A precision optical measurement method for trench size of GPP chip

Abstract

The application discloses a kind of GPP chip trench size precision optical measurement methods, it is related to the field of semiconductor optical wafer measurement, comprising: S1: acquisition low-power image is positioned with fixed position and position reliability score calculation, output center coordinates and azimuth angle;S2: based on positioning coordinates driving spectral confocal sensor acquisition height data, through waveform quality parameter generation depth calculation reliability, calculate trench depth value;S3: based on coordinates and azimuth angle driving high-power camera acquisition image, through edge feature parameter calculation profile linearity score, after optimal profile is selected, linear fitting is fitted, calculate trench width;S4: based on depth calculation reliability and profile linearity score, generate quality grade, combine trench depth and trench width to generate final measurement result.The reliability and environmental adaptability of measurement result are significantly improved while ensuring high precision and high efficiency by introducing multi-parameter quality evaluation and intelligent optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of semiconductor optical wafer measurement, specifically a precision optical measurement method for the trench dimensions of GPP chips. Background Technology

[0002] GPP chips are semiconductor wafers manufactured using glass passivation technology, primarily used in the production of discrete devices such as rectifier diodes, voltage regulator chips, and surge protectors. In the fabrication process, trenches are typically etched using wet etching after a photolithography mask. Once etched, passivation materials such as glass are filled in to improve the device's voltage withstand capability, reliability, and high-temperature stability. However, in actual operation, it is crucial to ensure that the trench etching is uniform and appropriate. This necessitates measuring the etched trenches to determine if their width and depth fall within standard specifications.

[0003] In the current industry, manual visual inspection is commonly used to calculate the depth and width of the die slots. Precision instruments such as scanning tunneling microscopes are used to measure the chips. Operators need to manually focus on the GPP chip to be inspected and visually estimate the width of the product using pre-marked scales. This process is affected by human eye error and is extremely inefficient. Even a skilled operator needs several minutes to complete the inspection of a single die. Furthermore, the number of dies in a GPP wafer is extremely large, making re-inspection difficult.

[0004] To address the numerous inconveniences of manual inspection, a precision optical measurement scheme for GPP groove dimensions is proposed, with a targeted design based on the two dimensions of groove dimensions—groove depth and groove width. Summary of the Invention

[0005] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a precision optical measurement method for the trench size of GPP chips, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a precision optical measurement method for the trench dimensions of a GPP chip, comprising:

[0007] S1: Acquire low-magnification visual images to locate the trench area and calculate the location confidence score, and output the center coordinates and azimuth parameters of the trench area;

[0008] S2: Based on the positioning coordinates of the trench area, drive the spectral confocal sensor to collect height data, extract waveform quality parameters to generate depth calculation reliability, and calculate the trench depth value when the depth calculation reliability meets the preset conditions.

[0009] S3: Based on the positioning coordinates and azimuth parameters of the trench area, a high-magnification camera is used to acquire high-magnification images for image segmentation. Edge feature parameters are extracted to calculate contour linearity score and direction consistency score. The optimal edge contour is selected through an optimization mechanism and a straight line is fitted. The distance between the fitted straight lines on both sides is calculated as the trench width.

[0010] S4: Based on the depth calculation reliability and contour linearity score, calculate the comprehensive reliability score to generate the quality level, and combine the trench depth and trench width values ​​to form the trench size and generate the final measurement result.

[0011] The present invention is further configured such that S1 includes:

[0012] A global image of a preset area of ​​a wafer containing multiple grains is acquired using a low-magnification industrial camera.

[0013] The global image is processed based on a pre-trained semantic segmentation model to obtain semantic segmentation results containing groove candidate regions;

[0014] From each candidate region of the semantic segmentation result, extract its edge distribution features, texture consistency features, and morphological regularity features.

[0015] The present invention is further configured to generate a location confidence score for characterizing the region confidence by fusing edge distribution features, texture consistency features and morphological regularity features.

[0016] By comparing the location confidence score with a preset groove determination threshold, the region with a location confidence score greater than the groove determination threshold is determined as a valid groove region;

[0017] For the effective trench area, the azimuth parameters are calculated using principal component analysis. The centroid of the area is calculated, and the centroid is mapped to the mechanical coordinate system through coordinate transformation using preset mechanical coordinate system calibration parameters to obtain the center coordinates.

