A method for detecting nanoscale defects in diodes

Abstract

This application provides a method for detecting nanoscale defects in diodes, comprising: acquiring initial three-dimensional scanning data of the internal structure of the diode, the data containing grayscale information and spatial coordinate information at the nanoscale; processing the initial three-dimensional scanning data using a preset noise filtering algorithm to filter out random noise signals introduced by environmental interference, obtaining denoised three-dimensional data; determining possible candidate defect regions and their spatial boundary coordinates based on the grayscale gradient changes of the denoised three-dimensional data in three-dimensional space; extending the influence range model of a single defect to the entire internal structure of the diode using a spatial interpolation algorithm to generate a complete three-dimensional heat map of defect distribution; and outputting the spatial location coordinates of the nanoscale defects and a comprehensive evaluation report based on the defect density and influence weight of each region in the three-dimensional heat map of defect distribution.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing, and more particularly to a method for detecting nanoscale defects in diodes. Background Technology

[0002] In the field of modern electronics, diodes are core components whose performance directly affects the stability and reliability of electronic devices. As device dimensions continue to shrink to the nanometer level, the impact of minute defects on diode function becomes particularly significant. Therefore, nanometer-level defect detection technology has become a crucial link in ensuring product quality and driving technological progress. Research in this field is not only related to the precision of device manufacturing but also plays a vital role in the future development of the entire semiconductor industry.

[0003] However, current mainstream defect detection methods often struggle to balance comprehensiveness and accuracy when dealing with nanoscale defects. Many techniques are susceptible to interference from complex environments during detection, leading to ambiguous judgments of defect locations, especially in three-dimensional space, where there is a lack of overall understanding of defect distribution. This limitation makes it difficult to accurately capture the deep features of defects, thus affecting the targeted nature of subsequent repairs or optimizations.

[0004] Focusing on the technical challenges, the core difficulty in nanoscale defect detection lies in achieving precise defect localization in three-dimensional space. Due to the extremely small internal structure of diodes, defects may exist as points, lines, or voids. Traditional planar scanning methods struggle to cover all spatial dimensions, resulting in incomplete defect location capture in three-dimensional space. As the spatial dimension increases, detection equipment needs to analyze each tiny region layer by layer at extremely small scales. This complexity makes the detection process time-consuming and error-prone. For example, in a tiny region with sides only a few nanometers long, the detection equipment may be unable to accurately distinguish the specific location of the defect due to resolution limitations, thus affecting the judgment of the defect type and its extent.

[0005] Therefore, how to accurately locate nanoscale defects inside a diode in three-dimensional space and comprehensively capture the defect distribution by refining the region division to the nanometer level has become a key problem that needs to be solved in this study. Summary of the Invention

[0006] This invention provides a method for detecting nanoscale defects in diodes, mainly comprising: Acquire initial three-dimensional scanning data of the internal structure of the diode, which includes grayscale information and spatial coordinate information at the nanoscale. The initial 3D scanning data is processed using a preset noise filtering algorithm to filter out random noise signals introduced by environmental interference, resulting in denoised 3D data. Based on the gray-level gradient changes of the denoised 3D data in the three-dimensional space, the possible candidate defect regions and their spatial boundary coordinates are determined. If the volume of the defect candidate region is smaller than the preset nanometer-level volume threshold, the high-resolution focusing scanning module is activated to perform secondary data acquisition on the region to obtain enhanced resolution data. By integrating enhanced resolution data and denoised 3D data, a refined 3D model of the defect area is constructed, which includes geometric features of point, line, and cavity shapes. Based on the geometric features of the refined 3D model and the material property database, the specific type of defect is determined to be a point defect, a line defect, or a volume defect. Based on the identified defect types and their spatial distribution in the refined 3D model, assess the extent of the physical impact of the defects on the conductive and insulating regions of the diode. A spatial interpolation algorithm is used to extend the influence range model of a single defect to the entire internal structure of the diode, generating a complete three-dimensional heat map of the defect distribution. Based on the defect density and influence weight of each region in the three-dimensional thermal map of defect distribution, the spatial location coordinates of nanoscale defects and a comprehensive evaluation report are output.

[0007] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a nanoscale detection method for diodes. This method addresses the problem in business scenarios where minute defects in the three-dimensional scanning data of the internal structure of nanoscale diodes are difficult to accurately characterize in the initial denoised data due to their volume being below a threshold. It solves this problem by integrating high-resolution focused scanning and data fusion techniques. After determining candidate regions based on grayscale gradients, this invention initiates enhanced resolution acquisition for subthreshold volume defects and fuses this data with denoised data to construct a refined three-dimensional model that clearly characterizes points, lines, voids, and other morphologies. This enables accurate identification of nanoscale defect types and assessment of their impact on conductive and insulating regions. Furthermore, spatial interpolation generates a complete three-dimensional heat map of defect distribution. The final technical effect is to achieve high-precision three-dimensional visualization analysis of the entire chain of nanoscale defects inside diodes, from detection and identification to impact assessment. Attached Figure Description

[0008] Figure 1 This is a flowchart of a diode nanoscale defect detection method according to the present invention.

[0009] Figure 2 This is a schematic diagram of a diode nanoscale defect detection method according to the present invention.

[0010] Figure 3 This is another schematic diagram of a diode nanoscale defect detection method according to the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0012] like Figures 1-3 This embodiment of a diode nanoscale defect detection method specifically includes: S101. Obtain initial three-dimensional scanning data of the internal structure of the diode, which includes grayscale information and spatial coordinate information at the nanoscale.

