Wafer surface defect detection system based on multi-channel sensor

Abstract

According to the wafer surface defect detection system based on the multi-channel sensor, multi-dimensional information of the to-be-detected wafer is collected through the data collection module, multiple types of defects can be detected, more complete information is provided for production quality control, and the chip performance problem and the yield reduction risk caused by the fact that the defects are not detected out are effectively reduced; the intelligent decision-making module can automatically learn and identify defect features, accurately judge defect types and severity, and reduce the influence of manual intervention and subjective factors. The data splicing and weighted fusion technology enables the model to make full use of multi-sensor information to adapt to a complex detection scene; the result storage and report generation module records detection data in detail and generates a report, the report contains defect position, type, size, image or signal feature and influence analysis, the detection result is visually displayed, a user is helped to rapidly know the wafer quality condition, and a powerful basis is provided for subsequent decision making.

Description

Technical Field

[0001] The present application relates to the field of wafer defect detection, and in particular to a wafer surface defect detection system based on a multi-channel sensor. Background Art

[0002] During the manufacturing process of wafers, due to various complex factors, various types of defects are prone to occur, which will affect the performance and yield of the chip. Common wafer surface defects include particle contamination, scratches and oxide layer defects; particle contamination mainly comes from dust in the production environment, debris generated by equipment wear, etc.; these tiny particles adhere to the surface of the wafer, and in subsequent process steps such as photolithography and etching, they will cause pattern transfer deviations, affecting the circuit structure and performance of the chip. Scratches are mostly caused by friction or collision with other objects during the transportation and processing of wafers. Scratches may destroy the circuit structure on the surface of the wafer, especially for advanced processes with increasingly smaller line widths. Even a slight scratch may cause the chip to short or break, thereby reducing the yield of the chip. In wafer manufacturing, the oxide layer is an important structure for insulation and protection. However, during the oxidation process, oxide layer defects such as uneven oxide layer thickness, pinholes, and cracks may occur. These defects will affect the insulation performance and reliability of the chip.

[0003] In the prior art, defect detection for wafers is mainly focused on determining whether the wafer has defects, but lacks detection of further detailed information on the wafer defects. Summary of the invention

[0004] In order to solve the problem that the wafer defect type and wafer defect degree cannot be detected in the prior art, a wafer surface defect detection system based on a multi-channel sensor is proposed;

[0005] Wafer surface defect detection system based on multi-channel sensors, including data acquisition module, defect feature extraction module and intelligent decision-making module;

[0006] The data acquisition module is used to collect optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested, and send the optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested to the defect feature extraction module;

[0007] The defect feature extraction module is used to extract features of the received optical image information of the wafer to be tested, the ultrasonic echo signal of the wafer to be tested, and the electromagnetic response signal of the wafer to be tested, respectively, to obtain an optical image feature vector, an ultrasonic feature vector, and an electromagnetic feature vector, and send the optical image feature vector, the ultrasonic feature vector, and the electromagnetic feature vector to the intelligent decision module;

[0008] The intelligent decision-making module is used to perform a splicing operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain splicing data, and is also used to perform a fusion operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain fusion data, and input the spliced ​​data and the fusion data into the wafer defect detection convolutional neural network to obtain the defect information of the wafer to be tested; the defect information of the wafer to be tested includes whether the wafer has defects, the type of wafer defects and the degree of wafer defects.

[0009] Beneficial Effects

[0010] The wafer surface defect detection system based on multi-channel sensors proposed in this application can collect multi-dimensional information of the wafer to be tested through the data acquisition module, detect various types of defects, provide more complete information for production quality control, and effectively reduce the risk of chip performance problems and yield reduction caused by undetected defects; the intelligent decision-making module can automatically learn and identify defect characteristics, accurately determine the type and severity of defects, and reduce the influence of human intervention and subjective factors. Data splicing and weighted fusion technology enable the model to make full use of multi-sensor information and adapt to complex detection scenarios; the result storage and report generation module records the detection data in detail and generates a report. The report contains defect location, type, size, image or signal characteristics and impact analysis, and intuitively displays the detection results, helping users quickly understand the wafer quality status and provide a strong basis for subsequent decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a structural diagram of a wafer surface defect detection system based on a multi-channel sensor;

[0012] Figure 2 This is the layout diagram of the data acquisition module;

[0013] Figure 3 Flowchart for data fusion and data splicing; DETAILED DESCRIPTION

[0014] Specific implementation method 1: The following will be combined with the attached embodiment of the present invention Figure 1 To Attachment Figure 3 , illustrate this implementation mode, and clearly and completely describe the technical solutions in the embodiments of the present invention:

[0015] Wafer surface defect detection system based on multi-channel sensors, including data acquisition module, defect feature extraction module, and intelligent decision-making module;

[0016] The data acquisition module is used to collect optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested, and send the optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested to the defect feature extraction module;