[0018] The present invention is further configured such that S2 includes:

[0019] Based on the center coordinates and azimuth parameters of the groove region output by S1, adjust the position of the plate support stage to be below the spectral confocal laser measurement device;

[0020] The driving spectral confocal laser measurement device scans along the vertical groove direction to obtain the original height data sequence;

[0021] Adaptive filtering based on sliding window and dynamic threshold is applied to the original height data sequence to suppress impulse noise;

[0022] From the filtered height data sequence, the high-level platform region and the low-level platform region are identified and separated.

[0023] The present invention is further configured to extract waveform quality parameters based on the high-plane plateau region and the low-plane plateau region, wherein the waveform quality parameters include: plateau stability coefficient, edge steepness factor and signal integrity index;

[0024] Calculate the platform stability coefficient based on the statistical variance of the high-surface platform area data;

[0025] The edge steepness factor is calculated based on the first-order differential features of the height data sequence in the jump intervals between platform regions.

[0026] Based on the proportion and continuity of valid data points in the original height data sequence, the signal completeness index is calculated;

[0027] The platform stability coefficient, edge steepness factor, and signal completeness index are weighted and fused to generate the depth computing credibility.

[0028] The present invention is further configured to compare the depth calculation confidence with a preset depth calculation enable threshold;

[0029] When the depth calculation confidence level is greater than or equal to the depth calculation enable threshold, the difference between the average value of the data in the high-surface platform region and the average value of the data in the low-surface platform region is calculated as the final trench depth value.

[0030] The present invention is further configured such that S3 includes:

[0031] Based on the center coordinates and azimuth parameters of the trench area output by S1, the position of the plate-bearing stage is adjusted to the position of the high-magnification industrial camera, and the high-magnification industrial camera is driven to acquire high-magnification images of the trench area.

[0032] Based on the azimuth parameter, the high-magnification image is oriented and the region is cropped to obtain a first width image containing the first side edge of the groove and a second width image containing the second side edge of the groove, respectively.

[0033] Edge feature parameters are extracted from the first width image and the second width image. The edge feature parameters include: gray-level transition slope representing gray-level change characteristics, texture heterogeneity index representing texture complexity, and phase consistency clustering degree representing edge structure consistency.

[0034] The present invention is further configured such that, based on the first width image and the second width image, for each side edge, an initial edge contour point set is extracted based on the phase consistency clustering degree;

[0035] Within the neighborhood of the initial edge contour point set, enhanced features are extracted based on the gray-level transition slope and texture heterogeneity index to obtain the enhanced edge contour point set.

[0036] The contour linearity score and orientation consistency score of the initial edge contour point set and the enhanced edge contour point set are calculated respectively. The contour linearity score is determined by calculating the residual between the contour point set and the fitted straight line, and the orientation consistency score is determined by calculating the angle between the main direction of the contour point set and the theoretical direction of the trench.

[0037] The present invention is further configured to obtain a comprehensive contour score by weighted summation of contour linearity score and direction consistency score, and select one of the initial edge contour point set and the enhanced edge contour point set as the optimal edge contour of that side based on the comprehensive contour score.

[0038] For the edges on both sides of the groove, a reverse scanning method based on symmetrical regions is used to determine the optimal edge contours of the first width image and the second width image respectively.

[0039] The optimal edge contours on both sides are fitted with straight lines, and the distance between the two fitted lines is calculated as the groove width value.

[0040] The present invention is further configured such that S4 includes:

[0041] Based on the depth calculation confidence score from S2 output and the contour linearity score from S3 output, a comprehensive confidence score is generated through fusion calculation.

[0042] The overall confidence score is compared with a predefined quality level threshold to determine the quality level of the measurement result.

[0043] The trench depth, trench width, overall reliability score, and quality level are correlated and output as the final measurement result.

[0044] This invention provides a precision optical measurement method for the trench size of a GPP chip. The method comprises: S1: acquiring low-magnification visual images to locate the trench region and calculate its positional reliability score, outputting the center coordinates and azimuth parameters of the trench region; S2: based on the trench region's location coordinates, driving a spectral confocal sensor to acquire height data, extracting waveform quality parameters to generate depth calculation reliability, and calculating the trench depth value when the depth calculation reliability meets preset conditions; S3: driving a high-magnification camera to acquire high-magnification images based on the trench region's location coordinates and azimuth parameters, performing image segmentation, extracting edge feature parameters to calculate contour linearity and orientation consistency scores, selecting the optimal edge contour through an optimization mechanism and performing straight line fitting, calculating the distance between the fitted lines on both sides as the trench width; S4: based on the depth calculation reliability and contour linearity score, calculating a comprehensive reliability score to generate a quality level, and combining the trench depth and trench width values ​​to generate the final measurement result for the trench size. The beneficial effects include:

[0045] This method achieves fully automated, high-precision groove dimension measurement, significantly improving inspection efficiency and consistency. By constructing a fully automated closed-loop measurement process of "low-magnification visual positioning - spectral confocal depth measurement - high-magnification visual width measurement," it completely replaces inefficient and subjective manual visual inspection. This method not only reduces the time for a single measurement from several minutes to seconds, but also completely eliminates human eye error through objective data acquisition from optical sensors and algorithm processing, ensuring high accuracy and repeatability of the measurement results.

