CN-121982707-A - Wafer defect detection method and system

CN121982707ACN 121982707 ACN121982707 ACN 121982707ACN-121982707-A

Abstract

The application discloses a method and a system for detecting wafer defects, which relate to wafer detection and comprise the steps of collecting a holographic image IH and a digital image reconstruction distance parameter d on the surface of a wafer; the method comprises the steps of carrying out digital reconstruction on a holographic image IH by using a digital image reconstruction distance parameter d to obtain a reconstruction intensity image, inputting the reconstruction intensity image into a pre-trained convolutional neural network to carry out feature extraction to obtain a feature vector V, calculating the similarity between the feature vector V and a pre-calibrated defect-free sample feature vector Vn, judging that a wafer has defects when the similarity is smaller than a preset similarity threshold value, carrying out principal component analysis on the feature vector V of the defect sample when the defect is judged to exist, obtaining dimension-reduction coordinates, carrying out K-means cluster analysis on the dimension-reduction coordinates, and determining defect types according to the cluster results. Aiming at the problem that the wafer detection precision in the prior art is low, the method and the device improve the detection precision based on the wafer defect detection of combining the feature vector similarity analysis and the principal component clustering by digital holographic imaging.

Inventors

ZHENG FEI
HE ZHENFEI

Assignees

合肥图迅电子科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251217

Claims (9)

1. The wafer defect detection method is characterized by comprising the following steps: s1, acquiring a holographic image IH and a digital image reconstruction distance parameter d of the surface of a wafer; S2, carrying out digital reconstruction on the holographic image IH by utilizing the digital image reconstruction distance parameter d to obtain a reconstruction intensity diagram; S3, inputting the reconstructed intensity map into a pre-trained convolutional neural network for feature extraction to obtain a feature vector V; S4, calculating a feature vector V and a pre-calibrated defect-free sample feature vector Similarity of (2); S5, when the similarity is smaller than a preset similarity threshold, judging that the wafer has defects; S6, when judging that the defect exists, carrying out principal component analysis on the feature vector V of the sample with the defect to obtain a dimension-reducing coordinate; And carrying out K-means cluster analysis on the dimension reduction coordinates, and determining defect types according to the cluster results.
2. The method for detecting wafer defects according to claim 1, wherein: S2, reconstructing a distance parameter d for the holographic image by using the digital image Performing digital reconstruction to obtain a reconstructed intensity map, including: based on angular spectrum propagation algorithm, reconstructing distance parameter d for holographic image Fourier transforming to generate frequency spectrum diagram of frequency domain space ; From the spectrogram Extracting data of a preset frequency component area, and carrying out notch filtering treatment on the extracted data to obtain frequency spectrum data after filtering treatment ; For spectrum data Performing inverse Fourier transform to obtain complex digital holographic reconstruction ; Reconstruction of images from digital holograms And (3) separating the real component to obtain a reconstructed intensity map of the wafer.
3. The method for detecting wafer defects according to claim 2, wherein: based on angular spectrum propagation algorithm, reconstructing distance parameter d for holographic image Fourier transform is performed using the following formula: Wherein, the method comprises the steps of, In the case of a fourier transform operator, For the reference wavefront, In the case of a hologram, lambda is the wavelength of the incident light, d is the reconstruction distance parameter, In the form of a spatial domain coordinate, Corresponding frequency domain coordinates.
4. The method for detecting wafer defects according to claim 2, wherein: notch filtering the extracted data, including: Identifying the frequency point position corresponding to the normal texture structure of the wafer in the frequency spectrum data of the preset frequency component area; setting a plurality of notch areas at the positions of the frequency points, and setting the spectrum gain in the notch areas to be a preset value which is more than or equal to 0 and less than 1; Setting the gain of the spectral region outside the notch region to 1; constructing a notch filter according to the set gain, and filtering the spectrum data of the preset frequency component area by using the notch filter to obtain the spectrum data after the filtering process 。
5. The method for detecting wafer defects according to claim 2, wherein: s3, inputting the reconstructed intensity map into a pre-trained convolutional neural network for feature extraction to obtain a feature vector V, wherein the method comprises the following steps: Inputting the reconstructed intensity map into a pre-trained convolutional neural network, and extracting image features layer by layer through a convolutional layer and a pooling layer of the network; Removing a last classified output layer of the convolutional neural network, and acquiring the output of a characteristic layer before the classified layer as an initial characteristic vector; And carrying out normalization processing on the initial feature vector to obtain a feature vector V with unit length.
6. The method for detecting wafer defects according to any one of claims 2 to 5, wherein: when the wafer is detected to be surface roughness, presetting a similarity threshold value range of 0.96-0.98; when the wafer is detected as the pattern integrity detection, the preset similarity threshold value is in the range of 0.93 to 0.95.
7. The method for detecting wafer defects according to claim 6, wherein: S6, when judging that the defect exists, carrying out principal component analysis on the feature vector V of the sample with the defect to obtain a dimension-reducing coordinate; k-means cluster analysis is carried out on the dimension reduction coordinates, and defect categories are determined according to the cluster results, wherein the method comprises the following steps: performing principal component analysis on the feature vectors V of a plurality of defective samples, and extracting a first principal component And a second main component As two orthogonal directions; projecting the feature vector V of each defective sample to the first principal component And a second main component On the two-dimensional plane, two-dimensional dimension-reducing coordinates of each sample are obtained ; For two-dimensional dimension-reducing coordinates K-means clustering is carried out, and two-dimensional dimension reduction coordinates are divided into a plurality of clustering clusters; Projecting a feature vector V of a sample with a defect to be detected onto a two-dimensional plane to obtain a two-dimensional dimension reduction coordinate of the sample to be detected, and judging whether the two-dimensional dimension reduction coordinate of the sample to be detected falls into a cluster range corresponding to a known defect type: if the sample falls within the range of the known cluster, classifying the corresponding sample as a corresponding known defect type; If it does not fall within any of the known cluster ranges, the corresponding sample is marked as a new defect type.
8. The method for detecting wafer defects according to claim 7, wherein: The pretrained convolutional neural network is any one of ResNet, VGG, or EFFICIENTNET.
9. A system for detecting wafer defects, comprising: The image acquisition module acquires holographic images on the surface of a wafer And reconstructing a distance parameter d from the digital image; the image reconstruction module is used for carrying out digital reconstruction on the holographic image IH by utilizing the digital image reconstruction distance parameter d to obtain a reconstruction intensity image; The feature extraction module inputs the reconstructed intensity map into a pre-trained convolutional neural network for feature extraction to obtain a feature vector V, wherein the pre-trained convolutional neural network is any one of ResNet, VGG or EFFICIENTNET; The similarity calculation module calculates a feature vector V and a pre-calibrated defect-free sample feature vector Similarity of (2); the defect judging module judges that the wafer has defects when the similarity is smaller than a preset similarity threshold value, wherein the preset similarity threshold value range is 0.96-0.98 when the wafer is detected as surface roughness detection, and the preset similarity threshold value range is 0.93-0.95 when the wafer is detected as pattern integrity detection; And the defect classification module is used for carrying out principal component analysis on the feature vector V with the defect sample to obtain a dimension reduction coordinate when the defect exists, carrying out K-means cluster analysis on the dimension reduction coordinate, and determining the defect type according to the cluster result.

