Non-destructive testing method for internal cracks in a wafer based on acoustic scanning microscopy

Abstract

The application discloses a wafer internal crack nondestructive detection method based on an acoustic scanning microscope, and relates to the field of wafer detection, and comprises the following steps: step 1: adopting a reflection pulse echo mode of the acoustic scanning microscope, emitting ultrasonic waves of an initial preset frequency, receiving returned reflection wave signals, and preliminarily estimating the position of a wafer internal crack based on the characteristics of the reflection waves; step 2: adjusting a time gate and an ultrasonic frequency according to the thickness of the wafer and the estimated position of the crack, focusing on a specific depth plane in the wafer, and performing tomographic imaging scanning to locate the depth position of all initial cracks in the wafer; the application can effectively eliminate signal crosstalk caused by sound wave diffusion and measurement differences between regions, restore more real original crack depth and topography through deconvolution operation, make the finally obtained crack three-dimensional position and size information more accurate and reliable, and reduce the risk of misjudgment and missed judgment.

Description

Technical Field

[0001] This invention relates to the field of wafer inspection technology, specifically to a non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy. Background Technology

[0002] As the feature size of integrated circuits continues to shrink to the nanometer level, any tiny crack within the wafer, the substrate of the chip, can become a critical hidden danger leading to device failure and reduced product yield. Traditional physical slicing inspection methods are destructive and cannot meet the full inspection requirements of large-scale production. Acoustic scanning microscopy, due to its ability to image the interior of materials using ultrasound, has become one of the mainstream methods for non-destructive testing of wafers. However, wafers have complex structures, diverse morphologies and causes of internal cracks, and multilayer thin-film structures cause complex reflections, refractions, and attenuations of acoustic signals.

[0003] While existing acoustic scanning technology can achieve internal imaging, the crack signals it acquires are easily distorted by various factors. When ultrasound propagates in a multilayer heterostructure of a wafer, it will undergo complex reflection, refraction and mode conversion at the interface, causing crack signals from different depths or adjacent regions to superimpose and diffuse, resulting in deviations between the initially detected crack morphology, boundary and depth location and the actual physical state. Existing technologies generate tomographic images that require manual interpretation by experienced engineers to guess the possible causes of cracks. It is difficult to accurately reconstruct the original depth, morphology, and other key features of cracks from distorted detection signals. It is also difficult to automatically and accurately pinpoint the specific process steps or parameter anomalies that cause defects from massive amounts of detection data. This cannot meet the needs of modern semiconductor manufacturing for real-time data feedback for process window monitoring and rapid root cause correction. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the prior art, the present invention provides a non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy, which can effectively solve the problems of the prior art.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This invention discloses a non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy, comprising the following steps: Step 1: Using the reflected pulse echo mode of an acoustic scanning microscope, an ultrasonic wave of an initial preset frequency is emitted, and the returned reflected wave signal is received. Based on the characteristics of the reflected wave, the location of the internal crack of the wafer is initially estimated. The estimation process is as follows: the propagation path length of the sound wave is calculated based on the propagation speed of the sound wave in the wafer material and the arrival time difference of the returned sound wave, the depth range of the crack is initially determined, the amplitude attenuation of the returned sound wave is analyzed, and based on the energy loss law caused by the reflection and scattering of the sound wave at the crack interface, the existence and relative size of the crack are determined. The waveform characteristics of the returned sound wave are extracted, and waveform distortion, frequency component changes or multiple echo phenomena caused by the crack interface are identified to cross-verify the estimated location of the crack and improve the reliability of the location. Step 2: Based on the wafer thickness and the estimated crack location, adjust the time gate and ultrasonic frequency. Adjusting the acoustic frequency and time gate is an adaptive adjustment based on the sound velocity and attenuation characteristics of the sound wave in the wafer material to ensure optimal acoustic resolution at the focusing plane. Focus on a specific depth plane inside the wafer and perform tomographic imaging scanning. By adjusting the time gate width and delay time, focusing scanning of different depth planes can be achieved to locate the depth of all initial cracks inside the wafer. Step 3: Divide the wafer scanning space involving frequency adjustment into a unified standard size to form multiple standard regions; the division of the unified standard size is based on the lateral resolution limit of the acoustic scanning microscope and the typical characteristic size of the wafer internal structure, and the size of each standard detection region is no greater than 1.5 times the diameter of the acoustic beam focal point. Step 4: Based on the amplitude, phase, frequency attenuation spectrum of the fused reflected wave and the fluctuation state of the transmitted wave data during the adjustment process, evaluate the adjustment importance of each standard region one by one, and merge or reduce the region edges according to the evaluation results to form a specific region. Step 5: Construct an edge association model using a deep learning algorithm, taking the initial crack edge position and depth of adjacent specific regions as input, and outputting the mutual interference coefficient of edge cracks in adjacent regions; Step 6: Based on the mutual interference coefficient, perform deconvolution calculation on the initial crack depth location data to eliminate crack signal diffusion caused by acoustic wave mutual interference and material boundary effects, and restore and identify the original crack location depth of the wafer. Step 7: Based on the original crack location and depth data, match within the preset process factor library to identify one or more related process factors that cause the crack depth and location state.