[0013] A three-dimensional scanning result of the diode structure is obtained. Initial data is acquired using a high-resolution imaging device, covering nanoscale grayscale information and spatial coordinates to obtain a complete scanning dataset. For the scanning dataset, a preprocessing tool is used to standardize the grayscale information. If the grayscale values ​​exceed a preset threshold range, they are smoothed to determine the processed grayscale dataset. Based on the processed grayscale dataset and spatial coordinate information, a three-dimensional model of the diode's internal structure is constructed. A volumetric data reconstruction method is used to obtain a preliminary three-dimensional structural representation. For the preliminary three-dimensional structural representation, the geometric distribution of the internal structure is analyzed. If missing data dimensions are detected, interpolation methods are used to fill in the missing parts, thus determining the complete structural model.

[0014]

[0015] The specific formula is as follows: This represents the grayscale value after the missing area is filled in. The spatial coordinates of the missing location, This represents the grayscale value of nearby known points. Represents the spatial coordinates of nearby known points. Indicates the number of neighboring points. This represents distance and is used for inverse distance weighted interpolation to fill in missing data.

[0016] By using a complete structural model, key structural features at the nanoscale are extracted. Considering scanning accuracy requirements, the location distribution and grayscale variation trends of the feature points are determined. Based on the location distribution and grayscale variation trends of the feature points, a support vector machine algorithm is used to classify the structural analysis results, obtaining refined hierarchical information about the diode's internal structure.

[0017] This represents the output of the classification decision function. A vector representing the location distribution and grayscale variation trend of feature points. Represents the support vector weights. Indicates the labels of the training samples. Represents the kernel function. Represents support vectors, Indicates the number of support vectors. The bias is used by the support vector machine to classify the structural analysis results to obtain hierarchical information.

[0018] To obtain initial three-dimensional scanning data of the diode's internal structure, including nanoscale grayscale and spatial coordinate information, the following complete process can be achieved. First, the diode sample is scanned using a high-resolution transmission electron microscope (TEM). An accelerating voltage of 200 kV is set, achieving a resolution of 0.1 nm to capture nanoscale structural details. The scanning area is set to a three-dimensional space of 100 nm × 100 nm × 50 nm, generating an initial data matrix containing 10^6 voxels. The grayscale value of each voxel ranges from 0 to 255, reflecting differences in material density. Next, a three-dimensional reconstruction algorithm, such as back projection, is used to reconstruct a three-dimensional model from the multi-angle two-dimensional projection data. Specifically, a weighted average of the grayscale values ​​of each voxel is calculated, with the weights based on the sine of the projection angle. The error is controlled within ±0.5 grayscale units to ensure reconstruction accuracy. Then, using spatial coordinate mapping technology, each voxel is bound to a three-dimensional spatial coordinate system with a coordinate accuracy of 0.01 nm. Geometric distortions during the scanning process are corrected using matrix transformation algorithms (such as Euclidean transformation), resulting in a coordinate error of less than 0.02 nm after correction. Finally, noise filtering is applied to the grayscale data using a Gaussian filtering algorithm with a kernel size of 3×3×3 and a standard deviation of 1.0, effectively reducing the impact of random noise. The standard deviation of the grayscale values ​​after filtering is reduced from the initial 5.2 to 2.1, improving data quality. This entire process is integrated into an automated system, with the data processing module seamlessly integrated with the scanning equipment. This ensures that the entire process from scanning to final data output is completed without human intervention. The generated grayscale and coordinate information can be directly used for subsequent structural analysis, such as defect detection or material property simulation, forming a complete technical closed loop.

[0019] S102. The initial three-dimensional scanning data is processed using a preset noise filtering algorithm to filter out random noise signals introduced by environmental interference and obtain denoised three-dimensional data.

[0020] The initial 3D scan data is processed using a pre-defined noise filtering method to eliminate interference signals introduced by the external environment, resulting in a denoised 3D dataset. Based on this denoised dataset, a layered extraction tool is used to decompose the data layer by layer, separating structural information at different depths to determine the layered data set. For each layered data set, a spatial mapping method is used to calibrate the position of the data in each layer, addressing the offset caused by the scanning angle, and obtaining a calibrated data set. If there are missing data areas in the calibrated data set, a neighborhood comparison tool is used to fill in the missing data, thus determining the complete data structure.

[0021]

[0022] This represents the data points after the missing regions have been filled in. The first neighborhood comparison represents the... Neighboring points, Indicates missing points The neighborhood set, Indicates missing points Location, Representing neighborhood points The position is α, which represents the distance weighting index; this formula supplements missing data in local areas and determines the complete data structure by using an inverse weighted average of the distances to neighboring points.

[0023] Based on the complete data structure, cluster analysis is used to group the structural information. The grouped data are then categorized by feature to obtain classified structural units. Using these classified structural units and spatial location information, boundaries are defined for each unit to determine the spatial distribution relationships between them. For these spatial distribution relationships, a preset threshold range is used to filter the connection strength between units. If the connection strength is below the threshold, the boundaries of the relevant units are adjusted to obtain the final structural partitioning result.

[0024] When processing 3D scanning data of the internal structure of a diode, random noise signals introduced by environmental interference in the initial data can be processed using a preset noise filtering algorithm to obtain denoised 3D data. Specifically, the initial 3D scanning data is first analyzed in the frequency domain. The Fast Fourier Transform (FFT) algorithm is used to convert the data from the spatial domain to the frequency domain, with a sampling frequency of 100 data points per nanometer. The analysis results show that high-frequency noise is mainly concentrated in the part exceeding 50Hz, accounting for approximately 12.3% of the original signal. Subsequently, a low-pass filter is designed with a cutoff frequency of 40Hz and a filter order of 5. The Butterworth filtering algorithm is used to process the frequency domain data, filtering out high-frequency noise components. During the calculation, the amplitude of each frequency domain point is weighted and adjusted. The weighting coefficient is based on the ratio of the frequency to the cutoff frequency, with the error controlled within ±0.01 amplitude units. Next, the processed frequency domain data is transformed back to the spatial domain using Inverse Fourier Transform (IFFT) to generate a preliminary denoised 3D data matrix with dimensions of 200×200×100, containing voxel data with gray values ​​ranging from 0 to 200. To further optimize the denoising results, an adaptive median filtering algorithm is introduced. Local statistical analysis is performed on the neighborhood (5×5×5) of each voxel, calculating the median and variance of the gray values ​​within the neighborhood. If the variance exceeds a threshold of 8.5, the gray value of the central voxel is replaced with the median; otherwise, the original value is retained. This process is automated and batch-processed to ensure data consistency. Finally, the denoised data is compared with the original data, calculating the variation in gray values ​​for each voxel. The average variation is 3.2, and the standard deviation is 1.8, indicating that the noise signal is effectively suppressed while retaining key structural information. The entire process is automatically executed by data processing software and interfaces with the subsequent material stress analysis module to form a complete data processing chain.

