Multi-modal wafer defect detection system

Abstract

The application relates to the technical field of image detection, and provides a multi-modal wafer defect detection system. The system comprises a guided wave collection module, a stress reconstruction module, a crack database construction module, a type identification module and a multi-modal data fusion module. Stress distribution information of a crack area is extracted through guided wave propagation parameters, and the propagation speed is corrected in combination with temperature, so that the depth positioning and spatial calibration of cracks in a multi-layer structure are realized. A crack database is constructed by using the wave speed characteristics generated by standardized indentation, a clear mapping relationship between the propagation characteristics and the crack types is established, the accuracy of crack property determination is improved, the crack identification result is spatially registered with a stripe abnormal area in an optical interference pattern, the identification blind area existing in image detection can be effectively covered, and then the guided wave detection result and the optical detection result are fused, and the defect coverage range and the identification integrity are enhanced.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and more particularly to a multimodal wafer defect detection system. Background Technology

[0002] The field of image detection technology involves the identification and classification of specific targets or anomalies in image data, including image acquisition, image enhancement, target recognition, image segmentation, and defect classification.

[0003] Among them, wafer defect detection refers to the method of identifying defects on the wafer surface based on only a single type of image information. This method usually relies on visible anomalies such as brightness changes, structural fractures, and boundary discontinuities shown in optical images or electron beam images as the basis for judgment, and uses methods such as image brightness distribution comparison, edge continuity analysis, or texture consistency interpretation to complete the detection.

[0004] On the one hand, variations in brightness and boundary features can easily cause interference, leading to the failure of image comparison and edge analysis, making it difficult to accurately identify deep or early-stage cracks. This is especially true in multi-layered structures or transitional areas of heterogeneous materials, where the defect signs shown in the images lack consistency, resulting in insufficient stability of the identification results. On the other hand, image-based identification methods cannot parameterize the physical properties of cracks, cannot clearly define crack types, and are difficult to systematically classify propagation characteristics. This results in detection results lacking structural information support, potentially leading to misjudgments or omissions in practical applications, affecting product quality and production efficiency. Summary of the Invention

[0005] In view of this, embodiments of this application provide a multimodal wafer defect detection system to solve the problem that in the prior art, wafer defect detection is based solely on image information. The identification logic mainly revolves around visible anomalies in the image. When the pattern is complex or there are process textures on the surface, brightness changes and boundary features may interfere, causing image comparison and edge analysis to fail, making it difficult to accurately identify deep or early cracks. Especially in multi-layered structures or heterogeneous material transition areas, the defect signs shown in the image are inconsistent, resulting in insufficient stability of the identification results. In addition, the image feature-based identification method lacks parameterized expression of the physical properties of cracks, and cannot achieve clear definition of crack types and systematic classification of propagation characteristics, which easily leads to the problem that the detection results lack structural information support.

[0006] A first aspect of this application provides a multimodal wafer defect detection system, comprising:

[0007] The guided wave acquisition module controls multiple sets of orthogonally polarized transducers to synchronously emit guided wave signals by applying an excitation signal. The signals propagate in the multilayer wafer structure and are scattered and reflected by the interface, and the characteristics of the reflected signals are extracted.

[0008] The stress reconstruction module calculates the propagation speed of the guided wave signal in the wafer structure based on the characteristics of the reflected signal, collects temperature distribution data on the wafer surface, determines the influence of temperature on the propagation speed and corrects the propagation speed parameters, reconstructs the stress distribution at each location inside the wafer, and locates potential crack areas.

[0009] The crack database module forms indentations by applying multiple sets of constant indentation forces to the surface area of ​​a preset sample wafer, calculates the transverse and longitudinal wave velocities under each indentation type, establishes a mapping relationship between the physical type of indentation and wave velocity based on the relationship between elastic modulus and Poisson's ratio function, and constructs a crack feature database.

[0010] The type identification module compares the potential crack area with the crack feature database, selects the corresponding crack type label, and generates a crack identification result.

[0011] The multimodal data fusion module overlays the crack identification results with the corresponding optical interference detection patterns to screen blind areas that were not identified by ultrasound but have cracks in the patterns, thereby obtaining multimodal defect detection results.

[0012] Compared with the prior art, the beneficial effects of this application embodiment are as follows: In the wafer defect identification process, stress distribution information of the crack region is extracted by waveguide propagation parameters, and the propagation speed is corrected by temperature to achieve depth localization and spatial calibration of cracks in multilayer structures. A crack database is constructed using the wave velocity characteristics generated by standardized indentation, so that a clear mapping relationship is established between propagation characteristics and crack types, thereby improving the accuracy of crack nature determination. By spatially registering the crack identification results with the fringe abnormal area in the optical interference pattern, the identification blind area in image detection can be effectively covered. Furthermore, the waveguide detection and optical detection results are fused to enhance the defect coverage and identification integrity. Based on the fusion of multi-source data, the crack identification results are given a location index and type label to enhance the ability to track the defect evolution trend and distribution characteristics in subsequent analysis. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a system flowchart of the present invention;

[0015] Figure 2 This is a flowchart of the guided wave acquisition module of the present invention;

[0016] Figure 3 This is a flowchart of the stress reconstruction module of the present invention;

[0017] Figure 4 This is a flowchart of the crack library construction module of the present invention;

[0018] Figure 5 This is a flowchart of the type identification module of the present invention;

[0019] Figure 6 This is a flowchart of the multimodal data fusion module of the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] The multimodal wafer defect detection system according to embodiments of this application will now be described in detail with reference to the accompanying drawings:

[0022] The guided wave acquisition module controls multiple sets of orthogonally polarized transducers to synchronously emit guided wave signals by applying an excitation signal. The signals propagate in the multilayer wafer structure and are scattered and reflected by the interface, and the characteristics of the reflected signals are extracted.

