A method for detecting and analyzing concentric circle defects of N-type monocrystalline silicon wafer

Abstract

A method for detecting and analyzing concentric circle defects in N-type monocrystalline silicon wafers belongs to the field of monocrystalline silicon quality inspection technology. This method collects the V / G ratio parameters of the crystal pulling process, the radial distribution data of oxygen content in Fourier transform infrared spectroscopy, and the two-dimensional minority carrier lifetime spectrum obtained by microwave photoconductivity attenuation method of the N-type monocrystalline silicon wafer under test, constructing a multi-dimensional dataset. The minority carrier lifetime image is then processed sequentially with circle center localization, polar coordinate transformation, defect ring detection, frequency domain feature extraction, and texture feature extraction to obtain image feature vectors containing defect morphology and distribution patterns. After preprocessing the three types of heterogeneous data, multi-source feature fusion is performed to construct a three-dimensional feature association model, establishing the intrinsic relationship between process parameters, material properties, and defect morphology. The fused features are input into a pre-trained machine learning classification model, outputting the concentric circle defect level judgment result. Simultaneously, key process parameter deviations are located through feature importance analysis, completing the process tracing of defect causes. A configurable output interface adapts to various silicon wafer quality evaluation standards, feeding back the detection and tracing conclusions to the crystal pulling process end, achieving closed-loop optimization of process parameters. This invention effectively improves the detection accuracy and traceability of concentric circle defects in N-type monocrystalline silicon wafers. It has high detection efficiency and strong standard adaptability, and can support quality control and process optimization in large-scale silicon wafer production.

Description

Technical Field

[0001] This invention belongs to the field of monocrystalline silicon, and specifically relates to a method for detecting and analyzing concentric circle defects in N-type monocrystalline silicon wafers. Background Technology

[0002] In the prior art, CN109273377A discloses a pre-processing method for detecting concentric circles and black corners in monocrystalline silicon wafers. This method achieves defect visualization through annealing, alkaline boiling, acid immersion, and minority carrier lifetime testing, with a detection accuracy exceeding 90%. However, this method only applies to general monocrystalline silicon wafers, failing to differentiate between N-type and P-type wafers, and as a pre-processing method, it does not involve intelligent defect classification or process traceability. CN101460665A discloses a method for determining the cause of COP (Crystal Optical Parts) in monocrystalline silicon wafers. It distinguishes between COP caused by crystallization and COP caused by amorphous processes by dividing the region into concentric circles and calculating the COP density uniformity index. However, this method only addresses a single defect type, lacks a multi-dimensional feature correlation model, and lacks machine learning-driven automated classification capabilities. CN120558988A discloses an FPD (Flawless Processing) testing method for micro-viewpoint defects in monocrystalline silicon wafers, constructing an "imaging-correction-noise reduction-segmentation-intelligent classification" algorithm chain. However, since the detection object is micro-viewpoint defects, it cannot effectively characterize the annular distribution pattern of concentric circle defects.

[0003] The existing technology has the following main defects: (1) It lacks a systematic quantitative characterization method for concentric circle defects in N-type single crystal silicon wafers. Most of the existing methods are general-purpose and have not been optimized for the special defect behavior caused by the phosphorus doping characteristics of N-type silicon wafers; (2) The existing detection methods cannot effectively correlate crystal pulling process parameters (especially V / G ratio) with the defect formation mechanism, which makes it difficult to trace the source of defects and cannot provide effective feedback for process parameter optimization; (3) The existing methods lack the ability to fuse multi-source data. The amount of information of a single detection method (such as μ-PCD or FTIR) is insufficient to support the accurate classification of defect levels; (4) Different customers have different silicon wafer quality standards. The existing detection system lacks the ability to flexibly adapt to different customer standards. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for detecting and analyzing concentric circle defects in N-type monocrystalline silicon wafers, thereby solving the technical problems of insufficient detection accuracy, difficulty in tracing the source, and poor standard compatibility in the prior art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers, comprising the following steps: S1. Multidimensional feature data acquisition: Collect crystal pulling process parameters, Fourier transform infrared spectrum oxygen content radial distribution data, and microwave photoconductivity attenuation method minority carrier lifetime two-dimensional map corresponding to the N-type single crystal silicon wafer to be tested, to form a multidimensional dataset. S2. Multi-dimensional image feature extraction: For the preprocessed minority carrier lifetime two-dimensional image, perform circle center localization, polar coordinate transformation, defect ring detection, frequency domain feature extraction and texture feature extraction in sequence to obtain an image feature vector containing defect morphology, distribution pattern and texture attributes. S3. Construction of a three-dimensional feature association model: After preprocessing the crystal pulling process parameters, oxygen content radial distribution data and image feature vectors, multi-source feature fusion is performed to construct a three-dimensional feature association model and establish the intrinsic relationship between process parameters, material properties and defect morphology. S4. Defect Classification and Process Tracing: Input the fused feature vector into the pre-trained machine learning classification model and output the level judgment result of concentric circle defects; locate the key process parameter deviations that lead to the formation of defects through feature importance analysis, and complete the process tracing of the cause of defects. S5. Standard Adaptation and Process Feedback: Through a configurable output interface, it adapts to various silicon wafer quality evaluation standards, feeds back defect detection results and process traceability conclusions to the crystal pulling process, and realizes closed-loop optimization of process parameters.