[0006] Furthermore, the ultrasonic wave with the preset frequency in step 1 is a broadband high-frequency sound wave with a center frequency range of 10MHz to 300MHz; the characteristics of the reflected wave include amplitude, phase and time of flight.

[0007] Furthermore, the process of adjusting the importance assessment of each region in step 4 includes the following steps: Acquire the fused acoustic dataset for each standard detection area during the acoustic frequency modulation process. This dataset includes the amplitude sequence, phase change sequence, frequency attenuation spectrum sequence, and transmitted wave energy sequence of the reflected wave. Calculate the volatility index of each sequence, which includes at least the variance of the amplitude sequence, the root mean square error of the phase change sequence, the slope change rate of the frequency decay spectrum sequence, and the number of abrupt change points of the transmitted wave energy sequence. Based on preset weighting coefficients, the volatility indicators are weighted and summed to calculate the initial importance score for each standard detection area; Based on the scanning path and sound field distribution characteristics of the acoustic scanning microscope, the initial importance score is spatially correlated to obtain the final modulated importance evaluation value.

[0008] Furthermore, the formation process of the specific region in step 4 is as follows: Set a first evaluation threshold and a second evaluation threshold, wherein the first evaluation threshold is higher than the second evaluation threshold; Standard detection areas with adjustment importance assessment values ​​higher than the first assessment threshold are marked as high-concern core areas; The region whose adjustment importance assessment value is between the first assessment threshold and the second assessment threshold is marked as the transition region to be merged; the region whose adjustment importance assessment value is below the second assessment threshold is marked as the low attention background region. Using the high-attention core area as the seed point, search and merge consecutive transition areas to be merged within the spatial neighborhood to form an initial specific region; The edges of the initial specific region are detected. If the evaluation value of the edge portion of the region is lower than the second evaluation threshold and the continuous area is less than the preset area threshold, the region is reduced and merged into the adjacent specific region. If the evaluation value of the edge portion of the region is continuously lower than the second evaluation threshold and the continuous area exceeds the preset area threshold, the region is separated from the specific region and classified into the low-interest background region. The processed high-attention core area and the merged transition area to be merged together constitute a specific area with targeted detection.

[0009] Furthermore, the process of constructing the edge association model in step 5 includes the following steps: Step 51, Data Collection: Collect the true crack edge location and depth data obtained from historical wafer acoustic scanning detection. The data includes crack edge location and depth data obtained from multiple sets of adjacent detection areas after initial tomographic imaging, as well as true crack edge location and depth data verified and confirmed by destructive physical analysis or high-resolution imaging technology after eliminating mutual interference. Step 52, Framework Construction: A neural network with an encoder and decoder structure is adopted. The encoder part consists of multiple cascaded convolutional layers, which are used to extract spatial features and correlation patterns in the crack depth data of adjacent regions. The decoder part includes fully connected layers and regression output layers, which are used to map the extracted features into mutual interference coefficients. Step 53, Framework Construction: Using the initial crack edge location and depth data of adjacent regions in the historical data as the model input, and the mutual interference coefficient calculated based on the actual crack edge location and depth data as the training label, the mean squared error loss function and adaptive moment estimation algorithm are used to supervise the training of the model until the model converges; Step 54, Framework Construction: Use an independent validation dataset to evaluate the performance of the trained model. After successful validation, apply the model to the detection of the current wafer. Input the initial crack edge position and depth of the current adjacent specific region, and the model will output the corresponding mutual interference coefficient.

[0010] Furthermore, the formula for calculating the mutual interference coefficient of edge cracks in adjacent regions in step 5 is as follows: ; In the formula, Represents a specific area and adjacent specific areas The mutual interference coefficient between the crack edge locations and depths. and Representing regions and The set of pixels at the crack edge and They belong to and The pixels in the set represent, and Representing pixels and The initial crack depth value. represent and The absolute value of the depth difference between two points. An empirical attenuation constant representing the depth of influence range. Representative point and points The phase difference of the reflected sound wave received at the location, Representative based on point and points The weighting coefficient for spatial distance decreases as the Euclidean distance between two points increases.

