Wafer detection system and detection method thereof

Abstract

The invention is suitable for the technical field of wafer detection, and provides a wafer detection system and a detection method thereof.The wafer detection system comprises a multi-axis motion platform, an optical imaging system and a control and processing unit, and the multi-axis motion platform is used for bearing a wafer and driving the wafer to rotate, translate and move in the focusing direction; the optical imaging system is used for imaging the surface of a wafer and supports switching of multiple illumination modes, the illumination modes comprise bright field illumination, dark field illumination and oblique illumination, and the control and processing unit is electrically connected with the multi-axis motion platform and the optical imaging system. According to the system, multi-angle and multi-dimensional image data of the same suspected defect point can be obtained through wafer attitude regulation and control of the multi-axis motion platform and cooperative switching of multiple illumination modes of the optical imaging system; and on the basis, four types of quantitative characteristics of motion adhesiveness, three-dimensional contour, multi-modal optical response and multi-view morphological stability for representing the physical essence of the defect are extracted, so that the real defect and the false defect are distinguished.

Description

Technical Field

[0001] This invention relates to the field of wafer inspection technology, and more specifically, to a wafer inspection system and inspection method thereof. Background Technology

[0002] In semiconductor manufacturing processes, wafer surface defect detection is a core step in ensuring chip yield and controlling production quality. As semiconductor processes continue to upgrade to the micro-nano level, the requirements for wafer defect detection in terms of detection accuracy, ability to distinguish between genuine and counterfeit defects, and detection efficiency are becoming increasingly stringent. At the same time, the detection system needs to be able to output quantitative characteristic data of defects to provide support for process optimization and defect root cause tracing.

[0003] However, existing wafer defect detection technologies are mostly based on two-dimensional image recognition schemes with single bright-field illumination. They can only extract shallow visual information such as grayscale, contrast, and simple shape features, and cannot distinguish between defects that are actually attached to the wafer and false defects introduced by the system or environment, resulting in a high false positive rate. At the same time, they are not good at identifying defects that depend on three-dimensional morphology and motion adhesion, such as cracks, micro-protrusions, and dents, which are prone to missed detection and affect the accuracy of yield judgment. This paper proposes a wafer inspection system and its detection method to improve the existing problems. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a wafer inspection system and inspection method.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a wafer inspection system, comprising: A multi-axis motion platform is used to carry the wafer and drive it to rotate, translate, and move in the focusing direction.

[0006] An optical imaging system for imaging a wafer surface and supporting switching between multiple illumination modes, including bright field illumination, dark field illumination, and oblique illumination.

[0007] The control and processing unit is electrically connected to the multi-axis motion platform and the optical imaging system, respectively. The control and processing unit includes a data acquisition module, a feature extraction module, a model calculation module and a control drive module.

[0008] The control and processing unit is configured to: control the multi-axis motion platform and the optical imaging system to work together to perform wafer full-domain scanning, suspected defect location and multi-condition imaging operations, extract multi-dimensional quantitative features of suspected defects, run a decision model to classify and determine suspected defects, and output classification results.

[0009] The present invention is further configured such that: the control drive module is used to control the multi-axis motion platform and the optical imaging system to coordinate their actions, and to perform wafer full-area scanning, suspected defect point localization and multi-condition imaging operations.

[0010] The data acquisition module is used to synchronously receive image data acquired by the optical imaging system and real-time position feedback data from the multi-axis motion platform.

[0011] The feature extraction module is used to extract four types of multi-dimensional quantitative features of suspected defect points: motion adhesion, three-dimensional contour, multimodal optical response, and multi-view morphological stability.

[0012] The model operation module has a built-in multi-evidence fusion network decision model, which is used to classify and determine suspected defects based on the multi-dimensional quantitative features of the input, and output the defect type classification result.

[0013] The core of the multi-evidence fusion network decision model is to dynamically allocate the weights of each feature and integrate multi-dimensional evidence to distinguish between different types of defects.

[0014] A wafer inspection method, using the wafer inspection system described above, includes the following steps: S1. The optical imaging system scans the wafer, and the control and processing unit performs initial screening based on the acquired wafer images, marking suspected defect points. Specifically, this includes: S11. Under the control of the control and processing unit, the optical imaging system performs a full-area scan of the wafer according to a preset path.

[0015] S12, The data acquisition module synchronously receives scanned image data and position feedback data from the multi-axis motion platform.

[0016] S13. The feature extraction module performs initial screening based on grayscale and contrast features of the scanned image, marks suspected defect points, and records the coordinates of each suspected defect point in the global coordinate system.

[0017] S2, the control and processing unit performs multi-condition verification imaging and feature analysis on suspected defect points to distinguish different types of defects, specifically including: S21. The control and drive module controls the multi-axis motion platform according to the coordinates of the suspected defect point, positioning the suspected defect point on the wafer to the center of the field of view of the optical imaging system. Then, it simultaneously performs angle adjustment and illumination mode switching operations, and acquires multiple sets of imaging data. S211, The control drive module controls the multi-axis motion platform to rotate the wafer to the first angle, switches to the first illumination mode, and acquires the first image of the suspected defect point at the angle and under the illumination mode through the optical imaging system.

[0018] S212, The control drive module controls the multi-axis motion platform to rotate the wafer to a second angle different from the first angle, switches to the second illumination mode, and acquires a second image of the suspected defect point at this angle and under this illumination mode through the optical imaging system.

[0019] S22. The feature extraction module extracts motion adhesion quantization features, three-dimensional contour quantization features, multimodal optical response quantization features, and multi-view morphological stability quantization features of suspected defect points based on the first image, the second image, and the position data of the multi-axis motion platform.

[0020] S23. The model operation module inputs four types of quantitative features—motion adhesion quantification features, three-dimensional contour quantification features, multimodal optical response quantification features, and multi-view morphological stability quantification features—into the multi-evidence fusion network decision model. The model dynamically allocates the fusion weights of each feature, integrates multi-dimensional evidence, performs classification operations, and outputs the defect type classification results of suspected defect points.

