Defect fine classification system based on decision tree and suitable for non-graphic wafer detection

Abstract

The invention discloses a defect fine classification system based on a decision tree and suitable for non-graphic wafer detection, and the system comprises a data preprocessing module which is used for carrying out the batch automatic analysis of cross-batch and cross-wafer historical KLARF files, extracting defect basic attributes and user-defined attributes, screening out high-importance features according to the importance of the attributes, and sending the high-importance features to a database; constructing training data; the model training module is used for training a decision tree model through training data and issuing a generated lightweight model file to a wafer detection equipment terminal, and the decision tree model is constructed by adopting a self-adaptive splitting mechanism and a double pruning strategy; and the visualization module is used for visually displaying the decision tree topological graph generated by training, and marking key classification features, split thresholds and sample sizes reaching the nodes on each node of the decision tree. The invention provides a set of end-to-end, automatic, high-precision and high-interpretability defect fine classification solution for non-graphic wafer detection.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing, and more particularly to a defect fine classification system suitable for the inspection of patternless wafers. Background Technology

[0002] In the semiconductor manufacturing industry, wafer inspection plays a decisive role in ensuring chip quality.

[0003] As chips continue to evolve towards miniaturization and high integration, the importance of unpatterned wafer inspection is becoming increasingly prominent.

[0004] Patternless wafer inspection primarily focuses on silicon wafer outgoing inspection and front-end wafer manufacturing processes. Its core task is to detect and locate defects such as particles, crystal craters (COPs), slip lines, and scratches on silicon wafers or epitaxial wafers, thereby helping to locate problems that occur in the process.

[0005] Currently, the technology for classifying defects in patternless wafers is developing rapidly. However, existing technical solutions still have many shortcomings when faced with the complex requirements of patternless wafer inspection.

[0006] I. Rule-Based Binning (RBB) Classification Method. This method is the earliest fundamental technical solution in the field of patternless wafer defect classification, and its core logic relies on the experience of senior process engineers. For example... Figure 1 As shown in (a), engineers first need to manually analyze defect characteristics, write a series of IF-THEN classification rules, and then embed the rule set into the classification software of the detection equipment to complete the defect classification through preset conditions. For example, for particulate defects, the rule IFSize≥0.3μm THEN can be set to classify as Large Particle; IF Size<0.3μm THEN can be classified as Small Particle. However, this method has significant limitations: First, it relies heavily on human experience. Simple rules cannot distinguish complex defects with fine granularity, while complex rules lead to a sharp drop in universality, making it difficult to adapt to different process scenarios. Second, it has extremely poor adaptability. When the manufacturing process is adjusted, new materials are introduced, or new types of defects appear, engineers need to manually screen a massive number of defect samples and repeatedly debug the rule thresholds and logic. The whole process is time-consuming and cumbersome (often taking several days to several weeks), which is completely unable to match the hourly iteration rhythm of modern semiconductor production lines.

[0007] II. Classification Methods Based on Traditional Machine Learning (ML). To address the automation shortcomings of RBB, this method introduces a machine learning classifier to automate part of the process, as shown in Figure 1-(b): First, engineers manually design and extract multi-dimensional features of defects based on domain knowledge, constructing feature vectors. These feature vectors are then input into traditional classifiers such as SVM and Random Forest for training, ultimately completing the classification. However, this method still fails to overcome the core bottleneck of reliance on manual intervention: on the one hand, the effectiveness of features is limited by the engineers' capabilities; manually designed features may miss essential differences in defects, and the process is prone to subjective bias due to differences in personal experience. On the other hand, scalability is insufficient; when detection equipment is updated or new defects appear, the entire feature engineering process needs to be redesigned, making it difficult to quickly adapt to production line changes, significantly weakening the automation advantages.

[0008] III. Classification Methods Based on Deep Learning (DL). This method attempts to bypass manual feature engineering through end-to-end learning to achieve a higher degree of automation. Traditional deep learning classification (such as CNN) requires defect images as input, but natural image data is lacking in imageless wafer inspection scenarios. Therefore, "pseudo-images" need to be constructed first, reorganizing the multi-dimensional features of defects into two-dimensional matrix or point cloud data structures. After annotation, these images are then input into CNN and other models for training, as shown in Figure 1-(c). However, this approach faces multiple challenges in industrial scenarios: First, it consumes extremely high resources. Model training requires massive amounts of labeled samples (especially rare defect labeling, which is very costly), and the inference process demands large computational resources, making it difficult to meet the real-time requirements of high-speed wafer scanning (hundreds of frames per second). Second, the construction of "pseudo-images" has a threshold, requiring prior knowledge to design data reorganization logic. If the structure is unreasonable, key features may be lost. Third, interpretability is lacking. The model decision-making process is a "black box," making it impossible to intuitively present the classification criteria. Engineers find it difficult to trace the causes of misjudgments and conduct root cause analysis of the process, severely restricting its application in high-reliability semiconductor inspection scenarios.