[0011] Furthermore, the deconvolution calculation process in step 6 includes the following steps: Based on the calculated mutual interference coefficients between adjacent specific regions, a system point spread function matrix is ​​constructed to characterize the spatial diffusion and coupling of acoustic signals during propagation inside the wafer. Each element of this matrix reflects the contribution weight of the crack signal at a specific spatial location to the observed signal value in its neighboring region. Its core is determined by the mutual interference coefficients and the known acoustic attenuation model. The initial crack depth location data, which includes signal diffusion caused by acoustic interference and material boundary effects, is used as the fuzzy observation result under the action of the diffusion function matrix at that point. An iterative deconvolution algorithm based on maximum a posteriori probability estimation is adopted. The fuzzy observation result is used as input, and the system point spread function matrix is ​​used as the known degradation model. Through constrained iterative solution, the crosstalk components between adjacent crack signals and the signal tailing effect caused by material interface reflection are gradually stripped and removed, and the intermediate crack depth distribution data after eliminating interference factors is output.

[0012] Furthermore, the process of restoring and identifying the original crack location and depth on the wafer in step 6 includes the following steps: Step 61: Perform threshold segmentation and edge enhancement processing on the intermediate crack depth distribution data obtained by deconvolution calculation to highlight the real crack boundary; Step 62: Using the connected component analysis algorithm, based on the enhanced data, identify and separate individual cracks, and calculate the geometric center depth of each crack and its contour distribution in the depth direction. Step 63: Compare this profile distribution with the standard acoustic impedance profile of the wafer material, perform iterative optimization, and correct the slight deviation in depth measurement caused by material inhomogeneity. Step 64: Output the original location coordinates of the crack and its corresponding depth value after eliminating signal diffusion interference.

[0013] Furthermore, the working logic of the connected component analysis algorithm in step 62 is as follows: The deblurred crack depth data after threshold segmentation and edge enhancement is converted into a binary image, where pixels above a preset depth threshold are marked as foreground points, representing possible crack areas. The binary image is traversed using a connected component labeling algorithm based on eight or four neighborhoods. A unique identifier is assigned to the set of spatially connected foreground points, and all connected pixel blocks in the image are identified and separated into several independent candidate crack individuals. For each marked individual crack, calculate the set of coordinates and depth values ​​of all its constituent pixels in three-dimensional space; The geometric center depth of the crack individual is obtained by calculating the arithmetic mean of all the depth values ​​constituting the pixel; its contour distribution in the depth direction is obtained by slicing along the depth axis direction perpendicular to the wafer surface with a preset depth interval as the step size. The number of pixels or projected area belonging to the crack individual on each depth slice is counted to generate a distribution histogram representing the morphological expansion of the crack in the depth direction.

[0014] Furthermore, the matching process within the preset process factor library in step 7 includes the following steps; Extract the depth distribution features of the original crack, including crack depth gradient, spatial clustering, and interface correlation; Based on the crack depth distribution characteristics, matching combinations of process parameters are retrieved from a preset process factor library. The process factor library includes at least: wafer manufacturing process data: cutting speed, grinding pressure, polishing fluid composition; packaging process data: bonding temperature, molding material shrinkage rate, hot pressing parameters; and reliability test data: temperature cycle number, mechanical impact strength. A similarity matching algorithm is used to calculate the correlation weight between crack depth distribution characteristics and combinations of various process factors, and one or more key process factors whose weights exceed a preset threshold are selected. Based on key process factors and the depth, location, and state of the crack, the crack initiation analysis results are generated.

[0015] (III) Beneficial Effects Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: 1. By introducing specific region division and importance assessment, as well as a deep learning-based edge correlation model, proactive compensation for systematic errors in the acoustic detection process is achieved. By calculating and utilizing the mutual interference coefficient, signal crosstalk caused by sound wave diffusion and measurement differences between regions can be effectively removed. Through deconvolution operation, the original crack depth and morphology that are closer to reality are restored, making the final obtained three-dimensional crack position and size information more accurate and reliable. This lays a high-quality data foundation for subsequent accurate analysis and reduces the risk of misjudgment and missed judgment.

[0016] 2. After obtaining high-fidelity raw crack depth data, the system automatically matches it with a preset process factor library. The system can directly output one or more potential related process factors that cause this type of crack, directly upgrading the detection process to the intelligent diagnosis process. This reduces the over-reliance on the human experience of senior engineers, enabling process engineers to quickly focus on the source of the problem. It greatly shortens the cycle from discovering the problem to locating the root cause of the process, providing a direct quantitative basis for rapid decision-making and process adjustment on the production line.

