A semiconductor processing defect detection method based on artificial intelligence

Abstract

The application discloses a kind of semiconductor processing defect detection methods based on artificial intelligence, it is related to semiconductor processing defect detection technical field.The method includes: obtaining wafer position coordinates, optical image and process equipment dynamic parameter data;Utilize multiscale Gaussian filter model and asymmetric convolution algorithm to generate microdefect feature matrix;Through space-time mapping and correlation matrix algorithm, process parameters are converted into position-related process parameter matrix;Radial pooling and multi-expansion rate hollow convolution algorithm are executed to fusion feature tensor, and global and local feature vector are extracted;Adopt polar coordinate dynamic weighted fusion algorithm to generate defect probability distribution graph;Through adaptive threshold algorithm, output defect coordinates and type identification, and generate process equipment control signal according to predefined logic.The application realizes defect accurate identification and process closed-loop control by innovative multi-source data fusion calculation model, improves detection precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of semiconductor processing defect detection technology, and more specifically, to an artificial intelligence-based semiconductor processing defect detection method. Background Technology

[0002] In the semiconductor manufacturing industry, accurate detection of wafer processing defects is a core element in ensuring product yield and reliability, and its detection accuracy and efficiency have a significant impact on the development quality of the semiconductor industry.

[0003] Traditional defect detection methods rely on single optical image acquisition and basic image processing algorithms, generally suffering from low defect recognition rates, inaccurate location, and weak correlation with process parameters. To improve this situation, the industry has seen the emergence of detection technologies that combine multi-source data, achieving preliminary fusion of images with some process parameters and basic defect recognition functions. However, in practical use, these technologies still have some shortcomings: Firstly, existing technologies lack effective mathematical models to establish high-precision spatiotemporal correlation between process parameters and wafer positions, resulting in insufficient depth of multi-source data fusion. Secondly, in feature extraction, most methods employ single image processing operators, failing to construct complex multi-scale feature extraction algorithms and thus unable to fully capture the multi-dimensional features of micro-defects.

[0004] On the other hand, existing technologies lack efficient computational frameworks for feature fusion and defect analysis. The generation of defect probability distributions relies on simple weighting and does not employ dynamic weighted fusion algorithms based on spatial coordinates. Furthermore, the defect classification threshold is fixed and lacks adaptive decision-making algorithms. In addition, there is a lack of closed-loop feedback mechanisms based on data-driven logic between detection results and process control. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides an artificial intelligence-based semiconductor processing defect detection method, which solves the problems mentioned in the background art through the following scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a semiconductor processing defect detection method based on artificial intelligence, comprising the following steps:

[0007] S1: Obtain wafer position coordinate dataset, wafer surface optical image dataset, and process equipment dynamic parameter dataset;

[0008] S2: Perform a mathematical transformation algorithm based on multi-scale Gaussian filtering and asymmetric convolution kernel on the optical image dataset to generate a micro-defect feature matrix; perform a modeling algorithm based on spatiotemporal mapping and correlation matrix on the dynamic parameter dataset to generate a process parameter matrix associated with the wafer position;

[0009] S3: The micro-defect feature matrix and the process parameter matrix are tensor concatenated to construct a fused feature tensor; the fused feature tensor is subjected to a dual-path feature extraction algorithm, including extracting the wafer global feature vector through radial pooling algorithm and extracting the local defect feature vector through multi-expansion rate hole convolutional neural network algorithm.

[0010] S4: Based on the wafer position coordinate dataset, a dynamic weighting algorithm related to polar coordinates is used to fuse the global feature vector and the local defect feature vector, and a defect probability distribution map is output by calculating through a convolutional layer classifier;

[0011] S5: The defect probability distribution map is analyzed using an adaptive threshold algorithm, and the defect coordinates and defect type identifiers are output. The process equipment control signals are generated according to the predefined mapping logic between defect types and control signals.