[0006] Preferably, in step S1: The crystal pulling process parameter is the V / G ratio, which is the ratio of crystal pulling speed to the solid-liquid interface temperature gradient. The radial distribution data of oxygen content is the distribution curve of interstitial oxygen content along the radius of the silicon wafer, which is obtained by multi-point measurement using Fourier transform infrared spectroscopy. The two-dimensional minority carrier lifetime map is a two-dimensional topological distribution of minority carrier lifetime on the silicon wafer surface, which is acquired by microwave photoconductivity attenuation method.

[0007] Preferably, step S2 specifically includes the following sub-steps: S2.1 Center point positioning: The outer contour of the silicon wafer is detected by Hough circle transformation, and the coordinates of the center of the outer contour circle are used as the origin of the polar coordinate transformation. S2.2 Polar coordinate transformation: Taking the center of the silicon wafer as the origin, the minority carrier lifetime image in the Cartesian coordinate system is transformed into a polar coordinate image, and the annular distribution characteristics of concentric circle defects are transformed into a horizontal strip-shaped periodic signal. S2.3 Defect ring detection: Apply Hough circle transform to the minority carrier lifetime image in the original Cartesian coordinates, set the detection radius range to 10~90mm, set the accumulator threshold to 60% of the peak value, set the minimum circle spacing to 5mm, and output the center coordinates, radius and lifetime difference amplitude of each defect ring. S2.4 Frequency Domain Feature Extraction: Extract the average lifetime distribution curve along the radial direction of the polar coordinate image, perform a fast Fourier transform on the one-dimensional curve, and extract the frequency value, amplitude, and half-width of the main frequency components to form a frequency domain feature vector. S2.5 Texture Feature Extraction: Taking the detected defect ring region as the region of interest, calculate the gray-level co-occurrence matrix, extract four texture features: contrast, energy, correlation, and uniformity, and take the average of the calculation results in multiple directions to obtain the texture feature vector. All the features extracted in the above sub-steps are concatenated sequentially to form a complete image feature vector.

[0008] Preferably, step S3 specifically includes: Preprocessing operations were performed on the three types of heterogeneous data respectively: A sliding window average is performed on the V / G ratio time series data to eliminate instantaneous fluctuation interference, and then maximum-minimum normalization is performed to map it to the [0,1] value range; For discrete measurement points of the radial distribution of oxygen content, cubic spline interpolation is used to generate continuous distribution curves. The average value of the distribution curves in the four directions of 0°, 45°, 90° and 135° is taken to obtain the representative radial distribution vector. Bilinear interpolation to unify resolution, median filtering for noise reduction, and silicon wafer center coordinate registration based on edge detection are performed on the two-dimensional minority carrier lifetime image. The preprocessed V / G ratio parameter, oxygen content radial distribution vector, and image feature vector are concatenated and fused to generate a high-dimensional feature vector. Principal component analysis is used to reduce the dimensionality of high-dimensional feature vectors, retaining low-dimensional features with a variance ratio of not less than 95%, which are used as the output of the three-dimensional feature association model.

[0009] Preferably, in step S4, the concentric circle defects are classified into four levels according to the ratio of the lifespan difference between the defect area and the normal area: No defects: Lifetime variation ratio less than 5%; Minor defects: lifespan variation ratio of 5%~20%; Moderate defects: lifespan variation ratio of 20%~40%; Severe defects: Lifetime variation ratio greater than 40%; The training and validation process for the machine learning classification model is as follows: Collect labeled samples of N-type monocrystalline silicon wafers covering different V / G ratio ranges and different oxygen content ranges, and divide them into training set, validation set and test set according to a preset ratio; The model was trained using 5-fold cross-validation, with classification accuracy, F1 score at each level, and macro-average F1 score as evaluation metrics. Model training is complete when the overall classification accuracy of the model on the independent test set is not less than 90% and the recall rate of severe defects is not less than 95%.

