CN118822960A - A wafer surface defect detection method, device and computer equipment - Google Patents

Abstract

The present invention relates to wafer surface defect detection technology, and in particular, to a wafer surface defect detection method, apparatus, and computer device; the method comprises the steps of acquiring a target image and a reference image; extracting image features of the target image; the image features comprise gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, fourier spectrum cross-correlation features, hu invariant moment features and low-frequency concentration features; the image cross-correlation feature and the fourier spectrum cross-correlation feature are determined based on the reference image; inputting the image features of the target image into a pre-trained machine learning model, performing defect detection processing on the image features, and outputting a defect detection result of the target image. The invention integrates 8 feature points, realizes the concentration and extraction of the image features of the wafer, and thereby greatly avoids the difficulties of extremely large calculation amount and data amount required by the existing detection schemes of deep learning and machine learning.

Description

A wafer surface defect detection method, device and computer equipment

Technical Field

The present invention relates to a wafer surface defect detection technology, and in particular to a wafer surface defect detection method, device and computer equipment.

Background Art

Integrated circuit wafers are the foundation of the semiconductor industry, and they directly determine the quality and performance of chips and electronic components. The production and manufacturing of wafers are extremely complex, with up to hundreds of process steps, and the processing of each process may lead to surface defects. Allowing defective wafers to continue to participate in the subsequent manufacturing process of chips will not only produce chips with poor performance and quality, but also waste manpower and material resources meaninglessly. Improving and improving production process technology can certainly reduce the probability of defects, but it cannot be completely avoided, and with the accumulation of marginal benefits, the cost of this method will be very huge. Therefore, it is particularly critical to carry out corresponding inspections in each step of wafer production and manufacturing to improve the quality and reliability of chips and electronic components.

Most existing detection methods are based on visual inspection by human eyes, and the basis for detection mainly relies on experience and judgment. This method has obvious disadvantages in terms of detection efficiency and reliability. In addition, with the development of image processing and deep learning theory, some automated detection solutions based on machine vision have also been derived. However, traditional image processing detection algorithms require a lot of prior knowledge, and usually only detect a single type of wafer, and their generalization and general capabilities are relatively limited. Deep learning-based detection algorithms and models have outstanding advantages in generalization capabilities, but they require a large number of samples for training. The amount of calculation and data volume puts high demands on hardware equipment and costs, limiting the application and deployment of this solution in actual detection scenarios.

Summary of the invention

The purpose of the present invention is to provide a wafer surface defect detection method, device and computer equipment. The present invention proposes a new solution for detecting wafer surface defects, which greatly improves the problems of low efficiency and poor reliability of existing manual detection, and excessive calculation and data volume of deep learning detection algorithms. On the basis of automated detection, it focuses on improving the accuracy of detection results and greatly reduces the calculation overhead and hardware cost in the detection process.

In a first aspect of the present invention, the present invention provides a method for detecting wafer surface defects, the method comprising:

Acquire a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected;

Extracting image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image;

The image features of the target image are input into a pre-trained machine learning model, defect detection processing is performed on the image features, and the defect detection result of the target image is output.

In a second aspect of the present invention, the present invention further provides a wafer surface defect detection device, the device comprising:

An acquisition module acquires a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected;

An extraction module, extracting image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image;

The prediction module inputs the image features of the target image into a pre-trained machine learning model, performs defect detection processing on the image features, and outputs the defect detection results of the target image.

In the third aspect of the present invention, a computer device is proposed, comprising: a processor, a memory and a bus, the memory storing machine-readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the wafer surface defect detection method as described in the first aspect of the present invention.

The present invention is a method, device and computer equipment for detecting wafer surface defects. The embodiment of the present application provides a method with high detection accuracy and low computational overhead for detecting wafer surface defects. The method consists of three parts: wafer image acquisition, feature point calculation, classification model training and test data verification. In this embodiment, classification model training and test data verification are both considered as defect detection processes.

The core of this method is to use a specific pattern to extract features from the wafer surface image, thereby concentrating hundreds or even thousands of pixels of the entire image into eight highly representative feature points. The calculation results of these eight feature points come from a variety of feature analyses of the sample image, mainly including: grayscale co-occurrence matrix, histogram statistics, image cross-correlation, Fourier spectrum cross-correlation, Hu invariant moment, and low-frequency concentrated features. The calculation of feature points is actually a highly concentrated and accelerated extraction of feature information of the wafer image. In addition, due to the sparse distribution of defects on the wafer image, the calculation process of feature points also realizes the data dimensionality reduction of the entire image, thereby filtering out a large amount of redundant data and extracting effective information.