[0009] In summary, existing technologies exhibit a clear contradiction: as the level of automation in the classification process increases, the interpretability of the model continues to decline. In the fine-grained classification of defects in patternless wafers, existing solutions consistently fail to simultaneously meet the two core requirements of high-efficiency automation and transparent interpretability, making it difficult to satisfy the demands of modern semiconductor manufacturing for accurate defect detection and support for process optimization. Summary of the Invention

[0010] The purpose of this invention is to provide a decision tree-based defect classification system suitable for patternless wafer inspection, so as to achieve high-precision, interpretable classification of defect types and meet the requirements.

[0011] To address this, the present invention provides a decision tree-based defect fine classification system suitable for imageless wafer inspection, comprising: a data preprocessing module for automatically parsing batches of historical KLARF files across batches and wafers, extracting basic defect attributes and user-defined attributes, and selecting high-importance features based on attribute importance to construct training data; a model training module for training a decision tree model using the training data and publishing the generated lightweight model file to the wafer inspection equipment terminal, wherein the decision tree model is constructed using an adaptive splitting mechanism and a dual pruning strategy; and a visualization module for visually displaying the topology of the trained decision tree and labeling each node of the decision tree with key classification features, splitting thresholds, and the number of samples reaching that node.

[0012] The technical solution provided by this invention, through its innovative modular design and collaborative working mechanism, brings significant technological progress and practical value to the field of defect detection in patternless wafers. Specific beneficial effects are as follows: 1. This invention achieves fully automated, high-representational-capability feature engineering, significantly improving efficiency and reducing costs. Through the automatic parsing and feature integration capabilities of the preprocessing module, the system can extract and derive high-discriminative features from standard KLARF files in batches without manual intervention. This not only completely eliminates the tedious and subjective manual feature design process in traditional methods, reducing feature preparation time from hours to minutes, but also effectively eliminates redundant and noisy features through an automatic feature filtering mechanism, reducing model computational complexity and laying a solid foundation for high-precision classification. It directly solves the problem of feature extraction relying on manual labor and incurring high costs.

[0013] 2. This invention endows the defect classification model with powerful adaptive and continuous optimization capabilities. The core innovation of the adaptive decision tree engine module lies in the introduction of an adaptive splitting mechanism. This mechanism can dynamically select the optimal binary or multi-split strategy based on the inherent distribution characteristics of defect data, and automatically learn the most accurate classification rules from historical data. Once the manufacturing process changes or new defect types appear, the model can automatically adjust its internal structure simply by retraining with new data, without the need for experts to manually revise the rules. This greatly improves the system's agility and sustainability in responding to production line changes, and solves the problems of rigidity and poor adaptability of rule-based systems.

[0014] 3. This invention offers industry-leading interpretability, greatly enhancing the credibility and practicality of the results. Through an interpretability visualization module, this invention maximizes the "white-box" advantage of decision trees. Engineers can clearly trace the complete decision path of any defect classification result and understand its judgment logic. Combining confusion matrices and feature importance ranking not only quickly locates the model's misjudgment points but also reveals the key detection parameters for distinguishing defects, providing direct and quantitative data support for optimizing processes and equipment settings. This effectively establishes user trust in the model and solves the problem of poor interpretability of "black-box" models such as deep learning.

[0015] 4. This invention deeply mines historical data assets, realizing a data-driven intelligent decision-making closed loop. The system automatically integrates and utilizes massive amounts of historical KLARF data across batches and wafers, transforming it from simple record files into the core fuel driving model optimization. Combined with the interactive feedback optimization function provided by the visualization module, a continuous learning closed loop of "analysis-intervention-validation" is formed. Engineers can inject domain knowledge (such as labeling difficult examples and adjusting weights) based on model performance, enabling the model to become increasingly adapted to specific production line environments through continuous iteration. Ultimately, this efficiently transforms data assets into improvements in productivity and product quality, solving the problem of underutilization of historical data value.

