Intelligent defect detection system based on laser diffraction imaging

Abstract

The invention discloses an intelligent defect detection system based on laser diffraction imaging and relates to the technical field of optical detection. The invention aims to solve the problems that a detection light beam is easily absorbed by optical elements such as a lens and the like and the optical resolution is low along with the shortening of a detection wavelength in optical detection. According to the intelligent defect detection system based on laser diffraction imaging, an optical diffraction unit is used for collecting diffraction information of the surface of a detected wafer and sending the diffraction information to a defect identification and classification unit driven by deep learning; and the deep learning driven defect identification and classification unit is used for performing feature extraction and classification on defects in the diffraction information by using a deep learning network model so as to realize detection of the detected wafer.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical detection. Background Art

[0002] The semiconductor industry has been continuously developing in the past few decades. More compact, faster, and more economical integrated circuits have become the key for countries to lead in the field of information science. As the feature size of semiconductor devices continues to shrink, surface micro-defects such as scratches, cracks, and defects are inevitably generated during the production and processing. If defective products enter the next process, it will not only have a great impact on the overall quality of the products, but also cause a huge waste of production costs. Wafer defects are mostly in the micron order of magnitude, and complex microscopic systems are needed for detection and classification. However, with the increase in the size of the wafer and the expansion of the scanning field of view, the detection speed of traditional microscopic systems is difficult to meet the requirements. Therefore, wafer automatic optical detection equipment with an ultra-large field of view and high detection efficiency has extremely high application prospects.

[0003] Currently, the most advanced optical detection can stably detect wafer defects smaller than 14 nm. Increasing the numerical aperture of the optical lens can improve the resolution. For future defect detection on wafers with smaller sizes, extreme ultraviolet light is required to further improve the optical resolution. However, as the detection wavelength shortens, the detection beam is easily absorbed by optical components such as lenses. Therefore, on the one hand, the detection system needs to minimize the participation of optical components, and on the other hand, it is extremely difficult to maintain extremely high optical resolution. Summary of the Invention

[0004] The present invention is to solve the problem that in optical detection, as the detection wavelength shortens, the detection beam is easily absorbed by optical components such as lenses, and the optical resolution is low. Now, a wafer defect detection system based on the combination of optical diffraction patterns and deep learning is provided, which improves the accuracy and sensitivity of wafer defect detection and can quickly classify wafer defects by identifying diffraction images.

[0005] An intelligent defect detection system based on laser diffraction imaging includes: an optical diffraction unit and a defect recognition and classification unit 7 driven by deep learning.

[0006] The optical diffraction unit is used to collect the diffraction information on the surface of the wafer to be measured and send it to the defect recognition and classification unit 7 driven by deep learning.

[0007] The defect recognition and classification unit 7 driven by deep learning is used to extract features and classify the defects in the diffraction information by using a deep learning network model, so as to realize the detection of the wafer to be measured.

[0008] Furthermore, the above-mentioned optical diffraction unit includes: a laser light source 1, a lens group 2, a beam splitter 3, an objective lens 4, and a detector 6.

[0009] The laser emitted by the laser light source 1 is incident on the beam splitter 3 through the lens group 2. The beam reflected by the beam splitter 3 is irradiated onto the surface of the wafer to be measured through the objective lens 4. The detector 6 is used to collect the diffraction information generated by the wafer to be measured under illumination.

[0010] Further, the above optical diffraction unit further includes a sample stage 5,

[0011] The sample stage 5 is a three-degree-of-freedom moving stage, which can drive the wafer to be measured to move in three mutually perpendicular directions.

[0012] Further, the above deep learning-driven defect identification and classification unit 7 includes: a deep learning network module 9,

[0013] The deep learning network module 9 performs feature extraction and classification on the defects in the diffraction information based on the deep learning network model.

[0014] Further, the above deep learning network model is trained using a training unit,

[0015] The training unit includes: an optical diffraction imaging simulation module and a deep learning network training module,

[0016] The optical diffraction imaging simulation module is used to generate a simulation data set,

[0017] The deep learning network training module is used to train the deep learning network model according to the simulation data set.

