Artificial intelligence-based etching process evaluation method and system

Abstract

This invention relates to the field of etching process evaluation technology, specifically disclosing an etching process evaluation method and system based on artificial intelligence. The method involves acquiring a sequence of top-view scanning electron microscope (SEM) images and a sequence of oblique-angle SEM images of the etched wafer surface. First, the top-view images undergo grayscale normalization and edge detection to extract an initial edge point set. Then, based on the grayscale gradient decay characteristics within the edge influence zone, the feathering morphology is divided and the feathering width index is calculated. Subsequently, the shrinkage step size and maximum iteration count are adaptively determined according to the morphology type, and the final boundary is output through a bidirectional boundary optimization loop. The final boundary is then mapped to the oblique image sequence to extract the sidewall grayscale profile, calculate the sidewall verticality deviation and sidewall roughness, and combine the distance between the final boundary and the mask design boundary to obtain the effective etching boundary offset. Finally, the above parameters are fused to output the over-etching damage index, thereby achieving a joint evaluation of the feathering edge, sidewall morphology, and over-etching degree.

Description

Technical Field

[0001] This invention belongs to the field of etching process evaluation technology, and relates to an etching process evaluation method and system based on artificial intelligence. Background Technology

[0002] In deep silicon etching and metal etching processes, as etching depth increases and sidewall perpendicularity requirements become more stringent, the etching results are no longer simply characterized by contour expansion or linewidth changes. Instead, they are accompanied by byproduct deposition, local redeposition, and uneven edge energy distribution, resulting in feathered morphologies with transition zone characteristics near the mask edges. Under a scanning electron microscope, these feathered edges typically exhibit gradual grayscale changes, boundary tailing, and localized burr-like expansion. While the surface may still maintain a planar width profile similar to the designed pattern, the actual sidewall perpendicularity, edge integrity, and local roughness have often changed. For high aspect ratio structures, even minor deviations in edge morphology can further affect hole wall stability, subsequent deposition uniformity, and device electrical consistency. Therefore, linewidth alone cannot fully reflect the etching quality.

[0003] Meanwhile, traditional etching inspection relies heavily on conventional linewidth measurement, fixed-threshold edge extraction, and single-view image judgment. It primarily focuses on whether the graphic boundary falls within the design dimensions, but rarely distinguishes between the true etched contour and the feathered transition zone formed by byproducts. In actual images, feathered edges often have similar spatial positions to the effective contour, and their grayscale changes are gradually distributed, making them easily identified as valid parts of the graphic by conventional algorithms. This makes the measurement results appear to meet dimensional requirements, while actually masking problems such as sidewall skew, boundary drift, and localized over-etching. In this situation, process evaluation tends to remain at the surface dimension level, failing to reflect edge distortion and over-etching trends in a timely manner, and making it difficult to provide a reliable basis for early adjustment of process parameters.

[0004] Furthermore, existing methods, when faced with complex feathered edges, typically lack a continuous judgment mechanism regarding the correlation between the degree of edge gradation and the risk of over-etching. They tend to directly output linewidth results after boundary identification, ignoring the gradation diffusion region that still exists outside the edge. As a result, while the detection result can provide a dimensional value, it cannot distinguish whether this value is formed by the actual etched contour or by the apparent boundary formed by feathering byproducts and grayscale tailing. Especially in deep silicon or metal etching, byproduct adhesion and local redeposition often precede dimensional deviations. Without more detailed identification and iterative judgment of feathered edges, it can easily lead to delayed detection of process conditions, affecting timely correction of over-etching and reducing the accuracy of subsequent yield analysis and process optimization. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an etching process evaluation method and system based on artificial intelligence to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides an artificial intelligence-based method for evaluating etching processes, the method comprising: Obtain top-view scanning electron microscope (SEM) image sequences and tilt-angle SEM image sequences of the etched wafer surface; The image sequence of top-view scanning electron microscope is subjected to grayscale normalization to obtain a standardized image, and edge detection is performed on the standardized image to extract the initial edge point set; Based on the initial edge point set, extend outwards along the normal direction by a preset pixel distance to construct an edge influence band. Calculate the grayscale gradient magnitude of each pixel within the edge influence band. Based on the attenuation characteristic of the grayscale gradient magnitude with distance from the edge, classify the feathered edge into a first morphological type, a second morphological type, or a third morphological type. Simultaneously calculate the feathering width index, which is the distance from the initial edge point set to the point where the grayscale gradient magnitude drops to the background noise threshold. After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, the bidirectional boundary optimization loop is entered and repeated until the boundary change is less than the convergence threshold, at which point the final boundary is output. The final boundary is mapped to a scanning electron microscope image sequence at an inclined angle, and the gray-scale profile of the sidewall at the corresponding position of the final boundary is extracted. The sidewall verticality deviation and sidewall roughness are calculated based on the sidewall gray-scale profile. Calculate the effective etch boundary offset, which is the distance between the final boundary and the mask design boundary; calculate the over-etch damage index based on the effective etch boundary offset, sidewall perpendicularity deviation, and sidewall roughness, and output the over-etch damage index.

[0007] A second aspect of the present invention provides an artificial intelligence-based etching process evaluation system, the system comprising: Edge point set acquisition module: acquires the top-view scanning electron microscope image sequence and the tilt angle scanning electron microscope image sequence of the etched wafer surface; performs grayscale normalization processing on the top-view scanning electron microscope image sequence to obtain a standardized image, and performs edge detection on the standardized image to extract the initial edge point set; Feathered edge segmentation module: Based on the initial edge point set, extend outward along the normal direction by a preset pixel distance to construct the edge influence band, calculate the gray-level gradient magnitude of each pixel in the edge influence band, and divide the feathered edge into the first morphological type, the second morphological type, or the third morphological type according to the attenuation characteristic of the gray-level gradient magnitude with the distance from the edge, and calculate the feathering width index at the same time. Etching boundary output module: After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, it enters a bidirectional boundary optimization loop and repeats the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, and then stops and outputs the final boundary. Etching damage assessment module: Maps the final boundary to a scanning electron microscope image sequence at an inclined angle, extracts the sidewall grayscale profile at the corresponding position of the final boundary, calculates the sidewall verticality deviation and sidewall roughness based on the sidewall grayscale profile, calculates the effective etching boundary offset, where the effective etching boundary offset is the distance between the final boundary and the mask design boundary, calculates the over-etching damage index based on the effective etching boundary offset, sidewall verticality deviation and sidewall roughness, and outputs the over-etching damage index.

[0008] As described above, the etching process evaluation method and system based on artificial intelligence provided by the present invention have at least the following beneficial effects: This invention also constructs an edge influence zone along the normal direction based on the initial edge point set, and classifies the feathered edges according to the attenuation characteristics of the gray-level gradient amplitude with edge distance. At the same time, it calculates the feathering width index, and then adaptively determines the shrinkage step size and maximum number of iterations according to the feathering morphology type and enters a bidirectional boundary optimization loop until the boundary change meets the convergence condition and outputs the final boundary. This dynamic optimization mechanism driven by feathering features, compared with the boundary update method with fixed threshold and fixed step size, can automatically adjust the boundary update intensity for different etching morphologies. It avoids the problem of premature convergence of boundary extraction when the feathering range is large, and also avoids the problem of repeated boundary oscillation when the feathering range is small. Therefore, this invention can more accurately lock the real etching boundary, improve the consistency between the final boundary and the actual process profile, and provide a higher precision spatial benchmark for subsequent over-etching damage evaluation.