[0016] In summary, by organically integrating the aforementioned key technologies, this invention forms an end-to-end, automated, high-precision, and highly interpretable defect classification solution, effectively overcoming many shortcomings of existing technologies and providing reliable technical support for improving the quality control level of semiconductor manufacturing.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a list of existing non-patterned wafer defect classification technologies; Figure 2 This is a schematic diagram of the system hardware platform architecture of the present invention; Figure 3 The interface showing the defect feature values ​​after preprocessing of the original defect data extracted by the present invention is shown. Figure 4 This is a flowchart of the adaptive model training module of the present invention; Figure 5This is a comparison chart of the classification effects of existing binary splitting and multi-way splitting; Figure 6 This is a schematic diagram of the decision tree path visualization of the present invention; Figure 7 This is a schematic diagram illustrating the visualization of the model performance evaluation results of this invention for process engineers. Detailed Implementation

[0019] Terminology Explanation Patternless wafer inspection technology: Patternless wafer inspection uses laser scanning and optical imaging as its core principles. A laser beam scans radially along a rotating wafer. Defects alter the intensity and angle distribution of scattered light, and the detector locates the defects by capturing these changes in optical signals. Since there is no circuit pattern reference on the surface, the shape, size, and distribution of defects are random, requiring multi-dimensional optical features to distinguish types. This is the technical prerequisite for defect classification.

[0020] KLARF data format: KLARF files are a common semiconductor inspection data format conforming to the SEMI E1079 standard. They are used to store critical defect information, including coordinates, dimensions, multi-channel light intensity values, wafer batch, and chip location. Their standardized parsing forms the basis for subsequent processing, providing raw data support for feature extraction.

[0021] Decision tree algorithm: A decision tree is a rule-based classification model that learns feature judgment rules from data and achieves classification in a tree structure. Internal nodes are feature conditions, branches are judgment results, and leaf nodes are classification conclusions. Traditional decision trees rely on manual feature selection and fixed splitting strategies, which have insufficient adaptability in imageless wafer inspection. However, their good "interpretability" provides a basis for improvement in this invention.

[0022] Feature engineering: Feature engineering is the process of extracting and filtering effective features from raw data. Its goal is to transform the raw signals of defects into features that can distinguish their types. Traditional methods rely on manual design, and the effectiveness of these features is limited by domain knowledge. This is the core bottleneck of existing technologies and a key area that this invention aims to overcome.

[0023] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] This invention provides a defect fine classification method and system based on adaptive decision tree suitable for patternless wafer inspection. Through modular design and algorithm innovation, it constructs an end-to-end closed-loop system of "data input - feature processing - model training - classification output - visualization optimization", which can efficiently adapt to the complex distribution characteristics of patternless wafer defect data, achieve high-precision and interpretable classification of defect types, and support continuous iterative optimization to meet the automation and intelligent requirements of semiconductor production line quality control.

[0025] The system hardware platform of this invention adopts a distributed architecture of a model development terminal and a wafer defect detection equipment terminal, such as... Figure 2 As shown, the two exchange data through the factory's internal network, with separated responsibilities, meeting the deployment requirements of industrial scenarios. After completing wafer scanning and inspection, the wafer inspection equipment terminal generates a KLARF file, stores it in the database, and uploads it to the model development terminal via the network for model training and optimization. After completing model training and evaluation, the model development terminal publishes the final lightweight model file to the wafer inspection equipment terminal.

[0026] The model development terminal, through the deep collaboration of the data preprocessing module, adaptive model training module, and visualization module, completes the entire process from raw data to a usable model. The functional boundaries of the three modules are clear, and the data flow is closed-loop. The specific design is as follows: I. Data Preprocessing Module This module serves as a crucial bridge connecting raw defect data with model training. Its core functionality involves the automatic batch parsing of KLARF files and the intelligent integration of defect features, addressing the issues of low efficiency and strong subjectivity in traditional manual feature processing and providing high-quality feature input for subsequent model training.

[0027] 1. Batch automatic parsing of KLARF files It automatically reads fields such as DefectRecord, extracting not only basic attributes (such as center coordinates, size, length, and area) but also user-defined attributes (such as defect area density and cross-channel light intensity ratio), forming a unified feature space. When loading files in batches, it automatically scans and integrates KLARF data from multiple lots and different wafers, and randomly divides them into large-scale training and test sets according to a preset ratio (such as 7:3).

[0028] 2. Automatic Feature Integration and Filtering The aforementioned feature space contains dozens to hundreds of features, which may contain redundant attributes (such as the defect area being related to the defect length × width and height) or attributes that are irrelevant to classification (such as the defect ID). Directly using these features for training will increase computational complexity, introduce noise, and may even lead to model overfitting.