[0018] Further, the above optical diffraction imaging simulation module includes:

[0019] An optical diffraction modeling sub-module: models different types of defects and generates simulated diffraction images,

[0020] A simulation data generation sub-module: performs parametric processing on defect features to generate a simulation data set, and the parametric processing includes adjustment of defect size, morphology, and distribution characteristics,

[0021] A data augmentation sub-module: simulates different detection environments of the simulation data set.

[0022] Further, the above training unit further includes: an optical dual-mode imaging module,

[0023] The optical dual-mode imaging module is used to collect real defect image data of the wafer, and the real defect image data includes diffraction images and defect microscopic images.

[0024] Further, the above optical dual-mode imaging module includes:

[0025] Diffraction imaging sub-module: used to obtain the diffraction image of the wafer surface defects,

[0026] Bright-field microscopy imaging sub-module: used to collect the defect microscopy image corresponding to the diffraction image obtained by the diffraction imaging sub-module,

[0027] Data synchronization and alignment sub-module: align the diffraction image obtained by the diffraction imaging sub-module with the defect microscopy image collected by the bright-field microscopy imaging sub-module in space and time.

[0028] Furthermore, the above-mentioned deep learning-driven defect recognition and classification unit 7 further includes: a data preprocessing module 8,

[0029] The data preprocessing module 8 is used to preprocess the received diffraction information, and the preprocessing includes: normalization, noise reduction, and enhancement.

[0030] Furthermore, the above-mentioned deep learning-driven defect recognition and classification unit 7 further includes: a real-time detection visualization module 10,

[0031] The real-time detection visualization module 10 is used to display the detection results, and the detection results include defect location information and classification results.

[0032] The intelligent defect detection system based on laser diffraction imaging provided by the present invention demonstrates significant technical advantages and effects in multiple aspects, mainly reflected in the following aspects:

[0033] 1. High-precision defect detection ability

[0034] The present invention combines two imaging modes of lensless diffraction imaging and lens bright-field imaging, and can make full use of the advantages of each technology to obtain high-quality and high-resolution wafer defect images. Through the lensless diffraction imaging technology, the system can detect tiny defects by using the subtle differences in diffraction images without optical lenses; while the lens bright-field imaging provides an intuitive standard image for diffraction imaging, ensuring the accuracy and reliability of defect detection. The deep learning algorithm further improves the ability of defect recognition, and can accurately detect and classify various wafer surface defects, including scratches, particle contamination, microcracks, etc.

[0035] 2. High detection speed and automation level

[0036] The present invention greatly improves the efficiency of defect detection through an automated image acquisition and light source switching mechanism. The light source switching system controlled by a servo motor can automatically switch between different imaging modes, avoiding manual intervention and reducing the time for data acquisition. The deep learning model can process a large amount of image data in a short time and quickly give defect detection results, achieving efficient real-time detection. This is of great significance for the demand of defect detection in the semiconductor manufacturing process, which can accelerate the production cycle and improve production efficiency.

[0037] 3. Good adaptability and scalability

[0038] The present invention adopts a defect recognition method based on deep learning, making the system have strong adaptability. Through continuous training and optimization, the deep learning model can adaptively detect on different wafers and different defect types. As the system data volume increases, the model can gradually improve the detection ability for various types of defects and can be extended to other types of defect detection through training. In addition, the hardware design of the system has good scalability and can adjust and upgrade the hardware configuration according to production requirements.

[0039] 4. Reduced cost and complexity

[0040] By adopting two imaging techniques, lensless diffraction imaging and lens bright field imaging, the present invention reduces the dependence on high-cost optical lenses and lowers the overall cost of the equipment. The automated control of the light source switching mechanism also simplifies the operation process, reduces the need for manual intervention, and thus reduces the operation complexity. The introduction of the deep learning model further optimizes the detection process, reduces the complexity of traditional image processing and manual analysis, making the overall system have a higher cost performance.

