Semiconductor part defect diagnosis method based on multi-modal data fusion
By using multimodal data fusion technology and visual and CMM data, a process potential heteroscedasticity index and adjustment factor are constructed, which solves the problem of missed detection in edge areas of traditional algorithms and realizes efficient and accurate detection and fault source diagnosis of semiconductor parts.
Patent Information
- Application Number
- CN202512019913.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-02-06
AI Technical Summary
Existing detection methods cannot find a balance between high efficiency and high accuracy. Traditional data fusion algorithms suffer from the smoothing effect, which leads to missed defects in edge areas. In particular, it is difficult to accurately detect the consistency of pore size and uniformity of distribution of micropores.
By fusing multimodal data, using visual full-field scanning data and CMM targeted sampling data, we construct process potential energy heteroscedasticity index and potential energy adjustment factor, correct the co-kriging interpolation model, and achieve accurate detection of micropores and fault source diagnosis across the entire field.
It achieves efficient and accurate detection of micropores across the entire field, eliminates systematic errors, improves detection accuracy in edge areas, can identify and locate faulty components of production equipment, and provides process optimization and equipment maintenance guidance for the production line.
Smart Images
Figure CN121482028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic data processing technology, and more specifically to a method for diagnosing defects in semiconductor components based on multimodal data fusion. Background Technology
[0002] In semiconductor wafer manufacturing equipment, the core gas distribution component, the gas spray head, is typically made of high-purity silicon, silicon carbide, or ceramic materials, with thousands of tiny vias distributed on its surface. The uniformity of these micro-holes in terms of pore size, distribution, and inner wall roughness directly determines the flow field distribution of the reactive gas within the vacuum chamber, thus affecting the uniformity of the etching rate on the wafer surface. For such components with massive micro-holes and high precision, existing inspection methods face a trade-off between efficiency and accuracy.
[0003] Currently, the industry mainly uses coordinate measuring machines (CMMs) or industrial inspection (AOI) for inspection. While CMMs offer extremely high accuracy, single-point measurements are too time-consuming, failing to meet the full inspection requirements of mass production environments. They typically only allow for very low-percentage sampling inspections, which easily leads to the omission of random defects in non-sampling areas. While industrial inspection can instantly cover the entire area, it is limited by the special reflective properties of the spray head surface, the graininess of ceramic surfaces, and the large aspect ratio of micropores. This often results in blurred image edges, leading to significant systematic errors in the measured values, making it difficult to directly use for sub-micron tolerance determination.
[0004] Existing improvement schemes attempt to use data fusion techniques, such as co-kriging, to calibrate visual data using CMM data. However, traditional algorithms are typically based on the second-order stationarity assumption, which assumes a uniform error distribution across the entire field. In actual machining, the spray head is affected by the CNC helical feed path, accumulated tool wear, and edge fixture stress. Error fluctuations in the edge region are much greater than in the center region, exhibiting significant spatial heteroscedasticity. Traditional algorithms forcibly smooth out these sharp edge fluctuations, causing continuous, small out-of-tolerance errors caused by tool chipping, distributed in a spiral pattern, to be masked by background noise, resulting in missed detections of critical functional defects. Summary of the Invention
[0005] This invention provides a semiconductor component defect diagnosis method based on multimodal data fusion to solve the problem of missed detection of edge region defects caused by the smoothing effect in traditional data fusion algorithms in the prior art.
[0006] The semiconductor component defect diagnosis method based on multimodal data fusion of the present invention includes the following steps: Acquire visual full-field scanning data and CMM targeted sampling data of the semiconductor component surface, and unify the visual full-field scanning data and the CMM targeted sampling data into the same mechanical coordinate system using a common reference point; The process potential energy heteroscedasticity index is calculated for each micro-hole location on the surface of a semiconductor component. The process potential energy heteroscedasticity index is determined by the radial distance of the micro-hole and the local visual variance, and is used to characterize the degree of error divergence in spatial location. Based on the process potential energy heteroscedasticity index, a potential energy adjustment factor is constructed. The weights and errors in the co-kriging interpolation model are corrected using the potential energy adjustment factor, and the predicted true value of each micropore in the entire field is calculated. The predicted true value is compared with the design tolerance of the micropore. Micropores that exceed the tolerance range are marked as defective holes. Based on the spatial distribution morphology of the defective holes on the two-dimensional plane, the processing failure source that causes the defect is diagnosed.