[0012] Preferably, the wafer position coordinate dataset includes:

[0013] Based on the photoelectric encoded pulse signal of the wafer stage and the mechanical positioning signal of the Notch angle, a coordinate mapping algorithm is executed to establish a polar coordinate grid model with the wafer center as the pole.

[0014] Preferably, the mathematical transformation algorithm based on multi-scale Gaussian filtering and asymmetric convolution kernels includes:

[0015] Calculate the pixel values ​​of the output image Where N is the number of scales and k is the scale index. Let be the weight of the k-th scale. Given the pixel value of the input image at coordinates (x, y), The variance is The Gaussian kernel function, where * denotes convolution operation; for the processed I out (x, y) Applying asymmetric convolution kernel operation Output the element values ​​of the micro-defect feature matrix, where K L It uses a 5x5 asymmetric convolution kernel.

[0016] Preferably, the modeling algorithm based on spatiotemporal mapping and correlation matrix includes:

[0017] For RF power timing data, the corresponding angular interval is determined by using an angular interval mapping function based on time t and wafer rotation speed. Where t is time, f rot N represents the wafer rotation speed. θ Number of angular partitions;

[0018] Construct a thermodynamic matrix of process parameters from the pressure differential data. Where α is the attenuation coefficient, The distance between intervals. Let M be the pressure difference between regions i and j, and D be the maximum correlation distance; let the thermodynamic matrix M be... pro (i, j) radial index Expanding it into a two-dimensional matrix forms a process parameter matrix.

[0019] Preferably, the radial pooling algorithm is expressed as follows:

[0020] Average pooling operation is performed on the fused feature tensor over radial intervals. ,in Radial index, For feature channel index, Radial index The total number of pixels within the corresponding interval is used to extract the global feature vector of the wafer.

[0021] Preferably, the multi-dilation rate dilated convolutional neural network algorithm adopts a three-layer dilated convolutional structure with a dilation rate sequence of [2, 4, 8], and introduces a position coding difference block after the convolution output. The input of the position coding difference block is the parameters of the polar coordinate grid model.

[0022] Preferably, the dynamic weighting algorithm related to polar coordinates calculates the fusion features as follows:

[0023] ,in For polar coordinate parameters, where the radial weight function is... R is the wafer radius of 50mm, and the local weighting function is... coordinate parameters ;

[0024] The convolutional layer outputs a probability defect distribution map. C 3*3 This is a 3x3 convolution operation, where c represents the defect category.

[0025] Preferably, the adaptive threshold algorithm has the following threshold calculation model:

[0026] Adaptive threshold , where c is the defect category, k is the sensitivity coefficient, μ is the probability mean corresponding to the defect category, and σ is the probability standard deviation corresponding to the defect category;

[0027] The defect probability distribution map is segmented based on the threshold, and the centroid coordinates of the defect connected domain Ω are calculated as the defect coordinates.

[0028] Preferably, the judgment logic for the defect type identifier includes:

[0029] When the probability value P1 of the defect region in channel 1 of the probability defect distribution map is greater than 0.7 and the average height h within the defect contour is greater than 100nm, it is identified as an etching residue.

[0030] When the probability value P2 of channel 2 is greater than 0.6 and the maximum width w is less than 500nm, it is identified as a metal bridge.

[0031] When the probability value P3 of channel 3 is greater than 0.5 and the ratio of the long side to the short side of the circumscribed rectangle of the defect is greater than 10, it is marked as a deep groove scratch.

[0032] If multiple defect type identifiers are satisfied simultaneously, the defect type identifier corresponding to the highest probability value is selected.

[0033] Preferably, the mapping logic between the predefined defect types and control signals includes:

[0034] When the defect is identified as an etching residue, an instruction to increase the RF power supply duty cycle and an instruction to increase the opening of the fluorocarbon gas valve are generated based on the functional relationship between the defect height and area.