[0041] 5. Large-scale data processing and real-time feedback ability

[0042] The automated data acquisition and processing system designed by the present invention can efficiently process large-scale data sets and provide real-time defect detection feedback. Through an optimized data processing flow and a deep learning model, the system can monitor and detect defects in real time on the wafer production line and output detection results in a timely manner. This is of great significance for improving the quality control of the production line and timely adjusting the production process.

[0043] 6. Adapt to a variety of industrial application scenarios

[0044] The present invention is not only applicable to the wafer defect detection in the semiconductor industry, but also can be extended to the defect detection applications in other fields. For example, industries such as microelectronics manufacturing and optical component production can utilize the detection system of the present invention to efficiently detect the surface defects of their products. In addition, the flexibility and scalability of the system enable it to adapt to production environments of different scales, and rapid and efficient defect detection can be achieved whether it is a large-scale production line or small-batch customized production. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the architecture of an intelligent surface defect detection system based on laser diffraction imaging.

[0046] Figure 2 It is a schematic diagram of the architecture of a defect detection algorithm training system based on deep learning for illustration.

[0047] Figure 3 It is an optical schematic diagram of a dual-mode imaging optical module in an embodiment for illustration.

[0048] Reference numerals in the figure: 1 laser light source, 2 lens group, 3 beam splitter, 4 objective lens, 5 sample stage, 6 detector, 7 deep learning-driven defect identification and classification unit, 8 data preprocessing module, 9 deep learning network module, 10 real-time detection visualization module, 11 adjustable laser emission source, 12 first industrial camera, 13 experimental lens, 14 tube lens, 15 second industrial camera, 16 automatic moving sample stage, 17 experimental objective lens, 18 first semi-reflective semi-transmissive lens, 19 second semi-reflective semi-transmissive lens, 20 LED light source. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0050] Bright-field lens imaging is mainly divided into three types: interference, diffraction, and scattering. The interference method detects defects through the interference induced by the phase difference between the reflected light and the reference light caused by the surface topography of the measured sample. The detection time increases for large-area detection and the engineering requirement indicators are very high; the diffraction method detects defects by using the diffraction image formed on the irradiated surface. The detection progress is affected by the established defect shape and size, and the applicable detection range is relatively limited; the scattering method utilizes the scattering characteristics of the defect part to the incident light to achieve defect detection. The scattering method is also used in the dark field. By removing the dark background formed by the direct reflected light and collecting the scattered light, defect information is obtained.

[0051] With the rapid development of deep learning and artificial intelligence technologies, it is necessary to explore new surface micro-nano defect detection technologies. An ideal technology should be able to automatically acquire and process images, automatically identify and classify defects, and should rely as little as possible on expensive optical devices and a large amount of computing resources. In physics, diffraction is a physical phenomenon when light passes through small holes or bypasses obstacles. Through theoretical derivation, the characterization of surface defect features in the diffraction image can be found, and then these features can be extracted using lensless imaging technology.

[0052] At the same time, deep learning technology has shown powerful capabilities in image recognition and classification. Using deep learning, diffraction images can be directly processed and defect classification can be carried out, thus avoiding the complex image reconstruction process and improving the recognition speed and accuracy at the same time. The combination of lensless imaging technology and deep learning technology has the potential to provide a new solution for surface defect detection. This method can reduce production costs and further improve the detection resolution, detection speed and intelligence level of automatic optical detection equipment.

[0053] Refer to Figure 1 and Figure 2 Specifically describing this embodiment, the intelligent defect detection system based on laser diffraction imaging described in this embodiment includes: an optical diffraction unit and a defect recognition and classification unit 7 driven by deep learning.