[0046] The innovative concept of parametric quality control ensures the reliability and traceability of measurement results. Creatively different from traditional direct calculation, this solution introduces "waveform quality parameters" and "edge feature parameters" in the depth and width measurement links, respectively, and generates "depth calculation reliability" and "contour linearity score" based on these parameters. Finally, by fusing these parameters, the solution outputs "overall reliability" and "quality level," providing a quantitative assessment of the reliability of each measurement result. This represents a qualitative leap in the measurement process, moving from simply "providing numerical values" to "evaluating the reliability of numerical values," providing crucial evidence for process decisions and problem traceability.

[0047] The intelligent decision-making mechanism, employing multi-feature fusion and bidirectional optimization, significantly improves measurement robustness under complex working conditions. Addressing challenges such as ambiguity and noise at trench edges, the solution utilizes a multi-feature fusion optimization mechanism. During the localization phase, multiple visual features are fused to intelligently identify valid trenches; during the width measurement phase, the optimal edge is automatically selected through bidirectional extraction and scoring of the "initial contour" and "enhanced contour." This decision-making process, with its introspective and optimization capabilities, enables the system to adapt to different surface conditions and imaging requirements, demonstrating excellent robustness and adaptability.

[0048] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in 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. In the drawings:

[0050] Figure 1 This is a flowchart illustrating a precision optical measurement method for the trench size of a GPP chip, as shown in an exemplary embodiment of the present invention. Detailed Implementation

[0051] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0052] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0053] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0054] Example:

[0055] A precision optical measurement method for the trench dimensions of GPP chips, such as Figure 1 As shown, it includes:

[0056] S1: Acquire low-magnification visual images to locate the trench area and calculate the location confidence score, and output the center coordinates and azimuth parameters of the trench area;

[0057] S2: Based on the positioning coordinates of the trench area, drive the spectral confocal sensor to collect height data, extract waveform quality parameters to generate depth calculation reliability, and calculate the trench depth value when the depth calculation reliability meets the preset conditions.

[0058] S3: Based on the positioning coordinates and azimuth parameters of the trench area, a high-magnification camera is used to acquire high-magnification images for image segmentation. Edge feature parameters are extracted to calculate contour linearity score and direction consistency score. The optimal edge contour is selected through an optimization mechanism and a straight line is fitted. The distance between the fitted straight lines on both sides is calculated as the trench width.

[0059] S4: Based on the depth calculation reliability and contour linearity score, calculate the comprehensive reliability score to generate the quality level, and combine the trench depth and trench width values ​​to form the trench size and generate the final measurement result.

[0060] The present invention is further configured such that S1 includes:

[0061] A global image of a preset area of ​​a wafer containing multiple grains is acquired using a low-magnification industrial camera.

[0062] The global image is processed based on a pre-trained semantic segmentation model to obtain semantic segmentation results containing groove candidate regions;

[0063] From each candidate region of the semantic segmentation results, edge distribution features, texture consistency features, and morphological regularity features are extracted. Specifically, firstly, a global image of a preset area of ​​a wafer containing multiple grains is acquired using a low-magnification industrial camera. The image resolution is 2048×2048 pixels, ensuring coverage of the wafer surface and complete grain information. Then, the acquired global image is input into a pre-trained semantic segmentation model, such as U-Net. The semantic segmentation model performs pixel-by-pixel classification of the image using a convolutional neural network, accurately identifying and labeling potential trench candidate regions. Next, multi-dimensional feature extraction is performed on each trench candidate region output by the semantic segmentation model: edge distribution features are extracted using the Canny edge detection algorithm to quantify the sharpness and directionality of its edge structure; texture consistency features are extracted using the local binary mode method to evaluate the uniformity of the texture within the region; simultaneously, the ratio of the region's area to the area of ​​its minimum bounding rectangle is calculated as a morphological regularity feature to measure the regularity of its shape.

[0064] The present invention is further configured to generate a location confidence score for characterizing the region confidence by fusing edge distribution features, texture consistency features and morphological regularity features.