[0021] The present invention is further configured such that the defect type includes type one defect and type two defect.

[0022] Type 1 defects refer to substantial defects present in the wafer itself and foreign matter defects attached to the wafer surface, including wafer cracks, dents, protrusions, and impurity attachments. Type 2 defects refer to interfering pseudo-defects not present in the wafer itself, including imaging noise, environmental dust interference, light reflection artifacts, and lens astigmatism interference.

[0023] The present invention is further configured such that: in step S21, angle settings and lighting mode combinations can be added according to defect detection requirements, and images of suspected defect points under different angles and lighting modes can be collected, and the timing of lighting mode switching is synchronized with the wafer angle adjustment action.

[0024] The present invention is further configured such that the motion adhesion quantification feature is obtained in the following manner: the feature extraction module calculates the motion vector of the suspected defect point between the first image and the second image, as well as the motion vector of different background regions in the same image. By comparing the difference in motion vectors and the consistency of motion direction between the suspected defect point and the background region, the motion adhesion quantification feature is obtained, thereby distinguishing between type one defects and type two defects.

[0025] The present invention is further configured such that the three-dimensional contour quantization feature is obtained through the following steps: At the first and second angles, the control drive module controls the multi-axis motion platform to move stepwise along the focusing direction, and the data acquisition module acquires a series of images of suspected defect points at different focusing positions.

[0026] The feature extraction module uses a 3D reconstruction algorithm to reconstruct the 3D height information of suspected defect points based on a series of images with different focal positions, and then extracts 3D contour quantification features that characterize the degree of protrusion and depression of the defect.

[0027] The present invention is further configured such that: the multimodal optical response quantization feature is based on image extraction under different illumination modes, specifically including characterizing the response characteristics of defects to different illuminations by quantifying the differences in parameters of defect regions under each illumination mode, thereby assisting in the determination of defect types.

[0028] The present invention is further configured such that: the multi-view morphological stability quantification feature is obtained by comparing the consistency of the contour shape, size, and edge features of suspected defect points in the first image and the second image; if the defect morphology remains stable under different angles, the probability of it being a type one defect is increased. If there are significant differences in morphology, the probability of it being a type two defect is increased.

[0029] The present invention is further configured such that: the specific configuration of the multi-evidence fusion network decision model is as follows: based on the initial screening information of suspected defect points, dynamic fusion weights are assigned to the motion adhesion quantification feature, three-dimensional contour quantification feature, multimodal optical response quantification feature and multi-view morphological stability quantification feature. The weight allocation coefficient is adaptively adjusted according to the difference of the initial screening features. After weighted fusion and integration of multi-dimensional evidence for each feature, the classification result of suspected defect points as type one defect and type two defect is output by the classifier, and the judgment confidence is output at the same time.

[0030] (1) This system can acquire multi-dimensional image data of the same suspected defect point under multi-angle and multi-modal illumination by controlling the wafer attitude of the multi-axis motion platform and switching the multi-illumination mode of the optical imaging system. Based on this, it extracts four types of quantitative features that characterize the physical nature of defects: motion adhesion, three-dimensional contour, multi-modal optical response, and multi-view morphological stability. It can distinguish between real defects and false defects from the root cause, and solves the limitation of traditional technology that relies only on shallow visual features for judgment.

[0031] (2) The multi-evidence fusion network designed in this system includes a weight allocation sub-network and a main classification sub-network. The weight allocation sub-network can dynamically allocate attention weights to the four core quantitative features based on the primary features in the initial screening stage, thus avoiding the rigidity problem of traditional fixed weight fusion. The main classification sub-network achieves multi-feature weighted fusion through a feature fusion layer that combines attention mechanism and convolution, and completes defect classification judgment.

[0032] (3) In the initial screening stage, the system performs pre-screening and spatial clustering, and combines precise primary features such as texture entropy and edge continuity to filter obvious false defects and merge duplicate markings in the same area to reduce the number of suspected defect points entering the fine inspection stage from the source. The four types of core features are prioritized according to computational cost and defect discrimination ability. A progressive verification and early termination strategy is adopted to extract only necessary features for suspected points to reduce invalid computation. Global path planning is performed on suspected points so that the multi-axis motion platform moves along the shortest path. At the same time, multi-illumination and multi-view images are acquired simultaneously under the same wafer posture to reduce the invalid time consumption of repeated positioning and mode switching. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of a wafer inspection system according to the present invention.

[0034] Figure 2 This is a flowchart of the wafer inspection system inspection method in this invention.

[0035] Figure 3 This is a flowchart illustrating the extraction process for four types of multi-dimensional quantitative features in this invention.

[0036] Figure 4 This is a schematic diagram of the structure of the multi-evidence fusion network in this invention.

[0037] Figure 5 This is a logic diagram of the illumination mode switching of the optical imaging system in this invention.

[0038] Figure 6 This is a flowchart illustrating the wafer inspection process in this invention. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0041] Please see Figures 1-6 The present invention provides the following technical solutions: Example 1, see Figure 1 A wafer inspection system includes: a multi-axis motion platform, an optical imaging system, and a control and processing unit.

[0042] Multi-axis motion platforms are used to carry wafers and drive them to rotate, translate, and move in the focusing direction.

[0043] The optical imaging system is used to image the wafer surface and supports switching between multiple illumination modes, including bright field illumination, dark field illumination, and oblique illumination.

[0044] The control and processing unit is electrically connected to the multi-axis motion platform and the optical imaging system, respectively. The control and processing unit includes a data acquisition module, a feature extraction module, a model calculation module and a control drive module.

[0045] The control and processing unit is configured to: control the multi-axis motion platform and the optical imaging system to work together, perform wafer full-domain scanning, suspected defect location and multi-condition imaging operations, extract multi-dimensional quantitative features of suspected defects, run the decision model to classify and determine suspected defects, and output the classification results.