[0054] The optical diffraction unit is used to obtain the diffraction image of the surface defects of the wafer to be measured and encode the characteristic information reflecting the surface structure of the wafer to achieve efficient surface defect detection. The optical diffraction unit includes: a laser light source 1, a lens group 2, a beam splitter 3, an objective lens 4, a sample stage 5 and a detector 6. The laser light source 1 is used to provide stable laser light to stimulate the diffraction effect on the wafer surface. The lens group 2 is used to regulate the beam shape to ensure the focusing accuracy of the optical system. The beam splitter 3 is used to reflect the laser light to the sensor to improve the imaging quality. The objective lens 4 is used to focus the diffracted beam onto the detector or imaging sensor to ensure obtaining a high-resolution diffraction image. The sample stage 5 is used to fix the wafer to be measured and support multi-angle detection to improve the system adaptability. The detector 6 is used to collect the diffraction information formed by the sample under laser irradiation for the deep learning model to identify and classify defects.

[0055] The specific light transmission process is as follows:

[0056] The laser light emitted by the laser light source 1 passes through the lens group 2 and is incident on the beam splitter 3. The beam splitter 3 reflects the incident light and focuses the beam onto the wafer to be measured through the objective lens 4. The detector 6 collects the diffraction information formed by the wafer to be measured under laser irradiation and sends it to the defect recognition and classification unit 7 driven by deep learning.

[0057] The deep learning-driven defect recognition and classification unit 7 is used to process the collected diffraction information and classify defects, and it includes: a data preprocessing module 8, a deep learning network module 9, and a real-time detection visualization module 10. The data preprocessing module 8 is used to perform processing such as normalization, noise reduction, and enhancement on the diffraction information to improve the signal quality. The deep learning network module 9 extracts features and classifies defects in the diffraction information based on the trained deep learning network model. The real-time detection visualization module 10 is used to display the detection results, provide defect location information, and output the classification results.

[0058] In order to improve the accuracy and robustness of wafer surface defect detection and classification based on deep learning, it is necessary to improve the applicability of the deep learning model in actual detection tasks by constructing a training unit that combines simulation data and real data. The training unit consists of an optical diffraction imaging simulation module, an optical dual-mode imaging module, and a deep learning network training module. Each module works together to form a complete training process.

[0059] The optical diffraction imaging simulation module is used to generate a simulation data set for training the deep learning network model. The simulation data provided by the optical diffraction imaging simulation module can be used to initially train the deep learning model to enable it to have basic defect recognition capabilities. The optical diffraction imaging simulation module includes:

[0060] An optical diffraction modeling sub-module: Based on the diffraction theory of light, mathematical models are established for different types of defects, and diffraction images are generated through numerical calculation methods.

[0061] A simulation data generation sub-module: Parametric processing is performed on defect features, including defect size, morphology, distribution characteristics, etc., to generate diverse diffraction image data.

[0062] A data enhancement sub-module: In response to factors such as noise interference and light changes, various detection environments are simulated to improve the applicability of the data set.

[0063] The optical dual-mode imaging module is used to collect real defect image data, which includes diffraction images and defect microscopic images. The real data provided by this module is used for the refined training and verification of the model to improve the accuracy of defect detection. The optical dual-mode imaging module includes:

[0064] A diffraction imaging sub-module: Based on the principle of optical diffraction, diffraction images of wafer surface defects are obtained.

[0065] A bright-field microscopic imaging sub-module: Using bright-field microscopy technology to obtain corresponding defect microscopic images to provide annotation references for the deep learning model.

[0066] Data synchronization and alignment sub-module: Ensure the precise alignment of diffraction images and microscopic images in space and time, providing high-quality data for subsequent model training.

[0067] The deep learning network training module is used to train and optimize deep learning-based defect detection and classification algorithms. The model trained by this module can be used for defect identification and classification, achieving high-precision defect detection and classification. The deep learning network training module includes:

[0068] Data preprocessing sub-module: Process the input images such as normalization, noise removal, contrast enhancement, etc., to improve the model training effect.

[0069] Model training sub-module: Use a neural network model to perform feature extraction and classification training on diffraction images.