[0007] The beneficial effects are as follows: First, by spatially registering high-efficiency visual data with high-precision CMM data, coordinate differences between different measuring devices can be eliminated, laying a unified and accurate data foundation for subsequent complementary use of multimodal data and elimination of systematic errors. Second, by introducing a process potential energy heteroscedasticity index containing physical laws, the reliability of data in edge regions and areas with high visual noise can be effectively assessed, accurately quantifying instability in the processing process, providing a scientific basis for non-stationary fusion, and solving the problem that traditional algorithms cannot adapt to spatial heteroscedasticity. Furthermore, by dynamically adjusting the correction intensity of CMM errors through a potential energy adjustment factor, the true value correction weight is amplified in high-risk edge regions, breaking the stationarity assumption of traditional algorithms. This enables precise capture of minute fluctuation defects masked by background noise, improving the detection accuracy in edge regions. Moreover, by analyzing the spatial morphological characteristics of defects, not only is virtual full inspection of micropores across the entire field achieved, eliminating the risk of random missed detections, but specific spiral distribution characteristics can also be identified to reverse-locate specific faulty components of the production equipment.
[0008] Preferably, the process potential heteroscedasticity index is calculated using the following formula: ; In the formula, for Heteroscedasticity index of process potential energy for micro-orifices at location Location of micropores Based on the basic process constants, The weighting coefficients for geometric location factors. The weighting coefficients for visual fluctuation factors. for The Euclidean distance from the micro-via to the center of the semiconductor component. The maximum radius of the semiconductor component. for Local visual variance of visual measurements near the micro-hole.
[0009] Preferred local visual variance The method of obtaining it is: Centered on the micropore, a neighborhood within a preset radius is selected, and the statistical variance of the visual measurements of all micropores within that neighborhood is calculated.
[0010] Preferably, the basic process constant is obtained by statistically analyzing the reference vibration error during the no-load operation of the equipment, and the basic process constant... ; and It is a non-negative real number.
[0011] Preferably, the predicted true value is calculated using the following formula: ; In the formula, for The true value of the predicted location of the micropore to be detected. The location of the micropore to be detected. for Visual measurements of the location of the micropore to be detected. The number of CMM sampling points. For the first The location of each CMM sampling point For the first Spatial weights of CMM sampling points For the first The true value of the CMM at each CMM sampling point For the first Visual measurements of each CMM sampling point It is the potential energy adjustment factor.
[0012] Preferably, the first The spatial weights of the CMM sampling points satisfy the unbiased estimation condition. And obtained by solving the system of variogram equations.
[0013] Preferably, the potential energy adjustment factor is used to adjust the correction effect of observation error on the result: when the micropore to be detected is located in the high potential energy region at the edge and the CMM sampling point is located in the low potential energy region at the center, the potential energy adjustment factor is greater than 1.
[0014] Preferably, the step of diagnosing the processing failure source causing the defect based on the spatial distribution morphology of the defective hole on a two-dimensional plane includes: All defective holes are plotted as heat maps on a two-dimensional plane, and the shape of the defective area is analyzed using image morphology algorithms. If the defective area exhibits the distribution characteristics of an Archimedean spiral, the fault source is determined to be tool chipping during the CNC finishing stage.
[0015] Preferably, the step of diagnosing the machining process fault source that causes the defect based on the spatial distribution morphology of the defective hole on the two-dimensional plane further includes: calculating the time point at which the tool chipping occurs based on the span of the spiral in the radial direction and the feed speed parameters of the CNC machine tool.
[0016] Preferably, the visual full-field scanning data is acquired by an area array camera equipped with dual telecentric lenses under bottom parallel backlight illumination; the sampling strategy for the CMM targeted sampling data is: sparsely distributed points in the central area, and densely distributed points in the edge area and at the tool entry and exit marks.
[0017] The beneficial effects of this invention are as follows: This invention breaks the stationarity assumption of traditional algorithms by introducing processing physical laws to construct an anisotropic covariance function, achieving precise weighting of high-risk areas. Without increasing the number of CMM measurement points, this invention uses mathematical methods to transfer the high-precision attributes of the CMM to the entire field of visual data, achieving virtual full inspection. Furthermore, by introducing process potential field correction, this invention solves the pain points of large edge processing errors in semiconductor parts and the easy failure of traditional algorithms, improving the detection rate of tiny spiral defects in edge regions. In addition, this invention can infer the specific fault source and occurrence time in the CNC machining process based on the spatial morphological characteristics of defects, providing data support for intelligent manufacturing process optimization. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating a semiconductor component defect diagnosis method based on multimodal data fusion, as provided in an embodiment of the present invention.