[0009] This invention further maps the final boundary onto a scanning electron microscope image sequence at an inclined angle, extracts the sidewall grayscale profile at the corresponding position, and calculates the sidewall verticality deviation and sidewall roughness accordingly. Simultaneously, it combines the distance between the final boundary and the mask design boundary to obtain the effective etching boundary offset. These parameters are then fused to calculate the over-etching damage index. This unified evaluation of boundary offset, sidewall tilt, and sidewall micro-undulations reflects the degree of over-etching damage from both macroscopic positional deviation and microscopic morphological defects. Compared to traditional evaluation methods that rely solely on a single dimensional offset, this approach is more effective in identifying complex process anomalies such as sidewall collapse, boundary expansion, and roughening. Therefore, the over-etching damage index output by this invention not only has stronger process interpretability but also provides a more direct and stable quantitative basis for etching parameter correction, process window adjustment, and abnormal wafer screening, thereby improving the accuracy and practicality of wafer etching quality control. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0012] Figure 2 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0013] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0014] In traditional wafer etching inspection systems, fixed edge determination rules and single-view image analysis methods cannot adapt to the non-uniform variations in sidewall morphology across different regions after etching. When the wafer edge exhibits a feathered transition in the top-view image, if the system extracts the boundary solely based on a static threshold, it easily blends the actual etched edge with background noise, causing the final boundary to deviate from the actual processing contour. Furthermore, without supplementary observation of the sidewalls using tilted-angle images, it cannot accurately reflect sidewall verticality deviation and sidewall roughness, thus lacking a reliable basis for over-etching damage assessment. This static determination mechanism reduces the stability of boundary extraction, causing subsequent damage index calculations to deviate from the actual process conditions, ultimately affecting the correction effect of etching process parameters.

[0015] For example, in a set of over-etched wafer samples, the top-view scanning electron microscope (SEM) image sequence shows a significant expansion in the attenuation range of the edge grayscale gradient from the boundary outwards, while the sidewall grayscale profile in the tilt-angle SEM image sequence exhibits obvious nonlinear fluctuations. If a uniform fixed shrinkage step size and a fixed number of iterations are still used for boundary updates, the system will misclassify the true boundary within the wider feathered area as a noise region, causing the final boundary to shift inwards. Simultaneously, the sidewall verticality deviation and sidewall roughness are underestimated because they are not fully mapped to the boundary position, making the over-etching damage index unable to accurately reflect process defects such as hole wall collapse, sidewall jaggedness, and boundary diffusion. Consequently, the output process evaluation result will incorrectly indicate that the etching state is normal, thus affecting subsequent process compensation judgments.

[0016] If the above problems are not addressed, the deviation between the final boundary and the actual etched contour will directly affect the calculation of the effective etched boundary offset, making the damage assessment unable to accurately reflect the spatial range of excessive etching. If the feathering morphology classification lacks a linkage with the boundary update strategy, it cannot automatically match the shrinkage step size and maximum iteration count for different attenuation types, easily causing the boundary optimization process to stop prematurely or be over-updated, thus leading to boundary oscillations. If the sidewall grayscale profile is not accurately mapped to the tilted image sequence, it is difficult to stably extract the sidewall verticality deviation and roughness from grayscale changes, ultimately causing the weight fusion of the over-etching damage index to lose its unified scale, reducing the ability to identify abnormal etching states.

[0017] When faced with the aforementioned problems, traditional boundary extraction methods use a fixed step size update, resulting in edges with large feathering ranges being quickly skipped, while edges with small feathering ranges are updated too finely. To address this, this application couples the feathering width index with the feathering morphology type, automatically generating the shrinkage step size and maximum iteration count based on the gray-level gradient decay characteristics within the edge influence zone, and performing bidirectional boundary correction both internally and externally through a bidirectional boundary optimization loop. Further analysis reveals that relying solely on top-view images cannot accurately reflect sidewall verticality deviation and roughness. It is necessary to map the final boundary to a scanning electron microscope image sequence at an inclined angle, extract the corresponding sidewall gray-level profile, and construct an over-etching damage index based on boundary offset, sidewall verticality deviation, and sidewall roughness, thereby solving the boundary mismatch and damage misjudgment problems under the static thresholding mechanism.

[0018] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Example 1: Please see Figure 1 As shown, the AI-based etching process evaluation method specifically includes the following steps: Obtain top-view scanning electron microscope (SEM) image sequences and tilt-angle SEM image sequences of the etched wafer surface.

[0020] In this embodiment of the invention, a preliminary image of a single frame of rapid scanning is extracted from each target area to be tested on the wafer surface. The grayscale standard deviation of the background area is calculated as an estimate of the environmental noise, and the average grayscale value of the target etched area is calculated using the formula... Calculate the signal-to-noise ratio of a single frame , The average grayscale value of the etched structure region in the initial frame image represents the signal intensity and is dimensionless. The standard deviation of grayscale values ​​for the patternless background region of the image is given, representing the intensity of random noise, which is dimensionless. Then, the optimal number of scan frames N required for each target area is calculated using the following formula: ,in This parameter is set as a dimensionless empirical constant to represent the preset target signal-to-noise ratio. After determining the optimal number of scan frames, the scanning electron microscope is controlled to perform N consecutive repeated scans of the current target area from a top-down angle, and a pixel-level cumulative averaging formula is used. Perform fusion computing, in which This represents the grayscale value at the two-dimensional pixel coordinates (x, y) in the output normalized top-view image. The original gray value at the position corresponding to the i-th scan frame is used to generate a denoised top-view scanning electron microscope image sequence by sequentially traversing all target areas. Next, in order to capture the three-dimensional morphological features of the etched sidewalls, the sample stage of the scanning electron microscope is rotated along a preset horizontal tilt axis (set as the X-axis of the image) to a tilt angle. During image acquisition, the tilted orientation causes a physical compression distortion in the projection of the imaging plane onto the pixel plane along a direction perpendicular to the tilt axis (i.e., the Y-axis), making circular holes appear as ellipses. Therefore, geometric perspective compensation must be performed on the acquired original tilted image. The compensation mapping formula is constructed as follows: ,in To compensate for the gray-level distribution function of the restored tilt angle image at coordinates (x, y), This represents the grayscale distribution data of the raw distorted image directly output from the microscope. x and y are both pixel coordinate grid indices of the image, in pixels. This represents the actual tilt angle of the sample stage, in degrees.