[0065] By comparing the location confidence score with a preset groove determination threshold, the region with a location confidence score greater than the groove determination threshold is determined as a valid groove region;

[0066] For valid trench regions, azimuth parameters are calculated using principal component analysis. The region centroid is then calculated and mapped to the machine coordinate system using a pre-defined calibration parameter, yielding the center coordinates. Specifically, after successfully extracting edge distribution features, texture consistency features, and morphological regularity features from each candidate trench region, the system enters the intelligent judgment and localization stage. First, the three feature values ​​are normalized to eliminate the influence of dimensions. Then, they are weighted and fused according to predetermined weight ratios to generate a localization reliability score within the range of 0 to 1, representing the region's reliability. The weight allocation prioritizes edge distribution features, set to 0.5 by default, as clear and oriented edges are the most stable and significant representation of trenches. Texture consistency features have a secondary weight, set to 0.3 by default, to help distinguish between real trenches and texture changes caused by noise. Morphological regularity features have a weight set to 0.2 by default to exclude obviously irregular interference regions. This weighting scheme is based on empirical values ​​derived from statistical analysis of a large number of training samples, ensuring the accuracy and robustness of the judgment. The specific weight ratio can also be fine-tuned according to the actual wafer process characteristics and imaging conditions. The weighted fusion calculation method is as follows: the normalized edge distribution feature is multiplied by 0.5, the texture consistency feature by 0.3, and the morphological regularity feature by 0.2. The three results are then added together to obtain the localization confidence score. Next, the calculated localization confidence score is compared with a preset trench judgment threshold. The default value of the trench judgment threshold is set to 0.65, which aims to ensure that subsequent measurement steps are only performed on areas with high confidence in alignment certainty, thereby saving system resources and improving overall process efficiency. After comparison, only areas with localization confidence scores greater than 0.65 are ultimately judged as valid trench areas, and the remaining areas are filtered out by the system. Finally, for each selected valid trench area, precise geometric parameter calculations are performed. The azimuth angle is calculated using Principal Component Analysis (PCA): The PCA algorithm takes the coordinates of all pixels in the effective trench region as input and calculates the first principal direction, which is the azimuth parameter of the trench. The center coordinates are determined in two steps: First, the centroid of all pixels in the region is calculated as its position in the image coordinate system. Then, using pre-obtained mechanical coordinate system calibration parameters obtained through procedures such as the nine-point calibration method, the centroid coordinates are transformed and precisely mapped to the mechanical coordinate system, thus obtaining the center coordinates in millimeters required to drive the stage movement. Finally, the center coordinates and azimuth parameters of the effective trench are output, providing a precise spatial reference and directional guidance for the subsequent spectral confocal depth measurement in step S2 and the high-magnification visual width measurement in step S3.

[0067] The present invention is further configured such that S2 includes:

[0068] Based on the center coordinates and azimuth parameters of the groove region output by S1, adjust the position of the plate support stage to be below the spectral confocal laser measurement device;

[0069] The driving spectral confocal laser measurement device scans along the vertical groove direction to obtain the original height data sequence;

[0070] Adaptive filtering based on sliding window and dynamic threshold is applied to the original height data sequence to suppress impulse noise;

[0071] From the filtered height data sequence, the high-level and low-level platform regions are identified and separated. Specifically, based on the center coordinates, the system adjusts the position of the support stage using a precision motion controller, accurately moving the trench to be measured directly below the probe of the spectral confocal laser measurement device. Subsequently, based on the azimuth parameters, the system controls the rotation or motion axis of the support stage to ensure that the trench direction is perpendicular to the scanning direction of the spectral confocal sensor. After preparation, the spectral confocal laser measurement device is driven to perform a one-dimensional linear scan along the direction perpendicular to the trench, acquiring the original height data sequence of the trench profile at a preset sampling frequency. The original height data sequence contains the structural information of the trench, but also contains impulse noise caused by electronic noise, environmental vibration, or surface contaminants. To suppress this noise, the original height data sequence is subjected to adaptive filtering based on a sliding window and dynamic threshold. Specifically, a sliding window with a length of 5 data points is defined. This default value achieves a good balance between noise smoothing and edge preservation, and the data sequence is traversed sequentially. For the data within the window, its mean and standard deviation are calculated in real time. A dynamic threshold is set, which is 2.0 times the current window standard deviation, to effectively distinguish noise from true edge transitions. The data at the window center point is then compared with the window mean. If the difference exceeds the dynamic threshold, it is identified as impulse noise, and the point value is replaced with the window mean; otherwise, the original data is retained. This process continues until the entire data sequence is filtered, effectively suppressing impulse noise while perfectly preserving the abrupt transition characteristics representing the trench edges. After filtering, the system identifies and separates high-level and low-level platform regions from the processed height data sequence. This process is based on the statistical characteristic of relatively small fluctuations in platform region data: first, the first-order difference of the data sequence is calculated, and potential platform regions are located by finding intervals with consistently small difference values; then, cluster analysis is performed on these candidate platform regions, and based on their average height values, the group with higher positions is identified as high-level platform regions, and the group with lower positions is identified as low-level platform regions.