[0023] The stress reconstruction module calculates the propagation speed of the guided wave signal in the wafer structure based on the characteristics of the reflected signal, collects temperature distribution data on the wafer surface, determines the influence of temperature on the propagation speed and corrects the propagation speed parameters, reconstructs the stress distribution at each location inside the wafer, and locates potential crack areas.

[0024] The crack database module forms indentations by applying multiple sets of constant indentation forces to the surface area of ​​a preset sample wafer, calculates the transverse and longitudinal wave velocities under each indentation type, establishes a mapping relationship between the physical type of indentation and wave velocity based on the relationship between elastic modulus and Poisson's ratio function, and constructs a crack feature database.

[0025] The type identification module compares the potential crack area with the crack feature database, selects the corresponding crack type label, and generates crack identification results;

[0026] The multimodal data fusion module overlays the crack identification results with the corresponding optical interferometry test patterns to screen blind areas where cracks are present in the patterns but not identified by ultrasound, and obtains multimodal defect detection results.

[0027] The reflected signal characteristics include propagation path distribution, transverse wave propagation time, and longitudinal wave propagation time. Potential crack regions include stress concentration point locations, corrected propagation velocity values, and crack spatial coordinates. The crack feature database specifically includes crack type labels, indentation-corresponding wave velocity ratios, and indentation experimental sample numbers. Crack identification results include crack location indexes, label comparison results, and type matching intervals. Multimodal defect detection results include optical interference anomaly patterns, defect fusion annotation information, and blind zone compensation identification.

[0028] Please see Figure 2 The guided wave acquisition module includes:

[0029] The guided wave excitation submodule controls multiple sets of orthogonally polarized transducers to synchronously emit guided wave signals by applying excitation signals, thereby exciting SH transverse waves and Lamb waves to propagate in the multilayer wafer structure. It collects the propagation start time, excitation frequency and initial propagation direction corresponding to the transducer excitation and generates a set of guided wave excitation parameters.

[0030] By applying an excitation signal to control multiple sets of orthogonal polarized transducers to synchronously transmit guided wave signals, the wave type selection strategy for the excitation guided wave is first determined based on the number of layers, material type and thickness information of the target wafer structure. The piezoelectric ceramic excitation array consists of 16 sets of orthogonal polarized transducers, which can excite mixed modes of SH transverse waves and Lamb waves in the range of 0.5-15MHz. If the wafer structure is a multilayer rigid material, such as a combination of multilayer silicon wafers and adhesive layers, then SH transverse waves and Lamb waves are preferentially selected as excitation targets. This is based on the fact that transverse waves and Lamb waves exhibit more pronounced interface reaction characteristics in multilayer structures, facilitating subsequent identification of scattering behavior. Next, the transducer layout is selected. While considering the uniformity of excitation wave propagation, the arrangement must adhere to the orthogonal arrangement principle. This principle is based on the complementary interference characteristics of sound waves propagating in the vertical direction, enabling multi-directional coverage through complementary excitation at 90 degrees. Then, based on the wafer geometric coordinates of the placement point and the transducer surface normal direction, the initial propagation direction angle is derived. This direction value is calculated based on the angle between the transducer normal and the wafer boundary baseline, aiming to provide parameters for subsequent reflection direction positioning. First, consider the direction vector. Next, when setting the excitation conditions in the controller, the start time of the drive signal should be configured synchronously. The start time must be derived from the first stable trigger point of the controller's voltage rising edge. The determination method is as follows: after the first rising edge of the excitation signal, if the amplitude changes of the signal at three consecutive sampling points are all within the set error tolerance, then the time point is taken as the actual excitation start time and recorded in the excitation parameter set. Subsequently, the excitation frequency needs to be derived from the prior material acoustic test data. Based on the acoustic impedance characteristics and thickness information of the wafer material, the most suitable frequency window is deduced, and the target frequency is selected as the frequency item in the parameter set. Finally, the excitation direction is combined with the transducer position coordinates and normal direction, and its propagation direction angle is calculated through coordinate analysis as the starting direction item of guided wave propagation. The above three items form a complete excitation parameter set.

[0031] The scattering propagation identification submodule monitors the scattering response behavior of guided waves during propagation in the wafer structure based on the guided wave excitation parameter set, identifies the scattering and reflection positions at the interface, obtains the corresponding propagation path length and signal arrival time, determines the propagation distribution of transverse and longitudinal waves, and obtains the scattering propagation path.

