Automatic SEM image analysis and process defect detection system based on full database

Abstract

The invention provides an automatic SEM image analysis and process defect detection system based on a full database, and relates to the technical field of semiconductor manufacturing, the system combines self-adaptive feature extraction and density anomaly detection of a high-performance calculation unit through dynamic feature fusion (Sp1) of a multi-scale SEM image and a multi-channel acquisition device, and realizes automatic analysis of the SEM image. Precise classification of defect types and automatic identification of new defects are achieved, the feature range and the fusion weight are dynamically adjusted, traditional static detection limitation is broken through, adaptive analysis of process context is supported, the detection precision is improved to 98%, the misjudgment and missing judgment rate is reduced by 30%-50%, especially in advanced processes such as 3nm, the new defects can be rapidly identified, rules can be updated, and the method is suitable for large-scale popularization and application. The process research and development efficiency and the product reliability are remarkably improved, and the flexible and intelligent detection capability is provided for semiconductor manufacturing.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to an automatic analysis and process defect detection system for SEM images based on a full database. Background Art

[0002] According to the semiconductor automatic cleaning monitoring system and method based on image processing disclosed in China with the publication number "CN118571794B", which belongs to the field of image processing, the present invention obtains the number of surface transistors in the cleaning sub-area and imports it into the cleaning sub-area importance evaluation strategy to perform cleaning sub-area importance evaluation, obtains the image abnormality evaluation result and the cleaning sub-area importance evaluation result to evaluate the semiconductor contamination level, judges the automatic cleaning time according to the semiconductor contamination level, makes a comprehensive judgment on the contamination level of the semiconductor wafer according to the finished product image of the semiconductor wafer and the number of surface transistors, and then conducts a comprehensive and accurate analysis of the cleaning time of the cleaning equipment, thereby improving the cleaning efficiency while avoiding damage to the semiconductor wafer caused by cleaning.

[0003] The above patent documents and prior art have the following technical problems when used:

[0004] Problem 1: In existing semiconductor manufacturing, defect detection often relies on fixed feature extraction and predefined classification models. These methods are unable to adapt to the diversity of multi-scale SEM images and are unable to identify unknown defects that arise in new processes. This leads to high rates of false positives and missed detections. This is especially true in advanced node processes such as 3nm, where new defects frequently occur, severely impacting R&D efficiency and yield.

[0005] Problem 2: In existing technologies, root cause analysis of defects is mostly offline statistical analysis, which is disconnected from process adjustments, has a long response time, and cannot intervene in high-risk process steps in a timely manner, resulting in large fluctuations in yield. Especially in high-production lines, delayed adjustments increase scrap rate and cost. Summary of the Invention

[0006] Technical problems solved

[0007] In view of the shortcomings of the existing technology, the present invention provides an automatic analysis and process defect detection system for SEM images based on a full database, which solves the following problems:

[0008] 1. The traditional defect detection methods lack dynamic adaptability and the ability to identify new defects;

[0009] 2. Address the problem of lack of real-time and closed-loop control in defect root cause analysis and process optimization.

[0010] Technical Solution

[0011] To achieve the above objectives, the present invention is implemented through the following technical solutions: an automatic analysis and process defect detection system for SEM images based on a full database, the automatic analysis and process defect detection system comprising the following steps:

[0012] Sp1: Adaptively extracts and fuses features from multi-scale scanning electron microscope (SEM) images to detect and classify defect types, and supports automatic identification of new defect types;

[0013] Sp2: Dynamically analyze the causal relationship between process steps and defects based on the time-dependent characteristics of defect type, process parameters, and equipment status, and isolate the interaction effects between process steps;

[0014] Sp3: Predict defect trends based on the causal relationships and multi-source data, collaboratively generate process parameter adjustment suggestions and equipment maintenance plans, and optimize the predictions and suggestions through feedback;

[0015] Sp4: Builds and updates in real time an interactive knowledge graph that includes defect types, process steps, and equipment status, allowing users to query defect root causes through path reasoning.

[0016] Sp5: Seamlessly integrate the above analysis results with the semiconductor manufacturing process to perform real-time closed-loop control from defect detection to process optimization during semiconductor manufacturing.

[0017] Preferably, when adaptively extracting and fusing features in step Sp1, comprehensive features suitable for different magnifications and process contexts are generated by dynamically adjusting the feature extraction range and fusion weights, and the fusion weights are calculated in real time based on the local statistical characteristics of the image and the type of process steps.

[0018] Preferably, when dynamically analyzing the causal relationship between the process steps and defects in step Sp2, historical data is encoded through a time decay mechanism to generate a time-varying feature vector, and the causal contribution and interaction effect strength of the process steps to the defects are calculated based on the time-varying feature vector.

[0019] Preferably, when predicting defect trends in step Sp3, the SEM image features, process parameters and equipment status are decomposed into long-term trends and periodic fluctuations, the probability of future defect occurrence is predicted based on the decomposition results, and the collaboratively generated recommendation priority is adjusted according to the predicted probability.

[0020] Preferably, when performing an interactive knowledge graph in step Sp4, it supports dynamic generation of the optimal causal path from defect type to process step through multi-layer path search and confidence assessment, and incrementally updates the graph structure and association strength based on new data. The interactive knowledge graph also includes an abnormal pattern recognition function, which identifies abnormal process patterns by analyzing changes in the association strength between defect types and process steps in the graph, and automatically triggers process adjustment suggestions or equipment inspection instructions.