[0072] The present invention is further configured to extract waveform quality parameters based on the high-plane plateau region and the low-plane plateau region, wherein the waveform quality parameters include: plateau stability coefficient, edge steepness factor and signal integrity index;

[0073] Calculate the platform stability coefficient based on the statistical variance of the high-surface platform area data;

[0074] The edge steepness factor is calculated based on the first-order differential features of the height data sequence in the jump intervals between platform regions.

[0075] Based on the proportion and continuity of valid data points in the original height data sequence, the signal completeness index is calculated;

[0076] The platform stability coefficient, edge kurtosis factor, and signal completeness index are weighted and fused to generate the depth calculation reliability. Specifically, after successfully separating the high-level platform region and the low-level platform region, the system enters the waveform quality analysis stage, extracting three core waveform quality parameters to evaluate the measurement quality. The platform stability coefficient is obtained by calculating the reciprocal of the statistical variance of the data in the high-level platform region. The platform stability coefficient characterizes the flatness of the measurement platform; a larger value indicates smaller platform fluctuations and more ideal measurement conditions. The edge kurtosis factor is calculated based on the first-order differential characteristics of the height data sequence in the transition intervals between platform regions. The verticality of the trench sidewall is evaluated by finding the differential extrema and analyzing their distribution characteristics. The signal completeness index is obtained by analyzing the proportion and continuity of effective data points in the original height data sequence. The signal completeness index reflects the completeness and reliability of the data acquisition process. After parameter extraction, the three waveform quality parameters are normalized and weighted and fused according to predetermined weights to generate the depth calculation reliability. In the weighting scheme, the edge kurtosis factor has the highest weight, set to 0.5 by default, because the integrity of edge features directly determines the accuracy of depth calculation. The platform stability coefficient has the second highest weight, set to 0.3 by default, as a stable measurement platform is fundamental to obtaining reliable data. The signal completeness index has a weight set to 0.2 by default, used to assess the impact of data acquisition quality on the results. This weighting allocation is based on statistical analysis of a large amount of experimental data to ensure the accuracy of the reliability assessment. The weighted fusion calculation method is as follows: the platform stability coefficient is multiplied by 0.3, plus the edge kurtosis factor multiplied by 0.5, plus the signal completeness index multiplied by 0.2, and finally the three weighted results are summed to obtain the depth calculation reliability ranging from 0 to 1.

[0077] The present invention is further configured to compare the depth calculation confidence with a preset depth calculation enable threshold;

[0078] When the depth calculation confidence level is greater than or equal to the depth calculation enabling threshold, the difference between the average value of the data in the high-level platform area and the average value of the data in the low-level platform area is calculated as the final trench depth value. Specifically, after obtaining the depth calculation confidence level, the system enters the depth calculation decision and execution phase. First, the depth calculation confidence level is compared with the preset depth calculation enabling threshold, which is set to a default value of 0.7. The setting of the depth calculation enabling threshold is based on a balance between measurement accuracy and efficiency; a value that is too high may lead to excessive filtering of qualified data, while a value that is too low may include unreliable measurement results. When the depth calculation confidence level is greater than or equal to 0.7, the system determines that the current data quality meets the requirements for precision measurement and then executes the depth calculation. The calculation process selects the verified high-level platform area and the low-level platform area, calculates the arithmetic mean of all height data in both areas, and then subtracts the average value of the low-level platform area from the average value of the high-level platform area. The difference obtained is the final trench depth value. This calculation method based on the average value of the platform area can effectively reduce the influence of random errors and ensure the accuracy and repeatability of the depth measurement results. If the depth calculation reliability is less than 0.7, the system will discard the current data and record the quality anomaly. At the same time, it can trigger the re-measurement process to ensure that the final output trench depth value meets the preset quality standard.