[0017] The defect feature extraction module is used to extract features from the received optical image information, the ultrasonic echo signal of the wafer to be tested, and the electromagnetic response signal of the wafer to be tested, respectively, to obtain an optical image feature vector, an ultrasonic feature vector, and an electromagnetic feature vector, and send the optical image feature vector, the ultrasonic feature vector, and the electromagnetic feature vector to the intelligent decision module;

[0018] The intelligent decision-making module is used to perform a splicing operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain splicing data, and is also used to perform a fusion operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain fusion data, and input the spliced ​​data and the fusion data into the wafer defect detection convolutional neural network to obtain the defect information of the wafer to be tested; the defect information of the wafer to be tested includes whether the wafer has defects, the type of wafer defects and the degree of wafer defects.

[0019] Specifically, the wafer to be tested is inspected through its optical image information, ultrasonic echo signal and electromagnetic response signal, so that multiple types of defects can be detected simultaneously on the wafer to be tested, thereby providing further information about the defects. At the same time, the intelligent decision-making module uses the wafer defect detection convolutional neural network to reduce the influence of human intervention and subjective factors.

[0020] Furthermore, the data acquisition module includes an optical sensor, an ultrasonic sensor and an electromagnetic sensor; the optical sensor is used to collect optical image information of the wafer to be tested, and send the optical image information of the wafer to be tested to the defect feature extraction module; the ultrasonic sensor is used to collect ultrasonic echo signals of the wafer to be tested, and send the ultrasonic echo signals of the wafer to be tested to the defect feature extraction module; the electromagnetic sensor is used to collect electromagnetic response signals of the wafer to be tested, and send the collected electromagnetic response signals of the wafer to be tested to the defect feature extraction module.

[0021] Specifically, the optical sensor uses a high-frame-rate area array camera with at least 50 million pixels, with a frame rate of more than 200fps, and is paired with a ring-shaped LED light source that can provide uniform, shadow-free lighting effects to ensure that the subtle features of the wafer surface can be clearly captured during high-speed shooting, and is used to collect optical image information of the wafer surface;

[0022] Ultrasonic sensor, using a phased array ultrasonic transducer with a frequency range of 5-20MHz, realizes multi-angle ultrasonic detection of the wafer detection area through electronic scanning, obtains ultrasonic echo signals, and analyzes the defects inside the wafer and at a certain depth below the surface;

[0023] The electromagnetic sensor is designed based on the eddy current principle. The operating frequency can be adjusted in the range of 10kHz-1MHz. It can sensitively detect changes in electromagnetic properties caused by defects on the wafer surface and collect corresponding electromagnetic response signals.

[0024] Further, the received optical image information is subjected to Gaussian filtering to obtain a processed image, and the gradient intensity of the processed image is calculated by the Canny algorithm, the pixel occurrence frequency of the processed image is obtained by gray-level co-occurrence matrix calculation, and the regional brightness and regional contrast of the processed image are calculated;

[0025] Extracting features according to the gradient intensity of the processed image, the pixel occurrence frequency of the processed image, the regional brightness and the regional contrast of the processed image to obtain an optical image feature vector;

[0026] The method of performing Gaussian filtering on the received optical image information to obtain the processed image is:

[0027] The Gaussian filter formula is: Wherein, σ is the standard deviation of the Gaussian filter, x and y are the horizontal and vertical coordinates of the optical image information, respectively, and G(x,y) is the filter value of the pixel point (x,y).

[0028] The method for calculating the gradient strength of the processed image by the Canny algorithm is:

[0029] The gradient strength calculation formula is: Among them, G is the gradient strength, G x is the gradient of the processed image in the x direction, G y is the gradient of the processed image in the y direction;

[0030] The Canny algorithm uses non-maximum suppression technology to retain the most significant edge points in the image and remove other weaker edges to obtain accurate edge locations. Finally, through double threshold processing and edge connection strategy, the edge points are connected into complete edges.

[0031] In wafer surface defect detection, the Canny algorithm is particularly suitable for scratch detection. Scratches usually appear as sudden changes in the edge of the image, and their shape is relatively regular. Using the Canny algorithm to extract the edge can accurately identify the contour features of the scratch, which helps with subsequent defect classification and location.

[0032] The texture analysis method based on gray-level co-occurrence matrix is ​​used to capture the texture features in the image, which is particularly suitable for the identification of particle defects. The gray-level co-occurrence matrix extracts the texture features of the image by calculating the spatial relationship between the gray values ​​of the image pixel pairs.