[0025] S103. Based on the gray-level gradient changes of the denoised 3D data in the three-dimensional space, determine the possible candidate defect regions and their spatial boundary coordinates.

[0026] By leveraging the distribution characteristics of denoised data in three-dimensional space, a pre-defined analysis tool is used to scan the grayscale gradient layer by layer, obtaining preliminary results of the gradient distribution. Based on these preliminary results, clustering analysis is used to group candidate defect regions for areas with significant variations, determining the clustering range of potential defects. Using the clustering range of the candidate defects, a spatial mapping tool is employed to refine the region identification, acquiring the spatial boundary information of each region. Based on this spatial boundary information, and considering the accuracy requirements for boundary positioning, data analysis methods are used to correct the coordinate acquisition, determining the accurate boundary coordinates. If the matching degree between the boundary coordinates and the gradient distribution is lower than a pre-defined threshold, a neighborhood comparison tool is used to locally adjust the region division, obtaining an optimized boundary range. Based on the optimized boundary range, and considering the completeness of the candidate defect regions, a data imputation method is used to fill in the missing parts, determining the final defect region distribution.

[0027] Based on the denoised 3D data, the 3D Sobel operator is first used to calculate the gray-level gradient of each voxel in the X, Y, and Z directions. The operator template size is set to 3×3×3. The gradient components Gx, Gy, and Gz are obtained through convolution operations. Then, the total gradient magnitude of each voxel is calculated using the formula M=√(Gx²+Gy²+Gz²). The calculated gradient magnitude matrix has the same dimensions as the original data. Subsequently, a region growing algorithm is applied to the gradient magnitude matrix to extract candidate defect regions. The initial seed point selection rule is set to voxels with gradient magnitudes greater than a threshold of 25. The growth criterion is that the gradient magnitude difference of neighboring voxels (26 neighbors) is less than 5 and the original gray-level value difference is within 15. The algorithm iterates until no new voxels meet the conditions, thereby generating a set of connected candidate regions. Next, the spatial boundary coordinates of each connected region are calibrated using an axially aligned bounding box method. All voxel coordinates within the region are traversed, and their minimum and maximum index values ​​on the X, Y, and Z axes are recorded. For example, the boundary coordinates of a region can be represented as Xmin=45, Xmax=78, Ymin=102, Ymax=135, Zmin=30, Zmax=55, thus defining its spatial range. To further filter high-probability defect regions, the statistical characteristics of the internal gradient magnitude are calculated for each candidate region, including the average gradient magnitude, gradient magnitude variance, and region volume. The filtering criteria are set as an average gradient magnitude greater than 18 and a volume between 100 and 5000 voxels. This condition automatically filters out normal tissue regions with gentle gradient changes and excessively small noise residue regions. The final output list of candidate defect regions and their spatial boundary coordinates can be directly imported into the subsequent defect classification and quantitative analysis module for automated processing.

[0028] S104. If the volume of the defect candidate region is smaller than the preset nanometer-level volume threshold, the high-resolution focusing scanning module is activated to perform secondary data acquisition on the region to obtain enhanced resolution data.

[0029] The volume value of the candidate region is compared with a preset threshold to determine if the volume value is less than the preset threshold. If the volume value is less than the preset threshold, a start command is sent to the scanning module. Based on the start command, the scanning module performs a secondary sampling of the candidate region using a focused scanning mode. This secondary sampling process acquires augmented data, which has higher resolution. A 3D reconstruction algorithm is then used to process the augmented data, generating a detailed 3D model of the candidate region.

[0030]

[0031] A precise geometric profile representing the defect. Represents points on the surface of a detailed 3D model. Representing a detailed 3D model, Indicates the change in surface curvature. This represents the threshold for curvature change.

[0032] The precise geometric profile of the defect is determined based on the surface curvature variations of the detailed 3D model.

[0033] In 3D data analysis, when the volume of a detected defect candidate region is smaller than a preset nanometer-level volume threshold, such as less than 50 cubic nanometers, the system automatically triggers a high-resolution focusing scanning module to perform secondary data acquisition on that region to obtain more refined, enhanced-resolution data. Specifically, the system first determines the number of voxels in the candidate region using a volume calculation algorithm. Assuming a region contains 30 voxels, each with a side length of 1 nanometer, the volume is 30 cubic nanometers, which does not meet the threshold requirement. Next, the system automatically transmits the spatial coordinate information of the region, such as the center point coordinates (X=60, Y=120, Z=40), to the high-resolution scanning control unit. The scanning resolution is adjusted from the original 1 nanometer / voxel to 0.5 nanometer / voxel, and the scanning range is set to a 10 nanometer × 10 nanometer × 10 nanometer cube region based on the center point, ensuring full coverage of the candidate region. During the scanning process, the system uses an improved 3D Gaussian filtering algorithm to preprocess the acquired data. The filter kernel size is set to 5 × 5 × 5, and the Gaussian standard deviation is 1.2, to reduce the small noise introduced by the high-resolution scanning. Subsequently, the system performs voxel-level grayscale analysis on the enhanced resolution data, calculating the local grayscale mean and standard deviation for each voxel. A grayscale mean threshold of 80 and a standard deviation threshold of 10 are set to filter out sub-regions with abnormal grayscale distributions. For example, a small cluster of 5 voxels with a grayscale mean of 85 and a standard deviation of 12 is found within the scanned area and marked as a potential defect core point. Finally, the system registers and fuses the enhanced data acquired a second time with the original data, employing a least-squares-based spatial transformation algorithm with errors controlled within 0.1 nanometers, forming a unified, fine-grained dataset. This dataset is automatically imported into the subsequent defect feature extraction module, providing data support for further analysis. Through this process, the system can accurately locate small-volume defect regions and improve data quality through algorithmic optimization, forming a complete technical processing chain.