[0079] The present invention is further configured such that S3 includes:

[0080] Based on the center coordinates and azimuth parameters of the trench area output by S1, the position of the plate-bearing stage is adjusted to the position of the high-magnification industrial camera, and the high-magnification industrial camera is driven to acquire high-magnification images of the trench area.

[0081] Based on the azimuth parameter, the high-magnification image is oriented and the region is cropped to obtain a first width image containing the first side edge of the groove and a second width image containing the second side edge of the groove, respectively.

[0082] Edge feature parameters are extracted from the first and second width images. These edge feature parameters include: gray-level transition slope (characterizing gray-level change characteristics), texture heterogeneity index (characterizing texture complexity), and phase consistency clustering degree (characterizing edge structure consistency). Specifically, based on the center coordinates and azimuth parameters of the groove region output by S1, the position of the stage is adjusted by a precision motion controller to accurately move the groove to be tested to the center of the field of view of the high-magnification industrial camera. The high-magnification industrial camera is driven to acquire high-magnification images of the groove region, ensuring that the micron-level groove edge structure can be clearly distinguished. For example, a 5-megapixel resolution and a 20x optical magnification can be used for image acquisition. The acquired high-magnification images are then oriented and cropped based on the azimuth parameters. A rotation correction algorithm is used to make the groove direction strictly parallel to the image coordinate axis. Subsequently, the region is segmented along the groove extension direction, and a first width image containing the first side edge of the groove and a second width image containing the second side edge of the groove are cropped respectively. An appropriate overlap area is retained between the two images to ensure edge integrity. Three key edge feature parameters are extracted from the first and second width images. The grayscale transition slope is obtained by calculating the image gradient magnitude, representing the drastic degree of grayscale change at the edge; the texture heterogeneity index is calculated using local binary mode variance to quantify the texture complexity of the edge region; and the phase consistency clustering is calculated using the response of a multi-scale, multi-directional Log-Gabor filter bank, representing the consistency and saliency of the edge structure. These three parameters describe edge features from different physical dimensions, providing a quantitative basis for subsequent edge optimization.

[0083] The present invention is further configured such that, based on the first width image and the second width image, for each side edge, an initial edge contour point set is extracted based on the phase consistency clustering degree;

[0084] Within the neighborhood of the initial edge contour point set, enhanced features are extracted based on the gray-level transition slope and texture heterogeneity index to obtain the enhanced edge contour point set.

[0085] The contour linearity score and orientation consistency score are calculated for the initial edge contour point set and the enhanced edge contour point set, respectively. The contour linearity score is determined by calculating the residual between the contour point set and the fitted straight line, and the orientation consistency score is determined by calculating the angle between the principal direction of the contour point set and the theoretical direction of the trench. Specifically, after obtaining the first width image and the second width image containing the edges on both sides of the trench, the system enters the edge contour extraction and quality assessment stage. For each edge, the following processing flow is performed: First, the initial edge contour point set is extracted based on the phase consistency clustering degree. In specific implementation, a phase consistency response threshold is set in the width image, with a default value of 0.75. This default value is selected based on experience from a large number of experiments, aiming to effectively screen out real edge points with high structural saliency while avoiding the introduction of too much noise. Pixels with phase consistency higher than this threshold are grouped by connected components, and the largest connected component point set is selected as the initial edge contour point set. This step mainly utilizes the robustness of phase consistency to changes in illumination and contrast to ensure the structural integrity of the initial contour. On both sides of the path determined by the initial edge contour point set, a strip search region with a fixed width is established, with a default width of 5 pixels. Within this search band, enhanced feature extraction is performed. The grayscale transition slope and texture heterogeneity index of each pixel are comprehensively evaluated. The grayscale transition slope weight is set to 0.6 by default, as it directly determines edge sharpness; the texture heterogeneity index weight is set to 0.4 by default to suppress non-edge regions with cluttered textures. A comprehensive edge intensity map for each pixel is obtained through weighted fusion. Then, non-maximum suppression and contour tracking algorithms are applied to this intensity map to obtain a precise, single-pixel-wide enhanced edge contour point set. Contour linearity scores and orientation consistency scores are calculated for both the initial and enhanced edge contour point sets. The contour linearity score is calculated by fitting a straight line using a random sampling consensus algorithm and calculating the average Euclidean distance from all points to this fitted line. This distance, after normalization, is inversely proportional to the linearity score. The orientation consistency score is calculated by using principal component analysis to calculate the principal direction of the point set, and then calculating the cosine of the angle between this direction and the theoretical trench direction. A larger cosine value indicates a higher orientation consistency score.