[0021] The image sequence of top-view scanning electron microscope is subjected to grayscale normalization to obtain a standardized image. Edge detection is then performed on the standardized image to extract an initial set of edge points. Specifically, this includes: Traverse all pixels in the top-view scanning electron microscope image sequence and extract the global highest and lowest gray values; The grayscale range is calculated using the highest and lowest global grayscale values. Based on the grayscale range, the initial grayscale values ​​of each pixel in the top-view scanning electron microscope image sequence are scaled proportionally and mapped to a grayscale range with a preset number of bits to generate a standardized image. Extract the gray-level change rate of each pixel in the standardized image in the horizontal and vertical dimensions, and combine the gray-level change rates in the horizontal and vertical dimensions to calculate the gray-level gradient magnitude and gray-level gradient direction of each pixel. The extreme values ​​of the gray-level gradient magnitude of each pixel are filtered along the gray-level gradient direction. Pixels whose gray-level gradient magnitude is greater than that of their neighboring pixels in the current gray-level gradient direction are retained as candidate edge points. Calculate the average gray-level gradient magnitude of all candidate edge points, and then spatially stitch together the candidate edge points whose gray-level gradient magnitude is greater than the average gray-level gradient magnitude to extract the initial edge point set.

[0022] In this embodiment of the invention, all two-dimensional spatial pixels in the top-view scanning electron microscope image sequence are traversed, and the global highest gray value representing the brightest physical feature and the global lowest gray value representing the darkest physical feature are extracted. The global gray range is calculated using the difference between the global highest and global lowest gray values. Considering that when scanning electron microscopes scan structures with high aspect ratios, the bottom of holes often exhibits an extremely dark gray distribution due to the difficulty of secondary electron escape, a nonlinear stretching variant model is used to proportionally scale and map the initial gray values ​​of each pixel in the top-view scanning electron microscope image sequence. The mapping formula is as follows: ,in This represents the pixel grayscale value at coordinates (x, y) of the normalized image output. These are the initial grayscale values ​​of each pixel in the image sequence. and These are the highest and lowest global grayscale values ​​extracted during the traversal, respectively. is the shadow feature enhancement coefficient, which is a dimensionless parameter, and n is the preset number of bits; Subsequently, the gray-level change rate of each pixel in the standardized image in the horizontal and vertical dimensions is extracted, and the gray-level partial derivatives in the horizontal direction are calculated using a discrete difference operator based on the principle of first-order partial derivatives. Partial derivative of grayscale with respect to the vertical direction Both parameters are measured in gray levels per pixel. The gray-level gradient magnitude and direction of each pixel are calculated by combining the gray-level change rates in the horizontal and vertical dimensions. The combined calculation formula is as follows: ,in This represents the grayscale gradient magnitude, with the dimension maintained as grayscale levels per pixel. The gray-level gradient direction is defined in radians. Then, along the calculated gray-level gradient direction, extreme values ​​of the gray-level gradient magnitude at each pixel are selected. Specifically, a bilinear interpolation algorithm is used to obtain the current pixel's position along the gray-level gradient direction. The gradient magnitudes of the virtual adjacent positions on both sides are compared strictly with the gray-level gradient magnitudes of the current pixel and the two virtual adjacent pixels. Only pixels whose gray-level gradient magnitudes are greater than those of their adjacent pixels in the current gray-level gradient direction are retained as candidate edge points. If the gray-level gradient magnitudes are not greater, their gray-level gradient magnitudes are set to zero and removed. This achieves non-maximum suppression localization at the single-pixel level. Finally, the arithmetic mean of the gray-level gradient magnitudes of all candidate edge points is calculated. The value of this arithmetic mean is used as an adaptive truncation threshold. Candidate edge points whose gray-level gradient magnitudes are greater than the average gray-level gradient magnitude are subjected to spatial connectivity component splicing and morphological closure processing to remove free pseudo-edge isolated points, thereby extracting a continuous and accurately localized initial edge point set.

[0023] Based on the initial edge point set, an edge influence band is constructed by extending outward by a preset pixel distance along the normal direction. The gray-level gradient magnitude of each pixel within the edge influence band is calculated. Based on the attenuation characteristic of the gray-level gradient magnitude with distance from the edge, the feathered edge is divided into a first morphological type, a second morphological type, or a third morphological type. At the same time, the feathering width index is calculated, which is the distance from the initial edge point set to the point where the gray-level gradient magnitude drops to the background noise threshold.

[0024] Preferably, the feathering edge is divided into corresponding morphological types, including a first morphological type, a second morphological type, and a third morphological type. Simultaneously, a feathering width index is calculated, including: Extract the gray-level gradient direction corresponding to each pixel in the initial edge point set, use the gray-level gradient direction as the normal direction, and extend the extension trajectory outward from the initial edge point set along the normal direction by a preset pixel distance to generate the extension trajectory. Combine all the extension trajectories to construct the edge influence zone. Extract the pixels contained in each extended trajectory within the edge influence zone, obtain the grayscale gradient magnitude of each pixel on the extended trajectory, and calculate the number of pixels between each pixel on the extended trajectory and the starting point of the initial edge point set. Use the number of pixels between them as the spatial distance. By combining the gray-level gradient magnitude of each pixel within the edge influence zone with the corresponding spatial distance, the ratio of the decrease in gray-level gradient magnitude to the spatial distance is calculated to obtain the gray-level gradient decay rate. The gray-level gradient decay rate is used as the decay feature of gray-level gradient magnitude with edge distance. Construct a mapping relationship between the numerical distribution range of gray-level gradient decay rate and morphological type, determine the numerical distribution range into which the gray-level gradient decay rate of each extended trajectory falls, and assign the corresponding morphological type to the feathering edge according to the numerical distribution range and the mapping relationship. The morphological types include the first morphological type, the second morphological type and the third morphological type. Collect background pixels in the standardized image that are not covered by edge influence bands, calculate the average gray-level gradient magnitude of all background pixels, and use the average gray-level gradient magnitude as the background noise threshold.

[0025] Along the extended trajectory, the grayscale gradient magnitude of each pixel is compared with the background noise threshold. Pixels whose grayscale gradient magnitude decreases to the background noise threshold are extracted as boundary cutoff points. The spatial distance corresponding to the boundary cutoff point is obtained, and this spatial distance is used as the feathering width index.

[0026] Preferably, the mapping relationship between the numerical distribution range of grayscale gradient decay rate and morphological type is constructed, including: Extract the gray-level gradient decay rate of all extended trajectories within the edge influence band, and arrange all gray-level gradient decay rates in ascending order to generate a gray-level gradient decay rate sequence. Calculate the mean and standard deviation of all gray-level gradient decay rates in the gray-level gradient decay rate sequence; The difference between the mean and the standard deviation is used as the first boundary node, and the sum of the mean and the standard deviation is used as the second boundary node. The overall numerical range of the gray-level gradient decay rate sequence is truncated and divided using the first boundary node and the second boundary node, and the first numerical interval, the second numerical interval and the third numerical interval are divided in sequence. The first numerical interval, the second numerical interval and the third numerical interval are collectively used as the gray-level gradient decay rate numerical distribution interval. Establish a correspondence between the first numerical interval and the first morphological type, establish a correspondence between the second numerical interval and the second morphological type, and establish a correspondence between the third numerical interval and the third morphological type. Combine all correspondences to generate a mapping relationship between the grayscale gradient decay rate numerical distribution interval and the morphological type.