Beneficial effects of the present invention:

The present invention proposes a method, device and computer equipment for wafer surface defect detection. The present invention no longer relies on the traditional detection mode of human eye observation, but is based on the research direction of machine vision, and has the characteristics of high automation, high reliability, high accuracy and high efficiency. The present invention uses a variety of image feature analysis methods to calculate the wafer surface image, integrates to form 8 feature points, realizes the concentration and extraction of sample image features, and greatly reduces the information of hundreds or even thousands of pixels in the entire image to 8 independent feature data points, thereby greatly avoiding the difficulties of the huge amount of calculation and data required by the existing deep learning and machine learning detection schemes. The method greatly optimizes the hardware overhead and deployment difficulty in the actual detection process without reducing the accuracy of the detection results. In addition, the feature extraction process in this method is applicable to a variety of machine learning models, and has shown good detection performance in the tests of a variety of machine learning models, with broad application prospects and high generalization capabilities.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is an overall block diagram of a detection scheme according to an embodiment of the present invention;

FIG2 is a partial wafer sample image of WM-811K according to an embodiment of the present invention;

FIG3 is a schematic diagram of a characteristic point calculation scheme according to an embodiment of the present invention;

FIG4 is a confusion matrix of test results under the random forest model according to an embodiment of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Wafers are the basis of semiconductor devices and integrated circuits, and their quality directly affects the performance of the final product. Through wafer image defect detection, it can be ensured that the wafer does not contain defects such as bad spots, cracks, and contamination during the production process, thereby ensuring the quality and reliability of the product.

Wafer defects can be mainly divided into surface defects, structural defects, impurity defects, processing defects, insulator adhesion defects, packaging defects, defects introduced during transportation and storage, etc. For wafer surface defects, surface defects usually include scratches, crystal points, oxide films, surface redundancy and other defects generated during wafer molding, cleaning, epitaxial growth and transportation: Scratches: Linear defects on the surface of the wafer, usually caused by improper cutting process or mechanical damage during cleaning. Scratches will affect the flatness of the wafer surface, and then affect the accuracy and stability of processes such as lithography and thin film deposition. Crystal points: Small point-shaped defects on the surface of the wafer, usually caused by impurities or foreign matter attached to the surface of the wafer during the manufacturing process. Crystal points will affect the electrical and optical properties of the wafer, reducing the reliability and availability of the wafer. Oxide film defects: Defects that do not meet the requirements on the oxide film on the surface of the wafer, usually caused by improper control during the oxidation process or the introduction of impurities. Oxide film defects will affect the insulation properties and dielectric constant of the wafer, thereby affecting the electrical performance and reliability of the device. Surface redundancy: including tiny particles, dust, residues from the previous process of wafer processing, etc. These redundancies will seriously affect the integrity of the wafer surface.

For the technical solution shown in FIG1 , taking the manufacturing process of semiconductor devices (such as memory chips, power devices, etc.) as an example, defect detection is usually performed on each wafer used as a manufacturing substrate, so that defect-free wafers or wafers with a small number of defects but no killer defects can be detected and screened to obtain wafers for subsequent semiconductor device manufacturing processes. Therefore, wafers are usually subjected to non-destructive testing, that is, the subsequent semiconductor device manufacturing process is not affected by the testing. In order to achieve non-destructive testing, an embodiment of the present invention preferably uses a circuit probe to perform electrical testing on each module of the wafer to obtain a visualized image as a wafer surface defect detection image. The defect detection process shown in FIG1 is completed. A wafer surface defect detection method according to an embodiment of the present invention, as shown in FIG1 , comprises:

101. Acquire a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected;

102. Extract image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features, and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image;

103. Input the image features of the target image into a pre-trained machine learning model, perform defect detection processing on the image features, and output the defect detection result of the target image.