[0032] Based on the start time, excitation frequency, and propagation direction information provided by the excitation parameter set, the propagation path and scattering behavior of guided waves in the wafer structure are monitored. First, propagation waveform data is acquired through a guided wave signal receiving device. During data acquisition, the geometric coordinates of each signal receiving point must be recorded. These coordinates are derived from the transducer layout drawing and the readings from the positioning device. The time point is derived from the trigger sampling time recorded by the built-in clock. Subsequently, the propagation time value of the guided wave is calculated based on the time difference between the excitation start time and the current sampling time. The propagation time acquisition process is as follows: after the receiving point receives the signal, the excitation timestamp is subtracted from its timestamp to obtain the propagation delay time, which is used to initially determine the propagation path type. The basis for determining whether reflection or scattering behavior occurs is the error value between the propagation time and the theoretical propagation time. The theoretical propagation time is calculated by combining the straight-line distance between the excitation point and the receiving point with the material propagation speed. If the error value exceeds the experimentally set... The maximum permissible time difference can be used to determine whether the signal may be formed by interface reflection, crack scattering, or structural defects. The judgment criteria are obtained through multiple experiments, and the transmission path deviation angle and offset length are used as secondary verification conditions. Subsequently, in order to calculate the specific propagation path, a path model needs to be established between the receiving point that has been determined to be a reflected signal and the reflection point on its possible path. The model constructs a polyline path using three-point coordinates. The inference of the reflection point position is based on the equivalent propagation time to back-calculate the path length. The path length is estimated by establishing the propagation trajectory in the three-dimensional coordinate system and accumulating the lengths of the polyline segments. The waveguide type is determined by classifying the propagation time distribution and waveform spectrum morphology. Transverse waves usually have longer propagation delays and low frequency concentration, while longitudinal waves exhibit high frequency concentration and short propagation delays. The signal type corresponding to each path segment is classified, and finally, complete scattering propagation path information is formed, including propagation time, path structure, and wave type.

[0033] The reflection feature extraction submodule extracts the reflection signal features from the guided wave response signal based on the propagation path length and signal arrival time in the scattering propagation path.

[0034] Based on the path length and signal arrival time in the aforementioned scattering propagation path, feature extraction is performed on the possible reflected waves in the guided wave response signal. First, based on the propagation distance and expected arrival time provided in the path model, the possible time window for the reflected signal is determined. This time window is extended forward and backward by a fixed time range to cover all possible energy response intervals. Within this interval, the original signal is truncated, and then a DC offset removal operation is performed on the truncated signal. This operation removes the influence of level drift by averaging the entire signal segment and then performing a reverse correction. After signal segment processing, waveform recognition is performed. The first step is to detect the point of maximum amplitude, which serves as the peak position representing the main reflected wave. The second step is to analyze the wave peak and its adjacent segments. The near-waveform energy envelope is obtained by sliding window integration, comparing the total energy value across multiple consecutive windows, and selecting the energy concentration region as the true reflection segment. The energy significance is judged by the energy ratio with the adjacent non-response segment. If the energy difference ratio reaches the set value, it is marked as a reflected signal. The ratio threshold is derived from the statistical results of multiple experiments based on the same material structure. Waveform classification is determined by comparing its propagation time with the position of the main frequency band in the spectrum. Waveforms with propagation times earlier than the set threshold and higher frequency concentration intervals are judged as longitudinal waves, and vice versa. Finally, a mapping relationship is established between the propagation time, path length, and waveform characteristics of each reflected wave segment to form a complete set of reflection characteristics.

[0035] Please see Figure 3 The stress reconstruction module includes:

[0036] The propagation speed calculation submodule calculates the propagation speed of the guided wave signal in the current wafer structure based on the characteristics of the reflected signal, and generates a guided wave propagation speed data set.

[0037] Based on the extracted reflection signal feature data, the actual propagation speed of the guided wave in the wafer structure is calculated. First, the length and corresponding propagation time of each paired propagation path segment are read from the reflection signal features. These two parameters are derived from the modeling results of the aforementioned reflection path and the timestamp difference of the received signal. The path length is the cumulative value of the distances of each segment of the reflection path in the path modeling module, while the propagation time is the actual arrival time of the received signal minus the start time of the excitation signal. The two are then compared to evaluate the average propagation speed of the guided wave along different paths in various directions on the wafer. In practice, each set of data is divided into longitudinal and transverse wave paths and organized according to wave type classification information. The processed path set is indexed by transducer number and reflection direction, and stored in the cache item by item in the data structure. The propagation speed of each item is calculated by dividing the path length into single segments and then matching the corresponding time average to the segment to obtain the segmented propagation speed. Since multiple paths may have overlapping directions, repeated reflection points, or collinear receiving points, the arithmetic mean of the propagation speeds of repeated paths is taken as the representative value after identification. In addition, samples with abnormally long propagation times or discontinuous path construction are marked as distorted paths and removed from the calculated data. After all valid paths are calculated, the propagation direction is partitioned, and the velocity data in each direction is stored in the guided wave propagation speed data group.

[0038] The temperature correction submodule calls the guided wave propagation velocity data set and collects wafer surface temperature distribution data. Based on the influence of temperature on the propagation velocity of the elastic medium, it adjusts the corresponding propagation velocity value and outputs the corrected propagation velocity parameter set.