[0010] Preferably, the specific implementation method of process traceability in step S4 is as follows: The Gini impurity reduction index of the random forest model is used to rank the importance of each feature dimension, identify key influencing features in V / G ratio, radial distribution gradient of oxygen content, and radius of main defect loop, and output the features with the highest importance as the basis for process traceability to locate the key process parameter deviations that lead to defect formation.

[0011] Preferably, step S5 specifically includes: A configurable threshold interface is set at the model output end to switch the judgment threshold of defect level according to various quality evaluation standards and generate a corresponding standard test report. The defect level determination results and process traceability conclusions are fed back to the crystal pulling process control system to guide the adjustment of parameters such as crystal pulling speed and thermal field temperature gradient, forming a closed-loop control system of detection-traceability-optimization.

[0012] A concentric circle defect detection and analysis system for N-type monocrystalline silicon wafers, employing the aforementioned concentric circle defect detection and analysis method for N-type monocrystalline silicon wafers, includes a data acquisition module, an image feature extraction module, a feature association and fusion module, a classification and tracing module, and a standard adaptation feedback module, which are sequentially connected via communication. The data acquisition module is used to collect crystal pulling process parameters, oxygen content radial distribution data and minority carrier lifetime two-dimensional spectrum of the N-type monocrystalline silicon wafer to be tested, forming a multi-dimensional dataset; The image feature extraction module is used to perform multi-dimensional feature extraction on the minority carrier lifetime image to generate an image feature vector containing morphology, distribution, and texture information. The feature association and fusion module is used to perform preprocessing and feature fusion on the three types of data to construct a three-dimensional feature association model; The classification and tracing module is used to output defect level results through a machine learning classification model and complete the process tracing of the cause of defects. The standard adaptation feedback module is used to adapt to various quality standards through a configurable interface and feed the results back to the crystal pulling process to achieve closed-loop optimization.

[0013] Preferably, the classification and tracing module has a built-in model training unit, a level determination unit, and a feature importance analysis unit; the standard adaptation feedback module has a built-in threshold configuration unit and a process feedback unit.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for detecting and analyzing concentric defects in N-type single-crystal silicon wafers.

[0015] The present invention can achieve the following beneficial effects: 1. This invention addresses the annular distribution characteristics of concentric circle defects under phosphorus doping in N-type single-crystal silicon wafers. It employs a multi-method collaborative feature extraction scheme combining polar coordinate transformation, Hough circle detection, frequency domain analysis, and texture feature extraction. This transforms the annular defects into easily analyzable strip signals and frequency domain features, achieving precise localization and quantitative characterization of concentric circle defects. This effectively solves the problem of insufficient characterization capability of traditional detection methods for annular defects. Combined with a machine learning classification model, it can achieve accurate determination of level four defects, with an overall classification accuracy of no less than 90% and a recall rate of no less than 95% for severe defects, significantly improving detection reliability.

[0016] 2. This invention constructs a three-dimensional feature correlation model covering the V / G ratio parameter of the crystal pulling process, the radial distribution characteristics of oxygen content, and the minority carrier lifetime image features. Through standardized preprocessing, it achieves effective fusion of three types of heterogeneous data, fully explores the intrinsic correlation between process conditions, intrinsic material properties and defect morphology, breaks through the bottleneck of insufficient information from a single detection method, and provides multi-dimensional data support for in-depth analysis and cause determination of defects.

[0017] 3. Based on feature importance analysis, this invention can accurately locate the key process parameter deviations that lead to concentric circle defects, identify core influencing factors such as V / G ratio and radial distribution gradient of oxygen content, and extend the detection results from simple quality judgment to cause tracing. The tracing conclusions can be directly fed back to the crystal pulling process to guide the optimization of parameters such as crystal pulling speed and thermal field temperature gradient, forming a complete quality control closed loop of "detection-tracing-optimization", which effectively helps to improve production yield and process iteration. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method; Figure 2 This is a schematic diagram of the process for extracting features from concentric circle defect images.