[0017] 3. By forming specific regions through standardized spatial division and dynamic evaluation based on multiple parameters, the detection strategy is no longer fixed but can be adaptively optimized according to the actual signal response of the wafer under test. This improves the fault tolerance rate of detection under complex and variable operating conditions, enabling the detection method to not only serve the failure analysis of single wafers but also be integrated into the production line, becoming a powerful tool for stability monitoring and early warning of key process steps. By continuously analyzing the correlation trend between crack type and process parameters, preventive quality control can be achieved, improving the overall production yield from the source. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart illustrating the process of constructing the edge association model in this invention; Figure 3 This is a schematic flowchart illustrating the process of restoring and identifying the depth of the original crack location on the wafer in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] The present invention will be further described below with reference to embodiments.

[0022] This embodiment presents a non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy, such as... Figure 1 - Figure 3 As shown, it includes the following steps: Step 1: Using the reflected pulse-echo mode of an acoustic scanning microscope, an ultrasonic wave of an initial preset frequency is emitted, and the returned reflected wave signal is received. Based on the characteristics of the reflected wave, the location of the internal crack of the wafer is initially estimated. The estimation process is as follows: the propagation path length of the sound wave is calculated based on the propagation speed of the sound wave in the wafer material and the arrival time difference of the returned sound wave, the depth range of the crack is initially determined, the amplitude attenuation of the returned sound wave is analyzed, and based on the energy loss law caused by the reflection and scattering of the sound wave at the crack interface, the existence and relative size of the crack are determined. The waveform characteristics of the returned sound wave are extracted, and waveform distortion, frequency component changes, or multiple echo phenomena caused by the crack interface are identified to cross-verify the estimated location of the crack and improve the reliability of the location. The ultrasonic wave of the preset frequency is a broadband high-frequency sound wave with a center frequency range of 10MHz to 300MHz. The characteristics of the reflected wave include amplitude, phase, and time of flight.

[0023] Step 2: Based on the wafer thickness and the estimated crack location, adjust the time gate and ultrasonic frequency. Adjusting the acoustic frequency and time gate is an adaptive adjustment based on the sound velocity and attenuation characteristics of the sound wave in the wafer material to ensure optimal acoustic resolution at the focusing plane. Focus on a specific depth plane inside the wafer for tomographic imaging scanning. By adjusting the time gate width and delay time, focusing scanning of different depth planes can be achieved to locate the depth of all initial cracks inside the wafer.

[0024] Step 3: Divide the wafer scanning space involving frequency adjustment into a unified standard size to form multiple standard regions; the division of the unified standard size is based on the lateral resolution limit of the acoustic scanning microscope and the typical characteristic size of the internal structure of the wafer, and the size of each standard detection region is no greater than 1.5 times the diameter of the acoustic beam focal point.

[0025] Step 4: Based on the amplitude, phase, and frequency attenuation spectrum of the fused reflected wave and the fluctuation state of the transmitted wave data during the adjustment process, the adjustment importance of each standard region is evaluated one by one, and the region edges are merged or reduced according to the evaluation results to form specific regions; the process of evaluating the adjustment importance of each region includes the following steps: Acquire the fused acoustic dataset for each standard detection area during the acoustic frequency modulation process. This dataset includes the amplitude sequence, phase change sequence, frequency attenuation spectrum sequence, and transmitted wave energy sequence of the reflected wave. Calculate the volatility index of each sequence, which includes at least the variance of the amplitude sequence, the root mean square error of the phase change sequence, the slope change rate of the frequency decay spectrum sequence, and the number of abrupt change points of the transmitted wave energy sequence. Based on preset weighting coefficients, the volatility indicators are weighted and summed to calculate the initial importance score for each standard detection area; Based on the scanning path and sound field distribution characteristics of the acoustic scanning microscope, the initial importance score is spatially correlated to obtain the final adjusted importance evaluation value. The evaluation value of the region located at the edge of the sound field or at the turning point of the scanning path will be increased according to a preset rule. The formation process of a specific region is as follows: Set a first evaluation threshold and a second evaluation threshold, wherein the first evaluation threshold is higher than the second evaluation threshold; Standard detection areas with adjustment importance assessment values ​​higher than the first assessment threshold are marked as high-concern core areas; The region whose adjustment importance assessment value is between the first assessment threshold and the second assessment threshold is marked as the transition region to be merged; the region whose adjustment importance assessment value is below the second assessment threshold is marked as the low attention background region. Using the high-attention core area as the seed point, search and merge consecutive transition areas to be merged within the spatial neighborhood to form an initial specific region; The edges of the initial specific region are detected. If the evaluation value of the edge portion of the region is lower than the second evaluation threshold and the continuous area is less than the preset area threshold, the region is reduced and merged into the adjacent specific region. If the evaluation value of the edge portion of the region is continuously lower than the second evaluation threshold and the continuous area exceeds the preset area threshold, the region is separated from the specific region and classified into the low-interest background region. The processed high-attention core area and the merged transition area to be merged together constitute a specific area with targeted detection.