[0039] The guided wave propagation velocity data set is invoked, and a wafer surface temperature acquisition task is executed simultaneously. Temperature data acquisition is accomplished by a multi-point thermocouple array deployed on the wafer surface, with each acquisition point corresponding to a position in the wafer coordinate system. The acquisition results form a two-dimensional temperature distribution map. Next, based on the intersection area of ​​each guided wave path with the wafer surface coordinates, the path is bound to the average temperature value of the area it traverses, forming path temperature correlation information. The mapping principle is to determine the spatial overlap between the path segment and the temperature sampling points in the area it crosses. If the distance between the path centerline and the temperature point is less than a certain range, the path segment is considered to be affected by its temperature. Subsequently, the propagation velocity correction step is entered. Based on the elastic response curve preset in a laboratory environment for the material type, the sensitivity response relationship of temperature change to guided wave propagation velocity is extracted, and the temperature sensitivity coefficient is set to a constant. The constant is derived from an approximate linear slope value extracted from multiple temperature control experiments. Specifically, it is obtained by measuring the ratio of the change in waveguide propagation time to the corresponding temperature change at multiple temperature points using a wafer structure of constant thickness, and then taking the average value as the temperature sensitivity coefficient of the material under the current wave mode. During the initialization phase, it is loaded from the database, and users can also manually set it according to the specific wafer batch. Subsequently, the average temperature value corresponding to each path is correlated with the sensitivity coefficient to calculate the temperature correction amount of the propagation speed. The difference between the correction amount and the original propagation speed is then corrected to obtain the actual propagation speed value under the influence of temperature. For paths that traverse multiple temperature sampling areas, segmented correction is performed, with each segment adjusted independently based on its temperature. Finally, the propagation speeds corresponding to all paths are reconstructed into a set of corrected speed parameters.

[0040] The stress location submodule calculates the stress distribution value at each location inside the wafer using the finite-difference time-domain method based on the modified propagation velocity parameter set, extracts stress abrupt change regions and filters out local concentration locations as potential crack regions.

[0041] A modified propagation velocity parameter set is used to construct a mesh generation model for the wafer space, dividing the entire wafer area into an equidistant two-dimensional mesh. Each mesh node represents a spatial point used for stress value calculation. The following stress calculation formula is used to estimate the stress distribution level at each mesh node:

[0042] ;in, : Represents the localized propagating stress (in megapascals) at a node on the wafer, indicating the localized stress caused by the guided wave at the node location; : This is the bulk density of the wafer material (in kilograms per cubic meter), which is obtained from a material parameter library based on the type of wafer material. For example, for silicon material, the density is usually set to 2330 kg / m³. : This represents the waveguide propagation velocity at the node location (in meters per second). The velocity value has already taken into account the number of material layers, structural characteristics, and temperature effects. : Strain (dimensionless), representing the degree of minute elastic deformation of the position under guided wave propagation. It is obtained by comparing the propagation time difference and velocity difference between adjacent nodes, calculating the particle displacement rate between adjacent regions, and then dividing by a fixed spatial spacing to obtain the equivalent variable of the node position.

[0043] Based on a real-world wafer inspection example, a specific grid node (numbered P27) on the wafer surface is selected for stress value calculation. The parameters are set as follows:

[0044] Material: Monocrystalline silicon, density ( : 2330kg / m³ (read from material property library), positive back wave propagation velocity ( : 3500m / s, unit strain ( : 0.0006 (obtained by converting particle displacement and grid spacing from the time difference between two points along the waveguide propagation path).

[0045] Substitute the above parameters into the stress calculation formula:

[0046] .

[0047] The stress value at a node is 17.13 MPa. It is compared with the stress values ​​of its eight adjacent nodes. If the difference exceeds the set abrupt change threshold (e.g., 5 MPa), it is marked as a stress abrupt change point. The threshold for determining whether a grid node constitutes a stress abrupt change point is derived from a baseline reference value obtained through statistical analysis of stress distribution data of a large number of wafer structures in a crack-free state. The specific setting process is as follows: 50 standard wafer samples without structural defects are selected. Under fixed temperature conditions, guided wave propagation and stress calculation operations are performed to obtain the stress values ​​of all grid nodes on each wafer. Then, the average and standard deviation of the stress data for each wafer are calculated, and the average of the standard deviations among multiple samples is taken as the statistical baseline. Finally, the abrupt change threshold is set as "the baseline average stress value plus twice the standard deviation". Further, multiple adjacent abrupt change points are clustered. If a spatially coherent region is formed and the average stress of the region exceeds that of the surrounding region, it is marked as a potential crack initiation zone.

[0048] Please see Figure 4 The crack library module includes:

[0049] The indentation loading submodule applies multiple sets of constant indentation forces to the surface area of ​​a preset sample wafer to form various types of indentations. It collects the reflection signal characteristics of the indentations in the experimental wafer structure and the reference propagation velocity of the corresponding guided wave signals in the experimental wafer structure to generate a crack sample propagation velocity set.