[0027] In this embodiment of the invention, the gray-level gradient direction of each pixel in the initial edge point set is first extracted. Since the gray-level gradient direction is naturally perpendicular to the image feature contour in a physical sense, the system directly uses it as the normal direction. Starting from each reference point in the initial edge point set, the system extends a preset pixel distance along the corresponding normal direction towards the outer edge of the image (i.e., the background area where the gray-level value decreases), thereby generating a series of radial extension trajectories. The preset pixel distance mentioned here is recommended to be set to an adjustable range of 20 to 50 pixel units. In order to ensure complete coverage of the sidewall gradient transition area under various etching conditions, and to avoid spatial overlap and data pollution with the gray-level edges of adjacent deep hole structures, a setting of 30 pixel units is preferred. By spatially combining the extension trajectories generated by all reference points on a two-dimensional plane, an edge influence band enveloping the initial edge can be constructed.

[0028] Next, the system sequentially extracts the pixel sequences contained in each extended trajectory within the edge influence zone, obtains the grayscale gradient magnitude of these pixels, and directly converts the pixel spacing between each pixel on the extended trajectory and the initial edge starting point of the trajectory into spatial distance by calculating the number of pixels between them. To quantify the blurring degree of the edge contour, the grayscale gradient decay rate is calculated based on the above data. The decay rate extraction formula is constructed as follows:

[0029] in The calculated gray-level gradient decay rate characterizes the decay characteristics of the gray-level gradient magnitude with edge distance. The grayscale gradient magnitude is the starting point of the extended trajectory (i.e., the initial edge point), with the dimension being grayscale level per pixel; The grayscale gradient magnitude of the current pixel on the extended trajectory is also measured in grayscale levels per pixel. This represents the spatial distance between the current pixel and the starting point, measured in pixels.

[0030] Based on this, the feathering morphology types of the edges are statistically refined and mapped. The system extracts the gray-level gradient decay rate of all extended trajectories within the edge influence zone and arranges them in ascending order to generate a gray-level gradient decay rate sequence. The arithmetic mean and standard deviation of all values ​​in the sequence are calculated using the standard deviation grading principle. The difference between the mean and the standard deviation is set as the first boundary node, and the sum of the two is set as the second boundary node. These two purely data-driven dynamic nodes objectively truncate the overall numerical range, successively dividing it into a first numerical interval smaller than the first boundary node, a second numerical interval between the two nodes, and a third numerical interval larger than the second boundary node. These three intervals together constitute the gray-level gradient decay rate numerical distribution interval.

[0031] Simultaneously, a mapping relationship based on the etching physics mechanism is established, corresponding to the first numerical range and the first morphological type. Because an extremely small attenuation rate falls within the first numerical range, it means the gradient amplitude decreases very slowly. Physically, this corresponds to the excessive polymer deposition during etching, resulting in a very gentle "wide slope" morphology on the sidewalls, thus defining it as the first morphological type (gentle feathering morphology). The second numerical range is then correlated with the second morphological type. Here, the attenuation rate is within the mean fluctuation range, corresponding to the "standard taper" sidewall morphology expected by the process, thus defining it as the second morphological type (standard transition morphology). The third numerical range is correlated with the third morphological type. A very large attenuation rate indicates a precipitous drop in gradient amplitude within a very short space, physically mapped to a "steep or even concave" edge caused by the transverse isotropic drilling of the etching gas, thus defining it as the third morphological type (steep truncation morphology). By traversing and determining the numerical distribution range of the grayscale gradient attenuation rate of each extended trajectory, and strictly following the above mapping relationship, the corresponding morphological type can be accurately and automatically assigned to each feathered edge on the image.

[0032] Finally, to determine the absolute boundary location of the sidewall gradient region, the system collects smooth background pixels not covered by the edge influence band in the standardized image and calculates the average gray-level gradient magnitude of all background pixels as the background noise threshold (in gray levels per pixel). The system compares the gray-level gradient magnitude of each pixel with the background noise threshold sequentially from the inside out along each extension trajectory, accurately capturing the pixel coordinates where the gray-level gradient magnitude first decreases to and no longer exceeds the background noise threshold, and using these as boundary cutoff points. The spatial distance between this boundary cutoff point and the initial edge starting point is extracted and defined as the feathering width exponent (in pixels). This exponent, together with the assigned morphology type, constitutes the core geometric parameter set characterizing the wafer over-etching state.

[0033] After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, the system enters a bidirectional boundary optimization loop and repeats the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, at which point the final boundary is output.

[0034] Preferably, the bidirectional boundary optimization loop includes: shrinking the current boundary inward along the normal direction by the shrinkage step size to obtain a shrunken boundary; calculating a first deviation between the average line width and the design line width within the area enclosed by the shrunken boundary; expanding the current boundary outward along the normal direction by the shrinkage step size to obtain an expanded boundary; calculating a second deviation between the average line width and the design line width within the area enclosed by the expanded boundary; comparing the first deviation and the second deviation; if the first deviation is less than the second deviation, selecting a shrinkage direction to update the boundary; if the second deviation is less than the first deviation, selecting an expansion direction to update the boundary; if the difference between the first deviation and the second deviation is less than a convergence threshold, marking the current boundary as a candidate boundary; repeating the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, and then outputting the final boundary.

[0035] Preferably, the shrinkage step size and maximum number of iterations are determined based on the feathering morphology type, including: Extract the morphology type corresponding to each extended trajectory within the edge influence zone, use the morphology type as the feathering morphology type, and obtain the feathering width index corresponding to each extended trajectory. Establish a mapping relationship between feathering morphology types and basic parameter configuration tables. In the basic parameter configuration tables, configure the corresponding basic step size coefficients and basic iteration coefficients for the first morphology type, the second morphology type, and the third morphology type, respectively. The basic step size coefficient and basic iteration coefficient corresponding to the feathering morphology type of each extended trajectory are obtained in the basic parameter configuration table through association mapping. Calculate the product of the feathering width index and the base step size coefficient to obtain the initial step size value. Round the initial step size value down to generate the shrinkage step size of the current extended trajectory. The feathering width index is divided by the shrinkage step size, and the quotient is extracted as the number of spatial segments. The product of the number of spatial segments and the basic iteration coefficient is calculated, and the product is rounded up. The rounded value is determined as the maximum number of iterations.

[0036] Preferably, the comparison between the first deviation and the second deviation in the bidirectional boundary optimization loop specifically includes: Extract the first deviation generated in the contraction direction and the second deviation generated in the expansion direction of the current boundary, subtract the first deviation from the second deviation, and obtain the absolute value of the calculation result as the absolute value of the difference. The absolute value of the difference is compared with the convergence threshold. If the absolute value of the difference is less than the convergence threshold, the evolution process of the current boundary is determined to have reached a stable state. The coordinates of all pixels contained in the current boundary are extracted, and the current boundary is marked as a candidate boundary. Under the condition that the absolute value of the difference is not less than the convergence threshold, the numerical values ​​of the first deviation and the second deviation are further compared, and the corresponding boundary update operation is performed according to the comparison result. Under the condition that the first deviation is less than the second deviation, it is determined that the current boundary has an inward evolution trend. The associated shrinkage step size is extracted, and the current boundary is moved along the shrinkage direction according to the value of the shrinkage step size to generate an updated boundary. If the second deviation is less than the first deviation, it is determined that the current boundary has a tendency to evolve outward. The current boundary is moved along the expansion direction according to the value of the contraction step size to generate an updated boundary.