[0035] When the defect is identified as a metal bridge, an instruction to decrease the RF power supply duty cycle and an instruction to increase the argon purge duration are generated based on the functional relationship between the reciprocal of the defect width and the number of defects.

[0036] When the defect is identified as a deep groove scratch, a command to increase the pressure in the reaction chamber is generated based on the logarithmic relationship of the defect's length-to-width ratio.

[0037] The technical effects and advantages of this invention are as follows:

[0038] 1. This invention constructs a feature extraction algorithm that combines multi-scale Gaussian filtering with asymmetric convolution, which can achieve high-precision mathematical representation of defect features at different scales in multi-source optical image data, breaking through the limitations of traditional single image processing operators;

[0039] 2. This invention achieves efficient fusion of non-image data and spatial location by establishing a spatiotemporal mapping model of process parameters and a correlation thermodynamic matrix algorithm, providing a core data foundation for the calculable tracing of defect causes;

[0040] 3. This invention constructs a complete, data-driven intelligent computing scheme for defect detection and classification by designing a computational framework that combines dual-path feature extraction and dynamic weighted fusion with an adaptive threshold decision algorithm, which significantly improves the automation level and decision accuracy of detection. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0042] Figure 2This is a schematic diagram of the feature processing and fusion structure of the present invention;

[0043] Figure 3 This is a schematic diagram of the defect decision-making and control structure of the present invention; Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] As attached Figure 1 -4 illustrates an artificial intelligence-based semiconductor processing defect detection method, which includes:

[0046] S1: Obtain wafer position coordinate dataset, wafer surface optical image dataset, and process equipment dynamic parameter dataset;

[0047] It should be further explained that the main reason why an artificial intelligence-based semiconductor processing defect detection method chooses wafers as the detection object is that, as the core carrier of semiconductor manufacturing, the defects of wafers directly determine the quality of the final chip, and their structural characteristics and manufacturing process enable efficient and accurate defect detection and process optimization by using wafers as the detection object.

[0048] It should be further noted that the wafer stage is equipped with an X / Y axis photoelectric encoder with a resolution of ≥0.1 micrometers, supports pulse signal output with a frequency of ≥1MHz, and a Notch angle mechanical positioning sensor with an accuracy of ±0.5°. The stage supports rotational motion with a default rotational speed of 5r / s. The wafer size is 100mm.

[0049] It should be specifically noted that the wafer position coordinate system is equipped with a real-time pulse signal acquisition system via an X / Y axis photoelectric encoder. Each pulse corresponds to a displacement of 0.1 micrometers, and the cumulative pulse value (X) is recorded. pulse Y pulse When the notch angle mechanical positioning sensor detects the wafer notch angle (positioning notch), it triggers a calibration command, adjusting the current coordinates (X... pulse ,Y pulseSet the temporary origin (X0, Y0); based on a wafer diameter of 100mm, the center coordinates are (X0 + 500000 pulses, Y0 + 500000 pulses); with the center as the pole, the radial radius r is in μm, the angular radius θ is in degrees, and the grid accuracy is 1μm * 0.1°, meaning each grid corresponds to an area of ​​(r ± 0.5μm, θ ± 0.05°). Rotate the wafer 360° using the stage, and record the number of times the notch angle passes the sensor. This number should match the number of rotations. If the deviation is > 1°, recalibrate the origin coordinates.

[0050] It should be specifically noted that the deployment method of the equipment for acquiring the optical scattering image of the wafer surface is as follows: a CCD camera is installed, with the lens vertically aligned with the center of the wafer, the focal length adjusted to achieve a clear image, the object distance is 200mm, and the depth of field is ≥5mm; the dark field illumination system is arranged with 4 light sources, symmetrically distributed around the wafer, at a 45° angle to the wafer surface to prevent direct light from entering the camera; the initial angle of the polarizer is set to 0°, and can be switched to 90°, 45°, and 135° through software control.