[0019] In the diagram: N-type: refers to a single-crystal silicon wafer with N-type conductivity; V / G ratio: refers to the ratio of crystal pulling speed V to solid-liquid interface temperature gradient G, where V represents crystal pulling speed and G represents solid-liquid interface temperature gradient. This ratio is used to characterize the matching relationship between thermal field and growth rate during crystal pulling. FTIR: Fourier transform infrared spectroscopy, used to measure the interstitial oxygen content and its radial distribution in silicon wafers; μ-PCD: refers to microwave photoconductivity attenuation method, used to obtain a two-dimensional distribution spectrum of minority carrier lifetime on the surface of a silicon wafer. "μ" indicates a symbol used in the context of microwave or micro-scale detection, and PCD stands for photoconductivity attenuation detection. Hough: refers to the Hough transform; the Hough circle transform is an image processing method used to detect the outer contour circle and concentric circle defect rings of silicon wafers. FFT: Fast Fourier Transform, used to transform the radial mean lifetime distribution curve to the frequency domain and extract frequency domain features such as main frequency, amplitude and full width at half maximum (FWHM). GLCM: Gray-level co-occurrence matrix, used to perform texture statistical analysis on defect ring regions and extract texture features such as contrast, energy, correlation, and uniformity; PCA: Principal Component Analysis, used to reduce the dimensionality of high-dimensional feature vectors after splicing and fusion, thereby reducing feature redundancy and retaining the main feature information; ROI: refers to the region of interest, which in this invention mainly represents the local image region to be extracted for texture features, as determined by defect ring detection; [Oi]: refers to interstitial oxygen content, used to characterize the concentration of oxygen in the silicon wafer in the form of interstitial oxygen; LT: refers to minority carrier lifetime; LT_ref refers to the reference minority carrier lifetime in the normal region; ΔLT refers to the difference in minority carrier lifetime between the defective region and the normal region; Accuracy: refers to classification accuracy, used to evaluate the overall classification performance of a machine learning classification model; F1: Refers to the F1 score, which is a classification evaluation metric that combines precision and recall. Macro-F1: refers to the macro average F1 score, which is used to evaluate the overall balance of multiple defect level classification results; p(i,j): refers to the probability that gray level i and gray level j co-occur in the gray-level co-occurrence matrix under a preset direction and distance; N: In the gray-level co-occurrence matrix, it represents the total number of gray levels, and N-1 represents the number of the largest gray level. Detailed Implementation

[0020] like Figure 1 , 2As shown, a method for detecting and analyzing concentric circle defects in N-type monocrystalline silicon wafers is presented. This method collects multi-source data, including crystal pulling process parameters, FTIR oxygen content, and μ-PCD minority carrier lifetime, to construct a three-dimensional feature correlation model. Combining polar coordinate transformation, Hough circle transformation, FFT frequency domain analysis, and GLCM texture feature extraction methods, machine learning algorithms are used to classify concentric circle defects and trace their origins. This method supports integration with various quality standards, forming a closed-loop control system of detection, traceability, and process feedback. It addresses the technical problems of insufficient detection accuracy, difficulty in traceability, and poor standard compatibility in existing N-type monocrystalline silicon wafer concentric circle defect technologies. The specific steps are as follows: Step 1: Multidimensional Feature Data Acquisition. The V / G ratio (the ratio of crystal pulling speed to temperature gradient) in the crystal pulling process parameters is acquired. The interstitial oxygen content ([Oi]) of the silicon wafer and its distribution curve along the wafer radius are measured using FTIR (Fourier Transform Infrared Spectroscopy). A two-dimensional topological map of the minority carrier lifetime on the silicon wafer surface is obtained using μ-PCD (Microwave Photoconductivity Attenuation Method), forming a multidimensional dataset covering both the process and material aspects.