[0026] Step 5: Construct an edge association model using a deep learning algorithm. The model takes the initial crack edge position and depth of adjacent specific regions as input and outputs the mutual interference coefficient of the edge cracks in adjacent regions.

[0027] Step 6: Based on the mutual interference coefficient, perform deconvolution calculation on the initial crack depth location data to eliminate crack signal diffusion caused by acoustic wave mutual interference and material boundary effects, and restore and identify the original crack location and depth of the wafer; the deconvolution calculation process includes the following steps: Based on the calculated mutual interference coefficients between adjacent specific regions, a system point spread function matrix is ​​constructed to characterize the spatial diffusion and coupling of acoustic signals during propagation inside the wafer. Each element of this matrix reflects the contribution weight of the crack signal at a specific spatial location to the observed signal value in its neighboring region. Its core is determined by the mutual interference coefficients and the known acoustic attenuation model. The initial crack depth location data, which includes signal diffusion caused by acoustic interference and material boundary effects, is used as the fuzzy observation result under the action of the diffusion function matrix at that point. An iterative deconvolution algorithm based on maximum a posteriori probability estimation is adopted. The fuzzy observation result is used as input, and the system point spread function matrix is ​​used as the known degradation model. Through constrained iterative solution, the crosstalk components between adjacent crack signals and the signal tailing effect caused by material interface reflection are gradually stripped and removed, and the intermediate crack depth distribution data after eliminating interference factors is output.

[0028] The process of reconstructing and identifying the location and depth of the original crack in a wafer includes the following steps: Step 61: Perform threshold segmentation and edge enhancement processing on the intermediate crack depth distribution data obtained by deconvolution calculation to highlight the real crack boundary; Step 62: Using a connected component analysis algorithm, based on the enhanced data, identify and separate individual cracks, and calculate the geometric center depth of each crack and its contour distribution along the depth direction; the working logic of the connected component analysis algorithm is as follows: The deblurred crack depth data after threshold segmentation and edge enhancement is converted into a binary image, where pixels above a preset depth threshold are marked as foreground points, representing possible crack areas. The binary image is traversed using a connected component labeling algorithm based on eight or four neighborhoods. A unique identifier is assigned to the set of spatially connected foreground points, and all connected pixel blocks in the image are identified and separated into several independent candidate crack individuals. For each marked individual crack, calculate the set of coordinates and depth values ​​of all its constituent pixels in three-dimensional space; The geometric center depth of the crack individual is obtained by calculating the arithmetic mean of all the depth values ​​constituting the pixel; its contour distribution in the depth direction is obtained by slicing along the depth axis direction perpendicular to the wafer surface with a preset depth interval as the step size. The number of pixels or projected area belonging to the crack individual on each depth slice is counted to generate a distribution histogram representing the morphological expansion of the crack in the depth direction. Step 63: Compare this profile distribution with the standard acoustic impedance profile of the wafer material, perform iterative optimization, and correct the slight deviation in depth measurement caused by material inhomogeneity. Step 64: Output the original location coordinates of the crack and its corresponding depth value after eliminating signal diffusion interference, that is, restore the original crack location and depth of the identified wafer.

[0029] Step 7: Based on the original crack location and depth data, match within the preset process factor library to identify one or more related process factors that cause the crack depth and location state; the matching process within the preset process factor library includes the following steps.

[0030] Extract the depth distribution features of the original crack, including crack depth gradient, spatial clustering, and interface correlation; Based on the crack depth distribution characteristics, matching combinations of process parameters are retrieved from a preset process factor library. The process factor library includes at least: wafer manufacturing process data: cutting speed, grinding pressure, polishing fluid composition; packaging process data: bonding temperature, molding material shrinkage rate, hot pressing parameters; and reliability test data: temperature cycle number, mechanical impact strength. A similarity matching algorithm is used to calculate the correlation weight between crack depth distribution characteristics and combinations of various process factors, and one or more key process factors whose weights exceed a preset threshold are selected. Based on key process factors and the depth, location and state of the crack, the crack initiation analysis results are generated. Precisely correlating crack depth data with process factors not only improves the accuracy of crack cause identification but also enhances the interpretability and practicality of the matching results, enabling them to directly guide process optimization.