[0051] It should be further explained that when acquiring dark field scattering images of the wafer surface, the stage rotates at 5 r / s, and the camera is triggered to acquire one frame of image every 1° of rotation, for a total of 360 frames. Each frame of image has a resolution of 1024*1024 pixels and a grayscale value range of 0-255. When storing, the images are associated with the corresponding polar coordinates (r, θ). After acquisition, Gaussian filtering (σ=1.0) is performed to reduce noise and preserve the image edge information.

[0052] When acquiring polarization-resolved bright-field images, the stage stops rotating, and bright-field illumination is switched to a wavelength of 550 nm. The polarizer angles are adjusted sequentially to 0°, 90°, 45°, and 135°, with one frame acquired for each angle, for a total of four frames. Histogram equalization is then performed on the images to enhance the contrast of micro-defects.

[0053] It should be further noted that if the proportion of noisy pixel grayscale values ​​>240 or <15 in a single frame image is >5%, the image at that angle should be re-acquired; if three consecutive frames are abnormal, the stability of the light source should be checked and the filter replaced.

[0054] It should be specifically noted that the deployment method of the dynamic parameter acquisition equipment of the process equipment is as follows: the radio frequency power sensor is connected to the radio frequency output terminal of the plasma etching machine, the sampling line length is ≤1m to reduce signal attenuation; six pressure sensors are evenly arranged on the inner wall of the reaction chamber at positions of 0°, 60°, 120°, 180°, 240° and 300°, 50mm away from the wafer surface, and one spare sensor is preset and arranged at position 30°.

[0055] It should be further noted that the sampling frequency of the RF power fluctuation timing data is 1kHz, and one data point is recorded every 1ms, including a timestamp t accurate to ms and a power value P.rf Unit: W; continuous acquisition for 10 seconds, covering two rotations of the wafer, ensuring synchronization between data and image acquisition time; storage format: [t, P] rf ] array.

[0056] The multi-zone pressure difference value of the reaction chamber was calculated by collecting the pressure values ​​of 6 zones every 5ms. There are 15 differences in total; the data is aligned with the RF power data via timestamps to ensure that parameters at the same time can be correlated.

[0057] It should be further noted that if the power value exceeds the normal range of 10-1000W at a certain moment, it is marked as an abnormal value and replaced with the average value of the previous 10ms; if the barometric pressure sensor data is missing, the backup sensor is activated.

[0058] S2: Perform a mathematical transformation algorithm based on multi-scale Gaussian filtering and asymmetric convolution kernel on the optical image dataset to generate a micro-defect feature matrix; perform a modeling algorithm based on spatiotemporal mapping and correlation matrix on the dynamic parameter dataset to generate a process parameter matrix associated with the wafer position.

[0059] It should be specifically noted that the method for generating the micro-defect feature matrix includes:

[0060] For the 364 frames of images acquired (360 frames of dark field + 4 frames of polarized bright field), the pixel value of the output image is calculated pixel by pixel. Where N is the number of scales and k is the scale index. Let be the weight of the k-th scale. Given the pixel value of the input image at coordinates (x, y), The variance is The Gaussian kernel function, where * denotes convolution operation;

[0061] For the processed I out (x, y) Applying asymmetric convolution kernel operation Output the element values ​​of the micro-defect feature matrix, where K L The convolution kernel is 5x5 asymmetric. During convolution, if the window exceeds the image boundary, mirror padding is used to avoid boundary effects. The M values ​​of 364 frames are associated by polar coordinates (r, θ). def (x, y) forms a three-dimensional micro-defect feature matrix M. def (r, θ, frame), where frame ranges from 1 to 364, and each element represents the intensity of the micro-defect feature at the corresponding position. The higher the value, the greater the probability that a defect may exist.

[0062] It should be further noted that the number of scales N=5, and the Gaussian kernel variance σ at each scale is... k: σ1=1.0, σ2=2.0, σ3=3.0, σ4=4.0, σ5=5.0; scale weight ω k Normalized calculations yielded ω1=0.067, ω2=0.133, ω3=0.2, ω4=0.267, and ω5=0.333; the 5x5 asymmetric convolution kernel K... L for It is used to enhance the edge features of micro-defects.