[0034] S105. By fusing enhanced resolution data with denoised 3D data, a refined 3D model of the defect area is constructed. This model includes geometric features of point, line, and cavity shapes.

[0035] By performing layered processing on the collected raw data, enhanced resolution data and denoised 3D data are separated to obtain a pre-processed data set. Based on the pre-processed data set, a preset segmentation method is used to locate the defect region and determine its boundary range. For the located defect region, high-precision information from the enhanced resolution data is obtained to acquire detailed feature data of the defect region. By overlaying the detailed feature data and the denoised 3D data, a geometric information dataset containing point-like, line-like, and cavity-like morphologies is constructed. If the amount of point-like morphology data in the geometric information dataset is lower than a preset threshold, this part of the data is supplemented to obtain a more complete description of the point-like morphology. Based on the supplemented geometric information dataset, a 3D reconstruction algorithm is used to construct a refined model of the defect region, resulting in a refined 3D model containing various morphologies. For the cavity morphology data in the refined 3D model, if incomplete boundaries of the cavity morphology are detected, interpolation methods are used to repair the boundaries and determine the final model structure.

[0036] In constructing a refined 3D model of the defect area, the system first fuses the enhanced resolution data and the denoised 3D data using a weighted average voxel value integration algorithm. The weight of the enhanced resolution data is set to 0.7, and the weight of the denoised 3D data is set to 0.3, ensuring the dominance of high-precision information while preserving the smoothness of the denoised data. It is assumed that the grayscale value of a voxel after fusion is adjusted from the original 90 to the calculated 87.5. Subsequently, the system uses a 3D meshing algorithm to perform geometric modeling on the fused data. The mesh unit size is set to 0.2 nm × 0.2 nm × 0.2 nm. By calculating the voxel density of each mesh unit, regions with a density greater than 0.8 are selected as candidate points for defect geometric features. For example, a region with a mesh density of 0.85 is marked as a potential point defect. Next, the system applies a gradient descent-based boundary tracking algorithm to analyze the voxel gradient changes around candidate points. A gradient threshold of 15 is set to track the continuous path of linear defects. Assuming a linear defect path spans 8 grid cells with a length of 1.6 nanometers, its direction vector is recorded as (0.5, 0.3, 0.2). For cavity morphology identification, the system uses a cavity filling algorithm to calculate the volume of low-density regions in the fused data. A volume threshold of 5 cubic nanometers is set. If a cavity region has a volume of 6.2 cubic nanometers, it is marked as a cavity defect, and its center coordinates are recorded as (X=75.4, Y=132.6, Z=45.8). Finally, the system integrates the geometric features of point, linear, and cavity morphologies into a unified, refined 3D model. Triangular facets are used to smooth the model surface, with three smoothing iterations to ensure the model boundary error is less than 0.05 nanometers. The generated model data is automatically stored in a standard format for subsequent analysis modules. Through this series of algorithms, the system can extract multi-form defect features from the fused data, forming a logically rigorous processing flow.

[0037] S106. Based on the geometric features of the refined 3D model and the material property database, determine the specific type of defect, whether it is a point defect, line defect, or volume defect.

[0038] Geometric morphological data is extracted from a refined 3D model, and automated tools are used to classify and organize the morphological data to obtain preliminary morphological classification results. These preliminary morphological classification results are then matched with a pre-established material property database to obtain material property information corresponding to the geometric morphological data, thus determining the preliminary range of defect types. If the preliminary range of defect types includes multiple possibilities, further extraction of local morphological data from the 3D model is performed, combined with material property information, for a secondary comparison to determine the specific defect type. Based on the defect type results after the secondary comparison, detailed classification data for point defects, linear defects, or volumetric defects are obtained to determine the final defect type label. For the final defect type label, a support vector machine algorithm is used to verify the morphological data in the 3D model, obtaining verified defect type confirmation results; the specific algorithm is as follows:

[0039] This indicates the confirmation result of the defect type after verification. Indicates the number of support vectors. Indicates the first The weights of the support vectors, Indicates the first Training sample labels, Represents the kernel function. Indicates the first Support vectors, Represents the morphological data of a three-dimensional model. This indicates the bias term; after verifying the defect type, the corresponding defect classification record is generated, and the association mapping between the defect type and the material property is determined.