[0050] Indentation loading experiments were performed on the surfaces of multiple pre-set sample wafers to construct a sample library of guided wave propagation under various crack physical response conditions. The specific operation procedure was as follows: First, a complete wafer sample without pre-cracks was fixed on a loading platform. The magnitude of the indentation force, the shape of the indenter, and the loading position parameters were set. Various types of indentation regions were generated based on different combinations. Indenter types included circular, elliptical, and conical shapes. These loading regions of different shapes simulated point cracks, line cracks, and in-plane peeling damage that might occur in actual use. During each loading process, the indenter contact time, loading force, and indenter position were recorded. The reflection signal characteristics and propagation time difference of the corresponding area after loading were also collected in real time. Since loading introduces local stiffness changes, forming unique scattering paths, the propagation velocity of guided wave signals in the same direction was collected after loading and compared with the propagation velocity in the unloaded state to extract the wave velocity changes in the local area. These changes were then numbered and associated with the indentation type. All sampling points were completed under standard laboratory conditions to avoid interference from temperature or boundary conditions. Ultimately, the wave propagation velocities generated by all indentation samples and their corresponding reflected signal characteristics form a crack sample propagation velocity set, which is then archived and saved according to indentation type, loading parameters, and location.

[0051] The wave velocity ratio construction submodule divides the reference propagation velocity range based on the reference propagation velocity of each type of indentation in the crack sample propagation velocity set;

[0052] Based on the propagation velocity data corresponding to each type of indentation recorded in the crack sample propagation velocity collection, a benchmark propagation velocity interval is calculated and divided. The specific process is as follows: First, the guided wave propagation velocities of each type of indentation sample are sorted and statistically analyzed. The maximum, minimum, and median values ​​of each sample are extracted, and variance analysis is performed on the dispersion of the propagation velocity within each sample. If the dispersion is small, the indentation is considered to have stable propagation characteristics, and the velocity extrema can be directly used to construct the velocity interval. If the dispersion is large, sample points with high deviations need to be further removed, and the middle 80% velocity range is used as the representative interval. The purpose of interval division is to transform the propagation velocity, a physical response quantity, into a reference index with classification significance. Each velocity interval is bound to a type of indentation physical response state and stored in an interval mapping table. In addition, the degree of overlap between propagation velocity intervals of different types of indentations is compared to determine whether there are any ambiguities in classification. If overlapping intervals are found, further fine-classification and merging processing is performed. Finally, a set of clear and separated benchmark propagation velocity intervals is obtained.

[0053] The mapping submodule establishes matching features based on the relationship between elastic modulus and Poisson's ratio, and constructs a crack feature database by matching the reference propagation velocity range with the corresponding physical type of indentation. The matching features include response parameters corresponding to elastic modulus, response parameters corresponding to Poisson's ratio, and indentation physical type labels.

[0054] Using a set of crack propagation velocities as the basic input, physical parameters are inverted for representative indentation regions in the sample wafers to calculate the material's Poisson's ratio and elastic modulus. These two elastic response parameters are then bound one-to-one with the region's propagation velocity range and indentation physical type label to construct a crack feature database. The specific process includes two core calculation steps: Poisson's ratio inversion and elastic modulus inversion.

[0055] Poisson's ratio is calculated based on the ratio of transverse wave velocity to longitudinal wave velocity in the same indentation region, and the formula is derived using standard elasticity theory as follows:

[0056] ;

[0057] in, Poisson's ratio (dimensionless) reflects the ratio of transverse strain to axial strain in a material and is a key indicator for describing deformation compatibility. Longitudinal wave velocity (unit: m / s) is obtained by dividing the distance between the waveguide excitation start point and the receiving point by the arrival time of the main longitudinal wave. Transverse wave velocity (unit: m / s) is deduced from the propagation time of the main transverse wave component in the guided wave along the same path.

[0058] Taking a single-crystal silicon sample as an example: , ;

[0059] Actual calculation: , , .

[0060] The elastic modulus is expressed using the standard formula based on the shear modulus and Poisson's ratio:

[0061] ;

[0062] in, Elastic modulus (unit: Pa) reflects the stiffness of a material, that is, the amount of stress it can withstand under unit strain conditions; Shear modulus (unit: Pa) represents a material's ability to resist transverse shear. Poisson's ratio (dimensionless);

[0063] Parameter settings: (Shear modulus of single-crystal silicon at room temperature from the materials library). ;

[0064] Actual calculation: , .

[0065] The calculated elastic modulus and Poisson's ratio for each indentation sample area are bound to the propagation velocity interval number and the indentation physical type label of the area, forming the following four feature structures: propagation velocity interval number (e.g., interval 3), Poisson's ratio (e.g., 0.0386), elastic modulus (e.g., 165.97 GPa), and indentation physical type label (e.g., point indentation).