[0031] Compared with existing technologies, this method not only achieves high-precision localization and imaging of internal cracks in wafers, but also effectively eliminates signal crosstalk and diffusion effects generated when sound waves propagate in multilayer media by introducing a deep learning-based edge correlation model and interference coefficient calculation. This allows for more accurate restoration of the original morphology and depth of cracks, enabling the detection results to be directly correlated with specific process factors with high confidence, providing a more reliable and direct basis for root cause analysis for process improvement.

[0032] At other levels, this embodiment also provides the process of constructing the edge association model, such as... Figure 2 As shown, it includes the following steps: Step 51, Data Collection: Collect the true crack edge location and depth data obtained from historical wafer acoustic scanning detection. The data includes crack edge location and depth data obtained from multiple sets of adjacent detection areas after initial tomographic imaging, as well as true crack edge location and depth data verified and confirmed by destructive physical analysis or high-resolution imaging technology after eliminating mutual interference. Step 52, Framework Construction: A neural network with an encoder and decoder structure is adopted. The encoder part consists of multiple cascaded convolutional layers, which are used to extract spatial features and correlation patterns in the crack depth data of adjacent regions. The decoder part includes fully connected layers and regression output layers, which are used to map the extracted features into mutual interference coefficients. Step 53, Framework Construction: Using the initial crack edge location and depth data of adjacent regions in the historical data as the model input, and the mutual interference coefficient calculated based on the actual crack edge location and depth data as the training label, the mean squared error loss function and adaptive moment estimation algorithm are used to supervise the training of the model until the model converges; Step 54, Framework Construction: Use an independent validation dataset to evaluate the performance of the trained model. After the validation is successful, apply the model to the detection of the current wafer. Input the initial crack edge position and depth of the current adjacent specific region, and the model will output the corresponding mutual interference coefficient. Unlike existing technologies that mainly rely on operator experience for qualitative judgment or use simple physical models for approximate correction, this model learns the complex mutual influence of crack signals between adjacent regions directly from historical data through deep learning, and outputs accurate mutual interference coefficients. This allows for a more accurate and automatic reconstruction of the true original morphology of the crack, improving the accuracy and objectivity of crack location and cause tracing.

[0033] The formula for calculating the mutual interference coefficient of edge cracks in adjacent regions is: ; In the formula, Represents a specific area and adjacent specific areas The mutual interference coefficient between the crack edge locations and depths. and Representing regions and The set of pixels at the crack edge and They belong to and The pixels in the set represent, and Representing pixels and The initial crack depth value. represent and The absolute value of the depth difference between two points. An empirical attenuation constant representing the depth of influence range. Representative point and points The phase difference of the reflected sound wave received at the location, Representative based on point and points The weighting coefficient for spatial distance decreases as the Euclidean distance between two points increases; For two adjacent specific detection areas, the positions and corresponding depth values ​​of all points on the crack edge identified in the initial scan are extracted respectively. The core of calculating this coefficient is to quantify the degree of influence of the crack depth signal of the edge point of one area on the measured depth value of the edge point of another adjacent area. The calculation process mainly considers three key factors: First, depth difference. The smaller the difference in crack depth between two edge points, the greater the possibility of mutual interference due to their similar acoustic characteristics. This factor participates in the calculation in the form of a negative exponential of the depth difference, and the smaller the difference, the greater its contribution. Second, phase consistency. The phase difference between the reflected sound waves received at two points. The closer the phases, the more similar the sound wave propagation paths and reflection characteristics, and the stronger the tendency for mutual interference. This factor is measured by the cosine value of the phase difference. Third, spatial proximity. The closer the actual physical distance between two edge points, the higher the probability of crosstalk and superposition of sound wave signals in space. Therefore, this effect is reflected by a weighting coefficient that decreases with increasing distance. Finally, the mutual influence of all edge points in the two regions calculated based on the above three factors is weighted and summed, and then normalized to obtain a coefficient value between 0 and 1. The closer the coefficient value is to 1, the more serious the mutual interference of the depth signals of the crack edges in the two adjacent regions is, and the stronger the correction is required when retrospectively deducing the original crack position.