In this example, the proposed method is tested and the test results are presented using the public wafer surface defect image dataset WM-811K. This dataset is derived from wafer surface images in a real inspection environment and is highly representative. It is one of the largest publicly accessible wafer image data. It includes information such as the physical properties of the wafer, process parameters in the production process, equipment status, and production environment, as well as measurement data of electrical, optical, and physical properties obtained during wafer testing. The dataset also includes wafer defect and flaw data, which helps analysts detect wafer quality problems and locate and resolve possible defect sources in the production process. As shown in Figure 2, images of some wafer samples in the WM-811K dataset. According to the data labels of the WM-811K dataset, the samples are divided into two cases: defective and intact in this example. Since the WM-811K dataset is collected from multiple batches and multiple resolution inspection scenes, the number of sample images of each type and resolution in the dataset is extremely unbalanced. In order to strike a balance between defective samples and defective sample images, this example selects images with a resolution of 25×27 in the data set as samples to be tested, and randomly selects 2500 defect-free images and 2500 defective images. Only one defect-free image is used as a reference image, and the other 2499 defect-free images and 2500 defective images can be used as target images to be tested.

The core of the present invention is to use feature extraction to concentrate and compress the pixel information of the entire image to 8 feature data points to reduce the computational overhead and hardware cost in the detection process; the feature extraction method mainly includes the measurement of gray-level co-occurrence matrix (GLCM), histogram statistics, image cross-correlation, Fourier spectrum cross-correlation, _Hu invariant moment and low-frequency concentrated features.

As shown in FIG3 , this is a specific selection and calculation scheme of the eight feature points described in the present invention:

Feature point 1 is derived from the gray level co-occurrence matrix of the image. The gray level co-occurrence matrix is a way to describe the texture features of an image. It mainly calculates and counts four components: image contrast, autocorrelation, energy, and uniformity. After obtaining the values of the four components, different weights are assigned to each component to calculate feature point 1. The calculation formula is as follows:

Point ₁ ＝a ₁ *GLCM ₁ +a ₂ *GLCM ₂ +a ₃ *GLCM _s +a ₄ *GLCM ₄

Wherein: Point ₁ represents the calculation result of feature point 1; a ₁ , a ₂ , a ₃ , a ₄ are the weights of each component, and the sum of the four weight values is 1; GLCM ₁ , GLCM ₂ , GLCM ₃ , GLCM ₄ are the calculation results of each component.

Feature point 2 is derived from the histogram statistics of the image. The histogram statistics is the calculation of the number and proportion of pixels occupied by the grayscale levels of the image, which can reflect the distribution of each grayscale of the image. Since the sample image in WM-811K has only three grayscale levels, the calculation result of feature point 2 comes from the statistics and weight calculation of the three grayscale components. The calculation formula is as follows:

Point ₂ ＝b ₁ *H ₁ +b ₂ *H ₂ +b ₃ *H ₃

Wherein: Point ₂ represents the calculation result of feature point 2; b ₁ , b ₂ , b ₂ are the weights of each component, and the sum of the three weight values is 1; H ₁ , H ₂ , H ₃ are the calculation results of each component.

Feature point 3 is the cross-correlation of the image. The calculation process is to first randomly select an image sample from all defect-free samples as the reference image, and then calculate the correlation between the sample to be tested and the reference sample in turn. Generally speaking, since the reference is taken from the defect-free sample image, the calculation result with the defect-free sample will have a higher correlation. Conversely, the presence of defects will lead to a lower correlation coefficient with the reference image. The correlation coefficient calculation process uses the Pearson correlation coefficient calculation method, and the calculation formula is as follows:

Among them, Point ₃ represents the calculation result of feature point 3; A and B represent the input components of the correlation coefficient calculation, namely the reference image and the target image respectively; μ _A and μ _B represent the means of the two components respectively; σ _A and σ _B represent the standard deviations of the two components respectively, and N represents the length of the input component, which can be the number of pixels of the image.

Feature point 4 is the cross-correlation of the Fourier spectrum of the image. Its calculation process is the same as that of feature point 3, except that the input component is replaced from the image to the Fourier spectrum of the image, which can be expressed as:

Among them, Point ₄ represents the calculation result of feature point 4; FA and FB represent the Fourier spectra of correlation coefficient calculation, that is, the Fourier spectra of the reference image and the target image respectively; μ _FA and μ _FB represent the means of the two components respectively; σ _FA and σ _FB represent the standard deviations of the two components respectively; N represents the length of the input component, which can be the Fourier spectrum length.

Feature points 5 to 6 are based on the analysis of the _Hu invariant moments of the image. The _Hu invariant moments have a high degree of concentration and centralization characteristics on the features of the image. The _Hu invariant moments are obtained by calculating multiple Hu _moments of different orders and normalizing them. These values constitute the feature vector of the image. _{The Hu} invariant moments usually contain 7 moment features. After analyzing the features of the WM-811K dataset and sorting the importance, this example chooses to assign the calculation results of the first invariant moment and the second invariant moment to feature points 5 and 6.