[0019] Figure 2 This image shows a comparison of the effectiveness of the present invention and existing technologies in detecting micropore defects in spray heads. The left image represents the detection results of the existing technology, while the right image represents the detection results of the present invention. Detailed Implementation
[0020] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0021] like Figure 1 As shown, the semiconductor component defect diagnosis method based on multimodal data fusion provided in this embodiment of the invention specifically includes the following steps: S1. Obtain visual full-field scanning data and CMM targeted sampling data of the semiconductor component surface, and use a common reference point to unify the visual full-field scanning data and the CMM targeted sampling data into the same mechanical coordinate system.
[0022] Furthermore, the visual full-field scanning data is acquired by an area array camera equipped with dual telecentric lenses under bottom parallel backlight illumination; the sampling strategy for the CMM targeted sampling data is: sparsely distributed points in the central area, and densely distributed points in the edge area and at the tool entry and exit marks.
[0023] Specifically, in this embodiment, an industrial vision inspection platform equipped with a 25-megapixel area array camera and dual telecentric lenses first performs a full-field scan of the gas spray head under illumination from a bottom parallel backlight. Image processing techniques and a Zernike moment-based sub-pixel edge detection algorithm are then used to extract the contours of all micro-holes in the full field, obtaining full-field visual scan data. This full-field visual scan data includes the position and visual aperture of each hole. For example, the image processing technique used could be filtering and noise reduction.
[0024] Simultaneously, a high-precision coordinate measuring machine (CMM) with a 0.3mm ruby stylus is used for targeted sampling inspection. The sampling strategy is as follows: sparse sampling is performed in the central area, for example, one point is taken every 50mm, while dense sampling is performed in the edge areas and at the tool entry and exit marks, for example, one point is taken every 10mm, thus obtaining CMM targeted sampling inspection data. For example, the edge area can be defined as an area with a radius greater than 80% of the maximum radius.
[0025] Finally, using the three locating pin holes on the sprinkler head design drawings as common reference points, a rigid body transformation matrix involving rotation and translation was calculated. This allowed for precise mapping of the visual coordinate system to the CMM coordinate system, ensuring that the coordinate error of the same hole in both sets of data was controlled within 0.01 mm, thus completing the spatial registration of multimodal data. In this way, high-precision coordinate unification ensured the spatial consistency of different modal data, providing an accurate positional reference for subsequent calibration of visual data deviations using CMM data.
[0026] S2, calculate the process potential energy heteroscedasticity index of each micro-hole position on the surface of the semiconductor component. The process potential energy heteroscedasticity index is determined by the radial distance of the micro-hole and the local visual variance, and is used to characterize the degree of error divergence in spatial position.
[0027] In this embodiment, considering that the edge areas are greatly affected by stress and have high visual noise in CNC machining, it is necessary to construct a process potential energy heteroscedasticity index to evaluate the instability of the data. The process potential energy heteroscedasticity index is calculated using the following formula: ; In the formula, for Heteroscedasticity index of process potential energy for micro-orifices at location Location of micropores Based on the basic process constants, The weighting coefficients for geometric location factors. The weighting coefficients for visual fluctuation factors. for The Euclidean distance from the micro-via to the center of the semiconductor component. The maximum radius of the semiconductor component. for Local visual variance of visual measurements near the micropore.
[0028] The inherent noise floor of the equipment is represented by the basic process constant, which is obtained by statistically analyzing the reference vibration error during the equipment's no-load operation. .For example, It is 0.01. and It is a non-negative real number. and Obtained by fitting historical data. Local visual variance. The method of obtaining it is: Centered on the micropore, a neighborhood within a preset radius is selected, and the statistical variance of the visual measurements of all micropores within that neighborhood is calculated.
[0029] For example, suppose , , , For micropores located at the center Assuming its Euclidean distance from the center of the semiconductor component is... ,and . micropores The local visual variance of visual measurements within a 5mm radius, and This indicates that micropores The fluctuations in surrounding visual measurements are minimal. The micropore located at the center was calculated using the formula for calculating the heteroscedasticity index of process potential energy. The heteroscedasticity index of the process potential energy is 0.261.
[0030] For micropores located at the edge Assuming its Euclidean distance from the center of the semiconductor component is... ,and , micropores The local visual variance of visual measurements within a 5mm radius, and This indicates that micropores The surrounding visual measurements fluctuated significantly. The micropores located at the edges were calculated using the formula for calculating the heteroscedasticity index of process potential energy. The heteroscedasticity index of the process potential energy is 90.309.