[0046] The control and drive module is used to control the coordinated movement of the multi-axis motion platform and the optical imaging system to perform wafer full-area scanning, suspected defect location and multi-condition imaging operations.

[0047] The data acquisition module is used to synchronously receive image data acquired by the optical imaging system and real-time position feedback data from the multi-axis motion platform.

[0048] The feature extraction module is used to extract four types of multi-dimensional quantitative features of suspected defect points: motion adhesion, three-dimensional contour, multimodal optical response, and multi-view morphological stability.

[0049] The model computation module has a built-in multi-evidence fusion network decision model, which is used to classify and determine suspected defects based on the multi-dimensional quantitative features of the input, and output the defect type classification results.

[0050] The core of the multi-evidence fusion network decision model is to dynamically allocate the weights of each feature and integrate multi-dimensional evidence to distinguish between different types of defects.

[0051] This system is a closed-loop process of perception, decision-making, and execution. It upgrades wafer defect detection from traditional passive two-dimensional image recognition to active multi-dimensional physical judgment. Its working principle consists of three stages: Phase 1: Full-area scanning and initial screening, which is the perception phase.

[0052] The system first controls a multi-axis motion platform to drive the wafer along a predetermined path, while the optical imaging system performs high-speed, continuous surface imaging under a specific illumination mode.

[0053] The control and processing unit performs real-time analysis on the acquired images, quickly identifies and marks all suspected defect points based on preset primary features (such as contrast and abnormal shapes), and accurately records their location coordinates.

[0054] The second stage: review and feature extraction, which is the review and quantification stage.

[0055] For each suspected defect identified in the initial screening, the system enters a deep verification mode. The control and drive module instructs the motion platform to precisely backtrack and position the wafer to that point, and then coordinates with the control platform to rotate the wafer to different angles and simultaneously switch the illumination mode (such as bright field and dark field) to acquire a series of high-definition images of the same defect point under multi-angle and multi-modal lighting conditions.

[0056] Based on this set of multidimensional image data, the feature extraction module will calculate multidimensional quantitative features that describe the physical nature of the defects. For example, it will calculate motion adhesion through image displacement (to determine whether it is truly attached to the wafer), reconstruct the three-dimensional contour through focus scanning, analyze the optical response under different illuminations, and compare the morphological stability at different angles.

[0057] like Figure 5 As shown in the figure, this diagram illustrates the illumination mode switching and optical path output of the optical imaging system, clearly demonstrating the hardware triggering and optical signal path for the three modes: bright field, oblique illumination, and dark field. The first step is to control the flow of commands.

[0058] Control driver module to lighting mode selector: The control driver module sends a mode selection command, and the lighting mode selector triggers the corresponding light source accordingly.

[0059] The lighting mode selector triggers different lighting through three branch instructions: Select brightfield to brightfield light source: Activate brightfield lighting mode.

[0060] Select oblique to oblique light source: Activate oblique lighting mode.

[0061] Select Dark Field to Ring Dark Field Light Source: Activate Dark Field Lighting Mode.

[0062] Secondly, there are the details of optical path transmission.

[0063] First transmission path: Bright field illumination optical path.

[0064] The sequence of light sources is as follows: bright field light source, condenser lens, semi-transparent mirror, objective lens, and wafer. The light emitted by the bright field light source is focused by the condenser lens, reflected by the semi-transparent mirror, and then focused onto the wafer surface by the objective lens.

[0065] The process involves sequential transmission of light through the wafer, objective lens, semi-transparent mirror, and camera: reflected light from the wafer surface is collected by the objective lens, passes through the semi-transparent mirror, and enters the camera to form an image.

[0066] Second transmission path: oblique illumination optical path.

[0067] The oblique light source, angle adjustment mirror, and wafer are transmitted in sequence: the light emitted by the oblique light source is adjusted at the incident angle by the angle adjustment mirror and then directly incident on the wafer surface at a specific angle.

[0068] The wafer, objective lens, semi-transparent mirror, and camera transmit light sequentially: scattered or reflected light from the wafer surface is collected by the objective lens and enters the camera along the same imaging path as the bright field.

[0069] Third transmission path: Dark field illumination optical path.

[0070] Ring-shaped dark field light source and wafer are transmitted sequentially: the light emitted by the ring-shaped dark field light source is incident on the wafer surface from the side in a ring manner. The background light does not enter the objective lens, and only the defect scattered light is collected.

[0071] The wafer, objective lens, semi-transparent mirror, and camera are transmitted sequentially: the scattered light from the defects is collected by the objective lens and enters the camera along the same imaging path.

[0072] Finally, the output results are linked with the illumination mode selector, meaning that the imaging data under different illumination modes will eventually be transmitted to the control and processing unit, where the detection results will be output after feature extraction and model calculation.

[0073] The third stage: intelligent fusion and classification judgment, which is the decision-making stage.

[0074] The extracted multi-dimensional feature data is fed into the decision model built into the model computation module. This model does not simply rely on a single feature, but dynamically integrates evidence from different physical dimensions and intelligently allocates weights.

[0075] After comprehensive analysis, the model finally outputs classification results, clearly distinguishing whether the point is a real defect (such as particles or scratches) or a false defect (such as noise or artifacts) introduced by the system or environment, and may further subdivide the real defects.

[0076] By actively manipulating wafer orientation and lighting conditions, the system can acquire physical characteristics such as three-dimensional morphology and adhesion that go beyond a single two-dimensional image, thereby fundamentally distinguishing between genuine and fake defects and significantly reducing the false positive rate that is difficult to avoid with traditional methods.

[0077] The system outputs not only the presence or absence of defects, but also multi-dimensional data such as the quantitative three-dimensional dimensions and optical properties of the defects, which can be used to trace the root cause of the defects.