[0070] Model optimization sub-module: Improve the generalization ability of the model through means such as hyperparameter adjustment and adversarial training, enabling it to adapt to different types of defect detection tasks.

[0071] Embodiment

[0072] This embodiment proposes an intelligent surface defect detection system based on laser diffraction imaging. This system combines advanced optical detection technology and deep learning algorithms to improve the detection efficiency and accuracy of micro-nano scale defects on the wafer surface.

[0073] In this embodiment, the optical schematic diagram of the dual-mode imaging module responsible for collecting diffraction images and defect microscopic images is as Figure 3 shown. The dual-mode imaging module consists of two sub-modules: a bright-field lens imaging sub-module and a lensless diffraction imaging sub-module. These sub-modules are combined on the same optical path through a compact design, thus realizing dual-mode imaging. These two imaging modes require different light sources. Diffraction imaging requires a laser light source, while bright-field lens imaging requires an LED light source. And in each imaging mode, only one light source can illuminate the sample surface at a time, and the interference from other light sources needs to be eliminated. Therefore, a motor can be used to control the switching of the light sources. Specifically, the dual-mode imaging module includes:

[0074] An adjustable laser emission source 11 of 638.2nm: Used for the diffraction imaging mode. After irradiating the sample surface, it generates diffraction patterns, which carry the micro-nano structure information of the sample surface.

[0075] The first industrial camera 12: This camera has high resolution and high frame rate, and can capture fine defects on the sample surface, ensuring that the image quality is sufficient to support subsequent deep learning analysis.

[0076] Experimental lens 13: Used to adjust the laser beam shape.

[0077] Endoscope 14: Used for optical focusing and imaging in bright-field imaging mode, ensuring that the optical system in bright-field mode has sufficient depth of field and resolution to capture the surface details of the sample.

[0078] Industrial camera 2: Records the light intensity distribution information diffracted by the sample.

[0079] Automatically moving sample stage 16: The sample stage has three degrees of freedom, and the motion accuracy is at the micron level, which is used for rapid and automated scanning of the sample.

[0080] Experimental objective lens 17: Focuses the reflected diffraction information to improve the optical imaging quality.

[0081] First semi-reflective and semi-transmissive lens 18: Used to switch between the laser light source and the LED light source.

[0082] Second semi-reflective and semi-transmissive lens 19: Used to implement the bright-field lens imaging sub-module.

[0083] LED light source 20: Provides uniform illumination in bright-field imaging mode.

[0084] In this embodiment, in order to enable the industrial camera to automatically capture images in two modes, it is necessary to perform secondary development on the camera's SDK. The camera can receive signals from the motor, complete image acquisition after the light source is switched, and then send signals to switch the light source again. By repeating this process in a loop, continuous bright-field lens imaging and diffraction imaging at the same wafer position can be achieved.

[0085] In the process of defect recognition and classification based on deep learning, the bright-field lens image set serves as the gold standard, while the lensless diffraction image set serves as the input for training and testing the deep learning network.

[0086] In this embodiment, the defect recognition and classification module based on deep learning consists of a data preprocessing module, a deep learning network module, and a real-time detection and visualization module, which is used for defect recognition and classification of lensless diffraction images.

[0087] The specific implementation method is carried out according to the following steps:

[0088] Step 1, preprocess the optical diffraction imaging images collected by the industrial camera 2. The preprocessing operations include: image scaling, normalization, color space conversion, data augmentation, noise suppression, target area cropping, normalization processing, equalization, image denoising, label generation, etc.

[0089] Step 2, use a deep learning network (such as ResNet, ShuffleNet) to automatically extract the key features in the image, and based on the learned features, detect and classify the surface defects.

[0090] Step 3: Display the detected defect information in an intuitive manner, including marking the defect area, providing information such as defect type and confidence level. Provide real-time defect detection results so that the operator can immediately view the detection situation and make corresponding adjustments. Record the detection data for subsequent analysis and quality traceability.