[0033] Setting gray-level co-occurrence matrix parameters, which include discrete levels of gray values, distances between pixel pairs, and directions of pixel pairs;

[0034] The calculation formula of the pixel occurrence frequency of the processed image calculated by the gray level co-occurrence matrix is:

[0035] Where P(i,j,d,θ) represents the frequency of occurrence of pixel pairs with gray levels i and j at distance d and direction θ, δ is the Kronecker function, which returns 1 when the condition is true and 0 otherwise; M and N are the width and height of the optical image information, respectively, g(x,y) is the gray value of the pixel point (x,y) in the optical image information; d represents the distance between the pixel pairs, and θ represents the direction angle between the pixel pairs;

[0036] By analyzing the gray-level co-occurrence matrix, we can extract the texture features of the image, such as contrast, correlation, energy and other indicators, which can effectively describe the texture characteristics of the particles. In particle defect detection, the texture analysis method based on the gray-level co-occurrence matrix can identify the existence of particles and distinguish particle defects from normal areas through the difference in texture features, especially in the judgment of particle size, shape and arrangement rules.

[0037] In specific application scenarios, the Canny algorithm is mainly used to detect obvious scratches with clear edges on the wafer surface. This type of defect usually has a more obvious contrast difference. The texture analysis method of the grayscale co-occurrence matrix is ​​suitable for the detection of particle defects, because particles on the wafer surface usually appear as areas with specific texture features. By extracting these texture features, particles can be effectively identified and other types of defects can be distinguished. The combination of the two can achieve accurate extraction of scratches, particles and other types of defects in wafer surface defect detection.

[0038] For the detection of pinhole defects, the brightness and contrast analysis method of the local area of ​​the image is used for detection. Pinhole defects are usually manifested as abnormal local brightness in the image, and there is a significant difference in contrast with the surrounding area. In order to accurately detect pinhole defects, the brightness and contrast characteristics of the local area of ​​the image can be calculated to effectively identify these abnormal areas.

[0039] The method for calculating the regional brightness and regional contrast of the processed image is;

[0040] The regional brightness calculation formula of the processed image is:

[0041] Among them, L R is the brightness of region R, |R| is the number of pixels in region R, and g(x,y) is the grayscale value of the pixel point (x,y) in the optical image information;

[0042] The regional contrast calculation formula of the processed image is:

[0043] Among them, C R is the contrast of region R.

[0044] This formula reflects the grayscale change of the area by calculating the difference between each pixel in the area and the regional brightness. For pinhole defects, there is usually an obvious local contrast change, because the pinhole defect is often a small and concentrated abnormal brightness area, and the contrast of the surrounding area is relatively high.

[0045] In practical applications, the brightness and contrast analysis algorithm is particularly suitable for the detection of pinhole defects. Pinhole defects are generally manifested as points in the image where the brightness changes suddenly, and the contrast is significantly different from that of the surrounding areas. By calculating the brightness and contrast of the local area of ​​the image, the abnormal area in the image can be located and the pinhole defect can be accurately identified. In terms of parameter selection, the detection accuracy can be further improved by setting a suitable local area size and combining the brightness and contrast thresholds. For example, when it is detected that the brightness value of a local area is significantly different from that of the surrounding area, and the contrast of the area is high, it can be determined that there is a pinhole defect in the area.

[0046] The advantage of this algorithm is that it is simple and effective. It can quickly locate and identify defect areas with drastic brightness changes in the image, especially in the case of low image quality or large background noise, which can effectively reduce the risk of false detection and missed detection. At the same time, the analysis of brightness and contrast also provides an important basis for subsequent defect classification and severity assessment.

[0047] For ultrasonic echo signals, the wavelet transform algorithm is used to decompose the signals at different scales and extract the characteristic frequency components reflecting the defects, including the depth, shape and other related information of the defects.

[0048] Specifically, wavelet transform can effectively analyze signal changes within different depth ranges and identify defect characteristics on the surface and inside of the wafer. Ultrasonic signals usually contain components of different frequencies, among which the characteristic frequency components of defects reflect information such as the location, depth and shape of the defects. Therefore, the application of wavelet transform can effectively extract these important features and provide useful input features for subsequent convolutional neural network (CNN) training.

[0049] Furthermore, extracting features from the received ultrasonic echo signal of the wafer to be tested to obtain an ultrasonic feature vector includes:

[0050] Performing wavelet transform and discrete wavelet transform on the received ultrasonic echo signal of the wafer to be tested to obtain wavelet transform coefficients and discrete wavelet transform coefficients; performing feature extraction based on the obtained wavelet transform coefficients and discrete wavelet transform coefficients to obtain an ultrasonic feature vector;

[0051] The wavelet transform calculation formula is:

[0052] where W ψ (a, b) are wavelet transform coefficients, x(t) is the received ultrasonic echo signal, ψ(t) is the mother wavelet, a is the scale factor, b is the translation factor, * represents the complex conjugate operation, and t is the time variable of the received ultrasonic echo signal;

[0053] By performing wavelet transforms at different scales and positions, changes in the signal within different frequency ranges can be captured. During the processing of ultrasonic signals, wavelet transforms can analyze the high-frequency and low-frequency components of the signal by adjusting the scale factor a and the position factor b respectively. When detecting wafer defects, the high-frequency components of the signal usually reflect surface defects (such as scratches, particles), while the low-frequency components may be related to internal defects (such as voids, cracks). Therefore, by performing multi-scale wavelet transforms on ultrasonic signals, frequency components reflecting defect characteristics can be extracted at different scales.