[0021] Preferably, the automatic analysis and process defect detection system includes a data alignment unit for aligning multi-scale SEM images with process parameters and equipment status data in time and space dimensions, and embedding process context identifiers in the aligned data to enhance the context relevance of the analysis.

[0022] Preferably, the automatic identification of new defect types in step Sp4 is achieved through density anomaly detection in the feature space, and the identified new defect types are associated with the trend of process parameter changes to generate a preliminary causal hypothesis of the new defect. After the causal hypothesis is verified by comparing with historical data, the system's defect classification rules are automatically updated.

[0023] Preferably, when dynamically analyzing the causal relationship between process steps and defects in step Sp2, the impact of small disturbances in process parameters is simulated to quantify the impact of disturbances on defect probability, and a priority adjustment sequence of process steps is generated based on the quantified results to minimize the defect occurrence rate.

[0024] Preferably, the collaborative generation of process parameter adjustment suggestions and equipment maintenance plans in step Sp3 includes: building a multi-objective optimization model based on defect trend prediction results and equipment operating status, while optimizing defect rate reduction and maintenance cost control, and dynamically adjusting model weights during the optimization process to adapt to changes in production demand.

[0025] Preferably, the hardware components of the automatic analysis and process defect detection system include:

[0026] Multi-channel SEM image acquisition device, equipped with multiple magnification lenses and real-time data transmission interface, used to collect multi-scale SEM images and transmit them to the full database;

[0027] A high-performance computing unit, integrating a multi-core GPU and a dedicated acceleration chip, is used to support real-time computing of the adaptive feature fusion, time-varying causal analysis, and defect trend prediction;

[0028] Process parameter and equipment status monitor, including a sensor array and time synchronization module, is used to collect process parameters and equipment operating status in real time and store them aligned with SEM image data;

[0029] An interactive display terminal equipped with a high-resolution touch screen and a graph rendering engine for presenting the interactive knowledge graph and supporting user path reasoning operations;

[0030] Adaptive storage server, equipped with a hierarchical cache structure and dynamic partitioning module, is used to dynamically allocate storage space based on the amount of SEM image data and process analysis requirements, and supports fast retrieval and update;

[0031] A closed-loop control interface unit integrates a programmable logic controller (PLC) and a process equipment communication protocol, and is used to transmit process parameter adjustment suggestions and equipment maintenance plans to semiconductor manufacturing equipment in real time, and receive execution feedback signals to optimize subsequent analysis.

[0032] Beneficial effects

[0033] The present invention provides an automatic analysis and process defect detection system for SEM images based on a full database.

[0034] It has the following beneficial effects:

[0035] 1. The present invention adopts a system that realizes accurate classification of defect types and automatic identification of new defects through dynamic feature fusion (Sp1) of multi-scale SEM images and multi-channel acquisition devices, combined with adaptive feature extraction and density anomaly detection of high-performance computing units, dynamically adjusts feature range and fusion weight, breaks through the limitations of traditional static detection, supports adaptive analysis of process context, improves detection accuracy to 98%, and reduces false positive and missed detection rates by 30%-50%. Especially in advanced processes such as 3nm, it can quickly identify new defects and update rules, significantly improving process R&D efficiency and product reliability, and providing flexible and intelligent detection capabilities for semiconductor manufacturing.

[0036] 2. The present invention adopts a system that integrates time-varying causal analysis (Sp2), defect trend prediction (Sp3) and real-time control (Sp5). Relying on monitors, storage servers and closed-loop control interface units, it forms a closed-loop system from root cause analysis to process optimization, disturbance simulation and multi-objective optimization, real-time adjustment of process parameters and feedback optimization, surpassing traditional offline analysis, with a response speed of seconds and a yield improvement of 5%-10%. For example, it can quickly reduce the defect rate from 5% to 2%, reduce production costs, ensure manufacturing stability, and provide efficient control for high-production lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a diagram of the system operation steps of the present invention;

[0038] Figure 2 This is a diagram of the system hardware architecture of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Specific embodiment one:

[0041] like Figures 1 to 2 As shown, the automatic analysis and process defect detection system of SEM images based on the full database includes the following steps:

[0042] Sp1: Adaptive feature extraction and defect detection classification: The system first adaptively extracts and fuses features from multi-scale scanning electron microscope (SEM) images to detect and classify defect types, and supports automatic recognition of new defect types. It starts with a multi-channel SEM image acquisition device acquiring images of different magnifications (such as low magnification for macro defect detection and high magnification for nano-level defect analysis). These images are transmitted to the full database for storage. Then, the system dynamically adjusts the feature extraction range and fusion weights to generate comprehensive features suitable for different magnifications and process contexts. Specifically, the feature extraction range is dynamically adjusted according to the local statistical characteristics of the image (such as grayscale variance and entropy value). For example, the range is narrowed in high-detail areas to capture tiny defects, and the range is expanded in smooth areas to identify large-scale anomalies. The fusion weight is based on the local statistical characteristics of the image and the type of process step (such as lithography or etching). It calculates the multi-scale features in real time and fuses them into a comprehensive feature vector through a weight distribution mechanism (such as weighted average based on process context). Based on this vector, the system classifies defects (such as bridging, fracture, and particle deposition) and identifies new defect types through density anomaly detection in the feature space. For example, if a feature cluster is clustered outside the predefined category and the density exceeds the normal range, the system marks it as a new defect and associates it with the trend of process parameter changes (such as temperature fluctuations) to generate preliminary causal hypotheses (such as "high temperature causes particle deposition"). After these hypotheses are verified by comparing with historical data, the defect classification rules are automatically updated to ensure the adaptability of the system and the efficient conversion from raw SEM images to defect types. The process reduces misjudgments through adaptability and dynamism, while providing accurate data for subsequent analysis. The workflow is manifested in image acquisition, feature extraction and fusion, defect classification and new type identification, and rule update.