The frequency of an image is an indicator of how drastically the grayscale changes in the image, and is the gradient of the grayscale in the plane space. For example, a continuous ocean or desert is an area with slow grayscale changes in the image, and the corresponding frequency value is very low; while undulating mountains or rivers are an area with drastic grayscale changes in the image, and the corresponding frequency value is higher. Generally speaking, the low-frequency part of an image represents the image's contour, structure, texture and other features, while the high-frequency part represents more detailed information. In addition, the Fourier spectrum has the characteristic of concentrated energy, which will concentrate most of the image information in the low-frequency part.

The surface of the wafer is closely arranged with identical small squares, which are called grains. Each of them is an independent unit, representing the etched electronic components. Although the surface of the wafer is distributed with grains and possible positioning structures, the surface of the wafer is relatively smooth as a whole, which leads to the concentration of the information of the wafer surface image in the low-frequency area. Therefore, the present invention processes the low-frequency information unique to the wafer image, that is, the two processes the low-frequency area between the center point of the two axes of the spectrum and the primary peak, and obtains the corresponding feature points 7 and 8.

Feature points 7 and 8 are calculated based on the following two formulas:

Among them, F represents the Fourier spectrum, that is, the spectrum of the current image; u and v are two axes in the frequency domain coordinate system, corresponding to the x and y directions in the Cartesian coordinate system in the time domain space, _um and _vn represent the coordinates of the peak on the two axes passing through the center point in the spectrum in the u and v directions. In the embodiment of the present invention, m and n respectively represent the coordinate distances of the peak on the two axes passing through the center point.

Among them, the formula of Fourier spectrum function is as follows:

Among them, f(x, y) represents the matrix element of the target image, and its coordinate system is the spatial domain. M and N represent the size of the target image, and the coordinate system of F(u, v) is the frequency domain.

In some preferred embodiments of the present invention, the present invention can also pre-process the target image. Since the wafer surface image acquired by the present invention is obtained by a circuit probe, and the electrical characteristics of the grains can be obtained in the circuit probe test, the electrical characteristics of these grains are visualized to obtain the wafer surface image of the present invention, so that its performance and quality can be evaluated. Therefore, the present invention can perform Gaussian pre-processing on the acquired wafer surface images including the target image and the reference image. Gaussian pre-processing can eliminate the noise in the wafer surface image as much as possible, make the wafer surface smooth, eliminate the influence of noise, and improve the accuracy of wafer surface image detection.

In some preferred embodiments of the present invention, it is considered that the acquired target image may have electrical connection problems, mechanical damage problems, material defects, pattern defects, size and position deviation problems, pollution and redundancy problems, and edge defects, etc., and these problems may appear in different degrees and forms in the image. For example, when mechanical damage occurs, it may appear in the form of irregular shapes, dark or light spots, lines, etc. For example, when pollution and redundancy problems occur, they may appear in the form of white or gray spots in the image, and the degree of expression of the spots in the two is different. When extracting image features, the extracted features may not be significant, affecting the subsequent detection results. Based on the above analysis, the present invention can also enhance the 8 feature points in the extracted image features to highlight the information of the target image; further, considering that the image information that needs to be enhanced for each target image may be different, the present embodiment can enhance the 8 feature points separately to obtain 8 enhanced images of the same target image, and input the target image and the corresponding 8 enhanced images into a pre-trained machine learning model, and output the defect detection results of each image, and integrate the defect detection results of these 9 images to obtain the final defect detection results.

In some preferred embodiments of the present invention, after the calculation of the feature points is completed, all samples are divided into two parts: a training set and a test set. Multiple models based on machine learning are used to classify whether the feature points corresponding to the training set samples have defects, thereby realizing the construction of the model. Then, each trained model is tested using the test set samples, and the detection performance of each model is analyzed and compared, and the model with the best detection performance is selected as the final model under the sample condition.

It is understandable that the embodiment of the present invention inputs the target image into a pre-trained machine learning model so that the pre-trained machine learning model can identify whether the target image has defects based on the extracted image features; the embodiment of the present invention can also record the image features corresponding to various defect types accumulated during the historical defect detection process, and train the machine learning model based on these recorded historical defect detection data until the set detection accuracy threshold is reached. After the machine learning model is trained, it can identify whether there are defects based on the input image features.