[0091] Although the present invention has been described herein with reference to particular embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Accordingly, it should be understood that numerous modifications can be made to the exemplary embodiments, and other arrangements can be designed, provided that they do not depart from the spirit and scope of the invention as defined by the appended claims. It should be understood that the features described in the different dependent claims and herein can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. An intelligent defect detection system based on laser diffraction imaging, characterized in that, Comprising: an optical diffraction unit and a deep learning-driven defect recognition and classification unit (7), the optical diffraction unit is used to collect the diffraction information on the surface of the wafer to be measured and send it to the deep learning-driven defect recognition and classification unit (7), the deep learning-driven defect recognition and classification unit (7) is used to extract features and classify the defects in the diffraction information using a deep learning network model, thereby realizing the detection of the wafer to be measured.

2. The intelligent defect detection system based on laser diffraction imaging according to claim 1, wherein The optical diffraction unit includes: a laser light source (1), a lens group (2), a beam splitter (3), an objective lens (4), and a detector (6), the laser emitted by the laser light source (1) is incident on the beam splitter (3) through the lens group (2), the beam reflected by the beam splitter (3) is irradiated on the surface of the wafer to be measured through the objective lens (4), and the detector (6) is used to collect the diffraction information generated by the wafer to be measured under illumination.

3. The intelligent defect detection system based on laser diffraction imaging according to claim 2, wherein, The optical diffraction unit further includes a sample stage (5), the sample stage (5) is a three-degree-of-freedom moving stage and can drive the wafer to be measured to move in three mutually perpendicular directions.

4. The intelligent defect detection system based on laser diffraction imaging according to claim 1, 2 or 3, characterized in that The deep learning-driven defect recognition and classification unit (7) includes: a deep learning network module (9), the deep learning network module (9) extracts features and classifies the defects in the diffraction information based on a deep learning network model.

5. The intelligent defect detection system based on laser diffraction imaging according to claim 4, wherein, The deep learning network model is trained using a training unit, the training unit includes: an optical diffraction imaging simulation module and a deep learning network training module, the optical diffraction imaging simulation module is used to generate a simulation data set, the deep learning network training module is used to train the deep learning network model according to the simulation data set.

6. The intelligent defect detection system based on laser diffraction imaging according to claim 5, characterized in that, The optical diffraction imaging simulation module includes: an optical diffraction modeling sub-module: modeling different types of defects and generating simulated diffraction images, a simulation data generation sub-module: performing parametric processing on defect features to generate a simulation data set, and the parametric processing includes adjusting defect size, morphology, and distribution characteristics, a data augmentation sub-module: simulating different detection environments of the simulation data set.

7. The intelligent defect detection system based on laser diffraction imaging according to claim 5 or 6, characterized in that, The training unit further includes: an optical dual-mode imaging module, the optical dual-mode imaging module is used to collect real defect image data of the wafer, and the real defect image data includes a diffraction image and a defect microscopic image.

8. The intelligent defect detection system based on laser diffraction imaging according to claim 7, wherein The optical dual-mode imaging module includes: a diffraction imaging sub-module: used to obtain a diffraction image of the defects on the wafer surface, a bright-field microscopic imaging sub-module: used to collect a defect microscopic image corresponding to the diffraction image obtained by the diffraction imaging sub-module, a data synchronization and alignment sub-module: aligning the diffraction image obtained by the diffraction imaging sub-module with the defect microscopic image collected by the bright-field microscopic imaging sub-module in space and time.

9. The intelligent defect detection system based on laser diffraction imaging according to claim 4, wherein The deep learning-driven defect recognition and classification unit (7) further includes: a data preprocessing module (8), the data preprocessing module (8) is used to preprocess the received diffraction information, and the preprocessing includes: normalization, noise reduction, and enhancement.

10. The intelligent defect detection system based on laser diffraction imaging according to claim 9, characterized in that, The deep learning-driven defect recognition and classification unit (7) further includes: a real-time detection visualization module (10), The real-time detection visualization module (10) is used to display the detection results, and the detection results include defect location information and classification results.

Images