[0043] Sp2: Dynamic process causal relationship analysis: After obtaining the defect type, the system dynamically analyzes the causal relationship between process steps and defects based on the time-dependent characteristics of defect type, process parameters and equipment status, and separates the interaction effects between process steps. Starting from the data alignment unit, the multi-scale SEM images are aligned with process parameters (such as temperature, pressure) and equipment status (such as vacuum degree, voltage stability) in time and space dimensions, and process context identifiers (such as "etching after lithography") are embedded in the aligned data to enhance the context relevance of the analysis. The aligned data enters the analysis stage, and the system encodes historical data through the time decay mechanism to generate time-varying feature vectors. Specifically, recent data are given higher weights to reflect the immediate impact of process changes, while long-term data provides long-term trend references. Based on these time-varying feature vectors, the system calculates The causal contribution and interaction effect strength of process steps on defects, such as determining the direct impact of a certain step (such as etching) on ​​bridging defects through covariance analysis, and separating the composite effects caused by the interaction between lithography and etching through tensor decomposition; in addition, the system simulates small perturbations of process parameters (such as increasing the etching temperature by 0.5°C) to quantify the impact of the perturbation on the defect probability (such as the probability increasing from 5% to 7%), and generates a priority adjustment sequence for process steps based on the quantified results (such as "prioritize the etching temperature") to minimize the defect incidence rate. Based on quantitative data, the adjustment is targeted. The function of step Sp2 is to reveal the dynamic causal chain between defects and processes. The process provides accurate root cause location through time-dependent analysis and perturbation simulation. The workflow is data alignment, time-varying feature generation, causal and interaction analysis, perturbation quantification and adjustment sequence generation.

[0044] Sp3: Defect trend prediction and optimization suggestion generation: Based on the causal relationship results of Sp2, the system predicts defect trends based on multi-source data, collaboratively generates process parameter adjustment suggestions and equipment maintenance plans, and optimizes predictions and suggestions through feedback. It starts with the integration of multi-source data, including SEM image features, process parameters and equipment status. The system decomposes these data into long-term trends (such as the defect rate slowly increasing over time) and periodic fluctuations (such as fluctuations caused by weekly equipment aging), which are achieved through dynamic frequency analysis (such as adaptive Fourier decomposition based on process cycles). Based on the decomposition results, the system predicts the probability of defect occurrence in the future (such as the next week). For example, it predicts that the probability of a certain defect type will increase from 3% to 5%. According to the predicted probability, the system adjusts the priority of collaboratively generated suggestions, and suggestions corresponding to high-probability defects are given priority. The collaborative generation process is achieved by constructing a multi-objective optimization model. The model simultaneously optimizes defect rate reduction (such as reducing the defect rate to below 2%) and maintenance cost control (such as limiting maintenance frequency), and dynamically adjusts model weights during optimization to adapt to changes in production demand (such as prioritizing defect rate during peak periods and taking cost into consideration during low periods). Specific suggestions include process parameter adjustments (such as "reducing photolithography exposure time by 5%") and equipment maintenance plans (such as "cleaning the etching chamber"). These suggestions are transmitted to the manufacturing equipment for execution through a closed-loop control interface unit. The feedback signal after execution (such as changes in defect rate) is received by the system and used to optimize the prediction model and suggestion generation logic. For example, if an adjustment does not achieve the expected effect, the system will adjust the prediction weight or optimization target, predict defects and provide optimization solutions. The process ensures the scientificity and practicality of the suggestions through decomposition prediction and multi-objective optimization. The workflow is data decomposition, trend prediction, optimization model construction, suggestion generation and execution, and feedback optimization.

[0045] Sp4: Interactive knowledge graph construction and update: The system then constructs and updates in real time an interactive knowledge graph containing defect types, process steps, and equipment status, supporting users to query the root cause of defects through path reasoning. Taking the analysis results of Sp1-Sp3 as input, the system uses defect types (such as bridging), process steps (such as etching), and equipment status (such as equipment temperature) as nodes to build a multi-layer association graph based on causal contribution and interaction effect strength; the graph supports dynamic generation of the optimal causal path through multi-layer path search and confidence evaluation, such as the path from "bridging defect" to "etching temperature is too high", and the confidence is quantified by the association strength (such as 0.9); the graph structure and association strength are incrementally updated according to new data (such as new defects or process adjustments) to avoid Global recalculation improves efficiency. In addition, the atlas has the function of abnormal pattern recognition. By analyzing the changes in the correlation strength between defect types and process steps (such as the correlation strength of a certain step suddenly rises to 0.95), it identifies abnormal process patterns (such as abnormalities caused by equipment aging) and automatically triggers process adjustment suggestions (such as "reduce etching power") or equipment inspection instructions (such as "check the vacuum pump"). Users can operate the atlas through the interactive display terminal, such as zooming in on a defect node to view related process paths, or screening high-confidence paths for decision-making. The function is to provide intuitive root cause queries and abnormality warnings. The process realizes the visualization and initiative of knowledge through dynamic construction and abnormality identification. The workflow is data input, atlas construction, path reasoning and update, abnormality identification and triggering.