[0078] Example 2, see reference Figure 6 Based on the above system setup, the system can perform detection through an operational process, which includes four steps: full-domain scanning and initial screening, multi-condition verification and intelligent classification, fusion decision-making, and experimental verification. The specific implementation process is as follows: After the wafer inspection begins, initialization settings are required. After the inspection process is triggered, the system completes hardware self-test, algorithm parameter loading, and communication link initialization to ensure that the equipment is in a ready state. The optical imaging system and multi-axis motion platform are calibrated to establish a global spatial reference to ensure the accuracy of subsequent positioning and imaging. The wafer to be inspected is loaded onto the multi-axis motion platform and its position is aligned so that the center of the wafer coincides with the center of the optical field of view.

[0079] After initialization, proceed directly to the preliminary screening stage.

[0080] Part 1: Full-area scanning and initial screening.

[0081] The system control platform drives wafer scanning, the optical system performs continuous imaging, and the feature extraction module performs real-time image differential processing. Any anomalies exceeding a low threshold are marked as suspected defect points, and their physical coordinates are recorded. The above operations can improve the defect capture rate.

[0082] After the initial scan is completed, it is necessary to determine whether a full-area scan has been completed. If not, a second scan and screening are required to improve the completeness of the detection.

[0083] Part Two: Multi-Condition Validation and Intelligent Classification.

[0084] First, positioning and imaging are performed. The multi-axis motion platform moves the suspected defect point to the center of the field of view and executes a collaborative imaging sequence of dark field, oblique illumination, and bright field.

[0085] Next, feature extraction is performed, including: motion adhesion quantification features, three-dimensional contour quantification features, optical response features, and morphological stability features.

[0086] Motion adhesion quantification features are used to determine whether suspected defects are actually adhered to the wafer surface. The extraction process includes the following steps: See Figure 3 Step 1: Synchronous acquisition of multi-view images.

[0087] The multi-axis motion platform is controlled to rotate the wafer to different angles. Under the same illumination mode, the optical imaging system acquires a first image of the suspected defect point at the first angle and a second image at the second angle.

[0088] Step 2: Calculate motion vectors.

[0089] Using a subpixel image matching algorithm, calculate respectively: The actual motion vector of the suspected defect point from the first image to the second image; the background motion vector of multiple background regions in the first image from the first image to the second image.

[0090] It should be noted that subpixel image matching algorithm is a digital image processing technology. Its core goal is to transcend the limitations of the physical pixel grid and measure the minute displacement of feature points or regions between images with an accuracy higher than one pixel. In other words, the subpixel image matching algorithm is used to calculate the actual motion vector and the background motion vector.

[0091] Step 3: Global background motion modeling.

[0092] Based on the background motion vector, the expected background motion vector generated by wafer movement at the location of suspected defect points is predicted by fitting a global motion transformation model.

[0093] Step 4: Quantification of motion consistency features.

[0094] By comparing the actual motion vector with the background expected motion vector, at least one motion consistency quantification index is calculated, including: motion vector difference magnitude; motion direction angle deviation; motion residual norm; and image matching confidence.

[0095] Step 5: Feature Output.

[0096] Combining one or more of the above indicators constitutes the quantitative characteristics of motion adhesion, which serves as the basis for distinguishing between real defects and non-adhesion artifacts.

[0097] Three-dimensional contour quantization features are used to characterize the physical height and three-dimensional shape of suspected defects. The extraction process is based on focused scanning and three-dimensional reconstruction technology, and includes the following steps: Step 1: Focused scanning image sequence acquisition.

[0098] For suspected defect points, the multi-axis motion platform is controlled to move in steps along the focusing direction of the optical imaging system with a preset step size; at the first angle and the second angle, a series of images at different focusing positions are acquired by the optical imaging system to form a focusing image sequence at two angles.

[0099] Step 2: Sharpness evaluation and height restoration.

[0100] For each focused image sequence from each angle, perform the following processing: For each image, calculate the sharpness rating of the local region centered on the suspected defect point.

[0101] A sharpness curve for each pixel is constructed along the focusing direction, and the focusing position corresponding to its peak value is located by curve interpolation.

[0102] The peak focus position is converted into a relative height value to generate a two-dimensional height map at that angle.

[0103] Step 3: Extraction of 3D contour features.

[0104] Based on the two-dimensional height map, quantitative features for characterizing the three-dimensional morphology of the defect are extracted, including at least one of the following: maximum height value and maximum depth value; defect volume; defect surface area; surface roughness or inclination.

[0105] Step 4: Multi-angle feature fusion.

[0106] The three-dimensional contour features extracted from the first angle and the second angle are fused or compared to obtain the final three-dimensional contour quantization feature vector, which is used for subsequent classification and determination.

[0107] The specific steps for extracting optical response features are as follows: Step 1: Image acquisition in illumination mode.

[0108] The optical imaging system is controlled to sequentially switch between at least two different illumination modes while keeping the wafer angle constant, and to acquire multiple illumination images of the same suspected defect point at the same angle. The illumination modes include at least two of bright field illumination, dark field illumination and oblique illumination.

[0109] Step 2: Defect region extraction and registration.

[0110] For the acquired multiple illumination images, defect regions are segmented separately, and image registration is used to align the defect regions in each image in spatial coordinates.

[0111] Step 3: Calculation of multimodal optical parameters.

[0112] For each registered defect region, a set of image optical parameters are calculated, including at least one of the following: region average gray level, gray level standard deviation, texture energy, and edge gradient intensity.

[0113] Step 4: Quantification of response difference characteristics.

[0114] By comparing the optical parameters of the same defect area under different lighting modes, their relative differences or ratios are calculated to form a multimodal optical response quantization feature vector.

[0115] Differences or ratios include: the average grayscale ratio of bright field to dark field images; the difference in contrast of defect areas under different lighting modes; and the standard deviation of texture features between different lighting modes.

[0116] Step 5: Feature Output.

[0117] The multimodal optical response quantization feature vector is output to the subsequent classification model for auxiliary determination of defect type.