[0063] It should be further explained that the variance of the Gaussian kernel is positively correlated with the filtering scale. The smaller the variance, the sharper the kernel function, and the stronger its ability to capture details of small defects such as tiny scratches and small assembly residues. Conversely, the larger the variance, the smoother the kernel function, which is more suitable for extracting the overall features of larger defects. Setting the variance in increments from 1.0 to 5.0 can achieve comprehensive coverage of defect features at different scales. In practical applications, the number of scales N=5 is a choice that balances detection accuracy and computational complexity. Too many scales will increase the amount of data processing and reduce real-time performance; too few scales may miss some defect features. The arithmetic progression from 1.0 to 5.0 can evenly cover the commonly used defect size range within a limited scale, ensuring effective response to micro-defects of different sizes.

[0064] The initial calculation of scale weights uses the Gaussian kernel variance σ corresponding to each scale. k Based on this, with the scale number N=5, the sum of variances is calculated first. It is 15, then according to the formula The weight values ​​for each scale are obtained.

[0065] It should be specifically noted that the method for generating the process parameter matrix includes:

[0066] For RF power timing data, the corresponding angular interval is determined by using an angular interval mapping function based on time t and wafer rotation speed. Where t is time, f rot N represents the wafer rotation speed. θ Let t*f be the number of angular partitions. rot *360 is the wafer rotation angle. mod360 ensures the angle is within the range of 0-360°. 360 / N θ The angular span of each interval;

[0067] Construct a thermodynamic matrix of process parameters from the pressure differential data. Where α is the attenuation coefficient, The distance between intervals. Let M be the pressure difference between regions i and j, and D be the maximum correlation distance; let the thermodynamic matrix M be... pro (i, j) radial index (0-50mm) is expanded into a two-dimensional matrix and matched with the wafer polar coordinates (r,θ) to form a process parameter matrix.

[0068] It should be further explained that the wafer rotation speed f rot The speed is 5 r / s, and the number of angular partitions is N. θ =36, intervals of 10°, standard interval mapping function. The pressure difference values ​​are normalized to [-1, 1]; the attenuation coefficient α is 0.1, which preserves the weak correlation across intervals while highlighting the strong correlation between adjacent intervals, consistent with the spatial gradient characteristics of process parameters; radial index It is 0-50mm.

[0069] S3: The micro-defect feature matrix and the process parameter matrix are concatenated into tensors to construct a fused feature tensor; the fused feature tensor is subjected to a dual-path feature extraction algorithm, including extracting the wafer-wide feature vector through radial pooling algorithm and extracting the local defect feature vector through multi-expansion-rate hole convolutional neural network algorithm.

[0070] It should be specifically noted that the fused feature tensor initially has two channels, and the micro-defect feature matrix M is divided according to the pixel position (x, y). def (x, y) and process parameter matrix M pro (x, y) are concatenated to form a fused feature tensor. Where c=1 and 2 correspond to micro-defects and process parameters, respectively; each channel is normalized separately, with micro-defect channel 1, Process parameter channel 2 M min M max These represent the maximum and minimum values ​​of the corresponding matrix.

[0071] It should be further explained that the reason for expanding to 32 channels using a 1*1 convolutional kernel is to enhance the feature dimensionality while preserving the original feature correlation, so as to adapt to the multi-scale extraction requirements of subsequent dilated convolution. During the expansion process, the convolutional kernel weights are optimized through training to ensure that channels 1-16 mainly retain the optical features of micro-defects, and channels 17-32 mainly retain the spatial distribution features of process parameters.