[0046] Sp5: Real-time Closed-Loop Control Integration with Semiconductor Manufacturing: Finally, the system seamlessly integrates the above analysis results with the semiconductor manufacturing process to perform real-time closed-loop control from defect detection to process optimization. Relying on a closed-loop control interface unit, the process parameter adjustment suggestions and maintenance plans from Sp3 are transmitted in real time to manufacturing equipment (such as lithography machines and etching machines) through a programmable logic controller (PLC) and communication protocol. At the same time, the abnormal trigger instructions from Sp4 are sent to the equipment management system. After the manufacturing equipment executes the instructions, feedback data (such as the adjusted defect rate and equipment status) is transmitted back to the system through the interface unit, updating the entire database and optimizing subsequent analysis. For example, if a suggestion reduces the defect rate to the target value, the system records the success case to strengthen the prediction model; if it does not meet expectations, the causal analysis or optimization weights are adjusted. The system supports this process through an adaptive storage server. The server's hierarchical cache structure and dynamic partitioning module allocate storage space based on data volume and analysis requirements, ensuring fast retrieval and real-time updates, and realizing seamless integration between analysis and manufacturing. The process forms a closed loop through real-time transmission and feedback. The workflow is result transmission, equipment execution, feedback collection, and system optimization, ultimately achieving continuous control from defect detection to process improvement in semiconductor manufacturing.

[0047] The system's operating plan uses the entire database as its data foundation, and forms a closed-loop process from SEM image analysis to process optimization through five steps: Sp1 extracts features from multi-scale images and detects defects, laying the foundation for subsequent analysis; Sp2 locates the root causes of defects through time-varying causal analysis, and provides directions for process improvement; Sp3 predicts trends and generates optimization suggestions to proactively prevent defects; Sp4 builds a knowledge graph to support queries and abnormality warnings; Sp5 integrates the results into the manufacturing process to achieve real-time control; the entire workflow is manifested as: image acquisition and data alignment, feature extraction and defect classification, causal analysis and root cause location, trend prediction and optimization suggestions, graph construction and abnormality triggering, manufacturing execution and feedback optimization, and achieves efficient defect management through dynamics (feature fusion, time-varying analysis), collaboration (prediction and optimization) and closed-loop (real-time control and feedback), significantly reducing manual intervention compared to traditional methods and improving the yield and efficiency of semiconductor manufacturing. Specific embodiment two:

[0049] like Figures 1 to 2 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0050] The hardware components of the automatic analysis and process defect detection system include:

[0051] Multi-channel SEM image acquisition device: The multi-channel SEM image acquisition device is equipped with multiple magnification lenses and real-time data transmission interfaces, which are used to acquire multi-scale SEM images and transmit them to the entire database. It is the starting point of the system operation and directly supports step Sp1 (adaptive feature extraction and defect detection classification). Its operation logic is based on multi-channel parallel acquisition. It simultaneously acquires multi-scale images through lenses with different magnifications (such as low-magnification lenses to capture macro defects of wafers and high-magnification lenses to focus on nano-level details). Relying on electron beam scanning and signal detection technology, it converts the surface morphology of the sample into a digital image. The operation process starts with the SEM scanning of the wafer in the manufacturing process. The system collects images in different channels according to preset magnifications (such as 500x, 5000x), and transmits the image stream to the entire database through a real-time data transmission interface (such as high-speed Ethernet or optical fiber), ensuring that Sp1 can obtain original data in real time. For example, when detecting bridging defects on the wafer surface, low-magnification images provide the overall distribution, and high-magnification images reveal the specific morphology. The collected multi-scale images provide diversified inputs for subsequent feature extraction and fusion. The process is manifested as sample scanning, multi-channel image acquisition, real-time transmission, and database storage. The efficient acquisition capability ensures the basis for dynamically adjusting the feature extraction range in Sp1, ensuring the comprehensiveness and accuracy of defect detection.

[0052] High-performance computing unit: Integrates multi-core GPU and dedicated acceleration chip to support real-time calculation of adaptive feature fusion in Sp1, time-varying causal analysis in Sp2 and defect trend prediction in Sp3. It is the core computing engine of the system. Based on parallel computing and task allocation, multi-core GPU handles large-scale matrix operations for image feature extraction and fusion, and dedicated acceleration chip optimizes complex model calculations in time-varying analysis and prediction. The working principle is to achieve high-throughput data processing through hardware acceleration. For example, in Sp1, GPU parallelly calculates local statistical characteristics of multi-scale images (such as grayscale variance), and the acceleration chip optimizes the real-time adjustment of fusion weights. ; In Sp2, the chip supports time decay coding and causal contribution calculation; in Sp3, the GPU drives the iteration of the multi-objective optimization model; the operation process is: receiving craters (SEM images, process parameters, equipment status) transmitted by the entire database, allocating computing tasks (feature fusion, causal analysis, trend prediction), parallel processing, outputting results (such as comprehensive feature vectors, causal strength, optimization suggestions), and transmitting to storage or display. High-performance computing capabilities ensure the real-time performance of Sp1-Sp3. For example, when Sp3 predicts defect trends, the unit can complete trend decomposition and optimization suggestion generation in seconds, laying the foundation for real-time closed-loop control.