[0054] The formula for discrete wavelet transform is:

[0055]

[0056] where D ψ (a, b) are discrete wavelet transform coefficients, x[n] are the sample points of the discrete signal, ψ is the discrete wavelet function, a and b are the scale factor and the position factor respectively, N is the length of the ultrasonic echo signal, and n is the sequence number of the sample points of the discrete signal.

[0057] By performing discrete wavelet transforms on ultrasonic signals, wavelet coefficients at different scales and positions can be obtained. These coefficients can reflect the changes in different frequency components in the signal, thus helping to identify different types of defects in the wafer.

[0058] In specific applications, the data of ultrasonic sensors usually contains information at multiple depth levels. Characteristics such as the depth, shape, and type of defects will all affect the frequency components of the signal. Through wavelet transforms, this information can be effectively extracted, providing rich input data for the CNN model, thereby improving the accuracy and robustness of defect detection. Wavelet transforms solve the problem of uneven distribution of frequency components in ultrasonic signals through multi-scale analysis, and can effectively extract subtle defect characteristics, enabling the CNN to accurately identify different types of defects on the surface and inside the wafer.

[0059] In terms of parameter setting, the selection of appropriate wavelet functions (such as Daubechies wavelet, Morlet wavelet, etc.) and scale factor a have an important impact on the multi-scale decomposition of the signal. Generally, it is necessary to select appropriate wavelet functions and scale ranges according to the frequency characteristics of the ultrasonic signal to maximize the retention of defect characteristics and reduce noise interference. These extracted characteristic frequency components can be directly used as inputs to CNN. Through training, CNN can automatically identify the defect type and provide information on the depth and shape of the defect.

[0060] Specifically, for optical image data, edge detection algorithms are used, such as the Canny algorithm combined with morphological operations, to accurately extract the edge features of scratches. The texture analysis method based on the grayscale co-occurrence matrix is ​​used to identify the presence of particles and their texture features. The location and size of pinhole defects are detected by analyzing the brightness and contrast of local areas of the image.

[0061] Electromagnetic sensors use the eddy current principle to detect defects on the surface and inside of wafers. Eddy current flaw detection technology mainly identifies defects by analyzing the distribution and changes of electromagnetic fields in materials. The phase and amplitude changes of electromagnetic response signals are important bases for identifying defects. By establishing an electromagnetic model, the phase and amplitude changes of the signal can be effectively analyzed to extract electromagnetic characteristic parameters related to the defect type (such as scratches, particles, pinholes). These features can provide strong data support for the identification and classification of defects, especially in the detection of tiny defects on the wafer surface.

[0062] The basic principle of eddy current signal is that when the probe generates alternating current, it will generate a magnetic field on the surface of the wafer to be tested, and this magnetic field interacts with the conductive properties inside the wafer. When there are defects on the surface or inside the wafer, it will affect the distribution of the magnetic field, thereby causing changes in the electromagnetic response signal. The amplitude and phase of the eddy current will change significantly due to the presence of defects, especially at scratches, particles or pinhole defects, where the phase and amplitude of the signal change more significantly.

[0063] Further, extracting features from the received electromagnetic response signal of the wafer to be tested to obtain an electromagnetic feature vector includes:

[0064] Calculating the eddy current impedance of the electromagnetic response signal and the phase change of the electromagnetic response signal according to the received electromagnetic response signal of the wafer to be tested, and extracting the features of the calculated eddy current impedance and the phase change of the electromagnetic response signal to obtain an electromagnetic feature vector;

[0065] The calculation formula of the eddy current impedance of the electromagnetic response signal is: Z(f)=R(f)+jX(f); where Z(f) is the eddy current impedance, R(f) is the resistance component, X(f) is the inductive reactance component, f is the electromagnetic response signal frequency, and j is the imaginary unit;

[0066] The amplitude of eddy current impedance is closely related to the conductivity of the wafer and the nature of the defect (such as size, shape, depth, etc.). When there is a defect on the wafer surface, the change in conductivity in the local area will cause the impedance amplitude of the eddy current to change, thus reflecting the characteristics of the defect. Especially under high-frequency signals, tiny defects will cause large amplitude changes. Therefore, amplitude analysis is particularly important for the detection of subtle defects.

[0067] The calculation formula of the phase change of the electromagnetic response signal is: Δφ(f)=arg(Z defect (f))-arg(Z reference (f)), where Δφ(f) is the phase change of the electromagnetic response signal, Z defect (f) is the eddy current impedance of the defective area, Z reference (f) is the eddy current impedance of the defect-free area, and arg(·) represents the phase angle of the complex number.