[0118] The specific steps for extracting morphological stability features are as follows: Step 1: Synchronous acquisition of multi-view images.

[0119] The multi-axis motion platform is controlled to sequentially locate the suspected defect points to the first and second angles; while maintaining the same illumination mode, the corresponding first and second images are acquired through the optical imaging system.

[0120] Step 2: Defect morphology parameter extraction.

[0121] From the first image and the second image respectively, extract the morphological description parameters of the suspected defect points. The parameters include at least one of the following: contour shape parameters; geometric size parameters: including area, equivalent diameter, and perimeter; edge feature parameters: including edge gradient direction histogram and edge curvature.

[0122] Step 3: Quantification of morphological consistency features.

[0123] By comparing the morphological description parameters extracted from the first image and the second image, their consistency measure is calculated, and a stability quantification feature is generated. The measure includes: the relative error or rate of change of similar parameters; the correlation coefficient or similarity score of similar parameters.

[0124] Step 4: Feature Output.

[0125] The stability quantification features are output as multi-view morphological stability quantification features to assist in determining the defect type.

[0126] See Figure 4 Part Three: Intelligent Integrated Decision Making.

[0127] The above four types of feature vectors are input into the multi-evidence fusion network, including 6-dimensional motion adhesion features, 8-dimensional three-dimensional contour features, 8-dimensional multimodal optical response features, and 6-dimensional multi-view morphological stability features.

[0128] The network first dynamically assigns weights to each feature based on the initial screening information, then performs weighted fusion and classification, and finally outputs the defect type result and the judgment confidence level. The maximum value of the two types of defect probabilities is taken from the judgment result.

[0129] Step 1: Detailed structure of the multi-evidence fusion network.

[0130] The weight allocation subnetwork consists of an input layer, a hidden layer, and an output layer, which are adapted to the dynamic weight output of the initial screening features.

[0131] The input layer has an 8-dimensional dimension, corresponding to the 8 primary features extracted in the initial screening stage. The primary features include average gray value, gray variance, edge gradient strength, shape complexity, area ratio, roundness, texture entropy, and positional deviation.

[0132] The hidden layers consist of two fully connected layers, with 64 neurons in the first layer and 32 neurons in the second layer. The activation function used is LeakyReLU (with a slope of 0.01) to address the vanishing gradient problem of ReLU in the negative region.

[0133] It should be noted that a single hidden layer may not be able to adequately represent the complex nonlinear relationships between multidimensional features; while an overly deep network (such as three or more layers) is prone to overfitting on small datasets and increases unnecessary computational burden. The two-layer structure achieves a good balance between model complexity and generalization ability, which is sufficient to complete the nonlinear mapping from multidimensional features to defect types.

[0134] The architecture adopts a progressively decreasing layer structure. The first layer, with its larger width (64 neurons), aims to fully decode and extract higher-order features from the input, pre-weighted, multi-dimensional quantized features (motion adhesion, 3D contour, optical response, and morphological stability). The second layer is reduced to 32 neurons, which serves to compress and refine the features. While retaining key discriminative information, the data dimensionality is gradually reduced to prepare for the final classification output layer.

[0135] If the standard ReLU function is used, its output and gradient will both be zero when the input is negative, which can easily lead to the problem of neuron death, that is, some neurons will permanently fail, which will damage the network capacity and training stability.

[0136] The output layer has a 4-dimensional dimension, corresponding to the attention weights of four types of quantized features, where W1 is the motion adhesion weight, W2 is the 3D contour weight, W3 is the multimodal optical response weight, and W4 is the multi-view morphological stability weight. These weights are normalized to [0,1] using the Sigmoid function and satisfy W1+W2+W3+W4=1.

[0137] It should be noted that the Sigmoid function is a smooth S-shaped curve function that maps any real number to the interval [0,1]. In this embodiment, the attention weights of the four types of quantized features can be normalized to [0,1] using the Sigmoid function.

[0138] The main classification sub-network includes an input layer, a feature fusion layer, a hidden layer, and an output layer, which is suitable for multi-feature fusion and classification.

[0139] The input layer has a dimension of 28, which is a concatenated vector of four types of quantized features.

[0140] The feature fusion layer combines attention mechanism with convolution, and sets a 1×1 convolution kernel (32 kernels) to perform dimensionality compression and information fusion on the weighted input features.

[0141] The hidden layer is a single fully connected layer with 48 neurons.

[0142] The output layer is 2-dimensional, corresponding to the probability distribution of type 1 and type 2 defects. The output uses the Softmax function, and the classification threshold is set to 0.5 (a probability ≥ 0.5 is judged as the corresponding defect type).

[0143] It should be noted that the Softmax function is a core function used for multi-class classification in neural networks and machine learning. It transforms a set of arbitrary real-valued scores into a probability distribution that sums to 1, thus intuitively representing the probability that an input sample belongs to each possible class.

[0144] Step 2: Network training process.

[0145] The training dataset contains 12,000 labeled samples, including 6,000 samples of type 1 defects (1,500 each of cracks, dents, protrusions, and impurities) and 6,000 samples of type 2 defects (1,500 each of imaging noise, environmental dust, lighting artifacts, and lens astigmatism).

[0146] All samples were labeled using a scanning electron microscope (resolution 0.1 μm), and wafer parameters (material, size, process stage) were recorded. Data augmentation was performed by rotation (±5°), scaling (0.9-1.1 times), and Gaussian noise addition (variance 0.005).

[0147] The optimizer for training parameters is AdamW, with an initial learning rate of 1e-4 and a weight decay coefficient of 1e-5; the loss function is weighted cross-entropy loss (1.2 for type 1 defects and 0.8 for type 2 defects); the total number of iterations is 800, with the learning rate linearly decaying to 5e-5 in the first 200 iterations and remaining unchanged in the next 600 iterations; an early stopping mechanism is set (the iteration stops if the accuracy on the validation set does not improve for 30 consecutive iterations).