[0053] Process parameter and equipment status monitor: It includes a sensor array and a time synchronization module, which is used to collect process parameters (such as temperature, pressure) and equipment operating status (such as voltage stability) in real time, and align and store them with SEM image data. It directly supports Sp2 (dynamic process causal relationship analysis) and Sp3 (defect trend prediction). Its operation logic monitors manufacturing equipment in real time through a distributed sensor network. The time synchronization module ensures the time consistency of data, based on physical quantity detection (such as thermocouple temperature measurement, pressure sensor pressure measurement) and signal digitization; the operation process is: the sensor array is deployed on the manufacturing equipment (such as lithography machine, etching machine), real-time acquisition Parameters and states are collected (such as etching chamber temperature of 45°C and vacuum degree of 10^-6Torr), and the time synchronization module aligns timestamps (such as synchronization with SEM image acquisition to the millisecond level), and transmitted to the entire database to provide Sp2 with input of time-dependent features. For example, the temperature change history is encoded through the time decay mechanism to generate a time-varying feature vector; in Sp3, the device state (such as voltage fluctuation) is decomposed into periodic fluctuations to predict potential defect trends. The monitor's high-precision acquisition and alignment capabilities ensure the contextual relevance of multi-source data. For example, when analyzing bridging defects, synchronized temperature and SEM image data reveal the correlation between high temperature and defects.

[0054] Interactive display terminal: Equipped with a high-resolution touch screen and a graph rendering engine, it is used to present the interactive knowledge graph in Sp4 and support user path reasoning operations. It is the key interface for user interaction with the system. Its operating logic is based on graphics rendering and touch response. The working principle is to convert the associated data of defect type, process steps and equipment status into a visual graph through the graph rendering engine. The touch screen supports user operations (such as zooming in and path selection). The operating process is as follows: receiving Sp4 analysis results (such as causal path and correlation strength), the rendering engine builds a multi-layer knowledge graph (such as the "bridge defect, etching temperature" path), displays it on the touch screen, user interaction (such as clicking on a node to query the root cause), and outputs the reasoning result (such as a confidence level of 0.9). In Sp4, the terminal presents a dynamically updated graph. Users can identify abnormal patterns (such as a sudden increase in correlation strength) through multi-layer path search and trigger adjustment suggestions. The interactivity supports users to make quick decisions. For example, after identifying an abnormal process pattern, the terminal automatically displays the suggestion of "reducing etching power", ensuring that Sp4's abnormal warning and root cause query functions are intuitive and efficient.

[0055] Adaptive Storage Server: Equipped with a hierarchical cache structure and a dynamic partitioning module, it dynamically allocates storage space based on the amount of SEM image data and process analysis requirements, and supports fast retrieval and updates. It is the data management core of Sp1-Sp5. Its operating logic is based on adaptive resource allocation. The hierarchical cache (high-speed SSD and low-speed HDD) is hierarchically stored according to data access frequency. The dynamic partitioning module adjusts the storage layout according to analysis tasks (such as Sp1 feature extraction and Sp3 trend prediction). The working principle is to optimize data throughput through intelligent scheduling. The operating process is as follows: receiving multi-source data (SEM images, process parameters, analysis results), hierarchical cache storage (such as storing frequently accessed feature vectors on SSD), dynamic partition adjustment (such as allocating more space for Sp3 prediction), supporting fast retrieval (such as millisecond-level extraction of historical defect data), and updating data (such as incremental updates of Sp4 atlases). This ensures fast access to multi-scale images in Sp1, efficient encoding of historical data in Sp2, real-time data support for prediction models in Sp3, and instant updates of feedback data in Sp5. For example, in closed-loop control, the server quickly stores execution feedback to optimize subsequent recommendations.

[0056] Closed-loop control interface unit: Integrates programmable logic controller (PLC) and process equipment communication protocol, used to transmit process parameter adjustment suggestions and equipment maintenance plans in Sp3 to semiconductor manufacturing equipment in real time, and receive execution feedback signals to optimize subsequent analysis. It is the execution bridge of Sp5 (real-time closed-loop control). Its operating logic is based on instruction conversion and feedback loop. The working principle is to convert the analysis results (such as "reduce exposure time by 5%) into equipment executable instructions through PLC, and the communication protocol (such as Modbus, OPC UA) ensures seamless connection with manufacturing equipment; the operating process is: receiving Sp3 suggestions and Sp4 trigger instructions, PLC generates control signals, transmits them to the equipment (such as adjusting parameters of the lithography machine), the equipment executes and feedbacks the results (such as the defect rate is reduced to 2%), and returns them to the system for update analysis to realize the closed-loop control of Sp5. For example, after executing "cleaning the etching chamber", the feedback data shows that the defects are reduced. The system optimizes the prediction model accordingly to ensure the continuity and real-time nature of process improvement.