[0068] The magnitude and direction of the phase change can reveal information about the depth, shape, and other aspects of the defect. For example, shallower surface defects usually result in larger phase changes, while deeper internal defects result in smaller phase changes. Therefore, combining phase changes with amplitude changes can more accurately determine the location and nature of the defect.

[0069] In application scenarios, electromagnetic response analysis is particularly suitable for the detection of wafer surface and shallow defects, such as scratches, particles and pinholes. Scratches usually cause significant changes in the eddy current impedance amplitude, while particles and pinholes mainly affect the phase of the signal. By analyzing the electromagnetic response data at different frequencies, different types of defects can be distinguished. Based on the frequency response characteristics of different defects, combined with the electromagnetic model, the electromagnetic characteristic parameters reflecting the defect type and depth can be effectively extracted, thereby providing accurate data for subsequent defect identification and classification.

[0070] Furthermore, the intelligent decision module performs a splicing operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain spliced ​​data in the following method:

[0071] X combined =[X optical ;X ultrasonic ;X electromagnetic ]; where X combined For splicing data, X optical is the optical image feature vector, X ultrasonic is the ultrasonic feature vector, X electromagnetic is the electromagnetic eigenvector;

[0072] The method by which the intelligent decision module fuses the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain fused data is as follows:

[0073] X fused =w optical ·X optical +w ultrasonic ·X ultrasonic +w electromagnetic ·X electromagnetic ; where X fused is the fused data, w optical is the optical image feature vector weight, w ultrasonic is the ultrasound feature vector weight, w electromagnetic is the electromagnetic eigenvector weight.

[0074] Specifically, this system uses two methods: data splicing and weighted fusion. Data splicing is to directly splice the feature data from different sensors along a certain dimension (usually the channel dimension) to generate a higher-dimensional input data. For the data of the three sensors, optical images are usually two-dimensional data, while ultrasonic signals and electromagnetic response data can be one-dimensional or two-dimensional feature vectors. The basic idea of ​​data splicing is to combine data from these different sources into a multi-channel input, so that the convolutional neural network can process information from multiple sensors at the same time.

[0075] The concatenated data dimensions can be written as:

[0076]

[0077] The optical image feature vector is a two-dimensional matrix X optical ∈R H×W , where H and W are the height and width of the image respectively; the ultrasound feature vector is X ultrasonic ∈R L , where L is the dimension of the ultrasonic signal feature; the electromagnetic feature vector is X electromagnetic ∈R M , where M is the dimension of the electromagnetic response feature.

[0078] Furthermore, the ultrasonic sensor is a phased array ultrasonic transducer with a frequency ranging from 5 Hz to 20 Hz.

[0079] Specifically, in wafer defect detection, the optical sensor provides a 256×256 pixel image, and the ultrasonic sensor and electromagnetic sensor provide 64-dimensional and 32-dimensional feature vectors, respectively. Through data concatenation, we can concatenate the optical image and the feature vectors of the two sensors to form an input data with a shape of 256×256×(1+64+32)=256×256×97 for the convolutional neural network to classify and identify defects.

[0080] Weighted fusion is to weight the feature data of different sensors according to their importance, and then fuse the weighted data. This method can make the network pay more attention to the sensor data that contributes to defect detection and reduce the impact of unimportant or noisy data. The process of weighted fusion can be expressed by the following formula:

[0081]

[0082] Among them, X fused is the weighted fusion data, X i is the data from the ith sensor (optical, ultrasonic or electromagnetic), w i is the weight of the ith sensor, and n is the total number of sensors (here 3). These weights w i It needs to be set according to the contribution of each sensor to defect detection. It can be dynamically adjusted through parameters learned during model training, or pre-set through domain knowledge.

[0083] Assume that the optical image feature vector is X optical ∈R H×W , the ultrasound feature vector is X ultrasonic ∈R L , the electromagnetic eigenvector is X electromagnetic ∈R M , and the corresponding weights are w optical 、w ultrasonic and w electromagnetic

[0084] Furthermore, the wafer defect detection convolutional neural network is a CNN convolutional neural network.

[0085] The main structure of the convolutional neural network includes three convolutional layers, a pooling layer, and a fully connected layer. In the convolutional layer, the convolution kernel extracts local features of the image or signal through a sliding window operation. The output of the convolutional layer is a feature map extracted by the convolution kernel, and each convolution kernel is responsible for extracting features of a specific scale or type. For optical images, the convolution kernel extracts the edge information of the scratch; for ultrasonic signals, the convolution kernel extracts the frequency component of the defect; and for electromagnetic signals, the convolution kernel extracts the amplitude and phase change information related to the defect.

