Wafer defect detection method, device, equipment and medium

Abstract

The invention relates to the technical field of wafer defect detection, in particular to a wafer defect detection method, device and equipment and a medium, and the method comprises the steps: obtaining real-time process data in a wafer manufacturing process; inputting the real-time process data into a pre-constructed digital twinning model, simulating wafer full-process production by the digital twinning model based on historical manufacturing data, and generating a defect prediction value; acquiring actual defect detection data of the wafer through non-contact detection equipment; comparing the actual defect detection data with the defect prediction value, and calculating a deviation value; judging whether the deviation value exceeds a 5% preset threshold value or not; when the deviation value exceeds a preset threshold value, a model updating program is automatically triggered; and performing defect verification on a subsequent wafer based on the updated digital twinborn model, outputting a defect identification result, and controlling the detection equipment to continue to operate when the verification is passed, otherwise, triggering a reinspection process. According to the invention, self-adaptive defect detection is realized, and the detection precision and the adaptability of the system in a dynamic production line are remarkably improved.

Description

Technical Field

[0001] This application relates to the technical field of wafer defect detection, and in particular to a wafer defect detection method, system, electronic device and storage medium. Background Technology

[0002] Digital twin technology is increasingly used in wafer defect detection, improving the accuracy of defect prediction and verification by constructing virtual models to simulate the manufacturing process. This technology, combined with real-time inspection data, aims to achieve intelligent quality control and error correction.

[0003] In related technologies, the common approach is to compare real-time detection data with the prediction results of a virtual model, and automatically trigger a re-inspection procedure when the deviation exceeds a threshold. This method utilizes a process and defect correlation model to dynamically adjust detection parameters, demonstrating the progress of digital twins in improving detection reliability.

[0004] However, the updates to digital twin models in related technologies lag behind the rapid changes in manufacturing processes, and model retraining relies heavily on computational resources and is not responsive in a timely manner. This lag leads to increased discrepancies between predictions and actual measurements, making it difficult to accurately identify defects, increasing the risk of misjudgment, and reducing the system's adaptability in dynamic production lines. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems, this application provides a wafer defect detection method, system, electronic device, and storage medium.

[0006] In a first aspect, this application provides a wafer defect detection method, employing the following technical solution, including: Real-time process data during wafer manufacturing is acquired, including etching rate and thin film thickness, and is collected by sensors at a sampling frequency of not less than 1 kHz. The real-time process data is input into a pre-built digital twin model, which simulates the entire wafer production process based on historical manufacturing data to generate defect prediction values, wherein the defect prediction values ​​include defect type, location and size. The actual defect detection data of the wafer is obtained by non-contact inspection equipment, and the actual defect detection data includes terahertz wave scanning signals and image reconstruction results; The actual defect detection data is compared with the predicted defect value, and the deviation value is calculated. The comparison includes data normalization and feature extraction. Determine whether the deviation value exceeds a preset threshold of 5%; When the deviation value exceeds a preset threshold, the model update program is automatically triggered. The model update program includes dynamically adjusting the parameters of the digital twin model based on real-time process data and actual defect detection data. The adjustment is achieved through an incremental learning mechanism, without the need for full retraining. Based on the updated digital twin model, subsequent wafers are subjected to defect verification, and the defect identification results are output. If the verification is successful, the detection equipment is controlled to continue running; otherwise, a re-inspection process is triggered.

[0007] By adopting the above technical solution, and combining real-time data acquisition with dynamic updates of the digital twin model, the problem of inaccurate defect identification caused by lagging model updates is solved, achieving adaptive defect detection and significantly improving detection accuracy and the system's adaptability in dynamic production lines. Its integrated real-time feedback and incremental learning mechanism non-obviously improves model response speed; significantly reduces the false positive rate, and enhances the stability of yield in semiconductor mass production.

[0008] In one possible implementation, the step of acquiring real-time process data further includes: The real-time process data is preprocessed by distributed edge nodes, redundant information is filtered and key features are extracted, and only deviation-related data is sent to the central processing unit. The preprocessing includes data compression and noise reduction, and uses multi-level caching technology to temporarily store data to ensure seamless data transmission. Dynamically allocate computing resources and automatically adjust task priorities based on the detected load to avoid processing delays; The edge nodes include dedicated hardware processors for locally performing data normalization and feature comparison.