Description

Wafer defect detection method and system Technical Field The present application relates to the field of wafer inspection, and more particularly, to a method and a system for inspecting wafer defects. Background The rapid real-time detection of wafer surface defects is a key element in the modern semiconductor manufacturing, and is directly related to yield control and productivity improvement. The development of wafer defect detection technology has always been around the balance between accuracy, speed and cost. The background technology can be mainly divided into three categories, namely an optical detection technology, an electron beam detection technology and detection based on deep learning. The optical detection technology is the most widely applied technology with the fastest speed, and is also the main force for realizing 'fast real-time' detection. The core principle is that defects are found by analyzing the change of optical signals after the interaction (such as scattering, reflection and interference) of light and substances. The surface of the wafer is irradiated vertically or nearly vertically by the light source through bright field optical detection, and the detector receives optical signals in the regular reflection direction. A flat, defect-free area will reflect light back like a mirror to form a bright image, while areas with defects (e.g., scratches, bumps, particles) will change the reflected path of light, appearing as dark spots or abnormal contrast in the image. However, the sensitivity of the existing bright field optical detection to tiny particles is low, because scattered light of the particles is difficult to enter the objective lens in the positive direction, and the scattered light is easily interfered by background noise caused by the color and thickness change of the surface film layer of the wafer. On the other hand, the specular reflection light is too strong, so that the detector is easily saturated, and weak defect signals are covered. Dark field optical inspection is performed by tilting the light source at a very high angle (even nearly horizontal) to the wafer surface. For a perfectly smooth surface, all light will be reflected off the same angle and will not enter the vertically oriented objective lens, so the image background is dark. Only when the light irradiates on defects (such as tiny particles, rough spots), scattering occurs, wherein a part of the scattered light enters the objective lens to form bright spots on a dark background. Dark field optical detection is insensitive to flat, non-scattering defects (e.g., water stains, certain residual films). While providing limited information on the depth, shape, etc. of the defect. Basically, the method is ' blind detection ', only coordinates and light intensity signals of the defects can be given, the morphology of the defects cannot be directly seen ', real-time classification of the defects cannot be realized, and the identification and classification capabilities of the defects are weaker. The electron beam detection technique uses a focused electron beam to scan the wafer surface, and the electron interacts with the sample to generate signals such as secondary electrons, backscattered electrons, etc., which are collected by a detector to form a high resolution surface topography image. The resolution is extremely high, can reach the nanometer level, and can identify the tiny defects which cannot be resolved by the optical microscope. However, the detection speed is extremely slow, and the real-time or full-film detection cannot be realized at all. Usually, the chip is required to be carried out in a vacuum environment, and electrified damage can be caused to the chip, so that the equipment is expensive, and the operation cost is high. Deep learning based algorithms can train neural network models (e.g., CNN, U-Net, yolo, etc.) using a large number of defective and non-defective samples. The complex characteristics of the defects can be learned, and the detection capability of weak and irregular defects is far superior to that of the traditional algorithm. And the defect types can be directly and automatically classified during detection, so that the efficiency of subsequent analysis is greatly improved. Of course, deep learning-based algorithms require extensive data training, but in industrial processes, defect sample data is sparse and unbalanced, most products are good, resulting in collection of enough defect samples, the cost of each type of defect sample becomes extremely high, and labeling requires expertise and a lot of manpower input. On one hand, the model is easy to generate misjudgment in the face of unknown new defect types in training, on the other hand, the generalization capability of the model is insufficient, the calculation cost is high, and the black box characteristic of deep learning also makes it difficult for engineers to locate whether the problem is derived from data or environmen