[0086] The size of the convolution kernel determines the scale of the extracted features. Generally speaking, smaller convolution kernels (such as 3×3 or 5×5) are suitable for extracting smaller-scale features, such as tiny scratches or particles on the surface, while larger convolution kernels (such as 7×7 or 11×11) are suitable for extracting larger-scale features, such as larger or deeper defects. The output of the convolution layer is processed by the pooling layer, which reduces the size of the feature map through downsampling operations, thereby reducing computational complexity and enhancing the robustness of the model. Commonly used pooling methods include maximum pooling and average pooling. Pooling operations help to retain important features, remove noise, and reduce overfitting.

[0087] After multiple convolution and pooling operations, the final stage of the convolutional neural network is the fully connected layer. The fully connected layer flattens the high-dimensional feature map after convolution and pooling into a one-dimensional vector, and then performs weighted summation through a series of weight matrices to finally output the probability distribution of each category. In the scenario of wafer defect detection, the output layer of the CNN will determine whether there are defects on the wafer surface based on the features, and classify the defect type (such as scratches, particles, pinholes) and severity (mild, moderate, severe).

[0088] For specific defect judgment, the cross entropy loss function can be used to train the model. The formula is as follows:

[0089]

[0090] Where L is the cross entropy loss, N is the number of categories, and y i is the true label (defect type or severity), is the predicted probability output by the CNN model. By minimizing the cross entropy loss, the model is able to optimize the parameters so that the difference between the predicted result and the true label is minimized, thereby improving the accuracy of defect classification.

[0091] This algorithm is particularly suitable for scenarios with complex defects and incomplete information. In wafer surface defect detection, due to the variety of defect types and possible overlap, traditional manual feature extraction methods are difficult to effectively distinguish different defects. By extracting features from data from optical, ultrasonic and electromagnetic sensors, fusing the feature vectors obtained after feature extraction, and inputting them into the convolutional neural network, the accuracy and robustness of defect recognition can be greatly improved.

[0092] Furthermore, the optical sensors, ultrasonic sensors and electromagnetic sensors are evenly distributed in the circumferential direction with the center of the wafer to be tested as the center.

[0093] Specifically, the three sensors are evenly distributed around the center of the wafer. During the inspection, the wafer is fully inspected by rotating the wafer around the center. All inspection data of the wafer is obtained. The data generated by the inspection is based on the optical sensor, and the position data in the other two sensor data are offset by 120 degrees and 240 degrees for position synchronization.

[0094] Furthermore, the wafer surface defect detection system based on multi-channel sensors also includes a result storage and report generation module; the result storage and report generation module is used to obtain and store optical image information of the wafer to be tested, ultrasonic echo signals of the wafer to be tested, electromagnetic response signals of the wafer to be tested, optical image characteristic parameters, ultrasonic echo signal characteristic parameters, electromagnetic response signal characteristic parameters, electromagnetic characteristic vectors and defect information of the wafer to be tested; the result storage and report generation module is also used to obtain the wafer number and detection date to be tested, and generate a wafer detection report according to the wafer number to be tested, the detection date and the detection information of the wafer to be tested.

[0095] Specifically, the result storage and report generation module saves the results generated by the convolutional neural network. The storage content includes the original collected data, processed feature data, defect judgment results, and corresponding wafer location information.

[0096] The report generation module automatically generates detailed inspection reports based on pre-set templates. The reports not only contain basic information such as defect location, type, and size, but also provide examples of defect images or signal features, as well as analysis and evaluation of the possible impact of the defects, presented to users in an intuitive and easy-to-understand manner.

[0097] Specifically, in the wafer surface defect detection system, the detection results generated by the convolutional neural network (CNN) need to be systematically stored and reported to ensure the transparency of the detection process and the ease of use of the results. The result storage module is responsible for saving multiple key data, including the original acquisition data, the processed feature data, the defect judgment results, and the wafer location information. Specifically, the original acquisition data includes optical images, ultrasonic signals, and electromagnetic response data, which are usually stored on a solid-state drive (SSD) to ensure efficient data reading and stability. The processed feature data includes various defect features extracted by the algorithm, such as the edge information of the image, the frequency domain features, and the phase and amplitude changes of the electromagnetic signal. The defect judgment results include the classification, location, type, and severity assessment of the defect by the CNN model, which will help further decision-making and operation. Finally, the wafer position information records the specific position and rotation angle of the wafer during the detection process to ensure that the detection results can correspond to the actual position of the wafer.