[0057] The operating logic and workflow of the hardware components work closely with Sp1-Sp5: the multi-channel SEM image acquisition device provides multi-scale images for Sp1, the high-performance computing unit drives the real-time analysis of Sp1-Sp3, the process parameter and equipment status monitor supports multi-source data input of Sp2 and Sp3, the interactive display terminal presents the map and interactive functions of Sp4, the adaptive storage server ensures the data management of Sp1-Sp5, and the closed-loop control interface unit realizes the real-time control of Sp5; the entire process is: image acquisition, computational analysis, data monitoring, map presentation, storage management, control execution and feedback, forming a closed-loop system from defect detection to process optimization, ensuring efficient and accurate semiconductor manufacturing optimization. Specific embodiment three:

[0059] like Figures 1 to 2 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0060] To further verify the feasibility of this application, a comparative analysis was conducted with existing commonly used systems or methods for semiconductor surface image defect detection, and experimental content was designed to verify the distinguishing features of this solution. The specific contents are as follows:

[0061] Existing commonly used systems or methods:

[0062] Deep learning-based defect detection systems (e.g., DeepSEM-Net, CN110672644A): use convolutional neural networks (CNN) or their variants (e.g., ResNet) to extract fixed features from SEM images and classify predefined defect types (e.g., bridging, fractures). These systems rely on images at a single magnification, have a fixed feature extraction range, require manual annotation of new defects, lack multi-scale feature fusion and dynamic adjustment capabilities, have low efficiency in identifying new defects, have a detection accuracy of approximately 90%, and respond slowly to process changes, making them suitable for routine defect detection in stable processes.

[0063] Statistical analysis and offline optimization systems (such as SEMVisionG10, US20200372635A1): Analyze SEM image defect distribution through statistical methods and generate reports for manual process adjustments. Root cause analysis is based on historical data, and optimization recommendations require additional processes. There is no real-time causal analysis and closed-loop control, resulting in long response times (hours to days), limited yield improvements (approximately 2%-5%), and no proactive intervention. These systems are suitable for low-frequency defect analysis and post-process optimization.

[0064] This technical solution (SEM image automatic analysis and process defect detection system based on the full database)

[0065] Distinguishing Feature 1: Dynamic Adaptive Defect Detection and Automatic Discovery of New Defects (Sp1): Through adaptive multi-scale feature fusion (dynamic adjustment of extraction range and fusion weight) and density anomaly detection, it supports automatic identification of new defect types and rule updates. Combined with a multi-channel SEM acquisition device and a high-performance computing unit, it breaks through the limitations of fixed feature extraction and adapts to process changes and new defects. The detection accuracy reaches 98%, and the false positives and missed positives are reduced by 30%-50%.

[0066] Distinguishing Feature 2: Real-time closed-loop control and process optimization (Sp2-Sp5): Integrating time-varying causal analysis (Sp2), trend prediction and optimization suggestions (Sp3), and closed-loop control (Sp5), relying on monitors, storage servers, and interface units, it achieves second-level response and dynamic optimization, surpassing offline analysis, improving yield by 5%-10%, significantly improving response speed, and building a closed-loop system from root cause analysis to process adjustment.

[0067] The differences in system characteristics are shown in Table 1 below:

[0068]

[0069] Table 1

[0070] Experimental objectives: To verify the two distinguishing features of this technical solution:

[0071] Dynamic adaptive defect detection and automatic discovery of new defects: Comparison of detection accuracy, new defect recognition rate, and false positive rate;

[0072] Real-time closed-loop control and process optimization: Compare response time, yield improvement, and process adjustment efficiency;

[0073] Subjects:

[0074] Sample: 3nm process wafer, containing known defects (such as bridging and particle deposition) and unknown new defects;

[0075] Comparison systems: DeepSEM-Net (representing deep learning methods), SEMVisionG10 (representing statistical analysis methods), and this system;

[0076] The experimental steps are as follows:

[0077] Step 1: Defect detection experiment:

[0078] Input: 1000 multi-scale SEM images (500 each at 500x and 5000x magnification), including known and new defects;

[0079] Operation: DeepSEM-Net uses a single-scale CNN for detection, SEMVisionG10 for statistical classification, and this system performs Sp1 dynamic feature fusion and density anomaly detection;

[0080] Record: detection accuracy (proportion of defects correctly classified), new defect recognition rate (proportion of new defects recognized), false positive rate (proportion of incorrectly classified defects);

[0081] Step 2: Real-time optimization experiment:

[0082] Input: defective wafer production line data (SEM images, process parameters, equipment status), simulation of high-risk process steps (such as etching temperature too high);

[0083] Operation: DeepSEM-Net only detects, SEMVisionG10 generates offline reports, and this system performs Sp2-Sp5 (causal analysis, prediction, closed-loop control) and adjusts parameters (e.g., lowering the temperature by 5°C);

[0084] Record: response time (from defect detection to adjustment completion), yield improvement (change in defect rate before and after adjustment), and adjustment efficiency (number of optimizations per unit time);

[0085] Experimental parameters:

[0086] Hardware environment: This system uses a multi-channel SEM acquisition device (lens magnification 500x-5000x), a high-performance computing unit (GPU + accelerator chip), a monitor, etc.; the comparison system uses standard SEM equipment and a general-purpose server;

[0087] Operating conditions: Production line simulation environment, processing 1,000 images / hour, and process parameter changes 10 times / hour;

[0088] Repeat 3 times and take the average value.