[0009] By adopting the above technical solution and introducing distributed edge preprocessing, the problem of real-time data processing latency was solved, and multi-level caching and dynamic resource allocation were realized. This significantly improved the speed of response, adapted to the pace of high-throughput production lines, and enhanced detection efficiency.

[0010] In one possible implementation, the step of triggering the model update procedure further includes: Construct a process-defect correlation model and dynamically optimize the parameters of the digital twin model based on real-time data stream; When the model health metric indicates that prediction errors have accumulated, the system will automatically switch to the backup model. Digital threading technology is used to connect the entire manufacturing process, ensuring that the virtual model and the physical entity are updated synchronously. The model update is triggered by a physical signal threshold without mathematical algorithms, thus avoiding complex calculations.

[0011] By adopting the above technical solutions, the accuracy of digital twins is enhanced through process and defect correlation and model health monitoring, and dynamic optimization and backup switching mechanisms are realized; the number of accidental re-inspections is significantly reduced, the adaptability of the system in diversified production lines is improved, and operational risks are reduced.

[0012] In one possible implementation, the step of acquiring actual defect detection data further includes: A sealed cavity was installed in the testing area to isolate the effects of temperature, humidity, and airflow. An integrated high-precision temperature control module maintains a constant temperature within the cavity; A reference calibration unit is embedded in the terahertz wave path to monitor the signal strength in real time and dynamically adjust the transmission parameters. Multi-band terahertz technology is used to switch frequency bands and offset environmental effects.

[0013] By adopting the above technical solution, through environmental adaptive shielding and compensation, the problem of accuracy fluctuation caused by interference of terahertz waves is solved, achieving sealed cavity and multi-band technology; significantly improving the signal-to-noise ratio and detection stability, and ensuring long-term accuracy maintenance.

[0014] In one possible implementation, the step of acquiring actual defect detection data via non-contact detection equipment further includes: Detection is performed using a reconfigurable optical array, which consists of multiple independently controlled microlens units; The lens position is adjusted using non-contact drive technology, reducing mechanical wear; The built-in optical encoder monitors lens displacement in real time and performs a self-calibration cycle to compare with a standard sample; Micro-motors are used for nanometer-level compensation to ensure accuracy of focal length and field of view.

[0015] By adopting the above technical solution and using a low-wear intelligent optical mechanism, the problems of mechanical wear and precision drift are solved, enabling non-contact driving and self-calibration; significantly extending component life, achieving seamless inspection of wafers of different sizes, and improving equipment durability.

[0016] In one possible implementation, the subsequent wafer defect verification step based on the updated digital twin model further includes: Integrated adaptive beamforming technology automatically optimizes detection parameters based on wafer size and material; The detection threshold is dynamically updated using a particle swarm optimization algorithm, and the system is automatically calibrated after every 20 wafers are inspected. The calibration process uses a standard defect template library for benchmark comparison.

[0017] By adopting the above technical solutions, through adaptive beamforming and dynamic calibration, the detection compatibility and accuracy are enhanced, and automatic parameter optimization and periodic calibration are achieved; the detection accuracy of the equipment is significantly maintained during continuous operation, and the yield stability is improved.

[0018] One possible implementation also includes: The detection system is decomposed into independent modules, including a terahertz emission module, a detection module, and an optical array module. Each module uses a standardized interface to enable quick plugging and unplugging. The built-in sensor network monitors the status of each module in real time, including temperature, vibration, and power consumption. When an anomaly is detected, the diagnostic program is automatically triggered or the system switches to a backup module. Predictive maintenance algorithms are used to warn of component degradation, and material optimization is combined to extend module life.

[0019] By adopting the above technical solutions, through modular design and self-maintenance mechanisms, the maintenance difficulties caused by high system complexity are solved, and standardized interfaces and predictive maintenance are achieved; the failure rate and operating costs are significantly reduced, and the reliability and scalability of the equipment are improved.