[0037] In this embodiment of the invention, a bidirectional boundary optimization evolutionary process based on adaptive step size is entered. First, based on the spatial uncertainty differences of different edge types, the shrinkage step size and maximum number of iterations for optimization evolution are dynamically determined. Specifically, the feathering morphology type corresponding to each extended trajectory within the edge influence zone and the previously calculated feathering width index (in pixels) are extracted, and the corresponding basic step size coefficient is matched in a pre-built basic parameter configuration table. and basic iteration coefficients Both parameters are dimensionless empirical constants; their underlying logic is that for a gently feathering morphology (first morphology type), due to the wide and large span of the edge blurring band, It is recommended to set the value to the range of 0.1 to 0.15 to avoid the optimization process getting stuck in a local stagnation; while for steep truncation patterns (third type), the edge uncertainty is small. The optimal value is set to the range of 0.02 to 0.05 for fine-tuning and positioning; the basic iteration coefficient... This serves as a safety redundancy factor to prevent the algorithm from looping infinitely, and is typically taken between 1.5 and 2.0. Next, the system constructs an adaptive step size calculation formula. ,in The shrinking step size of the current extended trajectory generated after rounding down, with the dimension being pixels; Let be the input feathering width exponent. Then, calculate the maximum number of iterations, using the formula: ,in The maximum number of iterations is the maximum number generated after rounding up, and is a dimensionless positive integer. This adaptive computing mechanism can dynamically adjust the allocation of computing resources according to the actual diffusion degree of local edges, which greatly improves the execution efficiency of the algorithm while ensuring convergence accuracy.

[0038] After determining the evolution parameters, the system constructs the current boundary based on the initial set of edge points and formally initiates a bidirectional boundary optimization loop. Utilizing morphological operation principles, the current boundary is extended along the normal direction towards both the inner and outer sides of the hole at distances of [missing information]. The system generates contraction and expansion boundaries through equidistant translation. It then calculates the feature deviation under the current boundary state using an equivalent linewidth variant model based on surface domain integrals. The first deviation evaluation formula is... ,in The first deviation is measured in pixels. For a two-dimensional closed integral region enclosed by a shrinking boundary, the double integral result represents the pixel area of ​​that region (in square pixels). The length of the center line of the skeleton feature of the current boundary (in pixels) is calculated by dividing the two to obtain the equivalent average line width (in pixels). The reference value for the design linewidth (in pixels) is obtained by parsing the layout design file and converting it according to the microscope resolution.

[0039] Similarly, the system uses the same surface integral model to calculate the second deviation generated by comparing the average linewidth of the region enclosed by the expansion boundary with the designed linewidth. After obtaining the bidirectional bias, the system extracts and calculates the absolute value of the difference between the first bias and the second bias. The difference is then compared to a preset convergence threshold, which represents the limit of the etching measurement system's tolerance error in pixels. If the absolute value of the difference is less than the convergence threshold, the system determines that the evolution of the current boundary has reached the optimal equilibrium stable state (i.e., a local minimum on the optimization surface) approaching the design goal. It then stops the current trajectory loop, extracts the coordinates of all pixels contained in the current boundary, and marks them as candidate boundaries. If the absolute value of the difference is not less than the convergence threshold, it indicates that the system still has room for optimization. The system further compares the absolute values ​​of the two deviations to determine the next direction of morphological evolution: if the first deviation is less than the second deviation, it determines that the current real physical boundary tends towards a contraction trend inside the hole, and the system extracts the associated contraction step size. The system pushes the current boundary inward along the contraction direction to generate the updated boundary for the next iteration. Conversely, if the second deviation is less than the first deviation, an outward expansion trend is identified, and the current boundary is expanded outward along the expansion direction with the same phase length to generate the updated boundary. The system continuously repeats the above evolutionary cycle of bidirectional deviation evaluation and unidirectional position update until the absolute value of the difference corresponding to the boundary change is less than the convergence threshold or the iteration count reaches the maximum number of iterations. The system is forced to stop at a certain time, thus accurately outputting the final boundary that excludes interference from the fuzzy feathering band.

[0040] The final boundary is mapped to a sequence of scanning electron microscope images at an inclined angle. The grayscale profile of the sidewall at the corresponding location of the final boundary is extracted. Based on the grayscale profile, the sidewall verticality deviation and sidewall roughness are calculated, specifically including: Obtain the planar coordinates of each pixel in the final boundary, extract the spatial tilt angle parameters of the tilt angle scanning electron microscope image sequence at the time of acquisition, construct a coordinate transformation matrix using the spatial tilt angle parameters, and transform the planar coordinates to the tilted three-dimensional coordinate system corresponding to the tilt angle scanning electron microscope image sequence through the coordinate transformation matrix to generate the tilted mapping boundary. Along the vertical extension direction of the tilted mapping boundary, the gray values ​​of the corresponding pixels are extracted layer by layer in the tilt angle scanning electron microscope image sequence. The gray values ​​of each corresponding pixel are arranged and combined according to the depth level to construct the gray profile of the sidewall at the corresponding position of the tilted mapping boundary. Extract global grayscale peak points and global grayscale valley points from the grayscale profile of the sidewall. Mark the depth level where the global grayscale peak points are located as the top feature position and the depth level where the global grayscale valley points are located as the bottom feature position. Calculate the horizontal offset distance and vertical depth difference between the top feature position and the bottom feature position. The horizontal offset distance is divided by the difference between the vertical depth, and the quotient is extracted as the sidewall verticality deviation. Construct an ideal sidewall contour line that is linearly connected between the top feature position and the bottom feature position, and obtain the reference gray value of the ideal sidewall contour line at each depth level. Calculate the absolute value of the difference between the actual gray value and the corresponding reference gray value of the sidewall gray profile at each depth level. Sum the absolute values ​​of the differences at all depth levels and calculate the average. Determine the average value as the sidewall roughness.

[0041] In this embodiment of the invention, the planar coordinates of each two-dimensional pixel in the final boundary under the top view plane are first obtained. Its dimension is pixels, and it also extracts the actual tilt angle of the sample stage set when the device acquires the tilt angle scanning electron microscope image sequence. The system utilizes the principle of affine geometric projection to construct a coordinate transformation matrix. The specific form of this matrix is ​​a main diagonal with elements 1 and 2. For a two-dimensional diagonal matrix, since the trigonometric function values ​​are dimensionless constants, the operator itself is dimensionless. This can be achieved through a matrix multiplication model. The coordinates of the two-dimensional orthogonal projection plane are stretched and transformed into the perspective space of the tilted viewpoint, thereby generating the corresponding coordinates of the tilted mapping boundary. .

[0042] Subsequently, using the tilted mapping boundary as the initial projection reference, the system traverses the tilted angle scanning electron microscope image sequence layer by layer downwards along the vertical extension direction representing the physical etching depth in the image. The system acquires the grayscale values ​​of each pixel at the corresponding coordinate position, and sets a depth level index z (in pixels) using the downward pixel interval as the depth reference. The extracted grayscale values ​​are then arranged and combined in ascending order of the depth levels to construct a sidewall grayscale profile sequence reflecting the mapping relationship between etching depth and the intensity of the secondary electron signal. The output grayscale values ​​are dimensionless grayscale levels.