In some embodiments, the machine learning technology used by the machine learning model may include one or more of a regression model (e.g., a set of statistical processes for estimating the relationship between variables), a classification model and/or a phenomenon model; in addition, the machine learning technology may include quadratic regression analysis, logistic regression analysis, support vector machine, Gaussian process regression, ensemble model or any other regression analysis; furthermore, in other embodiments, the machine learning technology may include decision tree learning, regression tree, boosted tree, gradient boosted tree, multilayer perceptron, one-to-one, naive Bayes, K nearest neighbor, association rule learning, neural network, deep learning, pattern recognition or any other type of machine learning; in other examples, the machine learning technology may include multivariate interpolation analysis.

In an embodiment of the present invention, the embodiment of the present invention further provides a wafer surface defect detection device, the device comprising:

An acquisition module acquires a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected;

An extraction module, extracting image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image;

The prediction module inputs the image features of the target image into a pre-trained machine learning model, performs defect detection processing on the image features, and outputs the defect detection results of the target image.

In an embodiment of the present invention, the present invention also provides a computer device, which includes: a processor, a memory and a bus, the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the wafer surface defect detection method.

In some embodiments, after the feature points of the selected 5000 samples are calculated, each image of the surface wafer will be simplified into a feature sequence consisting of 8 feature points. The 5000 feature sequences are divided into two parts, 80% of the data is used to train the machine learning model, and 20% of the data is used for testing after the model training is completed. This example selects 7 machine learning models (K nearest neighbor, support vector machine, random forest, logistic regression, neural network, decision tree, naive Bayes) to learn and train 4000 sequence data, and then uses 1000 reserved test data for testing. The test results are shown in Table 1.

Table 1. Machine learning model test results

As can be seen from the table, the final test accuracy of all machine learning used exceeds 99%, which shows that the feature extraction and defect detection method proposed in the present invention has good performance and strong generalization ability. Among them, the neural network model showed the best detection performance results, achieving a detection accuracy of 99.9% in the test data, with excellent performance. As shown in Figure 4, it is the confusion matrix of the detection results under the neural network model.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims (8)

1. A wafer surface defect detection method, characterized in that the method comprises: Acquire a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected; Extracting image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image; The image features of the target image are input into a pre-trained machine learning model, defect detection processing is performed on the image features, and the defect detection result of the target image is output. 2. The wafer surface defect detection method according to claim 1 is characterized in that the image cross-correlation feature is the Pearson correlation coefficient between the target image and the reference image. 3. The wafer surface defect detection method according to claim 1 is characterized in that the Fourier spectrum cross-correlation feature is the Pearson correlation coefficient between the Fourier spectrum diagram of the target image and the Fourier spectrum diagram of the reference image. 4. The wafer surface defect detection method according to claim 1 is characterized in that the Hu invariant moment feature includes a first invariant moment feature and a second invariant moment feature. 5. The wafer surface defect detection method according to claim 1 is characterized in that the low-frequency concentrated features include all frequency components between the center point and the primary peak on the two coordinate axes of the frequency domain of the target image. 6. The wafer surface defect detection method according to claim 5, characterized in that the calculation formula of the low-frequency concentrated feature is expressed as: Among them, F represents the Fourier spectrum, that is, the frequency spectrum function of the target image; u and v are the two axes in the frequency domain coordinate system, corresponding to the x and y directions in the Cartesian coordinate system in the time domain space; _um and _vn represent the coordinates of the peak on the two axes passing through the center point in the spectrum in the u and v directions. 7. A wafer surface defect detection device, characterized in that the device comprises: An acquisition module acquires a target image and a reference image; the reference image is a defect-free wafer surface image, and the target image is a wafer surface image with defects to be detected; An extraction module, extracting image features of the target image; the image features include gray level co-occurrence matrix features, histogram statistical features, image cross-correlation features, Fourier spectrum cross-correlation features, Hu invariant moment features and low frequency concentration features; the image cross-correlation features and the Fourier spectrum cross-correlation features are determined based on the reference image; The prediction module inputs the image features of the target image into a pre-trained machine learning model, performs defect detection processing on the image features, and outputs the defect detection results of the target image. 8. A computer device, characterized in that the device comprises: a processor, a memory and a bus, the memory stores machine-readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the wafer surface defect detection method as described in any one of claims 1 to 6.