[0020] Secondly, this application provides a wafer defect detection system, comprising: The data acquisition module is configured to acquire real-time process data during wafer manufacturing via a sensor network. The real-time process data includes etching rate and thin film thickness, and the sampling frequency is not less than 1 kHz. A digital twin processing module is configured to store and run a digital twin model, which simulates the entire wafer production process based on historical manufacturing data and generates defect prediction values. The detection data acquisition module is configured to acquire actual defect detection data of the wafer through a non-contact detection device. The actual defect detection data includes terahertz wave scanning signals and image reconstruction results. The comparison and analysis module is configured to receive actual defect detection data and defect prediction values, compare the data and calculate the deviation value to determine whether it exceeds a preset threshold of 5%. The model update module is configured to trigger a model update program when the deviation value exceeds a preset threshold, dynamically adjusting the parameters of the digital twin model. The adjustment is achieved through an incremental learning mechanism, eliminating the need for full retraining. The defect verification module is configured to perform defect verification based on the updated digital twin model, output the defect identification results, and control the detection equipment to continue running when the verification is successful; otherwise, it triggers the re-inspection process. The modules are connected via a bus to enable real-time data interaction.

[0021] By adopting the above technical solutions, the system solves the problem of lagging updates in traditional models through the tight integration of digital twin models and real-time data feedback, realizing modular design and adaptive update mechanism; it significantly improves detection accuracy and efficiency, supports the high-throughput requirements of large-scale production lines, and reduces maintenance costs.

[0022] Thirdly, this application provides an electronic device including a memory and a processor, wherein the memory is used to store computer program code, and the processor is used to execute the computer program code stored in the memory to implement the methods in the first aspect and any one of the first aspects, or in the second aspect and any possible implementation of the second aspect.

[0023] Fourthly, this application provides a computer-readable storage medium storing a computer program or instructions that, when executed, implement the methods described in the first aspect and any one thereof, or in the second aspect and any possible implementation thereof. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of a wafer defect detection method provided in an embodiment of this application.

[0025] Figure 2 This is a schematic diagram of the structure of a wafer defect detection system provided in an embodiment of this application.

[0026] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] The technical solutions in this application will now be described with reference to all the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0028] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Furthermore, in the description of the embodiments of this application, "plural" or "multiple" refers to two or more than two.

[0029] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this embodiment, unless otherwise stated, "a plurality of" means two or more.

[0030] The terminology used in the following embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to also include expressions such as “one or more,” unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of this application, “at least one” and “one or more” refer to one, two, or more than two.

[0031] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "one embodiment," "some embodiments," "another embodiment," "other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0032] This application provides a wafer defect detection method executed by an electronic device. This electronic device can be a standalone physical electronic device, a cluster of multiple physical electronic devices, a distributed system, or a cloud electronic device providing cloud computing services. This application does not impose limitations on this method. Figure 1 As shown, the method includes: S1. Obtain real-time process data during wafer manufacturing.

[0033] The real-time process data includes etching rate and film thickness, which are acquired by sensors at a sampling frequency of not less than 1 kHz.

[0034] Specifically, this step requires a high-precision sensor system to acquire real-time process data. The etching rate data is obtained through an optical emission spectroscopy sensor installed at the etching equipment exit. This sensor, based on the principle of plasma emission spectroscopy, monitors the changes in the intensity of specific wavelengths of light during etching in real time, thereby deriving the etching rate. Thin film thickness data is acquired through an ellipsometer or interferometer sensor, which uses changes in polarized light or interference fringes to measure the film thickness and refractive index.

[0035] To ensure that the sampling frequency is not less than 1kHz, the sensor system is equipped with a high-speed data acquisition card. This acquisition card is based on FPGA design, can process multiple signals in parallel, and is connected to the main control unit through the PCIe interface to achieve low-latency data transmission.

[0036] Meanwhile, the sensor network communicates with the central control host via the industrial Ethernet protocol. The central control host runs data acquisition software, which adopts a multi-threaded architecture to receive, buffer, and preprocess sensor data in real time to avoid data loss or congestion.