[0043] Based on the physical mechanism of morphological contrast in scanning electron microscopy, the top surface of a high aspect ratio porous structure exhibits extremely high brightness edge enhancement due to the large solid angle of secondary electron emission, while the deep bottom exhibits extremely dark grayscale distribution because electrons are largely absorbed and trapped by the sidewalls. The system systematically extracts grayscale profile sequences from the sidewalls. The location of the highest global grayscale value is taken as the global grayscale peak point, and its corresponding depth level is marked as the top feature location. (Unit: pixels); Similarly, extract the location of the lowest global grayscale value as the global grayscale valley point, and mark its corresponding depth level as the bottom feature location. (Dimensions are pixels). The system calculates the horizontal offset distances of the top and bottom feature positions in the transverse imaging plane. and the difference in vertical depth between the two in the direction of extension. Both of these parameters are measured in pixels. The system uses a variation of the tangent relationship of geometric projection to divide the horizontal offset distance by the vertical depth difference, and extracts the quotient as the sidewall verticality deviation. .

[0044] To further quantify the micro-geometric undulations of the sidewalls caused by etching plasma ion bombardment and uneven byproduct deposition, the system needs to establish a reference line for smooth sidewalls between the top and bottom feature locations. Since the secondary electron yield of an absolutely smooth and uniform sidewall structure should exhibit a relatively stable decay law with depth, the system constructs an ideal sidewall profile with a linear connection between the top and bottom feature locations and uses interpolation formulas... Calculate the reference grayscale value at each depth level z. Next, the system calculates the actual grayscale values ​​of the sidewall grayscale profile sequence at the actual depth level, layer by layer. and corresponding reference grayscale value The absolute value of the difference between them is calculated using the cumulative average formula. Perform a summary calculation on it, where To calculate the total number of depth levels contained within the interval, This refers to the total number of pixel layers contained from the top feature position to the bottom feature position; the calculated average is then determined as the sidewall roughness. .

[0045] Calculate the effective etch boundary offset, which is the distance between the final boundary and the mask design boundary; calculate the over-etch damage index based on the effective etch boundary offset, sidewall perpendicularity deviation, and sidewall roughness, and output the over-etch damage index.

[0046] Preferably, the calculated etching damage index includes: Obtain the layout design file of the target processing area, extract the theoretical outline of the target processing area from the layout design file, and use the theoretical outline as the mask design boundary. Traverse all boundary pixels contained on the final boundary, calculate the shortest plane distance from each boundary pixel to the mask design boundary, sum the shortest plane distances corresponding to all boundary pixels and calculate the average value, and use the calculated average value as the effective etch boundary offset. The calculated sidewall verticality deviation and sidewall roughness are extracted, and the effective etching boundary offset, sidewall verticality deviation and sidewall roughness are linearly mapped and transformed using a standardized numerical range with a preset number of bits, respectively, to generate normalized offset, normalized verticality deviation and normalized roughness that have eliminated the difference in dimensions. Obtain the pre-established process evaluation weight library, and extract the offset weight parameter that matches the normalized offset, the verticality weight parameter that matches the normalized verticality deviation, and the roughness weight parameter that matches the normalized roughness from the process evaluation weight library. The first damage score is obtained by multiplying the normalized offset by the offset weight parameter, the second damage score is obtained by multiplying the normalized perpendicularity deviation by the perpendicularity weight parameter, and the third damage score is obtained by multiplying the normalized roughness by the roughness weight parameter. The first damage score, the second damage score, and the third damage score are added together, and the sum of the values ​​is extracted as the over-etching damage index.

[0047] In this embodiment of the invention, a layout design file of the target processing area is first obtained. The theoretical contour lines of the target processing area are extracted from the layout design file. Based on the current magnification and physical pixel resolution of the scanning electron microscope (i.e., the actual nanometer size represented by each pixel), the physical coordinates of the theoretical contour lines are proportionally converted into image pixel coordinates, thereby generating a mask design boundary that is strictly aligned with the spatial scale of the microscope image. Next, the system iterates through all the boundary pixels contained on the final two-dimensional boundary, calculating the value of each boundary pixel one by one. To the mask design boundary The shortest plane Euclidean distance is calculated using a boundary integral mean variation model, and the formula is as follows: ,in The effective etch boundary offset is calculated, with the dimension being pixels; This represents the total number of boundary pixels contained on the final boundary. represents the two-dimensional Euclidean distance from the i-th boundary pixel to the smallest distance among all points on the design boundary, with the dimension of pixels; Subsequently, the effective etch boundary offset and the sidewall verticality deviation were extracted. and sidewall roughness Because these three characteristic parameters representing etching damage have completely different physical dimensions and numerical magnitudes, directly performing arithmetic fusion would cause the large numerical parameter to "overwhelm" the small numerical parameter. Therefore, the system adopts an adaptive linear mapping variant model based on historical process benchmarks to normalize and transform the three parameters to eliminate dimensional differences. The mapping formula is uniformly constructed as follows: Where X represents the original input feature parameters, and These represent the historical optimal lower limit and historical extreme upper limit of the parameter under normal process tolerance, respectively, extracted from the yield database in advance. The upper limit of the standardized numerical range with a preset number of digits is recommended to be set to 100 for intuitive mapping to a percentage-based evaluation system; this will generate the corresponding normalized offset. Normalized verticality deviation and normalized roughness .

[0048] Next, the system accesses a pre-established process evaluation weight library, which is generated offline using principal component analysis algorithm based on historical wafer electrical test yield data. The system extracts offset weight parameters from the library that match the normalized offset. Verticality weight parameters that match the normalized verticality deviation and roughness weight parameters that match the normalized roughness. These weighting parameters are all dimensionless coefficients greater than 0 and summing to 1; Finally, the system performs a weighted calculation, multiplying the normalized offset by the offset weight parameter to obtain the first damage score. The second damage score is obtained by multiplying the normalized verticality deviation by the verticality weight parameter. The normalized roughness is multiplied by the roughness weight parameter to obtain the third damage score. Ultimately, the system utilizes the formula The first damage score, the second damage score, and the third damage score are added together, and the sum is extracted as the over-etching damage index for a comprehensive evaluation of the current microstructure. It is then output to the process monitoring terminal.