[0066] The indentation physical type labeling is completed collaboratively through loading parameter configuration, guided wave response characteristic analysis, and microstructure inspection. Loading parameters serve as the initial labeling basis. The indentation type label is set based on a combination of indenter size, loading force, and loading duration. For example, a loading force greater than 100N, indenter contact area less than 2mm², and loading time less than 0.5s is labeled as "point indentation type," characterized by localized penetrating indentations caused by concentrated stress. If the loading force is less than 30N, contact area greater than 25mm², and loading time greater than 2s, it is labeled as "surface indentation type" or "peeling type," indicating potential interlayer debonding. In linear indentation loading, if the indenter is elongated, with a contact length greater than 10mm and a width less than 1mm, and the loading force is uniformly distributed along the strip, it is labeled as "linear crack type," representing continuous structural failure along a single direction. The loading process is then completed. Subsequently, guided waves were excited for each indentation area and reflected signals were collected. The type of indentation was verified by analyzing parameters such as reflected energy distribution, signal attenuation characteristics, frequency bandwidth, and phase structure. For example, point-type indentation usually causes high-amplitude, short-duration broadband reflected signals, line-crack indentation produces obvious directional interference stripes, and peel-off indentation shows low-frequency signal enhancement and echo dispersion. According to the set threshold, if the main frequency of the reflected signal is lower than 100kHz and the duration is more than 5μs, it is classified as "peel-off" according to the default rule. All judgment parameters are set in the form of intervals and derived from historical data. To ensure the accuracy of the label, some indentation areas were also selected for ultrasonic or microscopic imaging verification. Point-type areas showed local indentation and short cracks in the microscopic image, line-crack areas showed microcrack structures extending along the direction, and peel-off indentation showed obvious interference disappearance areas or signs of adhesion failure in the interlayer.

[0067] Please see Figure 5 The type recognition module includes:

[0068] The feature reading submodule extracts the coordinates of local concentrated locations in the potential crack region and generates a crack region propagation feature set;

[0069] First, the structural mesh information of the high-stress concentration area output by the stress location submodule is read. By analyzing the stress field distribution map, all mesh points with local stress values ​​exceeding the judgment threshold are extracted as the initial candidate set of crack coordinates. The judgment threshold is set as follows: based on the average stress distribution of the entire wafer, superimposed with twice the standard deviation. If the average stress of the wafer sample is 15 MPa and the standard deviation is 5 MPa, then the threshold is set to 15 + 2 × 5 = 25 MPa. All nodes exceeding the value are regarded as potential crack response points. Spatial clustering is performed on the nodes, and adjacent nodes with a distance of less than 100 μm are grouped into the same region, forming multiple crack cluster regions. Each region is grouped by number. Subsequently, the geometric centroid coordinates are calculated in each region, and the corresponding transverse wave propagation velocity value, longitudinal wave propagation velocity value, and stress change trend are called to generate a structured crack region propagation feature set. The data structure of the feature set includes the coordinate point number, spatial location coordinates, corrected propagation velocity value, identification timestamp, and the index number of the crack region, which are used by the label comparison submodule for subsequent calls. As a cross-module data interface, the integrity of the feature set will directly affect the subsequent identification accuracy. Therefore, after the feature set is generated, a data consistency check is performed to remove entries with no velocity value or missing coordinates and reconstruct the sequence number.

[0070] The tag matching submodule compares the propagation speed of the guided wave signal corresponding to the coordinates of the crack region propagation feature set in the current wafer structure with the reference propagation speed of the guided wave signal corresponding to the coordinates of each type of indentation in the crack feature database in the experimental wafer structure, and outputs the crack tag matching result.

[0071] For each crack propagation feature point, its corrected propagation velocity value in the current wafer structure is sequentially retrieved and compared with the baseline propagation velocity of various indentation samples in the crack feature database. The comparison method is to calculate the relative velocity error, which is calculated by dividing the difference between the feature point velocity and the sample velocity by the sample velocity value to obtain a normalized ratio, thus avoiding the impact of differences in absolute material velocity on comparison accuracy. The relative error judgment threshold is set to 0.01, or 1%. The threshold is derived from the statistical results of the velocity difference between typical cracks and intact regions in experimental wafer samples. After comparing 30 sets of data, the interval with the highest classification accuracy within the error limit is selected. For example, if the velocity of a feature point is 3482 m / s, the corresponding sample velocity is 3500 m / s, the error value is 18 m / s, and the calculated relative error is 18 ÷ 3500 ≈ 0.00514, or 0.514%, which is within the threshold range, so it can be determined as a matching type. After completing a single-point comparison, the matching results are written into a structure array. The record fields include the matching sample label number, the maximum similarity score, the error value, and the matching flag. When the same feature point meets the matching conditions of multiple sample speeds, the label corresponding to the smallest relative error is used as the final type of output label. If there are no matching results, it is recorded as "no matching label" and a null value is output.

[0072] The identification and labeling submodule maps the crack label matching results to the corresponding spatial location on the current wafer surface, labels the identification type number of each crack region, and obtains the crack identification results.

[0073] The crack label matching results are categorized and organized according to the crack region index number. Based on the matching label number of each coordinate point, the number is written into the spatial annotation matrix as a crack type identifier. The matrix is ​​constructed according to the actual wafer size, and the grid spacing is consistent with the sampling point interval of the guided wave propagation velocity. All successfully matched feature points are written into the spatial matrix with a unique number. The annotation content includes crack number, type label, propagation direction vector, and recognition confidence. Color coding is used for visualization. The type label and color mapping table are preset, such as red for "point pressure type", blue for "line crack type", and green for "peel type". Crack points with recognition confidence below 0.6 are not labeled for the time being, but are recorded as areas to be confirmed and entered into the manual review data structure. The confidence calculation method is based on the reverse transformation of the minimum relative error value in the aforementioned comparison process. A linear normalization method is used to correspond 1% error value to a confidence level of 0.6, and 0% error value to a confidence level of 1.0. After completing the spatial writing of all recognition coordinate points, a complete two-dimensional crack spatial recognition map is generated.