[0037] In addition, the hardware includes a signal conditioning module for amplifying and filtering the raw signal, reducing environmental noise interference, and ensuring data accuracy.

[0038] This step utilizes high sampling frequency and multiple sensor fusion to capture subtle fluctuations in process parameters through real-time high-frequency monitoring, providing a reliable basis for subsequent defect prediction.

[0039] Furthermore, from a technical perspective, the implementation of this step relies on real-time data stream processing and multi-source information fusion.

[0040] The raw data collected by the sensors first undergoes analog-to-digital conversion, transforming analog signals into digital signals. Then, the data acquisition software performs timestamp alignment and data format standardization to ensure that data from different sensors are synchronized in time. The central control unit uses a circular buffer to store the real-time data stream, preventing data overflow under high load conditions.

[0041] Meanwhile, the software logic also includes anomaly detection algorithms, such as sliding window-based statistical methods, to identify and mark abnormal data points, such as instantaneous jumps or sensor malfunctions, thereby improving data quality.

[0042] Furthermore, high-frequency sampling and real-time processing enable the system to respond quickly to changes in the manufacturing environment, avoiding detection errors caused by data lag. Hardware support also includes distributed sensor nodes, which perform local preprocessing through edge computing units, reducing the load on the central processing unit and ensuring system stability in high-throughput production lines.

[0043] In summary, firstly, high-frequency sampling of at least 1kHz ensures the real-time nature and completeness of process data, enabling the system to promptly capture rapid fluctuations in etching rate and film thickness. This provides accurate input for the digital twin model, reducing misjudgments of defects caused by inaccurate data. Secondly, multi-sensor fusion and signal conditioning technologies improve the signal-to-noise ratio of the data and reduce the impact of environmental factors such as temperature drift or electromagnetic interference on data quality, indirectly improving the accuracy of subsequent defect detection. Furthermore, this step, through the collaborative design of hardware and software, achieves low-latency data processing, adapting to the high-throughput requirements of large-scale production lines, shortening the single-wafer inspection cycle, and providing data support for process optimization through real-time monitoring, ultimately enhancing the yield stability of semiconductor mass production. All of these address the problem of data acquisition lag in traditional inspection methods, demonstrating the innovation of the adaptive inspection system.

[0044] In some embodiments, to achieve rapid response, adapt to the pace of high-throughput production lines, and improve testing efficiency, S1 further includes: S100 preprocesses real-time process data through distributed edge nodes, filters redundant information and extracts key features, and sends only deviation-related data to the central processing unit.

[0045] The preprocessing operations include data compression and noise reduction, and multi-level caching technology is used to temporarily store data to ensure seamless data transmission.

[0046] Specifically, in implementing this step, distributed edge nodes are deployed near the wafer fabrication equipment, such as in the control cabinet of an etching or deposition equipment. Each node consists of a dedicated hardware processor equipped with high-speed memory and local storage.

[0047] Preprocessing operations include data compression and noise cancellation. Data compression uses lossless compression algorithms to reduce data volume and is performed by hardware accelerators to reduce processing latency. Noise cancellation is achieved through digital filters, such as finite impulse response filters or wavelet transforms, to remove high-frequency noise and environmental interference from sensor signals.

[0048] Its multi-level caching technology adopts a hierarchical storage structure, including a level 1 cache (based on SRAM, used for temporary storage of raw data), a level 2 cache (based on DRAM, used for storing preprocessing intermediate results), and a level 3 cache (based on NVMe SSD, used for persistent critical data). Data flow is managed through a cache controller to ensure seamless data transmission.

[0049] Its edge nodes communicate with the central processing unit via industrial Ethernet or 5G modules, using lightweight protocols to transmit data, and only sending deviation-related data such as process parameters whose deviation from historical benchmarks exceeds a set threshold.

[0050] Its software logic includes an edge computing framework running on a real-time operating system, enabling pipelined operations for data acquisition, preprocessing, and transmission. This step employs a distributed architecture and multi-level caching to reduce central load through local preprocessing, achieving low-latency data processing.