[0040] The system matches the geometric features of a refined 3D model with a material property database to determine whether a defect is a point defect, line defect, or volume defect. The specific implementation method utilizes information technology to automatically process and form a rigorous logic. First, the system extracts geometric feature parameters from the refined 3D model. For example, a point feature has a volume of 0.3 cubic nanometers and coordinates (X=42.1, Y=88.7, Z=19.3), and its surface irregularity is recorded as 0.12. For line features, the system calculates its length as 2.4 nanometers and its radius of curvature as 0.5 nanometers. For volume features, the system measures its volume as 8.9 cubic nanometers and its shape factor as 0.75. Then, the system calls the material property database and loads the corresponding standard values ​​for defect features. For example, the volume threshold for point defects is set to below 0.4 cubic nanometers, the length threshold for line defects is set to above 1.0 nanometers and the radius of curvature to be less than 0.8 nanometers, and the volume threshold for volume defects is set to above 5.0 cubic nanometers. Next, the system employs a feature matching algorithm based on Euclidean distance to calculate the distance between the extracted geometric features and the standard values ​​in the database. For example, a point feature with a volume of 0.3 cubic nanometers is compared to the standard threshold of 0.4 cubic nanometers, with a distance value of 0.1, and is therefore classified as a point defect. A linear feature with a length of 2.4 nanometers and a radius of curvature of 0.5 nanometers both meet the standard, resulting in a combined distance value of 0.2, and is therefore classified as a line defect. A volumetric feature with a volume of 8.9 cubic nanometers has a distance value of 3.9, and is therefore classified as a volume defect. Finally, the system correlates the matching results with defect influence factors in the material property database. For example, the influence factor for point defects is 0.05, for line defects it is 0.15, and for volume defects it is 0.25. This generates a defect classification report and automatically stores it as structured data for subsequent use by the material performance analysis module, ensuring a rigorous logical process from geometric feature extraction to type determination.

[0041] S107. Based on the identified defect type and its spatial distribution in the refined 3D model, assess the extent of the physical impact of the defect on the conductive and insulating regions of the diode.

[0042] Defect types and their corresponding spatial locations are obtained from a refined 3D model. Using pre-established region division rules, the boundary information of conductive and insulating regions is extracted to obtain initial region division results. Based on these initial results, and combined with defect distribution and location data, automated tools are used to perform preliminary mapping of the spatial distribution of defects in conductive and insulating regions, determining the distribution of contact points between defects and each region. According to the contact point distribution obtained from the preliminary mapping, if the defect distribution is concentrated in the conductive region, a geometric distance calculation tool is used to analyze the physical range of the defect's effect on the conductive region, obtaining the specific regional action boundary. If the defect distribution is concentrated in the insulating region, the same geometric distance calculation tool is used to analyze the physical range of the defect's effect on the insulating region, determining the specific regional action boundary. For cases where the defect distribution involves both conductive and insulating regions, contact point distribution data for both regions is obtained. Combined with defect type information, a support vector machine algorithm is used to classify and predict the physical action range, determining the boundary division for cross-regional effects. The boundary division results obtained from the above classification and prediction are then used in conjunction with model analysis tools to verify the matching degree between the defect distribution and the region division, obtaining the final physical action range data.

[0043] The system automatically assesses the physical impact of identified defect types and their spatial distribution in a refined 3D model on the conductive and insulating regions of a diode using information technology, constructing a rigorous logical workflow. First, the system extracts the spatial location data of defects from the 3D model. For example, a point defect is located at coordinates (X=15.2, Y=33.4, Z=7.8), a line defect has a starting point coordinate of (X=22.5, Y=41.6, Z=9.1) and an ending point coordinate of (X=25.0, Y=43.2, Z=9.5), and a volume defect has a center point coordinate of (X=30.1, Y=50.3, Z=12.4). Then, the system calls the diode region partitioning database to obtain the boundaries of the conductive region (e.g., Z-axis range of 0 to 5.0 nm) and the insulating region (Z-axis range of 5.0 to 15.0 nm), and calculates the spatial distance between the defect and the region boundaries. For example, the distance from a point defect to the boundary of the conductive region is 2.8 nm, and the distance to the boundary of the insulating region is 2.2 nm. Next, the system employs an algorithm based on spherical influence range, setting defect influence radius parameters, such as 1.5 nm for point defects, 3.0 nm for line defects, and 5.0 nm for volume defects. Combining distance data with influence range overlap analysis, the system calculates the overlap rate of point defects' influence on the insulating region to be 0.73 and on the conductive region to be 0.27; the overlap rate of line defects' influence on the conductive region to be 0.85; and the overlap rate of volume defects' influence on the insulating region to be 0.92. Then, the system combines the influence range data with the influence weights of defect types, such as 0.1 for point defects, 0.3 for line defects, and 0.5 for volume defects, to comprehensively calculate the physical influence index. The resulting indexes are 0.073 for point defects on the insulating region and 0.027 for line defects on the conductive region, 0.255 for line defects on the conductive region, and 0.46 for volume defects on the insulating region. Finally, the system automatically generates an influence range analysis report, storing it as structured data for subsequent performance simulation modules, ensuring the logical rigor of the entire process from spatial distribution extraction to influence assessment.

[0044] S108. Using a spatial interpolation algorithm, the influence range model of a single defect is extended to the entire internal structure of the diode to generate a complete three-dimensional heat map of the defect distribution.

[0045] We acquire voxel mesh data of the diode's 3D structure and the influence intensity model of individual defects. Based on the influence model, we locate the spatial positions of all identified defect points in the voxel mesh. For each defect point, we calculate its initial influence value on the surrounding voxel mesh using its influence intensity model. We then use Kriging interpolation to fuse the initial influence values ​​of all defect points and calculate the interpolated influence intensity of the entire voxel mesh.

[0046]

[0047] Indicates position Kriging estimates at [location] Indicates the first The weight coefficients of the known points Indicates the first Observations at known points This represents the total number of known points.

[0048] If the interpolation effect intensity exceeds a preset intensity threshold, the voxel is marked as a high-influence region. Based on the marking results of all voxels and the interpolation effect intensity values, three-dimensional heatmap data is generated.