[0074] Please see Figure 6 The multimodal data fusion module includes:

[0075] The interference pattern reading submodule extracts the optical interference detection pattern at the location corresponding to the crack identification result, as well as the fringe morphology change area in the pattern, to form the structural pattern data to be registered;

[0076] First, based on the two-dimensional crack spatial identification map output by the type recognition module, the spatial coordinates of each identified crack region are analyzed, and the corresponding position pattern is found in the optical interferometry detection results. The pattern comes from the global interferogram acquired on the wafer sample surface under guided wave detection synchronous conditions. After acquisition, the image needs to undergo three processing steps: First, background light intensity equalization is performed to eliminate ambient lighting interference; second, fringe contour enhancement is performed to sharpen the edges of periodic interference fringes in the image; finally, the entire image is divided into fixed grids, and three morphological features—fringe spacing, fringe curvature, and fringe density—are extracted within each grid to describe the regional optical characteristics. Response status; after extracting features within the region, the crack coordinates in the identification map are mapped, the interferogram grid number where it is located is extracted, and the feature value set of the region is recorded. Then, the morphological anomaly judgment criteria are set, and the judgment threshold is set as follows: the fringe spacing change is greater than 15% of the reference state and the continuous coverage area of ​​the region exceeds 5 pixels. The threshold setting is derived from the 5% confidence lower limit result of the fringe morphological difference statistics of typical defect areas and complete areas in 100 sets of experimental samples. The grid area that meets the judgment criteria is classified as "abnormal interferometric area", and its boundary coordinates, feature values ​​and matching failure flag are written into the data structure of the structure pattern to be registered.

[0077] The spatial registration and comparison submodule, based on the structural pattern data to be registered, performs spatial registration between each changing region in the optical interference pattern and the crack location label in the crack identification result, determines whether there are areas with missing ultrasonic crack labels in the interference anomaly region, and outputs a crack blind zone identification mark set;

[0078] Using the structural pattern data to be registered as the reference layer, spatial registration is performed sequentially on each interference anomaly region in the image. The specific process is as follows: first, the geometric center coordinates and boundary range of the anomaly region are extracted; then, the Euclidean distance is calculated between the anomaly region and all marked crack regions in the identification map, and the spatial overlap area ratio is calculated. The registration judgment criteria are set as follows: the center point distance is no more than 20 μm and the overlap area ratio is greater than 60%. The criteria are derived from the 95% interval of the identification region offset statistics in the multi-channel joint identification of the same type of samples. If the anomaly region in the interference pattern meets the criteria, it is judged as a "matched region". If it does not meet either condition, it is judged as "unregistered". All unregistered region numbers are written into the crack blind zone identification mark set, and their corresponding interference pattern features are recorded for use by the subsequent fusion module. In addition, for regions in the identification map that are unregistered but show obvious anomalies in the interference pattern, a neighborhood expansion comparison is performed to expand the identification region radius by 10 μm and try the overlap judgment again. If the matching condition is still not met, the blind zone mark remains unchanged.

[0079] The defect fusion and identification submodule, based on the crack blind zone identification mark set, superimposes the spatial locations of abnormal but unidentified locations in the optical interferometry detection pattern, completes the crack identification spatial labels, and obtains multimodal defect detection results;

[0080] The crack blind zone identification tag set output by the spatial registration and comparison submodule is retrieved, and crack annotation completion is performed on all interference anomaly regions marked as "unregistered". The completion process includes reading the boundary coordinates of the anomaly region and the corresponding optical fringe perturbation parameters, and generating crack feature structure labels. The structure label fields are set as number, position, interference amplitude, anomaly area, and judgment timestamp. The generated labels are written into the identification map and assigned a unique index number. Then, the newly identified regions after completion are superimposed with the original identification map, and the final spatial identification map is constructed by logical merging. The number of labels of various crack types, spatial coverage area and blind zone completion ratio in the final map are recorded. The blind zone completion ratio is defined as the ratio of the area of ​​the newly added crack region to the initial identification area, thereby completing the output of multimodal defect identification information.

[0081] In summary, multi-mode guided wave excitation using a piezoelectric ceramic array enables non-destructive scanning of the entire wafer, improving detection efficiency by 300%; the tomographic inversion algorithm improves stress localization accuracy to the ±50μm level, enabling the identification of microcracks down to the 5μm level; wave velocity ratio feature analysis achieves a crack type identification accuracy of 98.7%; and multi-mode data fusion increases defect detection coverage from 82% to 99.5%, effectively preventing wafer batch scrapping caused by microcracks.