[0021] Step 2: Image Feature Extraction. Based on the preprocessed μ-PCD minority carrier lifetime 2D image, multi-method collaborative feature extraction is performed in the following order: (a) Center point determination: The outer contour of the silicon wafer is detected by Hough circle transformation, and the coordinates of the center of the detected outer contour circle are used as the origin of the polar coordinate transformation, with a positioning accuracy better than ±1 pixel. (b) Polar coordinate transformation: Taking the center of the silicon wafer as the origin, the 512×512 pixel Cartesian coordinate minority carrier lifetime image is transformed into a polar coordinate image (angular resolution 1°, radial resolution 0.5mm / pixel). The annular distribution characteristics of concentric circle defects are transformed into horizontal strip-shaped periodic signals in the polar coordinate image, which is convenient for subsequent frequency domain analysis. (c) Hough circle transformation defect ring detection: After the polar coordinate transformation is completed, the Hough circle transformation is further applied to the original Cartesian coordinate image. The detection radius range is set to 10-90mm (corresponding to the effective area of ​​the silicon wafer), the accumulator threshold is set to 60% of the peak value, the minimum circle spacing is set to 5mm, and the center coordinates, radius and corresponding lifetime difference amplitude of each defect ring are output to achieve accurate positioning of the defect ring. (d) FFT frequency domain analysis: Extract the average lifetime distribution curve from the polar coordinate image along the radial direction, perform Fast Fourier Transform (FFT) on the one-dimensional curve, and extract the amplitude and phase of the main frequency components. The periodic distribution of concentric circle defects is represented by peak values ​​at specific frequencies in the frequency domain. The peak frequencies, amplitudes, and full width at half maximum (FWHM) of each peak are used as frequency domain feature vectors. (e) GLCM Texture Feature Extraction: Using the defect ring regions determined by the Hough circle transform as regions of interest (ROIs), calculate the GLCM (Gray-Level Co-occurrence Matrix, step size 1 pixel, orientation 0° / 45° / 90° / 135°), and extract four texture features: contrast, energy, correlation, and homogeneity. The average value across the four orientations is taken to obtain a 4-dimensional GLCM texture feature vector. The features extracted by the above methods are sequentially concatenated to form a complete image feature vector for fusion in step three.

[0022] Step 3: Construction of a 3D Feature Correlation Model. The V / G ratio process parameters, FTIR oxygen content radial distribution data, and multidimensional features of the μ-PCD minority carrier lifetime spectrum are fused to construct a 3D feature correlation model, establishing the intrinsic relationship between process parameters, material properties, and defect morphology.

[0023] Multi-source data fusion preprocessing and feature association method: In the process of constructing the three-dimensional feature association model, the following preprocessing operations are performed on the three types of heterogeneous data respectively: (a) V / G ratio parameter preprocessing: The V / G ratio time series data of the same silicon wafer batch are averaged by a sliding window (window length of 5 sampling points) to eliminate the interference of instantaneous fluctuations in crystal pulling speed, and the maximum-minimum normalization is performed according to the interval [V / G_min, V / G_max], mapping to the range [0, 1]; (b) FTIR oxygen content radial distribution data preprocessing: Taking the center of the silicon wafer as the origin, along 0°, 45°, 90°, 1 Nine measurement points were collected in each of the four 35° directions (5mm spacing, coverage radius 0-40mm). Cubic spline interpolation was used to expand the sparse measurement points into a continuous radial distribution curve of oxygen content, and the average value of the data in each direction was taken to obtain a representative radial distribution vector; (c) μ-PCD minority carrier lifetime image preprocessing: the original two-dimensional minority carrier lifetime image was bilinearly interpolated to unify the resolution to 512×512 pixels, and the measurement noise was removed by median filtering (kernel size 3×3). Then, the silicon wafer center coordinates were determined based on the silicon wafer edge detection results (accuracy ±0.5mm) to complete spatial registration. After the three types of data were preprocessed, they were fused using a multimodal feature stitching method: the normalized V / G ratio value (1-dimensional), the radial distribution vector of oxygen content (9-dimensional), and the polar coordinate transformation features, Hough circle features, FFT frequency domain features, and GLCM texture features extracted from the μ-PCD image (total N dimensions) were stitched together into a unified high-dimensional feature vector, which was used as the input of the three-dimensional feature association model. To reduce feature redundancy, principal component analysis (PCA) is used to compress high-dimensional feature vectors into a low-dimensional representation that retains 95% of the variance, which is then used for subsequent machine learning classification.

[0024] Step 4: Machine Learning Classification and Source Tracing. Based on the multi-dimensional feature vectors extracted from the three-dimensional feature association model, machine learning algorithms (including but not limited to random forest, support vector machine, or deep neural network) are used to classify concentric circle defects into levels. Through feature importance analysis, the process source of defect causes is traced, and the key process parameter deviations that lead to defect formation are identified.