[0051] Furthermore, from a technical perspective, distributed edge nodes operate based on edge computing principles, offloading computing tasks from the central unit to the network edge, thereby reducing data transmission volume and latency.

[0052] The data compression removes redundant information through algorithms, such as differential coding using the correlation of time-series data; the noise cancellation is based on signal processing theory, using frequency domain analysis to isolate useful signals and noise.

[0053] This multi-level caching technology is based on the principle of computer storage hierarchy. It temporarily stores data through high-speed caching to avoid data loss due to network congestion, and uses the Least Recently Used algorithm to manage cache replacement.

[0054] The edge node communicates with the central unit using a publish-subscribe model, ensuring asynchronous data transmission and improving system reliability. This step decomposes complex data processing tasks across multiple edge nodes for parallel execution, thereby improving processing efficiency. Hardware support also includes temperature sensors and watchdog timers to monitor node health and ensure stable operation in harsh industrial environments.

[0055] In summary, firstly, distributed preprocessing reduces the data load on the central processing unit, avoiding bottlenecks caused by large data streams, thereby reducing data processing latency to the millisecond level, adapting to the fast pace of Qualcomm's mass production lines. Secondly, data compression and noise cancellation improve data quality, ensuring that only deviation-related data is transmitted, reducing network bandwidth usage, and improving transmission efficiency. Multi-level caching technology ensures the continuity and integrity of data during transmission, preventing data loss due to temporary interruptions. Furthermore, it solves the real-time data processing latency problem, achieving rapid response through edge intelligence, significantly improving detection efficiency, and ultimately contributing to the stability of semiconductor mass production yield and the reduction of production costs.

[0056] S101: Dynamically allocate computing resources and automatically adjust task priorities based on the detected load to avoid processing delays.

[0057] The edge nodes include dedicated hardware processors for performing data normalization and feature comparison locally.

[0058] Specifically, in implementing this step, the dynamic allocation of computing resources is performed by a resource management module, which is integrated into the edge node or central control system. The hardware includes heterogeneous computing units, multi-core processors combining CPUs and GPUs, and resource monitoring sensors such as power consumption and temperature sensors.

[0059] Its resource allocation is based on real-time load monitoring, such as using a load balancer to monitor data inflow rates and processing queue lengths, and automatically adjusting task priorities. For example, high-priority tasks like real-time defect comparison are allocated more computing cores and memory bandwidth, while low-priority tasks like historical data archiving are processed with a delay.

[0060] Its edge nodes have dedicated hardware processors for performing data normalization and feature matching locally. Data normalization is achieved through hardware-accelerated arithmetic logic units, such as scaling sensor data to a uniform range like 0-1. Feature matching uses a parallel processing architecture, such as using the SIMD instruction set to process multiple data points simultaneously to extract key features such as defect size or shape.

[0061] Its software logic includes real-time scheduling algorithms such as earliest deadline first or rate monotonic analysis, dynamically allocating CPU time and memory resources, and interacting with upper-layer applications via API. This step employs adaptive resource management and hardware acceleration, ensuring timely completion of critical tasks through intelligent scheduling.

[0062] Furthermore, from a technical principle perspective, this dynamic resource allocation is based on control theory and queue management principles. It adjusts resource allocation in real time through feedback loops. The resource management module periodically collects system indicators such as CPU utilization and memory usage, and uses PID controllers or machine learning models such as reinforcement learning to optimize task scheduling.

[0063] The dedicated processor at this edge node is based on parallel computing principles, employing fixed-point arithmetic or a floating-point coprocessor for data normalization to ensure both accuracy and speed. Feature comparison utilizes logic circuits implemented in a hardware description language to achieve high-speed pattern matching. This step enables the system to automatically optimize its performance based on load, reducing external intervention. Software support also includes a resource reservation mechanism to guarantee minimum resources for critical tasks and avoid latency caused by resource contention.