[0086] The present invention is further configured to obtain a comprehensive contour score by weighted summation of contour linearity score and direction consistency score, and select one of the initial edge contour point set and the enhanced edge contour point set as the optimal edge contour of that side based on the comprehensive contour score.

[0087] For the edges on both sides of the groove, a reverse scanning method based on symmetrical regions is used to determine the optimal edge contours of the first width image and the second width image respectively.

[0088] The optimal edge contours on both sides are fitted with straight lines, and the distance between the two fitted lines is calculated as the groove width value. Specifically, after calculating the contour linearity score and direction consistency score of both sides, the system enters the contour optimization and width calculation stage. A weighted sum of the contour linearity score and direction consistency score is obtained to obtain the comprehensive contour score, where the contour linearity score weight is set to 0.7 by default, and the direction consistency score weight is set to 0.3 by default. This weight allocation is based on a comprehensive consideration of edge geometric accuracy and direction consistency. The contour linearity weight is higher because it directly determines the edge fitting accuracy, while the direction consistency weight is appropriately retained to ensure that the edge direction conforms to the groove structure characteristics. Based on the comprehensive contour score, the set of points with the higher score is selected from the initial edge contour point set and the enhanced edge contour point set as the optimal edge contour for that side. For the edges on opposite sides of the groove, a reverse scanning method based on symmetrical regions is used to determine the optimal edge contours of the first width image and the second width image, respectively. In practice, the left edge contour is obtained by scanning from left to right in the first width image, and the right edge contour is obtained by scanning from right to left in the second width image. This symmetrical scanning strategy effectively eliminates systematic errors and improves the symmetry and accuracy of width measurement. The optimal edge contours determined on both sides are fitted with straight lines using a random sampling consensus algorithm. This algorithm, through iterative sampling and model verification, effectively eliminates outlier interference and obtains the straight line equation that best represents the edge direction. Finally, multiple sampling points are uniformly selected within the measurement area, and the perpendicular distance between the two fitted lines at each sampling point is calculated. The arithmetic mean of these distances is taken as the final groove width value. This calculation method based on multi-sampling point averaging effectively reduces local errors and ensures the accuracy and reliability of the width measurement results.

[0089] The present invention is further configured such that S4 includes:

[0090] Based on the depth calculation confidence score from S2 output and the contour linearity score from S3 output, a comprehensive confidence score is generated through fusion calculation.

[0091] The overall confidence score is compared with a predefined quality level threshold to determine the quality level of the measurement result.

[0092] The trench depth, trench width, overall reliability score, and quality level are correlated and output as the final measurement result. Specifically, after completing the trench depth and width measurements, the system enters the final measurement result quality assessment and output stage. Based on the depth calculation reliability score output by S2 and the contour linearity score output by S3, a weighted fusion calculation is used to generate the overall reliability score. The default weight for both the depth calculation reliability score and the contour linearity score is set to 0.5, reflecting the equal importance of depth and width measurements in the final result quality assessment. The calculated overall reliability score is compared with a predefined quality level threshold to determine the quality level of the measurement result. Three quality levels are set according to preset thresholds: Level A (default value 0.9, overall reliability score greater than or equal to Level A threshold is considered Level A); Level B (default value 0.7, score greater than or equal to Level B threshold but less than Level A threshold is considered Level B); and Level C (overall reliability score lower than Level B threshold is considered Level C). Grade A represents high-quality data that can be directly used for process control, Grade B represents usable data that needs to be monitored, and Grade C requires remeasurement. Ultimately, the system integrates trench depth, trench width, overall reliability score, and corresponding quality level to form a final measurement result containing complete measurement data and quality assessment.

[0093] 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.