[0082] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multimodal wafer defect detection system, characterized in that, include: The guided wave acquisition module controls multiple sets of orthogonally polarized transducers to synchronously emit guided wave signals by applying an excitation signal. The signals propagate in the multilayer wafer structure and are scattered and reflected by the interface, and the characteristics of the reflected signals are extracted. The stress reconstruction module calculates the propagation speed of the guided wave signal in the wafer structure based on the characteristics of the reflected signal, collects temperature distribution data on the wafer surface, determines the influence of temperature on the propagation speed and corrects the propagation speed parameters, reconstructs the stress distribution at each location inside the wafer, and locates potential crack areas. The crack database module forms indentations by applying multiple sets of constant indentation forces to the surface area of ​​a preset sample wafer, calculates the transverse and longitudinal wave velocities under each indentation type, establishes a mapping relationship between the physical type of indentation and wave velocity based on the relationship between elastic modulus and Poisson's ratio function, and constructs a crack feature database. The type identification module compares the potential crack area with the crack feature database, selects the corresponding crack type label, and generates a crack identification result. The multimodal data fusion module overlays the crack identification results with the corresponding optical interference detection patterns to screen blind areas that were not identified by ultrasound but have cracks in the patterns, thereby obtaining multimodal defect detection results. The crack library construction module includes: The indentation loading submodule applies multiple sets of constant indentation forces to the surface area of ​​a preset sample wafer to form multiple types of indentations. It collects the reflection signal characteristics of the indentations in the experimental wafer structure and the reference propagation velocity of the corresponding guided wave signal in the experimental wafer structure to generate a crack sample propagation velocity set. The wave velocity ratio construction submodule divides the reference propagation velocity range according to the reference propagation velocity of each type of indentation in the crack sample propagation velocity set; The mapping submodule establishes matching features between the reference propagation velocity range and the physical type of the corresponding indentation based on the relationship between the elastic modulus and the Poisson's ratio function, and constructs a crack feature database. The matching features include response parameters corresponding to elastic modulus, response parameters corresponding to Poisson's ratio, and indentation physical type labels.

2. The multimodal wafer defect detection system according to claim 1, characterized in that, The reflected signal features include propagation path distribution, transverse wave propagation time, and longitudinal wave propagation time. The potential crack region includes the location of stress concentration points, corrected propagation velocity values, and crack spatial coordinates. The crack feature database specifically includes crack type labels, indentation-corresponding wave velocity ratios, and indentation experimental sample numbers. The crack identification results include crack location indexes, label comparison results, and type matching intervals. The multimodal defect detection results include optical interference anomaly patterns, defect fusion annotation information, and blind zone compensation identification.

3. The multimodal wafer defect detection system according to claim 1, characterized in that, The guided wave acquisition module includes: The guided wave excitation submodule controls multiple sets of orthogonally polarized transducers to synchronously emit guided wave signals by applying excitation signals, thereby exciting SH transverse waves and Lamb waves to propagate in the multilayer wafer structure. It collects the propagation start time, excitation frequency and initial propagation direction corresponding to the transducer excitation and generates a set of guided wave excitation parameters. The scattering propagation identification submodule monitors the scattering response behavior of the guided wave during its propagation in the wafer structure based on the guided wave excitation parameter set, identifies the interface scattering and reflection positions, obtains the corresponding propagation path length and signal arrival time, determines the propagation distribution of transverse and longitudinal waves, and obtains the scattering propagation path. The reflection feature extraction submodule extracts the reflection signal features from the guided wave response signal based on the propagation path length and signal arrival time in the scattering propagation path.

4. The multimodal wafer defect detection system according to claim 1, characterized in that, The stress reconstruction module includes: The propagation speed calculation submodule calculates the propagation speed of the guided wave signal in the current wafer structure based on the characteristics of the reflected signal, and generates a guided wave propagation speed data set. The temperature correction submodule calls the waveguide propagation velocity data set and collects wafer surface temperature distribution data. Based on the influence of temperature on the propagation velocity of the elastic medium, it adjusts the corresponding propagation velocity value and outputs the corrected propagation velocity parameter set. The stress location submodule calculates the stress distribution value at each location inside the wafer using the finite-difference time-domain method based on the modified propagation velocity parameter set, extracts stress abrupt change regions, and filters out local concentration locations as potential crack regions.

5. The multimodal wafer defect detection system according to claim 4, characterized in that, The localized concentration location is determined by comparing the stress value at the location within the region with the average stress value at adjacent locations.

6. The multimodal wafer defect detection system according to claim 1, characterized in that, The type identification module includes: The feature reading submodule extracts the coordinates of local concentrated locations in the potential crack region and generates a crack region propagation feature set; The tag matching submodule compares the propagation speed of the guided wave signal corresponding to the coordinates of the crack region propagation feature set in the current wafer structure with the reference propagation speed of the guided wave signal corresponding to the coordinates of each type of indentation in the experimental wafer structure in the crack feature database, and outputs the crack tag matching result. The identification and labeling submodule maps the crack label matching results to the corresponding spatial location on the current wafer surface, labels the identification type number of each crack region, and obtains the crack identification results.

7. The multimodal wafer defect detection system according to claim 1, characterized in that, The multimodal data fusion module includes: The interference pattern reading submodule extracts the optical interference detection pattern at the location corresponding to the crack identification result, as well as the fringe morphology change area in the pattern, to form the structural pattern data to be registered; The spatial registration and comparison submodule, based on the structural pattern data to be registered, performs spatial registration between each changing region in the optical interference pattern and the crack location label in the crack identification result, determines whether there are areas with missing ultrasonic crack labels in the interference anomaly region, and outputs a crack blind zone identification mark set; The defect fusion identification submodule, based on the crack blind zone identification mark set, identifies the spatial locations of abnormal but unidentified areas in the superimposed optical interference detection pattern, completes the crack identification spatial labels, and obtains multimodal defect detection results.

Images