[0148] It should be noted that AdamW is an optimization algorithm used to train neural networks. By modifying the implementation of weight decay, it significantly improves the generalization ability and final performance of the model.

[0149] The validation set (3000 samples) for training results showed an accuracy of ≥98.2%, a recall of ≥97.8%, and a precision of ≥97.5%. Typical sample weighting results: W2 for suspected impurity attachment defects was ≥0.4, and W3 for suspected lighting artifact defects was ≥0.35.

[0150] Step 3: Verification of network innovation points.

[0151] Three sets of comparative models (fixed weight fusion model, single feature classification model, and traditional convolutional neural network model) were set up and tested on the same validation set: the false positive rate of the network of the present invention was reduced by 42% compared with the fixed weight model, 68% compared with the single feature model, and 53% compared with the traditional convolutional neural network model, proving the innovation and superiority of dynamic weight allocation.

[0152] By reviewing and extracting features, and then using fusion decision-making, suspected defects are examined and judged, and a detection report is generated.

[0153] Example 3: In the detection process of Example 2, a complete traversal of four types of features (motion adhesion, three-dimensional contour, multimodal optical response, and multi-view morphological stability) is performed on all suspected defect points. While this ensures the accuracy of the judgment, the rigid process leads to redundant consumption of calculations and mechanical actions, leaving room for efficiency improvement. This example adopts a strategy of hierarchical screening, dynamic adaptation, and hardware-software collaboration, which significantly improves the defect detection efficiency without sacrificing detection accuracy.

[0154] First, pre-screening in the initial screening stage reduces potential defects at the source.

[0155] During the full-domain scanning stage, precise primary features such as texture entropy and edge continuity are combined to filter out obvious imaging noise, lighting artifacts, and other false defects in advance, thereby reducing the number of suspected points entering the re-examination stage.

[0156] Spatial clustering is performed on suspected points identified in the initial screening to merge duplicate markers in the same region, thus avoiding multiple verifications of artifacts at the same location.

[0157] Second, feature-based hierarchical triggering dynamically selects features based on the degree of differentiation between components and features.

[0158] The four types of features are prioritized according to computational cost and defect discrimination ability. For each suspected point, a progressive verification and early termination strategy is adopted, which eliminates the need to traverse all features.

[0159] Table 1. Priority and Applicable Scenarios for Defect Detection Feature Hierarchy Execution logic: For each suspected point, first extract the features of priority 1. If it can be clearly determined to be a false defect (such as the optical response feature value exceeding the threshold), then terminate the subsequent feature extraction; if it cannot be clearly determined, then extract the features of higher priority in turn until a judgment result with sufficient confidence is obtained.

[0160] Third, optimize mechanical movements and data acquisition to reduce unnecessary time wastage.

[0161] Global path planning is performed on the locations of all suspected points to enable the multi-axis motion platform to move along the shortest path, avoiding repeated positioning and back-and-forth movement, and reducing the platform's movement time.

[0162] Under the same wafer orientation, images with multiple illumination modes and multiple viewing angles are acquired simultaneously, avoiding repeated wafer rotation and illumination mode switching, thus shortening the image acquisition time for single points.

[0163] Fourth, the model and algorithm are well-matched, supporting incremental decision-making.

[0164] The original multi-evidence fusion network structure was adjusted to support incremental feature input and real-time judgment. That is, after each type of feature is extracted, it is input into the model for a preliminary judgment. If the confidence level is ≥95%, the subsequent feature extraction is terminated in advance without waiting for all four types of features to be input.

[0165] In the initial screening stage, suspected points are pre-classified (e.g., suspected particles, suspected scratches, suspected artifacts). For different pre-classification results, only the corresponding features are triggered. For example: Pre-classified as a suspected lighting artifact, it can be determined simply by using multimodal optical response features.

[0166] The pre-classified defect is suspected to be a protrusion, and verification can be completed with only three-dimensional contour features.

[0167] The above settings can reduce the detection time of a single wafer, and the hierarchical screening and dynamic feature selection avoid over-reliance on a single feature, while the early termination mechanism does not sacrifice key discrimination information.

[0168] To verify the technical effect of the present invention, a comparative experiment was conducted, and the specific operation is as follows: Step 1: Set up the basic experimental conditions.

[0169] Experimental subjects: 30 8-inch silicon wafers (P-type single crystal silicon, thickness 775μm, surface roughness Ra≤0.1nm), of which 15 wafers contain real defects (each wafer contains 2-3 cracks, dents, protrusions, and impurity attachments, with defect size 0.5-10μm), and 15 wafers are interference wafers without real defects (artificially added environmental dust and lighting artifacts). Comparison scheme: Traditional detection scheme (single bright field illumination and two-dimensional image recognition, imaging resolution 1μm, feature extraction only includes grayscale and contrast); Experimental equipment: The system of this invention uses the same brand of optical imaging components (lens focal length 50mm, CCD pixel 12 million) and multi-axis motion platform (repeat positioning accuracy ±0.1μm) as the traditional solution; Experimental environment: temperature 23±2℃, humidity 45±5%.

[0170] Detection indicators: defect detection accuracy, false positive rate, single wafer detection time, defect quantification accuracy, and defect type differentiation accuracy.

[0171] Step 2: Quantify the experimental results.

[0172] Table 2 Comparison of Quantitative Experiment Results Step 3: Verify the stability of the effect.

[0173] Repeatability test: The same batch of 30 wafers was tested 5 times. The accuracy fluctuation was ≤ ±0.3% and the false positive rate fluctuation was ≤ ±0.2%, which proved the stability of the system.

[0174] Compatibility test: The test was performed using a 12-inch silicon wafer (725μm thick) and a 6-inch silicon carbide wafer (350μm thick). The accuracy rate was ≥98.5% and the false positive rate was ≤2.0%, which proves the compatibility of the solution.