[0089] The experimental results are shown in Table 2 below:

[0090]

[0091] Table 2

[0092] Result analysis:

[0093] Defect detection experiment:

[0094] Detection accuracy: Due to multi-scale feature fusion and dynamic adjustment, the accuracy of this system reaches 98.1%, which is better than DeepSEM-Net (90.2%) and SEMVisionG10 (87.5%).

[0095] New defect recognition rate: The system identified 85.3% of new defects (density anomaly detection), while the comparison system did not have this capability, verifying the breakthrough of Distinguishing Feature 1;

[0096] False positive rate: The system has a false positive rate of only 2.4%, much lower than the comparison systems (8.5% and 10.2%), thanks to its adaptability and the updating of new defect rules.

[0097] Real-time optimization experiments:

[0098] Response time: This system responds in seconds (3 seconds), far exceeding DeepSEM-Net (120 seconds) and SEMVisionG10 (3600 seconds), reflecting the advantages of closed-loop control;

[0099] Yield improvement: This system achieved a 7.5% improvement, outperforming the comparison system (2.1% and 3.8%), due to the synergistic effect of time-varying analysis and optimization suggestions;

[0100] Adjustment efficiency: This system makes 12 adjustments per hour, while the comparison system makes almost no real-time adjustments, verifying the efficiency of Distinguishing Feature 2.

[0101] By inputting multi-scale images and changing process parameters, we simulate the real production line environment to ensure the credibility of the results. The comparison system represents the current mainstream technologies (deep learning and statistical analysis) and is directly related to the distinguishing features of this system. Experimental results show that this system comprehensively surpasses existing methods in terms of accuracy, adaptability and response speed, and is particularly suitable for efficient defect management of advanced processes. Specific embodiment four:

[0103] like Figures 1 to 2 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0104] To further verify the application effect of this system, the following are two application cases:

[0105] Application Case 1: Detection and Optimization of Nanoscale Particle Deposition Defects in 3nm Process:

[0106] Background and Problem: During 3nm process development, a semiconductor foundry discovered unknown nanoscale particle deposition defects on wafer surfaces. The traditional DeepSEM-Net system, relying on a single-scale CNN, only identified predefined defects (such as bridging). Its detection accuracy remained at 90%, making it unable to identify new defects. This resulted in a missed detection rate of up to 15%, impacting yield and R&D progress.

[0107] Application technology solutions:

[0108] Sp1: Dynamic Adaptive Defect Detection: A multi-channel SEM image acquisition device captures wafer images at 500x and 5000x magnifications. The system dynamically adjusts the feature extraction range (high magnification focuses on particle details), and fusion weights are calculated in real time based on image entropy and process context (such as deposition steps). The high-performance computing unit detects density anomalies, identifies particle deposition as a new defect type, associates it with deposition temperature changes, and updates classification rules.

[0109] Sp2-Sp3: Causal Analysis and Optimization: The process parameter monitor collects the deposition chamber temperature (50°C). The system uses time-varying analysis to confirm that excessive temperature is the root cause. It predicts that the particle defect trend will increase to 10%, and generates the recommendation to "lower the temperature to 45°C."

[0110] Sp5: Closed-loop control: The closed-loop control interface unit transmits the recommendations to the deposition equipment, and after execution, the feedback defect rate is reduced to 2%;

[0111] Results: Dynamic adaptive detection identified 85% of new particle defects with 98% accuracy and a false positive rate of 2%, surpassing DeepSEM-Net's 0% new defect recognition rate and 15% missed detection rate. This shortened the R&D cycle by 30% and increased the yield by 7%, demonstrating the breakthrough of automatic new defect discovery.

[0112] Application Case 2: Real-time Control of Bridging Defects in Etching Steps in 5nm Processes

[0113] Background and Problem: During mass production of a 5nm process at a certain fab, the bridging defect rate after the etching step rose to 5%. The traditional SEMVisionG10 system only provided offline statistical reports with a response time of four hours. Manual adjustment of etching parameters only increased the yield by 3%, which was unable to meet the requirements of the high-volume production line.

[0114] Application technology solutions:

[0115] Sp2: Time-varying causal analysis: The process parameter monitor collects etching temperature (60°C) and pressure in real time. The system generates time-varying features through time decay encoding. Perturbation simulation (temperature rise of 1°C) quantifies that the bridging probability increases to 7%, confirming that temperature is the key factor.

[0116] Sp3: Trend Prediction and Optimization: The high-performance computing unit decomposes the data, predicts that the defect rate will rise to 8% within 24 hours, and collaboratively generates the recommendation to "lower the temperature to 55°C and clean the etching chamber." The adaptive storage server supports rapid data updates.

[0117] Sp5: Closed-loop control: The closed-loop control interface unit transmits the recommendations to the etching machine, and the adjustment is completed within 3 seconds. The feedback shows that the defect rate has dropped to 1.5%;

[0118] Sp4: Knowledge Graph: The interactive display terminal presents the "bridge → high temperature" path, allowing users to confirm the root cause and trigger subsequent inspections;

[0119] Results: The real-time closed-loop control response time is 3 seconds, the yield rate is increased by 7.5%, and the adjustment efficiency reaches 12 times / hour, far exceeding the 4-hour response and 0.5-time / hour adjustment of the SEMVisionG10. Production interruptions are reduced by 50% and costs are lowered by 10%, demonstrating the high efficiency of closed-loop optimization.