[0029] Therefore, this module is designed with a pre-screening mechanism based on feature importance.

[0030] We use the Gini Index, consistent with the subsequent decision tree splitting algorithm, as the feature importance evaluation criterion to ensure consistency between the selection logic and the model training logic. Before decision tree training, all samples are pre-split, and the purity improvement of each feature in distinguishing defect categories is calculated. These features are then sorted from highest to lowest, and the top K features with the highest purity improvement are retained (K can be manually configured or automatically determined by setting an importance threshold, such as retaining only features with importance ≥ 0.05). The selected feature subset contains only high-discriminative features, which can reduce the computational cost of subsequent model training and inference by 40%-60%, while effectively eliminating noisy features and improving the model's generalization ability.

[0031] Preprocessing results as follows Figure 3 As shown, the relevant feature values ​​of each defect will be displayed on the interface.

[0032] II. Adaptive Model Training Module This module is the core technology of this invention. It breaks through the bottlenecks of traditional binary decision trees in processing complex data distributions, such as excessive depth, low efficiency, and weak generalization ability. Through an adaptive multi-path splitting mechanism based on cluster analysis and a dual pruning strategy, it achieves a balance between accuracy, efficiency, and generalization ability.

[0033] The specific workflow is as follows: Figure 4 As shown: S11. Begin constructing the decision tree; S12. Select the best splitting rule for the current node: Traverse candidate features, generate splitting schemes, and calculate the Gini gain. Gini, Select Gini's largest solution; S13. Determine if the pre-pruning conditions are met. If not, recursively split the child nodes (return to S12). If they are met, stop splitting and mark them as leaf nodes, completing the construction of the entire tree. S14, Post-pruning (REP), to complete the final optimized model.

[0034] 1. Split candidate generation and evaluation For the decision tree node splitting process, a fully automated process of "candidate attribute traversal - purity evaluation - splitting scheme comparison - optimal scheme selection" is designed. The core innovation lies in the generation of multi-way splitting candidates and the adaptive selection of splitting schemes.

[0035] Candidate attribute traversal: Unlike traditional decision trees that traverse all features, this algorithm prioritizes traversing the highly important features selected by the data preprocessing module, reducing the number of invalid traversals and improving the efficiency of splitting rule search.

[0036] Purity metric: Gini gain ( Gini (Genius) is used as an indicator for evaluating splitting effectiveness. Gini = Gini index of the node before splitting minus the weighted sum of the Gini indices of all child nodes after splitting. The larger the Gini value, the higher the class discrimination of the samples after splitting.

[0037] Let the total number of nodes before splitting be... N After splitting, it produces k The nth child node, the th i The sample size of the child nodes is N i Then the first i The weights of the child nodes: .

[0038] The weighted sum of the Gini indexes of the child nodes after splitting is .in, It is the first i Gini index of each child node, , It is a child node i The Middle c The proportion of samples of each class. . Gini parent The Gini index is the index of the node before splitting.

[0039] Candidate splitting scheme generation: For discrete attributes, calculate the splitting schemes of all candidate values ​​into different child nodes. Gini. For continuous attributes, traditional decision trees typically only perform binary splits, taking the midpoint of each candidate value as a possible split point. However, for multimodal distributions, such as... Figure 5 As shown, the current split node data exhibits three distinct clusters in the Size attribute, precisely representing the three categories. If a traditional binary split is used, the result would be... Figure 5 As shown in (2), multiple splits are required to distinguish them.

[0040] The workflow of the adaptive splitting mechanism of this invention is as follows: S21. Using the traditional binary splitting method as candidates, after sorting the attribute values, take the midpoint of all adjacent values ​​as the candidate splitting threshold, and calculate the result of binary splitting for each threshold. Gini.

[0041] S22. Consider multi-way splitting as a candidate: Start the interval partitioning algorithm based on K-Means clustering. The specific process is as follows: S221. To balance efficiency and effectiveness, a maximum number of splits MaxK is preset (e.g., MaxK=5).

[0042] S222 The algorithm will attempt to cluster from k=3 to k=MaxK, that is, divide the sample value of the current node on this attribute into 3 to MaxK consecutive intervals, and then use the midpoint of the adjacent intervals as the split value. S223. Calculate the value corresponding to each k value. Gini, filtering out the maximum multi-way split. Gini. (e.g.) Figure 5 - (3) is shown.