[0049] Example 2: like Figure 2 As shown, the etching process evaluation system based on artificial intelligence includes an edge point set acquisition module, a feathered edge division module, an etching boundary output module, and an etching damage evaluation module. The various modules are connected via wired and / or wireless connections to enable data transmission between them; Edge point set acquisition module: acquires the top-view scanning electron microscope image sequence and the tilt angle scanning electron microscope image sequence of the etched wafer surface; performs grayscale normalization processing on the top-view scanning electron microscope image sequence to obtain a standardized image, and performs edge detection on the standardized image to extract the initial edge point set; Feathered edge segmentation module: Based on the initial edge point set, extend outward along the normal direction by a preset pixel distance to construct the edge influence band, calculate the gray-level gradient magnitude of each pixel in the edge influence band, and divide the feathered edge into the first morphological type, the second morphological type, or the third morphological type according to the attenuation characteristic of the gray-level gradient magnitude with the distance from the edge, and calculate the feathering width index at the same time. Etching boundary output module: After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, it enters a bidirectional boundary optimization loop and repeats the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, and then stops and outputs the final boundary. Etching damage assessment module: Maps the final boundary to a scanning electron microscope image sequence at an inclined angle, extracts the sidewall grayscale profile at the corresponding position of the final boundary, calculates the sidewall verticality deviation and sidewall roughness based on the sidewall grayscale profile, calculates the effective etching boundary offset, where the effective etching boundary offset is the distance between the final boundary and the mask design boundary, calculates the over-etching damage index based on the effective etching boundary offset, sidewall verticality deviation and sidewall roughness, and outputs the over-etching damage index.

[0050] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0051] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0053] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An etching process evaluation method based on artificial intelligence, characterized in that, include: Obtain top-view scanning electron microscope (SEM) image sequences and tilt-angle SEM image sequences of the etched wafer surface; The image sequence of top-view scanning electron microscope is subjected to grayscale normalization to obtain a standardized image, and edge detection is performed on the standardized image to extract the initial edge point set; Based on the initial edge point set, extend outwards along the normal direction by a preset pixel distance to construct an edge influence band. Calculate the grayscale gradient magnitude of each pixel within the edge influence band. Based on the attenuation characteristic of the grayscale gradient magnitude with distance from the edge, classify the feathered edge into a first morphological type, a second morphological type, or a third morphological type. Simultaneously calculate the feathering width index, which is the distance from the initial edge point set to the point where the grayscale gradient magnitude drops to the background noise threshold. After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, the bidirectional boundary optimization loop is entered and repeated until the boundary change is less than the convergence threshold, at which point the final boundary is output. The final boundary is mapped to a scanning electron microscope image sequence at an inclined angle, and the gray-scale profile of the sidewall at the corresponding position of the final boundary is extracted. The sidewall verticality deviation and sidewall roughness are calculated based on the sidewall gray-scale profile. Calculate the effective etch boundary offset, which is the distance between the final boundary and the mask design boundary; calculate the over-etch damage index based on the effective etch boundary offset, sidewall perpendicularity deviation, and sidewall roughness, and output the over-etch damage index.

2. The etching process evaluation method based on artificial intelligence according to claim 1, characterized in that, Edge detection is performed on the standardized image to extract an initial set of edge points, including: Extract the gray-level change rate of each pixel in the standardized image in the horizontal and vertical dimensions, and combine the gray-level change rates in the horizontal and vertical dimensions to calculate the gray-level gradient magnitude and gray-level gradient direction of each pixel. The extreme values ​​of the gray-level gradient magnitude of each pixel are filtered along the gray-level gradient direction. Pixels whose gray-level gradient magnitude is greater than that of their neighboring pixels in the current gray-level gradient direction are retained as candidate edge points. Calculate the average gray-level gradient magnitude of all candidate edge points, and then spatially stitch together the candidate edge points whose gray-level gradient magnitude is greater than the average gray-level gradient magnitude to extract the initial edge point set.

3. The etching process evaluation method based on artificial intelligence according to claim 1, characterized in that, The feathering edge is divided into corresponding morphological types, including the first morphological type, the second morphological type, and the third morphological type. Simultaneously, the feathering width index is calculated, including: Extract the gray-level gradient direction corresponding to each pixel in the initial edge point set, use the gray-level gradient direction as the normal direction, and extend the extension trajectory outward from the initial edge point set along the normal direction by a preset pixel distance to generate the extension trajectory. Combine all the extension trajectories to construct the edge influence zone. Extract the pixels contained in each extended trajectory within the edge influence zone, obtain the grayscale gradient magnitude of each pixel on the extended trajectory, and calculate the number of pixels between each pixel on the extended trajectory and the starting point of the initial edge point set. Use the number of pixels between them as the spatial distance. By combining the gray-level gradient magnitude of each pixel within the edge influence zone with the corresponding spatial distance, the ratio of the decrease in gray-level gradient magnitude to the spatial distance is calculated to obtain the gray-level gradient decay rate. The gray-level gradient decay rate is used as the decay feature of gray-level gradient magnitude with edge distance. Construct a mapping relationship between the numerical distribution range of gray-level gradient decay rate and morphological type, determine the numerical distribution range into which the gray-level gradient decay rate of each extended trajectory falls, and assign the corresponding morphological type to the feathering edge according to the numerical distribution range and the mapping relationship. The morphological types include the first morphological type, the second morphological type and the third morphological type. Collect background pixels in the standardized image that are not covered by edge influence bands, calculate the average gray-level gradient magnitude of all background pixels, and use the average gray-level gradient magnitude as the background noise threshold. Along the extended trajectory, the grayscale gradient magnitude of each pixel is compared with the background noise threshold. Pixels whose grayscale gradient magnitude decreases to the background noise threshold are extracted as boundary cutoff points. The spatial distance corresponding to the boundary cutoff point is obtained, and this spatial distance is used as the feathering width index.

4. The etching process evaluation method based on artificial intelligence according to claim 3, characterized in that, Constructing a mapping relationship between the numerical distribution range of grayscale gradient decay rate and morphological type, including: Extract the gray-level gradient decay rate of all extended trajectories within the edge influence band, and arrange all gray-level gradient decay rates in ascending order to generate a gray-level gradient decay rate sequence. Calculate the mean and standard deviation of all gray-level gradient decay rates in the gray-level gradient decay rate sequence; The difference between the mean and the standard deviation is used as the first boundary node, and the sum of the mean and the standard deviation is used as the second boundary node. The overall numerical range of the gray-level gradient decay rate sequence is truncated and divided using the first boundary node and the second boundary node, and the first numerical interval, the second numerical interval and the third numerical interval are divided in sequence. The first numerical interval, the second numerical interval and the third numerical interval are collectively used as the gray-level gradient decay rate numerical distribution interval. Establish a correspondence between the first numerical interval and the first morphological type, establish a correspondence between the second numerical interval and the second morphological type, and establish a correspondence between the third numerical interval and the third morphological type. Combine all correspondences to generate a mapping relationship between the grayscale gradient decay rate numerical distribution interval and the morphological type.

5. The etching process evaluation method based on artificial intelligence according to claim 4, characterized in that, The shrinkage step size and maximum number of iterations are determined based on the feathering morphology type, including: Extract the morphology type corresponding to each extended trajectory within the edge influence zone, use the morphology type as the feathering morphology type, and obtain the feathering width index corresponding to each extended trajectory. Establish a mapping relationship between feathering morphology types and basic parameter configuration tables. In the basic parameter configuration tables, configure the corresponding basic step size coefficients and basic iteration coefficients for the first morphology type, the second morphology type, and the third morphology type, respectively. The basic step size coefficient and basic iteration coefficient corresponding to the feathering morphology type of each extended trajectory are obtained in the basic parameter configuration table through association mapping. Calculate the product of the feathering width index and the base step size coefficient to obtain the initial step size value. Round the initial step size value down to generate the shrinkage step size of the current extended trajectory. The feathering width index is divided by the shrinkage step size, and the quotient is extracted as the number of spatial segments. The product of the number of spatial segments and the basic iteration coefficient is calculated, and the product is rounded up. The rounded value is determined as the maximum number of iterations.