[0098] The report generation module automatically generates a detailed inspection report based on a preset template by combining all stored data. The content of the report not only includes basic information such as the location, type, and size of the defect, but also provides examples of images or signal characteristics of the defect, as well as an analysis and evaluation of the possible impact of the defect. These contents are presented to users in a clear and concise manner to ensure the intuitiveness and operability of the inspection results. The location of the defects on the surface of the wafer is marked in detail in the report, usually in the form of an image, and the specific area where the defect occurs is marked in the image. In addition, the report will list the type of each defect (such as scratches, particles, pinholes, etc.) and give its size (such as length, width, depth, etc.) to help users fully understand the basic characteristics of the defect. For each type of defect, the report will also provide examples of images or signal characteristics to further help users understand the specific manifestations of the defect. In terms of the impact of the defect, the report will provide an assessment of the potential impact on product performance based on the analysis results, such as whether the defect may cause the electrical performance of the wafer to deteriorate, or whether it will affect the subsequent production process. Finally, the report will also provide a classification and severity assessment of the defect, such as whether the scratch is minor and whether the particle will affect the electrical performance.

[0099] Specifically, assuming that a wafer has been inspected, the report may be presented like this: The report first lists the basic information of the wafer, such as the wafer number, inspection date, and inspector. Then, the report lists in detail the three defects found: a slight scratch located in the center of the wafer, measuring 2.5mmx0.2mm, with a severity of slight, and no significant effect on the electrical performance of the wafer; a moderate particle located at the edge of the wafer, measuring 0.5mmx0.5mm, which may have an impact on electrical performance; and a severe pinhole located on the periphery of the wafer, measuring 0.1mm, which may cause electrical performance problems and is recommended to be processed or scrapped immediately. Next, the report is attached with relevant defect images and ultrasonic signal characteristics to help users intuitively understand the performance of the defects. The report also provides a detailed analysis of the impact of the defects, for example, scratches may affect the appearance but not the function, particles may cause electrical performance degradation, and pinholes will have a serious impact on the quality of the wafer and require urgent processing.

[0100] The last part of the report is the conclusion and suggestions, which provide follow-up operation suggestions according to the severity of different defects. For example, pinhole defects need to be treated first, particles can be repaired by cleaning, and scratches can be scrapped or repaired according to the situation. The entire report is generated by an automated report generation system, ensuring efficient data processing and report consistency.

[0101] Through this detailed and structured report, users can quickly understand the inspection results of wafers and evaluate the impact of defects on production and quality, so as to make more accurate decisions and ensure smooth production process and stable product quality.

[0102] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A wafer surface defect detection system based on a multi-channel sensor, characterized by: Including data acquisition module, defect feature extraction module and intelligent decision-making module; The data acquisition module is used to collect optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested, and send the optical image information of the wafer to be tested, ultrasonic echo signal of the wafer to be tested and electromagnetic response signal of the wafer to be tested to the defect feature extraction module; The defect feature extraction module is used to extract features of the received optical image information of the wafer to be tested, the ultrasonic echo signal of the wafer to be tested, and the electromagnetic response signal of the wafer to be tested, respectively, to obtain an optical image feature vector, an ultrasonic feature vector, and an electromagnetic feature vector, and send the optical image feature vector, the ultrasonic feature vector, and the electromagnetic feature vector to the intelligent decision module; The intelligent decision-making module is used to perform a splicing operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain splicing data, and is also used to perform a fusion operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain fusion data, and input the spliced ​​data and the fusion data into the wafer defect detection convolutional neural network to obtain the defect information of the wafer to be tested; the defect information of the wafer to be tested includes whether the wafer has defects, the type of wafer defects and the degree of wafer defects.

2. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: The data acquisition module includes an optical sensor, an ultrasonic sensor and an electromagnetic sensor; the optical sensor is used to collect optical image information of the wafer to be tested, and send the optical image information of the wafer to be tested to the defect feature extraction module; the ultrasonic sensor is used to collect ultrasonic echo signals of the wafer to be tested, and send the ultrasonic echo signals of the wafer to be tested to the defect feature extraction module; The electromagnetic sensor is used to collect the electromagnetic response signal of the wafer to be tested, and send the collected electromagnetic response signal of the wafer to be tested to the defect feature extraction module.

3. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: Extracting features from the received optical image information to obtain an optical image feature vector includes: Performing Gaussian filtering on the received optical image information to obtain a processed image, calculating the gradient intensity of the processed image by using the Canny algorithm, calculating the pixel occurrence frequency of the processed image by using the gray level co-occurrence matrix, and calculating the regional brightness and regional contrast of the processed image according to the processed image; Extracting features according to the gradient intensity of the processed image, the pixel occurrence frequency of the processed image, the regional brightness and the regional contrast of the processed image to obtain an optical image feature vector; The method of performing Gaussian filtering on the received optical image information to obtain the processed image is: The Gaussian filter formula is: Wherein, σ is the standard deviation of the Gaussian filter, x and y are the horizontal and vertical coordinates of the optical image information, respectively, and G(x,y) is the filter value of the pixel point (x,y); The method for calculating the gradient strength of the processed image by the Canny algorithm is: The gradient strength calculation formula is: Among them, G is the gradient strength, G x is the gradient of the processed image in the x direction, G y is the gradient of the processed image in the y direction; The calculation formula of the pixel occurrence frequency of the processed image calculated by the gray level co-occurrence matrix is: Where P(i,j,d,θ) represents the frequency of occurrence of pixel pairs with gray levels i and j at distance d and direction θ, δ is the Kronecker function, which returns 1 when the condition is true and 0 otherwise; M and N are the width and height of the optical image information, respectively, g(x,y) is the gray value of the pixel point (x,y) in the optical image information; d represents the distance between the pixel pairs, and θ represents the direction angle between the pixel pairs; The method for calculating the regional brightness and regional contrast of the processed image is; The regional brightness calculation formula of the processed image is: Among them, L R is the brightness of region R, |R| is the number of pixels in region R, and g(x,y) is the grayscale value of the pixel point (x,y) in the optical image information; The regional contrast calculation formula of the processed image is: Among them, C R is the contrast of region R.

4. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: The ultrasonic feature vector obtained by extracting the features of the received ultrasonic echo signal of the wafer to be tested includes: Performing wavelet transform and discrete wavelet transform on the received ultrasonic echo signal of the wafer to be tested to obtain wavelet transform coefficients and discrete wavelet transform coefficients; performing feature extraction based on the obtained wavelet transform coefficients and discrete wavelet transform coefficients to obtain an ultrasonic feature vector; The wavelet transform calculation formula is: Where W ψ (a, b) are wavelet transform coefficients, x(t) is the received ultrasonic echo signal, ψ(t) is the mother wavelet, a is the scale factor, b is the translation factor, * represents the complex conjugate operation, and t is the time variable of the received ultrasonic echo signal; The discrete wavelet transform calculation formula is: Among them, D ψ (a, b) are discrete wavelet transform coefficients, x[n] is the sample point of the discrete signal, ψ is the discrete wavelet function, a and b are the scale factor and position factor respectively, N is the length of the ultrasonic echo signal, and n is the serial number of the sample point of the discrete signal.

5. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: The electromagnetic feature vector obtained by extracting the features of the received electromagnetic response signal of the wafer to be tested includes: Calculating the eddy current impedance of the electromagnetic response signal and the phase change of the electromagnetic response signal according to the received electromagnetic response signal of the wafer to be tested, and extracting the features of the calculated eddy current impedance and the phase change of the electromagnetic response signal to obtain an electromagnetic feature vector; The calculation formula of the eddy current impedance of the electromagnetic response signal is: Z(f)=R(f)+jX(f); where Z(f) is the eddy current impedance, R(f) is the resistance component, X(f) is the inductive reactance component, f is the electromagnetic response signal frequency, and j is the imaginary unit; The calculation formula of the phase change of the electromagnetic response signal is: Δφ(f)=arg(Z defect (f))-arg(Z reference (f)), where Δφ(f) is the phase change of the electromagnetic response signal, Z defect (f) is the eddy current impedance of the defective area, Z reference (f) is the eddy current impedance of the defect-free area, and arg(·) represents the phase angle of the complex number.

6. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: The method in which the intelligent decision module performs a splicing operation on the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain spliced ​​data is as follows: X combined =[X optical ;X ultrasonic ;X electromagnetic ]; where X combined For splicing data, X optical is the optical image feature vector, X ultrasonic is the ultrasonic feature vector, X electromagnetic is the electromagnetic eigenvector; The method by which the intelligent decision module fuses the received optical image feature vector, ultrasonic feature vector and electromagnetic feature vector to obtain fused data is as follows: X fused =w optical ·X optical +w ultrasonic ·X ultrasonic +w electromagnetic ·X electromagnetic ; where X fused is the fused data, w optical is the optical image feature vector weight, w ultrasonic is the ultrasound feature vector weight, w electromagnetic is the electromagnetic eigenvector weight.

7. The wafer surface defect detection system based on a multi-channel sensor according to claim 2, characterized in that: The ultrasonic sensor is a phased array ultrasonic transducer with a frequency in the range of 5 Hz to 20 Hz.

8. The wafer surface defect detection system based on a multi-channel sensor according to claim 2, characterized in that: The optical sensor, ultrasonic sensor and electromagnetic sensor are evenly distributed in the circumferential direction with the center of the wafer to be tested as the center.

9. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: The wafer defect detection convolutional neural network is a CNN convolutional neural network.

10. The wafer surface defect detection system based on a multi-channel sensor according to claim 1, characterized in that: It also includes a result storage and report generation module; the result storage and report generation module is used to obtain and store the optical image information of the wafer to be tested, the ultrasonic echo signal of the wafer to be tested, the electromagnetic response signal of the wafer to be tested, the electromagnetic characteristic vector and the defect information of the wafer to be tested; the result storage and report generation module is also used to obtain the wafer number and the inspection date to be tested, and generate a wafer inspection report according to the wafer number to be tested, the inspection date and the inspection information of the wafer to be tested.

Images