[0043] S23. Select the optimal splitting scheme: Compare the maximum splitting rate of the binary splitting scheme. Gini and multi-way splitting are the largest Gini, Select Gini's larger scheme serves as the splitting rule for the current node.

[0044] If the data distribution is simple, binary splitting can meet the requirements and avoid complex calculations. If the data distribution is complex, multi-way splitting can distinguish multiple categories at once, reduce the number of node splits, and reduce the tree depth. Experiments have shown that it can reduce the tree depth by 30%-50%.

[0045] 2. Dual pruning strategy To prevent the model from overfitting to noise or anomalies in the training data and to ensure stable classification capabilities in new wafer inspection, this module designs a dual pruning mechanism of pre-pruning (early stopping strategy) + post-pruning (error rate reduction pruning).

[0046] Pre-pruning (early termination strategy). Termination conditions are checked before node splitting to control tree structure complexity from the source. Three core termination conditions are set: Minimum number of samples threshold: If the number of defective samples in the current node is less than or equal to a preset threshold (e.g., less than 10), the splitting stops to avoid rule overfitting caused by a small number of samples.

[0047] Minimum purity gain threshold: If the maximum of all candidate splitting schemes If Gini ≤ a preset threshold (e.g., ≤ 0.01), stop splitting to avoid meaningless splitting.

[0048] Maximum tree depth threshold: If the current tree depth has reached the preset maximum value (e.g., ≤10), stop splitting to avoid excessive tree depth leading to reduced inference efficiency.

[0049] Post-pruning (Reduced Error Pruning, REP): REP uses the hard criterion of not reducing the classification error rate on the validation set. It effectively prevents the model from overfitting to specific noise in the training data, ensuring high accuracy and stability when detecting new wafers. The pruning effect is verified using a pre-defined test set. The specific process is as follows: Starting from the bottom of the decision tree, replace non-leaf nodes with leaf nodes that represent the majority of defect types for that node.

[0050] Calculate the defect classification error rate of the model on the test set before and after the replacement.

[0051] If the error rate does not increase after replacement, pruning is performed; if the error rate increases, the original branch is retained.

[0052] Pruning Effect: This system can automatically integrate massive amounts of historical wafer data, providing a data foundation for reserving a validation set. Trading this minimal cost for high model reliability in actual production lines is an optimal choice. The resulting decision tree structure is streamlined, with an average reduction of 20%-30% in the number of leaf nodes and shorter decision paths. This not only improves inference efficiency but also makes defect classification rules easier to understand, facilitating process engineers' ability to trace the classification logic.

[0053] By combining the aforementioned adaptive splitting mechanism with a strict pruning strategy, this invention can generate a decision tree model with a simpler structure, stronger generalization ability, and higher interpretability while ensuring high classification accuracy, perfectly adapting to the complex characteristics of non-graphical wafer defect data.

[0054] III. Visualization Module This module transforms the "white-box" advantage of decision trees into an intuitive interface that engineers can operate, solving problems such as the poor interpretability of deep learning classifiers, and building an optimized closed loop of "analysis-intervention-verification".

[0055] 1. Visualization of decision tree paths.

[0056] Based on the training data, a complete decision tree topology visualization interface is generated, such as... Figure 6 As shown in (1), each node on the decision path is clearly labeled with its key detection features (such as Size), splitting threshold (such as Size < 0.518), and the number of training or testing defect samples reaching that node. This allows engineers to clearly understand why a defect is classified as a scratch rather than a particle, achieving complete transparency of the classification logic.

[0057] Clicking on any node in the topology graph will display a histogram showing the category distribution of all defects reaching that node and the sample size for each category, such as... Figure 6 -(2)and Figure 6 - (3) shows that it helps engineers understand the clustering of different defects under specific rules and assists in identifying “easily confused categories”.

[0058] 2. Model performance evaluation.

[0059] Confusion Matrix: Clearly displays the true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) for each defect type, and automatically calculates the accuracy and purity for each category. This quickly identifies weak points in the classification process. Accuracy describes the model's ability to capture specific defects, while purity describes the reliability of a specific predicted category. Higher accuracy indicates a lower false negative rate for that defect type; higher purity indicates a lower false positive rate for that defect type.

[0060] Feature Importance Matrix: Clearly displays the importance scores of the selected features for different defects, identifies key physical quantities that distinguish defect types, and provides direct data support for process engineers to optimize testing equipment parameters. For example... Figure 7 As shown.

[0061] 3. Interactive adjustments and optimizations.

[0062] It supports engineers in injecting domain knowledge based on performance evaluation results to achieve online iterative optimization of the model.