[0175] Example 4, see reference Figure 2 A wafer inspection method, using the aforementioned wafer inspection system, includes the following steps: S1. The optical imaging system scans the wafer, and the control and processing unit performs initial screening based on the acquired wafer images, marking suspected defect points. Specifically, this includes: S11. Under the control of the control and processing unit, the optical imaging system performs a full-area scan of the wafer according to a preset path.

[0176] S12, The data acquisition module synchronously receives scanned image data and position feedback data from the multi-axis motion platform.

[0177] S13. The feature extraction module performs initial screening based on grayscale and contrast features of the scanned image, marks suspected defect points, and records the coordinates of each suspected defect point in the global coordinate system.

[0178] S2, the control and processing unit performs multi-condition verification imaging and feature analysis on suspected defect points to distinguish different types of defects, specifically including: S21. The control and drive module controls the multi-axis motion platform according to the coordinates of the suspected defect point, positioning the suspected defect point on the wafer to the center of the field of view of the optical imaging system. Then, it simultaneously performs angle adjustment and illumination mode switching operations, and acquires multiple sets of imaging data. S211, The control drive module controls the multi-axis motion platform to rotate the wafer to the first angle, switches to the first illumination mode, and acquires the first image of the suspected defect point at the angle and under the illumination mode through the optical imaging system.

[0179] S212, The control drive module controls the multi-axis motion platform to rotate the wafer to a second angle different from the first angle, switches to the second illumination mode, and acquires a second image of the suspected defect point at this angle and under this illumination mode through the optical imaging system.

[0180] S22. The feature extraction module extracts motion adhesion quantization features, three-dimensional contour quantization features, multimodal optical response quantization features, and multi-view morphological stability quantization features of suspected defect points based on the first image, the second image, and the position data of the multi-axis motion platform.

[0181] S23. The model operation module inputs four types of quantitative features—motion adhesion quantification features, three-dimensional contour quantification features, multimodal optical response quantification features, and multi-view morphological stability quantification features—into the multi-evidence fusion network decision model. The model dynamically allocates the fusion weights of each feature, integrates multi-dimensional evidence, performs classification operations, and outputs the defect type classification results of suspected defect points.

[0182] Defect types include Type 1 defects and Type 2 defects. Type 1 defects are substantial defects present in the wafer itself and foreign matter defects attached to the wafer surface, including wafer cracks, dents, protrusions, and impurity attachments. Type 2 defects are interfering pseudo-defects not present in the wafer itself, including imaging noise, environmental dust interference, light reflection artifacts, and lens astigmatism interference.

[0183] In step S21, angle settings and lighting mode combinations can be added according to defect detection requirements to collect images of suspected defect points at different angles and under different lighting modes, and the timing of lighting mode switching is synchronized with the wafer angle adjustment action.

[0184] The motion adhesion quantification features are obtained as follows: the feature extraction module calculates the motion vector of the suspected defect point between the first image and the second image, as well as the motion vector of different background regions in the same image. By comparing the difference in motion vectors and the consistency of motion direction between the suspected defect point and the background region, the motion adhesion quantification features are obtained, thereby distinguishing between type one defects and type two defects.

[0185] The 3D contour quantization features are obtained through the following steps: At the first and second angles, the control drive module controls the multi-axis motion platform to move stepwise along the focusing direction, and the data acquisition module acquires a series of images of suspected defect points at different focusing positions.

[0186] The feature extraction module uses a 3D reconstruction algorithm to reconstruct the 3D height information of suspected defect points based on a series of images with different focal positions, and then extracts 3D contour quantification features that characterize the degree of protrusion and depression of the defect.

[0187] Multimodal optical response quantization features are based on image extraction under different illumination modes. Specifically, they include characterizing the response characteristics of defects to different illuminations by quantifying the differences in parameters of defect regions under various illumination modes, thus assisting in defect type determination.

[0188] The multi-view morphological stability quantification feature is obtained by comparing the consistency of the contour shape, size, and edge features of suspected defect points in the first and second images. If the defect morphology remains stable under different angles, the probability of it being classified as a type 1 defect increases. If there are significant differences in morphology, the probability of it being classified as a type 2 defect increases.

[0189] The specific configuration of the multi-evidence fusion network decision model is as follows: Based on the initial screening information of suspected defect points, dynamic fusion weights are assigned to the motion adhesion quantification feature, three-dimensional contour quantification feature, multimodal optical response quantification feature, and multi-view morphological stability quantification feature. The weight allocation coefficients are adaptively adjusted according to the differences in the initial screening features. After weighted fusion and integration of multi-dimensional evidence, the classifier outputs the classification results of suspected defect points as type 1 defects and type 2 defects, and outputs the judgment confidence level.

[0190] Obviously, the embodiments described above are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A wafer inspection system, characterized by: The application relates to a multi-axis motion platform, an optical imaging system, a control and processing unit, and a method for detecting defects on a wafer. The multi-axis motion platform is used for carrying a wafer and driving the wafer to rotate, translate and move in a focusing direction. The optical imaging system is used for imaging the wafer surface and supports switching of multiple illumination modes, including bright field illumination, dark field illumination and oblique illumination. The control and processing unit is electrically connected with the multi-axis motion platform and the optical imaging system and comprises a data acquisition module, a feature extraction module, a model operation module and a control driving module. The control and processing unit is configured to control the multi-axis motion platform and the optical imaging system to cooperatively perform wafer global scanning, suspected defect point positioning and multi-condition imaging operation, extract multi-dimensional quantitative features of the suspected defect points, run a decision model to classify and determine the suspected defect points and output a classification result.

