrix of the corresponding camera to perform homogeneous coordinate transformation, and obtain the second coordinates of the four sets of alignment marks in the machine coordinate system. S32. For each group of second coordinates, the difference between the second coordinates and the corresponding design coordinates is calculated. The linearized rigid body transformation is then expanded into a linear residual equation, and an overdetermined system of equations is constructed. The first substrate pose deviation vector is then obtained by solving the equations.

7. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 6, characterized in that, S32 includes: S321. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value. Substitute the coordinate difference value into the linearized rigid body transformation equation after performing a first-order Taylor expansion on the rotation angle. Expand it into a linear residual equation with translation and rotation angle as unknowns. Concatenate the four groups of linear residual equations row by row to construct an overdetermined equation set. S322. After transposing the coefficient matrix of the overdetermined equation system, multiply it by itself to obtain the normal equation coefficient matrix. After inverting the normal equation coefficient matrix, multiply it by the product of the transposed coefficient matrix and the difference vector to obtain the first substrate pose deviation vector.

8. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 7, characterized in that, S321 includes: S3211. Subtract the second coordinate from the corresponding design coordinate for each group to obtain the coordinate difference value of each alignment mark; S3212. Substitute the coordinate difference into the sine and cosine terms of the rotation angle and perform the linearized rigid body transformation equation after the first-order Taylor expansion. Expand each set of coordinate differences into a linear residual equation with the translation and rotation angle as unknowns and the corresponding design coordinate components as coefficients, to obtain four sets of linear residual equations. S3213. The coefficient rows and difference rows of the four sets of linear residual equations are concatenated row by row in the order of the groups to construct an overdetermined system of equations represented by a coefficient matrix and a difference vector.

9. The fully automated dual-station lithography machine alignment method based on a multi-camera vision system according to claim 1, characterized in that, S40 includes: S41. The translation and rotation angle in the first substrate pose deviation vector are superimposed on the stage target position register, the stage is controlled to perform compensation motion, and S20-S30 are iteratively executed after the stage is in place to obtain the second substrate pose deviation vector. S42. Convert the translation residual and rotation angle residual in the second substrate pose deviation vector into equivalent linear displacements and then calculate the norm to obtain the residual pose error. Compare the residual pose error with the convergence threshold. If the residual pose error is less than the convergence threshold, output the alignment ready signal; otherwise, return the second substrate pose deviation vector to S41 to continue iterative compensation.

10. A fully automated dual-station lithography alignment device based on a multi-camera vision system, characterized in that, A fully automated dual-station lithography machine alignment method based on a multi-camera vision system as described in any one of claims 1-9, comprising: The calibration module is used to jointly calibrate four cameras and calculate the target homography matrix from the pixel coordinates of each camera to the machine coordinate system. The positioning module is used to acquire images in parallel by the four cameras to obtain four images, and to perform normalized cross-correlation positioning and elliptic least squares fitting on the alignment marks in each image to obtain the first coordinates of the four sets of alignment marks. The pose solving module is used to map the first coordinates to the second coordinates of four sets of alignment marks in the machine coordinate system according to the target homography matrix, and to substitute the second coordinates and the corresponding design coordinates into the overdetermined equation system to solve the first substrate pose deviation vector. The output module is used to control the stage to perform compensation motion and iteratively execute the steps of the positioning module-pose solving module according to the first substrate pose deviation vector until the residual pose error is less than the convergence threshold, and output the alignment ready signal.

Images