[0072] It should be specifically noted that the method for extracting the wafer-wide feature vector includes:

[0073] Average pooling operation is performed on the fused feature tensor over radial intervals. ,in Radial index, For feature channel index, Radial index The total number of pixels within the corresponding interval is used to extract the wafer-wide feature vector; radial index. Divided into 10 intervals, each interval being 5mm in length: =1 (0-5mm) =2 (5-10mm), ... =10 (45-50mm), Feature Channel Index =1,2 (corresponding to the two channels of the fused tensor); the pooling results of the 10 radial intervals are then... =1 to The sequence of 10 is used to form a global feature vector with a dimension of 20 (10 intervals * 2 channels). The 2-channel fused feature tensor is expanded in dimension using a 1*1 convolution kernel with 32 kernels, mapping the number of channels from 2 to 32. This achieves matching with the input of subsequent dilated convolutions, and the expansion process preserves the feature correlation between micro-defects and process parameters.

[0074] It should be specifically noted that the method for extracting the local defect feature vector includes:

[0075] A dilated convolution sequence with an expansion ratio of [2, 4, 8] is used. After the dilated convolution output, a position encoding difference block is connected. The input is the polar coordinate grid parameters of the wafer position coordinate system.

[0076] It should be further explained that the dilated convolution sequence consists of 3 layers of 3*3 convolution kernels with dilation rates of 2, 4, and 8 respectively; the number of input channels is 32 and the number of output channels is 64. Each convolution layer is activated by ReLU.

[0077] The first layer pairs the fused feature tensor T fus Applying a 3x3 dilated convolution with a dilation rate of 2, the output channel count is 16. Where k is the core size, b1 is the expansion rate, and b1 is the bias term initialized to 0.01.

[0078] The second layer applies a 3x3 dilated convolution with a dilation rate of 4 to C1, resulting in 32 output channels. .

[0079] The third layer applies a 3x3 dilated convolution with an inflation rate of 8 to C2, resulting in 64 output channels. .

[0080] Adding polar coordinate encoding to the convolutional output enhances positional relevance. , for F loc Calculate the average value of each channel (x, y, c) to obtain the local defect feature vector V. loc The dimension is 64, which is consistent with the number of output channels.

[0081] S4: Based on the wafer position coordinate dataset, a dynamic weighting algorithm related to polar coordinates is used to fuse the global feature vector and the local defect feature vector, and a defect probability distribution map is output by a convolutional layer classifier.

[0082] It should be further explained that the dynamic weighting algorithm related to polar coordinates calculates the fusion features as follows:

[0083] ,in For polar coordinate parameters, where the radial weight function is... R is the wafer radius of 50mm, and the local weighting function is... coordinate parameters .

[0084] It should be further explained that the method for generating the defect probability distribution map includes:

[0085] F fusion feature F fusion Input a 3x3 convolutional layer and output a 3-channel defect probability distribution map. Where c=1 (etching residue), c=2 (metal bridging), and c=3 (deep groove scratch), each pixel value represents the probability that the location belongs to the corresponding defect, 0 indicates no defect, and 1 indicates that it is definitely the defect.

[0086] S5: The defect probability distribution map is analyzed using an adaptive threshold algorithm, and the defect coordinates and defect type identifiers are output. The process equipment control signals are generated according to the predefined mapping logic between defect types and control signals.

[0087] It should be specifically noted that the output process for the defect coordinates and defect type identifier is as follows:

[0088] Calculate the segmentation threshold for each defect category For Map pb Pixel values ​​in (x, y, c) > T c The regions are marked as candidate defect areas. Using the 8-neighborhood criterion, adjacent pixels that are both candidate areas are considered the same connected component. The defect connected component Ω is extracted, and the defect centroid coordinates are determined. , where P c (x, y) represents the defect probability value at that location.

[0089] It should be further explained that k is the sensitivity coefficient with a value of 1.2, µ is the mean probability of defect category, μ1=0.5 (etching residue), μ2=0.4 (metal bridging), μ3=0.3 (deep groove scratch), and σ is the probability standard deviation, σ1=0.1, σ2=0.1, σ3=0.1.