2. The wafer inspection system of claim 1, wherein: The control driving module is used for controlling the multi-axis motion platform and the optical imaging system to cooperatively perform wafer global scanning, suspected defect point positioning and multi-condition imaging operation. The data acquisition module is used for synchronously receiving image data collected by the optical imaging system and real-time position feedback data of the multi-axis motion platform. The feature extraction module is used for extracting four types of multi-dimensional quantitative features of the suspected defect points, including motion adhesion, three-dimensional profile, multi-modal optical response and multi-view shape stability. The model operation module internally comprises a multi-evidence fusion network decision model which is used for classifying and determining the suspected defect points based on the input multi-dimensional quantitative features and outputting a defect type classification result. The multi-evidence fusion network decision model is used for dynamically allocating feature weights and integrating multi-dimensional evidence to realize defect type differentiation.

3. A method for detecting a wafer using a wafer detection system according to claim 1 or 2, wherein: The method comprises the following steps. S1. The optical imaging system scans the wafer, and the control and processing unit performs preliminary screening based on the collected wafer image and marks suspected defect points, specifically comprising the following steps. S11. The optical imaging system performs global scanning on the wafer according to a preset path under the control of the control and processing unit. S12. The data acquisition module synchronously receives scanning image data and position feedback data of the multi-axis motion platform. S13. The feature extraction module performs gray scale and contrast feature preliminary screening based on the scanning image, marks suspected defect points and records coordinates of the suspected defect points in a global coordinate system. S2. The control and processing unit performs multi-condition verification imaging and feature analysis on the suspected defect points to distinguish different types of defects, specifically comprising the following steps. S21. The control driving module controls the multi-axis motion platform to position the suspected defect point on the wafer to the center of the field of view of the optical imaging system according to the coordinates of the suspected defect point, then synchronously performs angle adjustment and illumination mode switching operation and collects multiple sets of imaging data. S211. The control driving module controls the multi-axis motion platform to drive the wafer to rotate to a first angle, switches to a first illumination mode and obtains a first image of the suspected defect point at the angle and under the illumination mode through the optical imaging system. S212. The control driving module controls the multi-axis motion platform to drive the wafer to rotate to a second angle, switches to a second illumination mode and obtains a second image of the suspected defect point at the angle and under the illumination mode through the optical imaging system. S213. The control driving module controls the multi-axis motion platform to drive the wafer to rotate to a third angle, switches to a third illumination mode and obtains a third image of the suspected defect point at the angle and under the illumination mode through the optical imaging system. S214. The control driving module controls the multi-axis motion platform to drive the wafer to rotate to a fourth angle, switches to a fourth illumination mode and obtains a fourth image of the suspected defect point at the angle and under the illumination mode through the optical imaging system. S22. The model operation module receives the four sets of imaging data of the suspected defect point and runs the multi-evidence fusion network decision model to classify and determine the suspected defect point and output a defect type classification result. S23. The control and processing unit marks the classification result on the wafer image. S212, the control driving module controls the multi-axis motion platform to rotate the wafer to a second angle different from the first angle, switches to a second illumination mode, and acquires a second image of the suspected defect point under the angle and the illumination mode through the optical imaging system; S22, the feature extraction module extracts the motion adhesion quantitative feature, the three-dimensional profile quantitative feature, the multi-modal optical response quantitative feature, and the multi-view shape stability quantitative feature of the suspected defect point based on the first image, the second image, and the multi-axis motion platform position data; S23, the model operation module inputs the four types of quantitative features into a multi-evidence fusion network decision model, dynamically allocates a feature fusion weight by the model, integrates multi-dimensional evidence, and then performs classification operation to output a defect type classification result of the suspected defect point.

4. The method of claim 3, wherein: The defect types include a type one defect and a type two defect; The type one defect is a substantial defect existing in the wafer body and a foreign matter defect attached to the wafer surface, including wafer cracks, depressions, protrusions, and impurity adhesion. The type two defect is an interference false defect not existing in the wafer itself, including imaging noise, environmental dust interference, light reflection artifact, and lens astigmatism interference.

5. The method of claim 3, wherein: In step S21, the angle range and the illumination mode combination can be increased according to the defect detection requirement, the suspected defect point images under different angles and different illumination modes are collected, and the illumination mode switching time is synchronized with the wafer angle adjustment action.

6. The method of claim 3, wherein: The motion adhesion quantitative feature is obtained by comparing the motion vector difference and the motion direction consistency of the suspected defect point and the background region, and quantifying the motion adhesion quantitative feature to distinguish the type one defect and the type two defect.

7. The method of claim 3, wherein: The three-dimensional profile quantitative feature is obtained by the following steps: Under the first angle and the second angle, the control driving module controls the multi-axis motion platform to step along the focusing direction, and the data acquisition module collects a series of suspected defect point images at different focusing positions; The feature extraction module reconstructs the three-dimensional height information of the suspected defect point based on the series of images at different focusing positions by using a three-dimensional reconstruction algorithm, and then extracts the three-dimensional profile quantitative feature representing the defect protrusion and depression degree.

8. The method of claim 3, wherein: The multi-modal optical response quantitative feature is extracted based on the images under different illumination modes, which specifically includes quantifying the parameter difference of the defect region under each illumination mode to represent the response characteristics of the defect to different light, and assisting the defect type determination.

9. The method of claim 5, wherein: The multi-view shape stability quantitative feature is obtained by comparing the consistency of the contour shape, size, and edge feature of the suspected defect point in the first image and the second image. If the defect shape remains stable under different angles, the probability of being the type one defect is improved; If there is a significant difference in the shape, the probability of being the type two defect is improved.

10. The method of claim 5, wherein: The specific configuration of the multi-evidence fusion network decision model is: based on the screening information of the suspected defect point, dynamic fusion weights are assigned to the motion adhesion quantitative feature, the three-dimensional contour quantitative feature, the multi-modal optical response quantitative feature and the multi-view shape stability quantitative feature, the weight distribution coefficient is adaptively adjusted according to the difference of the screening features, and then the features are weighted and fused, the multi-dimensional evidences are integrated, and the classification results of the suspected defect point as type one defect and type two defect are output through the classifier, and the determination confidence is output.

Images