6. The etching process evaluation method based on artificial intelligence according to claim 1, characterized in that, The bidirectional boundary optimization loop includes: shrinking the current boundary inward along the normal direction by the shrinkage step size to obtain a shrunken boundary; calculating the first deviation between the average line width and the design line width within the area enclosed by the shrunken boundary; expanding the current boundary outward along the normal direction by the shrinkage step size to obtain an expanded boundary; calculating the second deviation between the average line width and the design line width within the area enclosed by the expanded boundary; comparing the first deviation and the second deviation; if the first deviation is less than the second deviation, selecting the shrinkage direction to update the boundary; if the second deviation is less than the first deviation, selecting the expansion direction to update the boundary; if the difference between the first deviation and the second deviation is less than the convergence threshold, marking the current boundary as a candidate boundary; repeating the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, and then outputting the final boundary.

7. The etching process evaluation method based on artificial intelligence according to claim 6, characterized in that, The comparison between the first and second deviations in the bidirectional boundary optimization loop specifically includes: Extract the first deviation generated in the contraction direction and the second deviation generated in the expansion direction of the current boundary, subtract the first deviation from the second deviation, and obtain the absolute value of the calculation result as the absolute value of the difference. The absolute value of the difference is compared with the convergence threshold. If the absolute value of the difference is less than the convergence threshold, the evolution process of the current boundary is determined to have reached a stable state. The coordinates of all pixels contained in the current boundary are extracted, and the current boundary is marked as a candidate boundary. Under the condition that the absolute value of the difference is not less than the convergence threshold, the numerical values ​​of the first deviation and the second deviation are further compared, and the corresponding boundary update operation is performed according to the comparison result. Under the condition that the first deviation is less than the second deviation, it is determined that the current boundary has an inward evolution trend. The associated shrinkage step size is extracted, and the current boundary is moved along the shrinkage direction according to the value of the shrinkage step size to generate an updated boundary. If the second deviation is less than the first deviation, it is determined that the current boundary has a tendency to evolve outward. The current boundary is moved along the expansion direction according to the value of the contraction step size to generate an updated boundary.

8. The etching process evaluation method based on artificial intelligence according to claim 1, characterized in that, Extract the grayscale profile of the sidewall at the location corresponding to the final boundary, and calculate the sidewall verticality deviation and sidewall roughness based on the grayscale profile, including: Obtain the planar coordinates of each pixel in the final boundary, extract the spatial tilt angle parameters of the tilt angle scanning electron microscope image sequence at the time of acquisition, construct a coordinate transformation matrix using the spatial tilt angle parameters, and transform the planar coordinates to the tilted three-dimensional coordinate system corresponding to the tilt angle scanning electron microscope image sequence through the coordinate transformation matrix to generate the tilted mapping boundary. Along the vertical extension direction of the tilted mapping boundary, the gray values ​​of the corresponding pixels are extracted layer by layer in the tilt angle scanning electron microscope image sequence. The gray values ​​of each corresponding pixel are arranged and combined according to the depth level to construct the gray profile of the sidewall at the corresponding position of the tilted mapping boundary. Extract global grayscale peak points and global grayscale valley points from the grayscale profile of the sidewall. Mark the depth level where the global grayscale peak points are located as the top feature position and the depth level where the global grayscale valley points are located as the bottom feature position. Calculate the horizontal offset distance and vertical depth difference between the top feature position and the bottom feature position. The horizontal offset distance is divided by the difference between the vertical depth, and the quotient is extracted as the sidewall verticality deviation. Construct an ideal sidewall contour line that is linearly connected between the top feature position and the bottom feature position, and obtain the reference gray value of the ideal sidewall contour line at each depth level. Calculate the absolute value of the difference between the actual gray value and the corresponding reference gray value of the sidewall gray profile at each depth level. Sum the absolute values ​​of the differences at all depth levels and calculate the average. Determine the average value as the sidewall roughness.

9. The etching process evaluation method based on artificial intelligence according to claim 8, characterized in that, The calculated etch damage index includes: Obtain the layout design file of the target processing area, extract the theoretical outline of the target processing area from the layout design file, and use the theoretical outline as the mask design boundary. Traverse all boundary pixels contained on the final boundary, calculate the shortest plane distance from each boundary pixel to the mask design boundary, sum the shortest plane distances corresponding to all boundary pixels and calculate the average value, and use the calculated average value as the effective etch boundary offset. The calculated sidewall verticality deviation and sidewall roughness are extracted, and the effective etching boundary offset, sidewall verticality deviation and sidewall roughness are linearly mapped and transformed using a standardized numerical range with a preset number of bits, respectively, to generate normalized offset, normalized verticality deviation and normalized roughness that have eliminated the difference in dimensions. Obtain the pre-established process evaluation weight library, and extract the offset weight parameter that matches the normalized offset, the verticality weight parameter that matches the normalized verticality deviation, and the roughness weight parameter that matches the normalized roughness from the process evaluation weight library. The first damage score is obtained by multiplying the normalized offset by the offset weight parameter, the second damage score is obtained by multiplying the normalized perpendicularity deviation by the perpendicularity weight parameter, and the third damage score is obtained by multiplying the normalized roughness by the roughness weight parameter. The first damage score, the second damage score, and the third damage score are added together, and the sum of the values ​​is extracted as the over-etching damage index.

10. An etching process evaluation system based on artificial intelligence, characterized in that, It is implemented based on the artificial intelligence-based etching process evaluation method according to any one of claims 1-9, and includes: Edge point set acquisition module: acquires the top-view scanning electron microscope image sequence and the tilt angle scanning electron microscope image sequence of the etched wafer surface; performs grayscale normalization processing on the top-view scanning electron microscope image sequence to obtain a standardized image, and performs edge detection on the standardized image to extract the initial edge point set; Feathered edge segmentation module: Based on the initial edge point set, extend outward along the normal direction by a preset pixel distance to construct the edge influence band, calculate the gray-level gradient magnitude of each pixel in the edge influence band, and divide the feathered edge into the first morphological type, the second morphological type, or the third morphological type according to the attenuation characteristic of the gray-level gradient magnitude with the distance from the edge, and calculate the feathering width index at the same time. Etching boundary output module: After determining the shrinkage step size and maximum number of iterations based on the feathering morphology type, it enters a bidirectional boundary optimization loop and repeats the bidirectional boundary optimization loop until the boundary change is less than the convergence threshold, and then stops and outputs the final boundary. Etching damage assessment module: Maps the final boundary to a scanning electron microscope image sequence at an inclined angle, extracts the sidewall grayscale profile at the corresponding position of the final boundary, calculates the sidewall verticality deviation and sidewall roughness based on the sidewall grayscale profile, calculates the effective etching boundary offset, where the effective etching boundary offset is the distance between the final boundary and the mask design boundary, calculates the over-etching damage index based on the effective etching boundary offset, sidewall verticality deviation and sidewall roughness, and outputs the over-etching damage index.

Images