[0090] It should be further explained that the defect type determination method is as follows:

[0091] Etching residue: P1>0.7 within the connected domain. The average defect height h>100nm is inferred from the dark field scattering intensity. The scattering intensity is positively correlated with h². The preset intensity>500 corresponds to h>100nm.

[0092] Metallic bridging: P2 > 0.6 within the connected domain, and the maximum width of the bounding box w < 0.0051 pixels.

[0093] Deep groove scratches: P3>0.5 within the connected domain, and the ratio of the long side to the short side of the circumscribed rectangle of the defect is >10.

[0094] If multiple conditions are met simultaneously, the category with the highest probability value is selected as the identifier.

[0095] It should be further explained that the average height h within the defect contour was measured by randomly selecting 10 sampling points within the connected domain of the defect using an atomic force microscope and then taking the average value; the maximum width w is the length of the short side of the circumscribed rectangle of the defect, which was calculated after extracting the contour using the Canny operator image edge detection algorithm.

[0096] It should be specifically noted that the generation process of the process equipment control signal includes:

[0097] When the received defect identifier is an etching residue identifier, an RF power supply duty cycle incrementing pulse signal is generated, and the duty cycle increment is positively correlated with the defect height; a fluorocarbon gas valve opening step command is output, and the opening increment is positively correlated with the defect area.

[0098] It should be further noted that the output power range of the RF power supply duty cycle increment pulse signal is 10-1000W. The increment calculation is based on the positive correlation function between the current power value and the defect height, ensuring that the adjusted power does not exceed the normal operating range.

[0099] When the received defect identifier is a metal bridging identifier, an RF power supply duty cycle decreasing pulse signal is generated, and the duty cycle decrease is positively correlated with the reciprocal of the defect width; an argon purge control command is triggered, and the purge duration is positively correlated with the number of defects.

[0100] When the defect identifier at the receiving site is a deep groove scratch, an output command to increase the pressure in the reaction chamber is issued, and the pressure increase is positively correlated with the logarithm of the defect's length-to-width ratio.

[0101] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0102] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting defects in semiconductor processing based on artificial intelligence, characterized by, include: S1: Obtain wafer position coordinate dataset, wafer surface optical image dataset, and process equipment dynamic parameter dataset; S2: Perform a mathematical transformation algorithm based on multi-scale Gaussian filtering and asymmetric convolution kernel on the optical image dataset to generate a micro-defect feature matrix; A modeling algorithm based on spatiotemporal mapping and correlation matrix is ​​executed on the dynamic parameter dataset to generate a process parameter matrix associated with the wafer location; S3: The micro-defect feature matrix and the process parameter matrix are tensor concatenated to construct a fused feature tensor; the fused feature tensor is subjected to a dual-path feature extraction algorithm, including extracting the wafer global feature vector through radial pooling algorithm and extracting the local defect feature vector through multi-expansion rate hole convolutional neural network algorithm. S4: Based on the wafer position coordinate dataset, a dynamic weighting algorithm related to polar coordinates is used to fuse the global feature vector and the local defect feature vector, and a defect probability distribution map is output by calculating through a convolutional layer classifier; The wafer position coordinate dataset employs a dynamic weighting algorithm related to polar coordinates, and its fusion feature calculation method is as follows: where is the polar coordinate parameter, where the radial weight function R is the wafer radius 50mm, the local weight function is the coordinate parameter , is the wafer global feature vector, is the local defect feature vector; Outputting a probability defect distribution map through a convolutional layer where C 3*3 is a 3*3 convolution operation and c is a defect category S5: The defect probability distribution map is analyzed using an adaptive threshold algorithm, and the defect coordinates and defect type identifiers are output. The process equipment control signals are generated according to the predefined mapping logic between defect types and control signals. 2.The AI-based semiconductor processing defect detection method of claim 1, wherein: The wafer location coordinate dataset includes: Based on the photoelectric encoded pulse signal of the wafer stage and the mechanical positioning signal of the Notch angle, a coordinate mapping algorithm is executed to establish a polar coordinate grid model with the wafer center as the pole. 3.The method of claim 1, wherein: The mathematical transformation algorithm based on multi-scale Gaussian filtering and asymmetric convolution kernels includes: Calculate the pixel values ​​of the output image Where N is the number of scales and k is the scale index. Let be the weight of the k-th scale. Given the pixel value of the input image at coordinates (x, y), The variance is The Gaussian kernel function, where * denotes convolution operation; for the processed I out (x, y) Applying asymmetric convolution kernel operation Output the element values ​​of the micro-defect feature matrix, where K L It uses a 5x5 asymmetric convolution kernel. 4.The method of claim 1, wherein: The modeling algorithm based on spatiotemporal mapping and correlation matrix includes: For the radio frequency power time sequence data, according to time t and wafer rotation speed, the corresponding angular interval is determined through an angular interval mapping function Wherein t is time, f rot is wafer rotation speed, N θ is the number of angular partitions; Construct a thermodynamic matrix of process parameters from the pressure differential data. Where α is the attenuation coefficient, The distance between intervals. Let M be the pressure difference between regions i and j, and D be the maximum correlation distance; let the thermodynamic matrix M be... pro (i, j) radial index Expanding it into a two-dimensional matrix forms a process parameter matrix.