[0034] Compared with existing technologies, this formula achieves accurate modeling of the mutual interference effect of acoustic signals under complex propagation paths inside wafers by quantifying the depth difference of crack edges, acoustic wave phase coherence, and spatial proximity. It not only overcomes the limitations of traditional methods that rely solely on single amplitude or depth data for linear correction, but also improves the accuracy and reliability of inverting and reconstructing the true location and depth of original independent cracks from mixed signals by introducing phase information and spatial weights. This lays a crucial data foundation for subsequent accurate attribution of crack causes.

[0035] In summary, based on the traditional tomographic imaging for locating initial cracks, this invention achieves intelligent optimization of detection resources through region division and importance assessment mechanisms, thereby improving the detection accuracy and efficiency for complex crack network regions.

[0036] By introducing a deep learning-based edge correlation model and calculating the mutual interference coefficient between crack edges, it is possible to effectively remove signal crosstalk and artifacts caused by multiple factors such as acoustic diffraction and material boundary reflection, thereby deducing the original crack morphology and depth that are closer to the physical reality, which greatly enhances the fidelity and reliability of the detection results. By intelligently matching the precisely located original crack features with a pre-set process factor library, not only can cracks be accurately detected, but also their causes can be diagnosed. This can effectively link and trace back the specific process steps that lead to the defects, providing direct and reliable data support for optimizing the manufacturing process and improving wafer yield.

[0037] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy, characterized in that, Includes the following steps: Step 1: Using the reflected pulse echo mode of an acoustic scanning microscope, an ultrasonic wave of an initial preset frequency is emitted, and the returned reflected wave signal is received. Based on the characteristics of the reflected wave, the location of the internal crack of the wafer is initially estimated. Step 2: Based on the wafer thickness and the estimated location of the crack, adjust the time gate and ultrasonic frequency, focus on a specific depth plane inside the wafer, and perform tomographic imaging scan to locate the depth of all initial cracks inside the wafer. Step 3: Divide the wafer scanning space involving frequency adjustment into a unified standard size to form multiple standard regions; Step 4: Based on the amplitude, phase, frequency attenuation spectrum of the fused reflected wave and the fluctuation state of the transmitted wave data during the adjustment process, evaluate the adjustment importance of each standard region one by one, and merge or reduce the region edges according to the evaluation results to form a specific region. Step 5: Construct an edge association model using a deep learning algorithm, taking the initial crack edge position and depth of adjacent specific regions as input, and outputting the mutual interference coefficient of edge cracks in adjacent regions; Step 6: Based on the mutual interference coefficient, perform deconvolution calculation on the initial crack depth location data to eliminate crack signal diffusion caused by acoustic wave mutual interference and material boundary effects, and restore and identify the original crack location depth of the wafer. Step 7: Based on the original crack location and depth data, match within the preset process factor library to identify one or more related process factors that cause the crack depth and location state.

2. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The ultrasonic wave with the preset frequency in step 1 is a broadband high-frequency sound wave; the characteristics of the reflected wave include amplitude, phase and time of flight.

3. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The process of adjusting the importance assessment of each region in step 4 includes the following steps: Obtain the fused acoustic dataset for each standard detection region during the acoustic frequency modulation process; Calculate the volatility index of each sequence, which includes at least the variance of the amplitude sequence, the root mean square error of the phase change sequence, the slope change rate of the frequency decay spectrum sequence, and the number of abrupt change points of the transmitted wave energy sequence. Based on preset weighting coefficients, the volatility indicators are weighted and summed to calculate the initial importance score for each standard detection area; Based on the scanning path and sound field distribution characteristics of the acoustic scanning microscope, the initial importance score is spatially correlated to obtain the final modulated importance evaluation value.

4. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The formation process of the specific region in step 4 is as follows: Set a first evaluation threshold and a second evaluation threshold; Standard detection areas with adjustment importance assessment values ​​higher than the first assessment threshold are marked as high-concern core areas; The region whose adjustment importance assessment value is between the first assessment threshold and the second assessment threshold is marked as the transition region to be merged; the region whose adjustment importance assessment value is below the second assessment threshold is marked as the low attention background region. Using the high-attention core area as the seed point, search and merge consecutive transition areas to be merged within the spatial neighborhood to form an initial specific region; The edges of the initial specific region are detected. If the evaluation value of the edge portion of the region is lower than the second evaluation threshold and the continuous area is less than the preset area threshold, the region is reduced and merged into the adjacent specific region. If the evaluation value of the edge portion of the region is continuously lower than the second evaluation threshold and the continuous area exceeds the preset area threshold, the region is separated from the specific region and classified into the low-interest background region. The specific area is composed of the processed high-concern core area and the merged transition area to be merged.

5. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The process of constructing the edge association model in step 5 includes the following steps: Step 51, Data Collection: Collect the actual crack edge location and depth data obtained from historical wafer acoustic scanning detection; Step 52, Framework Construction: A neural network with an encoder and decoder structure is adopted. The encoder part consists of multiple cascaded convolutional layers, which are used to extract spatial features and correlation patterns in the crack depth data of adjacent regions. The decoder part includes fully connected layers and regression output layers, which are used to map the extracted features into mutual interference coefficients. Step 53, Framework Construction: Using the initial crack edge location and depth data of adjacent regions in the historical data as the model input, and the mutual interference coefficient calculated based on the actual crack edge location and depth data as the training label, the mean squared error loss function and adaptive moment estimation algorithm are used to supervise the training of the model until the model converges; Step 54, Framework Construction: Use an independent validation dataset to evaluate the performance of the trained model. After successful validation, apply the model to the detection of the current wafer. Input the initial crack edge position and depth of the current adjacent specific region, and the model will output the corresponding mutual interference coefficient.

6. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The formula for calculating the mutual interference coefficient of edge cracks in adjacent regions in step 5 is as follows: ； In the formula, Represents a specific area and adjacent specific areas The mutual interference coefficient between the crack edge locations and depths. and Representing regions and The set of pixels at the crack edge and They belong to and The pixels in the set represent, and Representing pixels and The initial crack depth value. represent and The absolute value of the depth difference between two points. An empirical attenuation constant representing the depth of influence range. Representative point and points The phase difference of the reflected sound wave received at the location, Representative based on point and points The weighting coefficient for spatial distance decreases as the Euclidean distance between two points increases.

7. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The deconvolution calculation process in step 6 includes the following steps: Based on the calculated mutual interference coefficients between adjacent specific regions, a system point diffusion function matrix is ​​constructed; The initial crack depth location data is used as the fuzzy observation result under the action of the diffusion function matrix at that point; An iterative deconvolution algorithm based on maximum a posteriori probability estimation is adopted. The fuzzy observation result is used as input, and the system point spread function matrix is ​​used as the known degradation model. Through constrained iterative solution, the crosstalk components between adjacent crack signals and the signal tailing effect caused by material interface reflection are gradually stripped and removed, and the intermediate crack depth distribution data after eliminating interference factors is output.

8. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The process of restoring and identifying the location and depth of the original crack in the wafer in step 6 includes the following steps: Step 61: Perform threshold segmentation and edge enhancement processing on the intermediate crack depth distribution data obtained by deconvolution calculation; Step 62: Using the connected component analysis algorithm, based on the enhanced data, identify and separate individual cracks, and calculate the geometric center depth of each crack and its contour distribution in the depth direction. Step 63: Compare this profile distribution with the standard acoustic impedance profile of the wafer material, perform iterative optimization, and correct the slight deviation in depth measurement caused by material inhomogeneity. Step 64: Output the original location coordinates of the crack and its corresponding depth value after eliminating signal diffusion interference.

9. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 8, characterized in that, The working logic of the connected component analysis algorithm in step 62 is as follows: The deblurred crack depth data after threshold segmentation and edge enhancement is converted into a binary image, where pixels above a preset depth threshold are marked as foreground points, representing possible crack areas. The binary image is traversed using a connected component labeling algorithm based on eight or four neighborhoods. A unique identifier is assigned to the set of spatially connected foreground points, and all connected pixel blocks in the image are identified and separated into several independent candidate crack individuals. For each marked individual crack, calculate the set of coordinates and depth values ​​of all its constituent pixels in three-dimensional space; The geometric center depth of the individual crack is obtained by calculating the arithmetic mean of all the depth values ​​that make up the pixel. Its contour distribution in the depth direction is obtained by slicing along the depth axis direction perpendicular to the wafer surface, with a preset depth interval as the step size. The number of pixels or projected area belonging to the crack individual on each depth slice is counted to generate a distribution histogram representing the morphological expansion of the crack in the depth direction.

10. The non-destructive testing method for internal cracks in wafers based on acoustic scanning microscopy according to claim 1, characterized in that, The matching process within the preset process factor library in step 7 includes the following steps; Extract the depth distribution features of the original crack, including crack depth gradient, spatial clustering, and interface correlation; Based on the crack depth distribution characteristics, a matching combination of process parameters is retrieved from a pre-defined process factor library. A similarity matching algorithm is used to calculate the correlation weight between crack depth distribution characteristics and combinations of various process factors, and one or more key process factors whose weights exceed a preset threshold are selected. Based on key process factors and the depth, location, and state of the crack, the crack initiation analysis results are generated.

Images