[0025] Step Four Supplementary Explanation – Training Dataset Construction and Model Validation Scheme: (a) Defect level labeling rules: Based on the lifetime difference ratio (ΔLT / LT_ref) between defective and normal regions in the μ-PCD minority carrier lifetime map, four levels of defect labeling standards are defined: no defects (ΔLT / LT_ref<5%), slight defects (5%~20%), moderate defects (20%~40%), and severe defects (>40%). The labeling results are manually reviewed and confirmed by experienced process engineers. (b) Training sample construction: N-type single crystal silicon wafer samples under different crystal pulling process conditions were collected, covering the V / G ratio range of 0.10 to 0.22 mm² / (min·K) (divided into 5 intervals, with no less than 30 wafers in each interval), and the interstitial oxygen content [Oi] range of 8 × 10¹. 7 ~14×10¹ 7 A total of no less than 500 labeled sample datasets were constructed, with the training set: validation set: test set divided in a ratio of 7:1.5:1.5. (c) Model Validation Scheme: Five-fold cross-validation is used to evaluate the generalization performance of the model on the training set, with classification accuracy, F1 score at each level, and macro-F1 as evaluation metrics; finally, a one-time evaluation is performed on the independent test set, requiring an overall classification accuracy ≥90% and a severe defect recall rate ≥95%; (d) Feature importance analysis: The Gini impurity reduction index of the random forest model is used to rank the importance of each feature dimension, identify key traceability features such as V / G ratio, radial distribution gradient of oxygen content, and Hough circle radius distribution, and output the top 10 important features as the main basis for process traceability.

[0026] Step 5: Customer Standard Integration and Process Feedback. The model output is designed with a configurable interface to support integration with various silicon wafer quality evaluation standards, feeding back test results to the crystal pulling process to achieve closed-loop optimization of key process parameters such as V / G ratio.

[0027] Example 1: Detection of concentric circle defects in N-type monocrystalline silicon wafers (1) Data acquisition: Take a 300mm diameter N-type single crystal silicon wafer (phosphorus doped, resistivity 1~5Ω·cm) sample and record the V / G ratio parameters of the corresponding crystal pulling process; use FTIR to measure the interstitial oxygen content [Oi] of the silicon wafer, with the measurement points evenly distributed along the radial direction (5mm spacing) to obtain the radial distribution curve of oxygen content; use a μ-PCD device (such as SemilabWT-2000PVN) to obtain a two-dimensional distribution map of minority carrier lifetime on the silicon wafer surface.

[0028] (2) Feature extraction: Perform polar coordinate transformation on the minority carrier lifetime map and establish a polar coordinate system with the center of the silicon wafer as the origin; use the Hough circle transform algorithm to detect concentric circle structures in the image and determine the center coordinates and radius of each defect ring; perform FFT analysis on the radial lifetime distribution curve to extract the main frequency components; calculate the GLCM texture features (contrast, energy, correlation, uniformity) of each defect region.

[0029] (3) Model construction and classification: The extracted multidimensional features are fused with V / G ratio and oxygen content parameters, and input into the pre-trained machine learning classification model to output the concentric circle defect level (such as no defect, mild defect, moderate defect, severe defect) and the corresponding process traceability conclusion (such as a high V / G ratio leading to an increase in vacancy type defects).

[0030] (4) Results docking: Configure output thresholds according to customer quality standards, generate silicon wafer test reports, and feed back traceability conclusions to the crystal pulling process system to guide V / G ratio parameter adjustment.

[0031] Example 2: Application of Process Traceability Feedback Closed Loop This embodiment takes the batch testing of N-type monocrystalline silicon wafers from a photovoltaic silicon wafer manufacturer as an example to illustrate the specific implementation process of the process traceability feedback closed loop.

[0032] (1) Abnormal batch detection: For a batch of 50 N-type single crystal silicon wafers with a diameter of 300 mm (phosphorus doped, resistivity 2-4 Ω·cm), the detection process of steps one to four was performed. The μ-PCD minority carrier lifetime spectrum showed that the lifetime of this batch of silicon wafers was significantly lower than that of the normal region in the annular region 30-45 mm from the center (ΔLT / LT_ref=35%). The Hough circle transform detected a significant defect ring with a radius of about 38 mm. The FFT analysis showed that the defect ring was located at a spatial frequency of 0.026 mm. - A significant peak appears at point ¹, corresponding to a ring-shaped defect with a period of approximately 38 mm. The machine learning classification model outputs a defect level of "moderate defect" for this batch (92% confidence).

[0033] (2) Process traceability analysis: The feature importance analysis results of the random forest model show that the top three important features of this batch of anomalies are: normalized V / G ratio (importance weight 38%), radial distribution gradient of oxygen content (importance weight 27%), and radius of Hough circle main defect ring (importance weight 18%). The traceability conclusion is that the average V / G ratio of the crystal pulling process corresponding to this batch is 0.195 mm² / (min·K), which is higher than the upper limit of the normal range (0.185 mm² / (min·K)), resulting in a high vacancy concentration in the silicon wafer and the formation of COP defect ring in the region 30-45 mm from the center. At the same time, the radial distribution gradient of oxygen content in this region is too large (the oxygen content ratio at the edge / center is 1.18, while the normal value should be <1.10), which aggravates the formation of defects.