[0120] Case summary: Case 1 demonstrated dynamic adaptive detection and new defect discovery (Sp1), solving unknown defect problems through multi-scale fusion and density detection, significantly improving accuracy and adaptability. Case 2 verified real-time closed-loop control (Sp2-Sp5), optimizing the process through time-varying analysis and rapid intervention, significantly improving response speed and yield. The two cases, combined with hardware support, fully demonstrated the distinctive features and effects of the technical solutions, providing breakthrough solutions for semiconductor manufacturing.

[0121] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Automatic analysis and process defect detection system of SEM images based on the entire database, characterized by: The automatic analysis and process defect detection system includes the following steps: Sp1: Adaptively extracts and fuses features from multi-scale scanning electron microscope (SEM) images to detect and classify defect types, and supports automatic identification of new defect types; Sp2: Dynamically analyze the causal relationship between process steps and defects based on the time-dependent characteristics of defect type, process parameters, and equipment status, and isolate the interaction effects between process steps; Sp3: Predict defect trends based on the causal relationships and multi-source data, collaboratively generate process parameter adjustment suggestions and equipment maintenance plans, and optimize the predictions and suggestions through feedback; Sp4: Builds and updates in real time an interactive knowledge graph that includes defect types, process steps, and equipment status, allowing users to query defect root causes through path reasoning. Sp5: Seamlessly integrate the above analysis results with the semiconductor manufacturing process to perform real-time closed-loop control from defect detection to process optimization during semiconductor manufacturing.

2. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: When adaptively extracting and fusing features in step Sp1, comprehensive features suitable for different magnifications and process contexts are generated by dynamically adjusting the feature extraction range and fusion weights, and the fusion weights are calculated in real time based on the local statistical characteristics of the image and the type of process steps.

3. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: When dynamically analyzing the causal relationship between the process steps and the defects in step Sp2, historical data is encoded through a time decay mechanism to generate a time-varying feature vector, and the causal contribution of the process steps to the defects and the intensity of the interaction effect are calculated based on the time-varying feature vector.

4. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: When predicting defect trends in step Sp3, the SEM image features, process parameters and equipment status are decomposed into long-term trends and periodic fluctuations, the probability of future defect occurrence is predicted based on the decomposition results, and the priority of the collaboratively generated suggestions is adjusted according to the predicted probability.

5. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: When performing the interactive knowledge graph in step Sp4, it supports dynamic generation of the optimal causal path from defect type to process step through multi-layer path search and confidence assessment, and incrementally updates the graph structure and association strength based on new data. The interactive knowledge graph also includes an abnormal pattern recognition function, which identifies abnormal process patterns by analyzing the changes in the association strength between defect types and process steps in the graph, and automatically triggers process adjustment suggestions or equipment inspection instructions.

6. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: The automatic analysis and process defect detection system includes a data alignment unit for aligning multi-scale SEM images with process parameter and equipment status data in time and space dimensions, and embedding process context identifiers in the aligned data to enhance the contextual relevance of the analysis.

7. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: The automatic identification of new defect types in step Sp4 is achieved through density anomaly detection in the feature space, and the identified new defect types are associated with the trend of process parameter changes to generate a preliminary hypothesis of the cause of the new defect. After the hypothesis of the cause is verified by comparing with historical data, the defect classification rules of the system are automatically updated.

8. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: When dynamically analyzing the causal relationship between process steps and defects in step Sp2, the impact of small disturbances in process parameters is simulated to quantify the impact of disturbances on the defect probability, and a priority adjustment sequence of process steps is generated based on the quantified results to minimize the defect occurrence rate.

9. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: The collaborative generation of process parameter adjustment suggestions and equipment maintenance plans in step Sp3 includes: building a multi-objective optimization model based on defect trend prediction results and equipment operating status, optimizing defect rate reduction and maintenance cost control at the same time, and dynamically adjusting model weights during the optimization process to adapt to changes in production demand.

10. The automatic analysis and process defect detection system for SEM images based on a full database according to claim 1, characterized in that: The hardware components of the automatic analysis and process defect detection system include: Multi-channel SEM image acquisition device, equipped with multiple magnification lenses and real-time data transmission interface, used to collect multi-scale SEM images and transmit them to the full database; A high-performance computing unit, integrating a multi-core GPU and a dedicated acceleration chip, is used to support real-time computing of the adaptive feature fusion, time-varying causal analysis, and defect trend prediction; Process parameter and equipment status monitor, including a sensor array and time synchronization module, is used to collect process parameters and equipment operating status in real time and store them aligned with SEM image data; An interactive display terminal equipped with a high-resolution touch screen and a graph rendering engine for presenting the interactive knowledge graph and supporting user path reasoning operations; Adaptive storage server, equipped with a hierarchical cache structure and dynamic partitioning module, is used to dynamically allocate storage space based on the amount of SEM image data and process analysis requirements, and supports fast retrieval and update; A closed-loop control interface unit integrates a programmable logic controller (PLC) and a process equipment communication protocol, and is used to transmit process parameter adjustment suggestions and equipment maintenance plans to semiconductor manufacturing equipment in real time, and receive execution feedback signals to optimize subsequent analysis.

Images