Claims

1. A precision optical measurement method for GPP chip trench size, characterized in that, Comprise: S1: collect low-power visual image for positioning of groove area and calculation of position confidence score, output center coordinates and azimuth angle parameters of groove area; S2: based on the center coordinates of the groove area, drive the spectral confocal sensor to collect height data, extract waveform quality parameters to generate depth calculation reliability, and when the depth calculation reliability meets the preset condition, calculate the groove depth value, wherein the S2 comprises: based on the center coordinates and azimuth angle parameters of the groove area output by S1, adjusting the position of the film holding table to the lower side of the spectral confocal laser measuring device; drive the spectral confocal laser measuring device to scan along the vertical groove direction to obtain the original height data sequence; the original height data sequence is subjected to adaptive filtering based on sliding window and dynamic threshold to suppress pulse noise; from the filtered height data sequence, identify and separate the high and low platform regions; based on the high and low platform regions, extract waveform quality parameters, including platform stability coefficient, edge steepness factor and signal completeness index; based on the statistical variance of the high platform region data, calculate the platform stability coefficient; based on the first derivative characteristics of the height data sequence in the jump interval between the platform regions, calculate the edge steepness factor; based on the proportion and continuity of the effective data points in the original height data sequence, calculate the signal completeness index; the platform stability coefficient, edge steepness factor and signal completeness index are weighted and fused to generate the depth calculation reliability; S3: based on the center coordinates of the trench area and the azimuth angle parameter, driving the high-power camera to collect high-power images for image segmentation, respectively extracting edge feature parameters to calculate contour linearity score and direction consistency score, selecting the optimal edge contour through the optimization mechanism and performing linear fitting, calculating the distance between the two fitted straight lines as the trench width, wherein the S3 comprises: based on the center coordinates of the trench area and the azimuth angle parameter output by S1, adjusting the position of the wafer holder to the position of the high-power industrial camera, driving the high-power industrial camera to collect high-power images of the trench area; based on the azimuth angle parameter, the high-power image is directionally corrected and regionally cropped to obtain a first width image containing the first side edge of the trench and a second width image containing the second side edge of the trench; edge feature parameters are extracted from the first width image and the second width image, wherein the edge feature parameters include: gray transition slope representing gray level change feature, texture heterogeneity index representing texture complexity, and phase consistency aggregation degree representing edge structure consistency; based on the first width image and the second width image, for each side edge, the initial edge contour point set is extracted based on the phase consistency aggregation degree; within the neighborhood range of the initial edge contour point set, enhanced feature extraction is performed based on the gray transition slope and the texture heterogeneity index to obtain an enhanced edge contour point set; the contour linearity score and the direction consistency score of the initial edge contour point set and the enhanced edge contour point set are calculated respectively, wherein the contour linearity score is determined by calculating the residual between the contour point set and the fitted straight line, and the direction consistency score is determined by calculating the included angle between the main direction of the contour point set and the theoretical direction of the trench; S4: based on the depth calculation confidence and the contour linearity score, a comprehensive confidence score is calculated to generate a quality level, and combined with the trench depth and the trench width value to form the final measurement result.

2. The method of claim 1, wherein the GPP chip trench size is precisely measured by using the optical method. The S1 comprises: collecting a global image of a wafer preset area containing a plurality of dies using a low-power industrial camera; processing the global image based on a pre-trained semantic segmentation model to obtain a semantic segmentation result containing a trench candidate area; extracting edge distribution features, texture consistency features, and morphological regularity features from each candidate area of the semantic segmentation result.

3. The precision optical measurement method for the trench size of the GPP chip according to claim 2, characterized in that: fusing the edge distribution features, the texture consistency features, and the morphological regularity features to generate a fixed-position confidence score representing the confidence of the area; comparing the fixed-position confidence score with a preset trench determination threshold to determine the area with a fixed-position confidence score greater than the trench determination threshold as an effective trench area; for the effective trench area, calculating the azimuth angle parameter through principal component analysis, mapping the region centroid to the mechanical coordinate system through coordinate transformation by using the preset mechanical coordinate system calibration parameter to obtain the center coordinates.

4. The precision optical measurement method for the trench size of the GPP chip according to claim 1, characterized in that: comparing the depth calculation confidence with a preset depth calculation enabling threshold; When the depth calculation confidence is greater than or equal to the depth calculation enabling threshold, a difference between an average of the high-face platform region data and an average of the low-face platform region data is calculated as a final trench depth value.

5. The method of claim 1, wherein the GPP chip trench size is precisely measured by using the optical measurement method, and the method further comprises: The profile comprehensive score is obtained by weighted sum based on the profile linearity score and the direction consistency score, and one of the initial edge profile point set and the enhanced edge profile point set is selected as the optimal edge profile of the side according to the profile comprehensive score; For the edges of the opposite sides of the trench, the first width image and the second width image are determined by using the reverse scanning method based on the symmetric region, respectively; The optimal edge profiles of the two sides are subjected to linear fitting, and the distance between the two fitted straight lines is calculated as the trench width value.

6. The method of claim 1, wherein the GPP chip trench size is precisely measured by using an optical method. The S4 comprises: Based on the depth calculation confidence output by S2 and the profile linearity score output by S3, a comprehensive confidence score is generated by fusion calculation; The comprehensive confidence score is compared with a pre-defined quality level threshold to determine the quality level of the measurement result; The trench depth value, the trench width value, the comprehensive confidence score and the quality level are associated and output as the final measurement result.

Images