[0031] It can be seen that the heteroscedasticity index of the process potential energy of micro-holes at the edge is much larger than that at the center, accurately characterizing the fact that the processing quality in the edge region is unstable and the measurement noise is high. Thus, by constructing a process potential energy heteroscedasticity index that reflects physical laws, differentiated assessment of the reliability of data in different regions is achieved, effectively avoiding the drawbacks of traditional algorithms that uniformly process errors across the entire field.
[0032] S3. Based on the process potential energy heteroscedasticity index, a potential energy adjustment factor is constructed. The weights and errors in the co-kriging interpolation model are corrected using the potential energy adjustment factor, and the predicted true value of each micropore in the entire field is calculated.
[0033] In this embodiment, in order to accurately reproduce the true size, the following method is used: Improved co-kriging formula. True prediction value for every micropore across the entire field. Calculated using the following formula: ; In the formula, for The true value of the predicted location of the micropore to be detected. The location of the micropore to be detected. for Visual measurements of the location of the micropore to be detected. The number of CMM sampling points. For the first The location of each CMM sampling point For the first Spatial weights of CMM sampling points For the first The true value of the CMM at each CMM sampling point For the first Visual measurements of each CMM sampling point It is the potential energy adjustment factor.
[0034] in, Represents the sampling points in CMM Visual errors observed at the location; It is a spatial weight based on distance allocation. The first... The spatial weights of the CMM sampling points satisfy the unbiased estimation condition. The potential energy adjustment factor is obtained by solving the system of variogram equations. The potential energy adjustment factor is used to adjust the correction effect of observation error on the results: when the micropore to be detected is located in the high potential energy region at the edge and the CMM sampling point is located in the low potential energy region at the center, the potential energy adjustment factor is greater than 1.
[0035] For example, suppose we want to predict the micropores in the high-potential region at the edge. True value of prediction And use a CMM sampling point located in the central low potential energy region. (Right now This is used as a reference point for correction. Assume micropores... The original visual measurements The value is 0.520 mm; assuming CMM sampling points CMM Truth Value The value is 0.500mm, a visual measurement. If it is 0.505mm, then in The visual error of a point is At this point, the calculated potential energy adjustment factor is 18.6.
[0036] At this point, the error correction amount is This means that, due to micropores Located in the high potential energy region at the edge, micropores The visual data is highly unreliable, therefore the CMM sampling points located in the central low-potential region are... The error correction is amplified by 18.6 times. When the adjustment factor of the traditional algorithm is 1, its error correction is only -0.005. Compared with the traditional algorithm, the error correction is amplified by 18.6 times. Therefore, the present invention can strongly correct micropores in the edge region, greatly enhance the sensitivity of the edge region to errors, and make the real processing deviations that are masked by visual noise visible, thus improving the detection capability of edge defects.
[0037] S4. The predicted true value is compared with the design tolerance of the micropore. Micropores that exceed the tolerance range are marked as defective holes. Based on the spatial distribution morphology of the defective holes on the two-dimensional plane, the processing failure source that causes the defect is diagnosed.
[0038] In this embodiment, a virtual inspection report containing the true values of the full-field micropore prediction is automatically generated. The design tolerance is set to... If the true value of a certain micropore is If it is, then mark it as a defective hole.
[0039] Furthermore, all defective holes are plotted as heat maps on a two-dimensional plane, and the shape of the defective area is analyzed using image morphology algorithms. If the defective area exhibits the distribution characteristics of an Archimedean spiral, the fault source is determined to be tool chipping during the CNC finishing stage.
[0040] like Figure 2As shown in the right figure, the dark red highlighted spots represent severely out-of-tolerance apertures. These spots exhibit a discontinuous Archimedean spiral distribution, and the spiral trajectory closely matches the helical feed path of the CNC machine tool. Therefore, the fault source is determined to be tool chipping during the CNC finishing stage. Furthermore, based on the radial span of the spiral and the feed rate parameters of the CNC machine tool, the specific time point when the tool chipping occurred can be directly calculated, guiding operators to check the tool status records for the corresponding period. Thus, by analyzing the spatial morphology of the defective area, a process fault diagnosis process has been developed, moving beyond simple defective product screening to in-depth analysis, providing direct data guidance for production line process optimization and equipment maintenance.
[0041] like Figure 2 As shown, the effect comparison between the present invention and the prior art is illustrated. In the prior art shown in the left figure, the edge deviation is smoothed, the boundary is unclear, and it is easy to miss the detection. In the present invention shown in the right figure, the dark red area indicates that the aperture is seriously out of tolerance, showing obvious processing path characteristics. The background noise is effectively suppressed, and the defect characteristics of the Archimedean spiral are clearly revealed, achieving accurate diagnosis.