[0063] Sample weight adjustment: If the confusion matrix shows that the two types of defects are continuously confused, the training weights of the two types of defect samples can be manually increased to enhance the model's ability to distinguish them.

[0064] Difficult sample labeling: For defective samples that the model misclassifies, the correct category can be manually labeled and added to the training set to trigger model retraining.

[0065] Feature intervention: If the feature importance ranking shows that a certain feature is irrelevant noise, the feature can be manually removed to simplify the model structure.

[0066] Optimization closed loop: All adjustment operations will generate adjustment logs. After retraining, the model performance before and after adjustment can be compared (e.g., accuracy improved by 3%), forming a closed loop of "analysis-intervention-validation-re-optimization", so that the model can continuously adapt to the changes in the defect characteristics of specific production lines.

[0067] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A decision tree-based defect fine classification system suitable for patternless wafer inspection, characterized in that, include: The data preprocessing module is used to automatically parse historical KLARF files across batches and wafers, extract basic defect attributes and user-defined attributes, and filter out high-importance features based on attribute importance to construct training data. The model training module is used to train the decision tree model using training data and publish the generated model file to the wafer inspection equipment terminal. The decision tree model is constructed using an adaptive splitting mechanism and a dual pruning strategy. The visualization module is used to visualize the topology of the decision tree generated during training, and to label the key classification features, split threshold, and number of samples reaching each node on the decision tree.

2. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 1, characterized in that, The construction of the decision tree model includes the following steps: S11. Begin constructing the decision tree; S12. Adopt an adaptive splitting mechanism to select the best splitting scheme for the current node: Traverse candidate features, generate binary splitting schemes and multi-way splitting schemes respectively, and calculate the maximum Gini gain of each scheme. Gini, select Gini's larger plan; S13. Determine whether the pre-pruning conditions are met. If they are met, stop splitting and mark it as a leaf node to complete the construction of the entire decision tree. Otherwise, recursively split the child nodes. S14. After the entire decision tree is constructed, post-pruning is performed to obtain the final optimized model.

3. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 2, characterized in that, The candidate features are the highly important features selected by the data preprocessing module.

4. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 2, characterized in that, Generating a multi-way splitting scheme includes the following steps: S221. Preset a maximum number of splits MaxK; S222. Use the K-Means clustering algorithm to try clustering from k=3 to k=MaxK respectively, and then use the midpoint of the adjacent interval as the split value; S223. Calculate the corresponding value for each K value. Gini, filtering out the maximum multi-way split. Gini.

5. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 2, characterized in that, Gini is the Gini index of the node before splitting minus the weighted sum of the Gini indices of all child nodes after splitting, where The larger the Gini value, the higher the class discrimination of the samples after splitting.

6. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 2, characterized in that, The pre-pruning is used to determine termination conditions before node splitting. The termination conditions include: a minimum sample size threshold, a minimum purity gain threshold, and a maximum tree depth threshold. Specifically, splitting stops if the number of defective samples in the current node is less than or equal to a preset threshold; otherwise, splitting stops if the maximum number of defective samples in all candidate splitting schemes is less than or equal to a preset threshold. If Gini is less than or equal to the preset threshold, the splitting will stop. If the current tree depth has reached the preset maximum value, the splitting will stop.

7. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 2, characterized in that, The post-pruning includes the following steps: S131. Starting from the bottom of the decision tree, replace the non-leaf nodes with leaf nodes that represent the majority of defect types of that node in turn. S132. Calculate the defect classification error rate of the model before and after replacement on the test set; S133. If the error rate does not increase after replacement, perform pruning; if the error rate increases, retain the original branch.

8. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 1, characterized in that, The process of selecting high-importance features from the parsed defect information includes the following steps: Before training the decision tree, all samples are pre-divided, the purity improvement of each feature in distinguishing defect categories (Gini index) is calculated, and the samples are sorted from high to low. The top K features with the highest purity improvement are retained, or the K features are automatically determined by setting an importance threshold.

9. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 1, characterized in that, The visualization module is also used to display the model's performance evaluation results, which include a confusion matrix and a feature importance matrix. The confusion matrix includes true positives, false positives, true negatives, and false negatives for each type of defect, and calculates the accuracy and purity for each type.

10. The decision tree-based defect fine classification system for patternless wafer inspection according to claim 9, characterized in that, The performance evaluation results also include domain knowledge injected by engineers to enable online iterative optimization of the model. This domain knowledge includes sample weight adjustment and hard example labeling.

Images