[0049] The system automatically implements a spatial interpolation algorithm using information technology to extend the influence range model of a single defect to the entire internal structure of the diode, generating a complete three-dimensional heat map of the defect distribution and forming a rigorous logical process. First, the system extracts core data for a single defect from the database. For example, the center coordinates of a defect at a certain point are (X=10.5, Y=20.7, Z=8.3), and its initial influence radius is 2.0 nanometers. Then, the system calls the Kriging Interpolation algorithm, using the defect center as a reference point and combining the radius data, to calculate the influence intensity of surrounding grid points. For example, at a grid point 1.0 nanometers from the center (X=11.5, Y=20.7, Z=8.3), the algorithm predicts its influence value to be 0.75 of the original intensity. At a grid point 2.0 nanometers away, the influence value decays to 0.25, forming a continuous influence gradient distribution. Next, the system extends the interpolation results to the entire grid of the diode's internal structure. For example, the entire model has dimensions of 0 to 50.0 nm on the X-axis, 0 to 60.0 nm on the Y-axis, and 0 to 20.0 nm on the Z-axis, with a grid resolution of 0.5 nm. Influence values ​​are calculated and filled point-by-point to ensure coverage of all areas. Then, the system performs overlay analysis on multi-defect scenarios. For example, at the same grid point, a point defect has an influence value of 0.3, while a line defect has an influence value of 0.4. Using a weighted average algorithm, the weight of the point defect is set to 0.2, and the weight of the line defect is set to 0.5, ultimately calculating a combined influence value of 0.35 for that point. Finally, based on the interpolation and overlay data, the system generates a 3D heatmap, using color gradients to represent influence intensity. For example, influence values ​​of 0.1 to 0.3 are mapped to blue, 0.3 to 0.6 to yellow, and 0.6 to 1.0 to red. The heatmap data is stored in a standard format for subsequent structural analysis modules to use, ensuring the logical rigor of extending from a single defect to a full-domain distribution.

[0050] S109. Based on the defect density and influence weight of each region in the three-dimensional thermal map of defect distribution, output the spatial location coordinates of nanoscale defects and a comprehensive evaluation report.

[0051] A three-dimensional heatmap of defect distribution is acquired. For each region, based on the recorded defect density and influence weight information, a weighted influence value is obtained using a pre-defined weighting calculation rule. Based on the weighted influence value, a region division method is used to classify the regions in the three-dimensional heatmap according to their influence level, determining the distribution of high-influence and low-influence regions. For the defect distribution data within high-influence regions, spatial positioning technology is used to extract the specific location information of nanoscale defects, obtaining the corresponding spatial coordinate set. Based on the spatial coordinate set, combined with the distribution analysis results of defect density and influence weight, if the weighted influence value corresponding to a certain coordinate point exceeds a pre-defined threshold, that point is marked as a critical defect point. For the marked critical defect points, their neighboring region data in the three-dimensional heatmap is acquired. A comprehensive evaluation method is used to calculate the associated influence range of each critical defect point, determining its intensity of influence on the surrounding area. Based on the intensity of influence and the spatial coordinate set, data generation technology is used to integrate the distribution of critical defect points and their associated influence range into structured data, obtaining the final evaluation dataset.

[0052] The system automatically processes 3D heatmap data of defect distribution using information technology, spatially locating and generating comprehensive assessment reports for nanoscale defects, forming a rigorous logical process. First, the system extracts defect density data for each region from the constructed 3D heatmap. For example, within a certain region (X=15.0 to 20.0 nm, Y=25.0 to 30.0 nm, Z=5.0 to 10.0 nm), the defect density is 0.12 defects per cubic nanometer, and the influence weight of this region is 0.45. Next, the system uses a density peak clustering algorithm to analyze high-density areas in the heatmap and locate the coordinates of key defects. For instance, the algorithm calculates the center location of a significant defect as (X=17.3, Y=27.8, Z=7.2), with an average density of 0.18 within a 1.5 nm radius around it, exceeding the regional average by more than 50%, thus marking it as a high-risk point. Subsequently, the system performs multi-dimensional feature analysis on the located defects, combining defect density and impact weight to calculate a comprehensive risk index. For example, the risk index for this defect point is 0.67 (calculated by multiplying the density value of 0.18 by the weight of 0.45 and then adding a weighted adjustment coefficient of 0.2). The results are then compared with defect points in other areas to generate a ranking list; for example, this point ranks 3rd in risk among all defects in the entire region. Next, based on all located defect data, the system automatically generates a comprehensive assessment report, including statistics on the spatial distribution of defects. For example, a total of 12 high-risk defect points are identified, with 58% located within the 15.0 to 25.0 nanometer range on the X-axis. The system also analyzes potential impact areas, such as predicting that a certain defect point may affect the structural stability within a 2.2-nanometer radius. Finally, the system integrates the location coordinates and assessment report into a standard data format, such as storing the coordinates (X=17.3, Y=27.8, Z=7.2) and the risk index of 0.67 in XML format for subsequent modules to access, ensuring the logical integrity from data extraction to report generation.

[0053] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting nanoscale defects in diodes, characterized in that, The method includes: Acquire initial three-dimensional scanning data of the internal structure of the diode, which includes grayscale information and spatial coordinate information at the nanoscale. The initial 3D scanning data is processed using a preset noise filtering algorithm to filter out random noise signals introduced by environmental interference, resulting in denoised 3D data. Based on the gray-level gradient changes of the denoised 3D data in the three-dimensional space, the possible candidate defect regions and their spatial boundary coordinates are determined. If the volume of the defect candidate region is smaller than the preset nanometer-level volume threshold, the high-resolution focusing scanning module is activated to perform secondary data acquisition on the region to obtain enhanced resolution data. By integrating enhanced resolution data and denoised 3D data, a refined 3D model of the defect area is constructed, which includes geometric features of point, line, and cavity shapes. Based on the geometric features of the refined 3D model and the material property database, the specific type of defect is determined to be a point defect, a line defect, or a volume defect. Based on the identified defect types and their spatial distribution in the refined 3D model, assess the extent of the physical impact of the defects on the conductive and insulating regions of the diode. A spatial interpolation algorithm is used to extend the influence range model of a single defect to the entire internal structure of the diode, generating a complete three-dimensional heat map of the defect distribution. Based on the defect density and influence weight of each region in the three-dimensional thermal map of defect distribution, the spatial location coordinates of nanoscale defects and a comprehensive evaluation report are output.

2. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The initial three-dimensional scanning data of the diode's internal structure is obtained, which includes nanoscale grayscale information and spatial coordinate information, including: To obtain the three-dimensional scanning results of the diode structure, initial data is acquired using a high-resolution imaging device, covering nanoscale grayscale information and spatial coordinates, to obtain a complete scanning dataset; For the scanned dataset, a preprocessing tool is used to standardize the grayscale information. If the grayscale value exceeds the preset threshold range, it is smoothed to determine the processed grayscale dataset. Based on the processed grayscale dataset and combined with spatial coordinate information, a three-dimensional model of the internal structure of the diode is constructed. A volume data reconstruction method is used to obtain a preliminary three-dimensional structural representation. Based on the preliminary three-dimensional structural representation, the geometric distribution of the internal structure is analyzed. If missing regions of data dimensions are detected, the missing parts are filled in using interpolation methods to determine the complete structural model. ； The specific formula is as follows: This represents the grayscale value after the missing area is filled in. The spatial coordinates of the missing location, This represents the grayscale value of nearby known points. Represents the spatial coordinates of nearby known points. Indicates the number of neighboring points. Represents distance, used for inverse distance-weighted interpolation to fill in missing data; By using a complete structural model, key structural features at the nanoscale are extracted, and the location distribution and grayscale variation trend of feature points are determined in combination with scanning accuracy requirements. Based on the location distribution and grayscale variation trend of feature points, the support vector machine algorithm is used to classify the structural analysis results to obtain refined hierarchical information of the internal structure of the diode. ； This represents the output of the classification decision function. A vector representing the location distribution and grayscale variation trend of feature points. Represents the support vector weights. Indicates the labels of the training samples. Represents the kernel function. Represents support vectors, Indicates the number of support vectors. The bias is used by the support vector machine to classify the structural analysis results to obtain hierarchical information.

3. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The initial 3D scan data is processed using a preset noise filtering algorithm to filter out random noise signals introduced by environmental interference, resulting in denoised 3D data, including: The initial data of the 3D scan is processed by a preset noise filtering method to eliminate interference signals introduced by the external environment, resulting in a denoised 3D dataset. Based on the denoised 3D dataset, a hierarchical extraction tool is used to decompose the data layer by layer, separate structural information at different depths, and determine the hierarchical dataset. For the layered dataset, the spatial mapping method is used to calibrate the position of the data in each layer, handle the offset problem caused by the scanning angle, and obtain the calibrated data set; If there are missing data in local areas of the calibrated data set, the missing parts are supplemented by a neighborhood comparison tool to determine the complete data structure. ； This represents the data points after the missing regions have been filled in. The first neighborhood comparison represents the... Neighboring points, Indicates missing points The neighborhood set, Indicates missing points Location, Representing neighborhood points The position is where α represents the distance weighting index; this formula supplements missing data in local areas and determines the complete data structure by using a weighted average of inverse distances to neighboring points. Based on the complete data structure, cluster analysis is used to group the structural information, and feature classification is performed on the grouped data to obtain the classified structural units. By classifying the structural units and combining them with spatial location information, the boundaries of each unit are divided to determine the spatial distribution relationship between the units. Based on the spatial distribution relationship, the connection strength between units is filtered using a preset threshold range. If the connection strength is lower than the threshold, the boundary of the relevant units is adjusted to obtain the final structural division result.

4. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The step of determining potential defect candidate regions and their spatial boundary coordinates based on the gray-level gradient changes of denoised 3D data in three-dimensional space includes: By analyzing the distribution characteristics of the denoised data in three-dimensional space, a preset analysis tool is used to scan the gray-level gradient layer by layer to obtain preliminary results of the gradient distribution. Based on the preliminary results of the gradient distribution, cluster analysis was used to group the candidate defect regions for areas with significant changes, thereby determining the clustering range of potential defects. By analyzing the clustering range of defect candidates, spatial mapping tools are used to refine the region identification and obtain the spatial boundary information of each region. Based on the spatial boundary information and in accordance with the accuracy requirements of boundary positioning, data analysis methods are used to correct the coordinate acquisition and determine the accurate boundary coordinates. If the matching degree between the boundary coordinates and the gradient distribution is lower than the preset threshold, the region division is locally adjusted using a neighborhood comparison tool to obtain the optimized boundary range. Based on the optimized boundary range, and considering the integrity of the defect candidate region, a data filling method is used to supplement the missing parts, thereby determining the final defect region distribution.

5. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, If the volume of the defect candidate region is smaller than a preset nanometer-level volume threshold, the high-resolution focusing scanning module is activated to perform secondary data acquisition on the region to obtain enhanced resolution data, including: The volume value of the candidate region is compared with a preset threshold to determine whether the volume value is less than the preset threshold. If the volume value is less than the preset threshold, a start command is sent to the scanning module; According to the start command, the scanning module uses a focused scanning mode to perform a second sampling of the candidate area; The enhanced data is obtained through a secondary sampling process, and the enhanced data has higher resolution. A 3D reconstruction algorithm is used to process the enhancement data and generate a detailed 3D model of the candidate region. ； A precise geometric profile representing the defect. Represents points on the surface of a detailed 3D model. Representing a detailed 3D model, Indicates the change in surface curvature. Indicates the threshold of curvature change; The precise geometric profile of the defect is determined based on the surface curvature variations of the detailed 3D model.

6. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The fusion of enhanced resolution data and denoised 3D data constructs a refined 3D model of the defect region. This model includes geometric features of point, line, and cavity morphology, including: By performing layered processing on the collected raw data, enhanced resolution data and denoised 3D data are separated to obtain a pre-processed data set. Based on the pre-processed data set, the defect area is located using a preset segmentation method to determine the boundary range of the defect area; For the located defect area, high-precision information is obtained from the enhanced resolution data to obtain the detailed feature data of the defect area; By overlaying detailed feature data and denoised 3D data, a geometric information dataset containing point-like, line-like, and cavity-like shapes is constructed. If the amount of point-like morphological data in the geometric information dataset is lower than a preset threshold, then the data in that part will be supplemented by sampling to obtain a more complete description of the point-like morphology. Based on the geometric information dataset after supplementary sampling, a three-dimensional reconstruction algorithm is used to construct a fine model of the defect area, resulting in a fine three-dimensional model containing various shapes. For the cavity morphology data in the detailed 3D model, if the boundary of the cavity morphology is found to be incomplete, the boundary is repaired by interpolation method to determine the final model structure.

7. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The process involves matching the geometric features of a refined 3D model with a material property database to determine the specific type of defect, whether it is a point defect, line defect, or volume defect, including: By extracting geometric morphological data from a refined 3D model and using automated tools to classify and organize the morphological data, preliminary morphological classification results are obtained. Based on the preliminary morphological classification results, the data is matched with a pre-established material property database to obtain material property information corresponding to the geometric morphology data and determine the preliminary range of defect types. If the initial range of defect types includes multiple possibilities, then the specific defect type can be determined by further extracting local morphological data from the 3D model and comparing it with material property information. Based on the defect type results after secondary comparison, detailed classification data of point defects, linear defects, or volumetric defects are obtained to determine the final defect type label. For the final defect type label, the support vector machine algorithm is used to verify the morphological data in the 3D model to obtain the verified defect type confirmation result. The specific algorithm is as follows: ； This indicates the confirmation result of the defect type after verification. Indicates the number of support vectors. Indicates the first The weights of the support vectors, Indicates the first Training sample labels, Represents the kernel function. Indicates the first Support vectors, Represents the morphological data of a three-dimensional model. Indicates the bias term; Based on the verified defect type confirmation results, corresponding defect classification records are generated, and the association mapping between defect types and material properties is determined.

8. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The assessment of the physical impact of defects on the conductive and insulating regions of a diode, based on the identified defect types and their spatial distribution in the refined 3D model, includes: Defect types and corresponding spatial location data are obtained from a refined 3D model. By using pre-established region division rules, the boundary information of conductive and insulating regions is extracted to obtain the initial region division results. Based on the initial region division results, and combined with defect distribution and location data, automated tools are used to perform preliminary mapping of the spatial distribution of defects in conductive and insulating regions, and to determine the distribution of contact points between defects and each region. Based on the contact point distribution obtained from the initial mapping, if the defect distribution is concentrated in the conductive area, the physical range of the defect's effect on the conductive area is analyzed using a geometric distance calculation tool to obtain the specific regional action boundary. If the defects are concentrated in the insulation area, the same geometric distance calculation tool is used to analyze the physical range of the defects on the insulation area and determine the specific boundary of the area. For cases where the defect distribution involves both conductive and insulating regions, the contact point distribution data of the two regions is obtained. Combined with the defect type information, the support vector machine algorithm is used to classify and predict the physical action range, and determine the boundary division of cross-regional action. The boundary division results obtained from the above classification predictions are used to verify the matching degree between defect distribution and region division using model analysis tools, thus obtaining the final physical action range data.

9. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, The spatial interpolation algorithm is used to extend the influence range model of a single defect to the entire internal structure of the diode, generating a complete three-dimensional heat map of the defect distribution, including: Acquire voxel mesh data of the three-dimensional structure of the diode and the intensity model of the effect of a single defect; Based on the action model, locate the spatial positions of all identified defect points in the voxel mesh; For each defect point, the initial impact value on the surrounding voxel mesh is calculated using its effect intensity model; The Kriging interpolation method is used to fuse the initial influence values ​​of all defect points and calculate the interpolation influence intensity of the global voxel mesh; the algorithm is as follows: ； Indicates position Kriging estimates at [location] Indicates the first The weight coefficients of the known points Indicates the first Observations at known points This represents the total number of known points; If the interpolation effect intensity exceeds the preset intensity threshold, the voxel is marked as a high-influence region. Based on the labeling results of all voxels and the interpolation influence intensity values, three-dimensional heat map data is generated.

10. The method for detecting nanoscale defects in diodes according to claim 1, characterized in that, Based on the defect density and influence weight of each region in the three-dimensional thermal map of defect distribution, the system outputs the spatial location coordinates of nanoscale defects and a comprehensive evaluation report, including: Obtain three-dimensional heat map data of defect distribution, and obtain the weighted influence value of each region based on the defect density and influence weight information recorded in each region and the preset weight calculation rules. Based on the weighted impact value, the region division method is used to classify the regions in the three-dimensional heat map according to the degree of impact, and to determine the distribution of high-impact and low-impact regions; For defect distribution data in high-impact areas, spatial positioning technology is used to extract the specific location information of nanoscale defects and obtain the corresponding set of spatial coordinates. Based on the set of spatial coordinates and the distribution analysis results of defect density and influence weight, if the weighted influence value corresponding to a certain coordinate point exceeds the preset threshold, then the point is marked as a critical defect point. For the marked critical defect points, obtain their neighboring area data in the three-dimensional heat map, and calculate the associated influence range of each critical defect point through a comprehensive evaluation method to determine the intensity of its effect on the surrounding area. Based on the intensity of action and the set of spatial coordinates, data generation technology is used to integrate the distribution of key defect points and the associated impact range into structured data, resulting in the final evaluation dataset.

Images