5. The method of claim 1, wherein the method is based on artificial intelligence. The radial pooling algorithm is expressed as follows: Performing average pooling operation on fusion feature tensor by radial interval wherein is a radial index, is a feature channel index, denotes a pixel index variable, is a radial index the total number of pixels in the corresponding interval, thereby extracting the wafer global feature vector.

6. The method of claim 1, wherein the method is based on artificial intelligence. The multi-dilation rate dilated convolutional neural network algorithm adopts a three-layer dilated convolutional structure with a dilation rate sequence of [2, 4, 8], and introduces a position coding difference block after the convolution output. The input of the position coding difference block is the parameters of the polar coordinate grid model.

7. The method of claim 1, wherein the method is based on artificial intelligence. The adaptive threshold algorithm uses the following threshold calculation model: Adaptive threshold where c is a defect class, k is a sensitivity coefficient, μ is a probability mean corresponding to the defect class, and σ is a probability standard deviation corresponding to the defect class. The defect probability distribution map is segmented based on the threshold, and the centroid coordinates of the defect connected domain Ω are calculated as the defect coordinates.

8. The semiconductor processing defect detection method based on artificial intelligence according to claim 7, characterized in that: The judgment logic for the defect type identifier includes: When the probability value P1 of the defect region in channel 1 of the probability defect distribution map is greater than 0.7 and the average height h within the defect contour is greater than 100nm, it is identified as an etching residue. When the probability value P2 of channel 2 is greater than 0.6 and the maximum width w is less than 500nm, it is identified as a metal bridge. When the probability value P3 of channel 3 is greater than 0.5 and the ratio of the long side to the short side of the circumscribed rectangle of the defect is greater than 10, it is marked as a deep groove scratch. If multiple defect type identifiers are satisfied simultaneously, the defect type identifier corresponding to the highest probability value is selected. 9.The AI-based semiconductor processing defect detection method of claim 1, wherein: The predefined mapping logic between defect types and control signals includes: When the defect is identified as an etching residue, an instruction to increase the RF power supply duty cycle and an instruction to increase the opening of the fluorocarbon gas valve are generated based on the functional relationship between the defect height and area. When the defect is identified as a metal bridge, an instruction to decrease the RF power supply duty cycle and an instruction to increase the argon purge duration are generated based on the functional relationship between the reciprocal of the defect width and the number of defects. When the defect is identified as a deep groove scratch, a command to increase the pressure in the reaction chamber is generated based on the logarithmic relationship of the defect's length-to-width ratio.

Images