[0042] The implementation principle of the semiconductor component defect diagnosis method based on multimodal data fusion in this invention is as follows: This invention constructs an anisotropic covariance function by introducing processing physical laws, achieving precise weighting of high-risk areas. Furthermore, without increasing the number of CMM measurement points, this invention uses mathematical methods to transfer the high-precision attributes of the CMM to the full-field visual data, achieving virtual full inspection. Moreover, by introducing process potential field correction, this invention solves the pain points of large edge processing errors and easy failure of traditional algorithms in semiconductor components, improving the detection rate of small spiral defects in edge areas. In addition, this invention can infer the specific fault source and occurrence time during CNC machining based on the spatial morphological characteristics of the defects, providing data support for intelligent manufacturing process optimization.
[0043] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for defect diagnosis of semiconductor components based on multimodal data fusion, characterized in that, Includes the following steps: Obtain visual full-field scanning data and CMM targeted sampling data of the semiconductor component surface, and use a common reference point to unify the visual full-field scanning data and the CMM targeted sampling data into the same mechanical coordinate system; The process potential energy heteroscedasticity index is calculated for each micro-hole location on the surface of a semiconductor component. The process potential energy heteroscedasticity index is determined by the radial distance of the micro-hole and the local visual variance, and is used to characterize the degree of error divergence in spatial location. Based on the process potential energy heteroscedasticity index, a potential energy adjustment factor is constructed. The weights and errors in the co-kriging interpolation model are corrected using the potential energy adjustment factor, and the predicted true value of each micropore in the entire field is calculated. The predicted true value is compared with the design tolerance of the micropore. Micropores that exceed the tolerance range are marked as defective holes. Based on the spatial distribution morphology of the defective holes on the two-dimensional plane, the processing failure source that causes the defect is diagnosed.
2. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The process potential energy heteroscedasticity index is calculated using the following formula: ； In the formula, for Heteroscedasticity index of process potential energy for micro-orifices at location Location of micropores Based on the basic process constants, The weighting coefficients for geometric location factors. The weighting coefficients for visual fluctuation factors. for The Euclidean distance from the micro-via to the center of the semiconductor component. The maximum radius of the semiconductor component. for Local visual variance of visual measurements near the micro-hole.
3. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 2, characterized in that, Local visual variance The method of obtaining it is: Centered on the micropore, a neighborhood within a preset radius is selected, and the statistical variance of the visual measurements of all micropores within that neighborhood is calculated.
4. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 2, characterized in that, The basic process constants are obtained by statistically analyzing the reference vibration error of the equipment during no-load operation, and the basic process constants ; and It is a non-negative real number.
5. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The predicted true value is calculated using the following formula: ； In the formula, for The true value of the predicted location of the micropore to be detected. The location of the micropore to be detected. for Visual measurements of the location of the micropore to be detected. The number of CMM sampling points. For the first The location of each CMM sampling point For the first Spatial weights of each CMM sampling point For the first The true value of the CMM at each CMM sampling point For the first Visual measurements of each CMM sampling point It is the potential energy adjustment factor.
6. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 5, characterized in that, The first The spatial weights of the CMM sampling points satisfy the unbiased estimation condition. And obtained by solving the system of variogram equations.
7. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 5, characterized in that, The potential energy adjustment factor is used to adjust the correction effect of observation error on the results: when the micropore to be detected is located in the high potential energy region at the edge and the CMM sampling point is located in the low potential energy region at the center, the potential energy adjustment factor is greater than 1.
8. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The method of diagnosing the processing failure source causing the defect based on the spatial distribution morphology of the defective hole on a two-dimensional plane includes: All defective holes are plotted as heat maps on a two-dimensional plane, and the shape of the defective area is analyzed using image morphology algorithms. If the defective area exhibits the distribution characteristics of an Archimedean spiral, the fault source is determined to be tool chipping during the CNC finishing stage.
9. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 8, characterized in that, The method of diagnosing the machining process fault source that causes the defect based on the spatial distribution morphology of the defect hole on the two-dimensional plane also includes: calculating the time point when the tool chipping occurs based on the span of the spiral in the radial direction and the feed speed parameters of the CNC machine tool.
10. The semiconductor component defect diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The full-field visual scanning data is acquired by an area array camera equipped with dual telecentric lenses under bottom parallel backlight illumination; the sampling strategy for the CMM targeted sampling data is: sparsely distributed points in the central area and densely distributed points in the edge area and at the tool entry and exit marks.