[0034] (3) Process feedback and adjustment: Based on the source tracing conclusion, the following adjustment suggestions are fed back to the crystal pulling process system: reduce the crystal pulling speed from the current 1.05 mm / min to 0.98 mm / min (a reduction of about 6.7%), and at the same time appropriately increase the solid-liquid interface temperature gradient (achieved by adjusting the thermal field structure) so that the V / G ratio falls back to the normal range of 0.175~0.185 mm² / (min·K).

[0035] (4) Verification of adjustment effect: After adjusting the process parameters as described above, the testing process was re-executed for subsequent batches (50 wafers). The μ-PCD minority carrier lifetime spectrum showed that the lifetime difference ratio ΔLT / LT_ref in the region 30-45mm from the center decreased to 3.2%, which is lower than the mild defect threshold (5%). The machine learning model output the defect level as "no defect" (confidence level 96%), and the defect elimination rate reached 100%, which verified the effectiveness of the process traceability feedback closed loop.

[0036] (5) Customer standard alignment: For high-end customer A, which requires ΔLT / LT_ref < 10%, all batches are now qualified after adjustment; for standard customer B, which requires ΔLT / LT_ref < 20%, the batch before adjustment (moderate defects, ΔLT / LT_ref = 35%) was judged as unqualified, and all batches after adjustment are qualified. Various judgment thresholds can be automatically switched through configurable interfaces, realizing flexible alignment of test results.

[0037] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers, characterized in that, Includes the following steps: S1. Collect the crystal pulling process parameters, Fourier transform infrared spectral oxygen content radial distribution data, and microwave photoconductivity attenuation method minority carrier lifetime two-dimensional spectrum corresponding to the N-type single crystal silicon wafer to be tested, and form a multi-dimensional dataset. S2. For the preprocessed minority carrier lifetime two-dimensional image, perform circle center localization, polar coordinate transformation, defect ring detection, frequency domain feature extraction and texture feature extraction in sequence to obtain an image feature vector containing defect morphology, distribution pattern and texture attributes. S3. After preprocessing the crystal pulling process parameters, oxygen content radial distribution data and image feature vectors respectively, multi-source feature fusion is performed to construct a three-dimensional feature association model and establish the intrinsic relationship between process parameters, material properties and defect morphology. S4. Input the fused feature vector into the pre-trained machine learning classification model and output the level judgment result of the concentric circle defect. By analyzing the importance of features, we can pinpoint the key process parameter deviations that lead to the formation of defects and complete the process tracing of the cause of defects. S5. Through a configurable output interface, it adapts to various silicon wafer quality evaluation standards, feeds back defect detection results and process traceability conclusions to the crystal pulling process end, and realizes closed-loop optimization of process parameters.

2. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, In step S1: The crystal pulling process parameter is the V / G ratio, which is the ratio of crystal pulling speed to the solid-liquid interface temperature gradient. The radial distribution data of oxygen content is the distribution curve of interstitial oxygen content along the radius of the silicon wafer, which is obtained by multi-point measurement using Fourier transform infrared spectroscopy. The two-dimensional minority carrier lifetime map is a two-dimensional topological distribution of minority carrier lifetime on the silicon wafer surface, which is acquired by microwave photoconductivity attenuation method.

3. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S2.1 Center point positioning: The outer contour of the silicon wafer is detected by Hough circle transformation, and the coordinates of the center of the outer contour circle are used as the origin of the polar coordinate transformation. S2.2 Polar coordinate transformation: Taking the center of the silicon wafer as the origin, the minority carrier lifetime image in the Cartesian coordinate system is transformed into a polar coordinate image, and the annular distribution characteristics of concentric circle defects are transformed into a horizontal strip-shaped periodic signal. S2.3 Defect ring detection: Apply Hough circle transform to the minority carrier lifetime image in the original Cartesian coordinates, set the detection radius range to 10~90mm, set the accumulator threshold to 60% of the peak value, set the minimum circle spacing to 5mm, and output the center coordinates, radius and lifetime difference amplitude of each defect ring. S2.4 Frequency Domain Feature Extraction: Extract the average lifetime distribution curve along the radial direction of the polar coordinate image, perform a fast Fourier transform on the one-dimensional curve, and extract the frequency value, amplitude, and half-width of the main frequency components to form a frequency domain feature vector. S2.5 Texture Feature Extraction: Taking the detected defect ring region as the region of interest, calculate the gray-level co-occurrence matrix, extract four texture features: contrast, energy, correlation, and uniformity, and take the average of the calculation results in multiple directions to obtain the texture feature vector. All the features extracted in the above sub-steps are concatenated sequentially to form a complete image feature vector.

4. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, Step S3 specifically includes: Preprocessing operations were performed on the three types of heterogeneous data respectively: A sliding window average is performed on the V / G ratio time series data to eliminate instantaneous fluctuation interference, and then maximum-minimum normalization is performed to map it to the [0,1] value range; For discrete measurement points of the radial distribution of oxygen content, cubic spline interpolation is used to generate continuous distribution curves. The average value of the distribution curves in the four directions of 0°, 45°, 90° and 135° is taken to obtain the representative radial distribution vector. Bilinear interpolation to unify resolution, median filtering for noise reduction, and silicon wafer center coordinate registration based on edge detection are performed on the two-dimensional minority carrier lifetime image. The preprocessed V / G ratio parameter, oxygen content radial distribution vector, and image feature vector are concatenated and fused to generate a high-dimensional feature vector. Principal component analysis is used to reduce the dimensionality of high-dimensional feature vectors, retaining low-dimensional features with a variance ratio of not less than 95%, which are used as the output of the three-dimensional feature association model.

5. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, In step S4, concentric circle defects are classified into four levels based on the ratio of the lifespan difference between the defect area and the normal area: No defects: Lifetime variation ratio less than 5%; Minor defects: lifespan variation ratio of 5%~20%; Moderate defects: lifespan variation ratio of 20%~40%; Severe defects: Lifetime variation ratio greater than 40%; The training and validation process for the machine learning classification model is as follows: Collect labeled samples of N-type monocrystalline silicon wafers covering different V / G ratio ranges and different oxygen content ranges, and divide them into training set, validation set and test set according to a preset ratio; The model was trained using 5-fold cross-validation, with classification accuracy, F1 score at each level, and macro-average F1 score as evaluation metrics. Model training is complete when the overall classification accuracy of the model on the independent test set is not less than 90% and the recall rate of severe defects is not less than 95%.

6. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, The specific implementation method of process traceability in step S4 is as follows: The Gini impurity reduction index of the random forest model is used to rank the importance of each feature dimension, identify key influencing features in V / G ratio, radial distribution gradient of oxygen content, and radius of main defect loop, and output the features with the highest importance as the basis for process traceability to locate the key process parameter deviations that lead to defect formation.

7. The method for detecting and analyzing concentric circle defects in N-type single-crystal silicon wafers according to claim 1, characterized in that, Step S5 specifically includes: A configurable threshold interface is set at the model output end to switch the judgment threshold of defect level according to various quality evaluation standards and generate a corresponding standard test report. The defect level determination results and process traceability conclusions are fed back to the crystal pulling process control system to guide the adjustment of parameters such as crystal pulling speed and thermal field temperature gradient, forming a closed-loop control system of detection-traceability-optimization.

8. A concentric circle defect detection and analysis system for N-type single-crystal silicon wafers, characterized in that, The method for detecting and analyzing concentric circle defects in N-type monocrystalline silicon wafers according to any one of claims 1-7 is adopted, comprising a data acquisition module, an image feature extraction module, a feature association and fusion module, a classification and tracing module, and a standard adaptation feedback module connected in sequence: wherein: The data acquisition module is used to collect crystal pulling process parameters, oxygen content radial distribution data and minority carrier lifetime two-dimensional spectrum of the N-type monocrystalline silicon wafer to be tested, forming a multi-dimensional dataset; The image feature extraction module is used to perform multi-dimensional feature extraction on the minority carrier lifetime image to generate an image feature vector containing morphology, distribution, and texture information. The feature association and fusion module is used to perform preprocessing and feature fusion on the three types of data to construct a three-dimensional feature association model; The classification and tracing module is used to output defect level results through a machine learning classification model and complete the process tracing of the cause of defects. The standard adaptation feedback module is used to adapt to various quality standards through a configurable interface and feed the results back to the crystal pulling process to achieve closed-loop optimization.

9. The concentric circle defect detection and analysis system for N-type single-crystal silicon wafers according to claim 8, characterized in that, The classification and tracing module has a built-in model training unit, a level determination unit, and a feature importance analysis unit; the standard adaptation feedback module has a built-in threshold configuration unit and a process feedback unit.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for detecting and analyzing concentric circle defects in an N-type single-crystal silicon wafer as described in any one of claims 1 to 7.

Images