[0064] In summary, dynamic resource allocation ensures system stability under high loads, avoids processing delays by prioritizing critical tasks, and thus improves detection response speed. The use of dedicated hardware processors accelerates local data normalization and feature matching, reduces reliance on the central unit, and further improves efficiency. Furthermore, it solves the bottleneck problem caused by traditional fixed resource allocation, enhances the system's adaptability to changing production line environments, and ultimately achieves a significant improvement in detection efficiency and a reduction in false detection rate.

[0065] Based on this, by introducing distributed edge preprocessing, the problem of real-time data processing latency was solved, and multi-level caching and dynamic resource allocation were realized; significantly, rapid response was achieved, adapting to the pace of high-throughput mass production lines and improving detection efficiency.

[0066] In this embodiment, the method further includes: S2. Input real-time process data into a pre-built digital twin model.

[0067] The digital twin model simulates the entire wafer production process based on historical manufacturing data, generating defect prediction values, which include defect type, location, and size.

[0068] Specifically, in implementing this step, the digital twin model is deployed on a high-performance computing server equipped with a multi-core CPU and a large amount of memory, which in this embodiment is more than 64GB, to handle complex simulation calculations.

[0069] Its model is built on a machine learning framework and trained using historical manufacturing data, which includes past wafer process parameters, defect records and environmental conditions. The data is stored in a distributed database and is preprocessed through data cleaning and feature engineering to ensure the consistency of the model input.

[0070] This model simulates the entire wafer manufacturing process, covering key process stages such as deposition, photolithography, and etching. By combining physical equations and data-driven methods, it generates defect prediction values. For example, it identifies defect patterns through convolutional neural networks and outputs defect types such as particle contamination or scratches, their locations in coordinate form, and micron-level quantized dimensions.

[0071] The central control unit sends real-time process data to the model server via an API interface, and the data transmission uses an encrypted protocol to ensure security. Hardware facilities also include GPU accelerator cards to improve model inference speed and ensure real-time response. This step utilizes the deep fusion of digital twins and historical data to accurately map the physical world through a virtual model, enabling early defect prediction.

[0072] Furthermore, from a technical principle perspective, the digital twin model operates based on multiphysics simulation and statistical learning principles. The model first receives real-time process data as input, then uses a built-in finite element analysis module to simulate the stress distribution and material properties inside the wafer, thereby predicting the probability of defect occurrence. Simultaneously, the model employs time series analysis to process the continuous data stream, capturing the impact of process trends on defects.

[0073] Its software logic includes model version management, allowing dynamic loading of different model versions to adapt to production line changes, and reducing latency from redundant calculations through a caching mechanism. This step achieves closed-loop feedback between prediction and reality by replicating the real manufacturing process through digital twins. Hardware support also includes a network storage system for rapid access to model parameters and historical data, ensuring stable model operation under high-pressure environments.

[0074] In summary, digital twin models, based on historical data simulation, can identify potential defects in advance, reduce blind spots in actual detection, and thus lower the false detection rate and missed detection rate. By generating detailed defect prediction values ​​of type, location, and size, they provide a benchmark for subsequent comparisons and improve the accuracy of defect identification.

[0075] In addition, the high-performance computing support of the model ensures real-time performance, adapts to the rapid changes in dynamic production lines, solves the problem of lagging updates in traditional models, enhances the system's adaptability, helps improve the yield stability of semiconductor mass production, and reduces scrap costs caused by defects.

[0076] S3. Obtain actual defect detection data of wafers through non-contact inspection equipment.

[0077] The actual defect detection data includes terahertz wave scanning signals and image reconstruction results.

[0078] Specifically, in implementing this step, the non-contact detection equipment mainly includes a terahertz wave emitter, a superconducting detector, and an image reconstruction unit.

[0079] Its terahertz wave transmitter generates terahertz waves based on a femtosecond laser-excited array structure. Its hardware includes a laser source, an optical modulator, and a waveguide array, and it can emit electromagnetic waves with frequencies in the range of 0.1-10THz, which can penetrate wafer materials.

[0080] Its superconducting detector operates using a cryogenic cooling system such as liquid helium cooling to receive transmitted or reflected terahertz signals with high sensitivity and convert analog signals into digital signals.

[0081] Its image reconstruction unit is based on FPGA or ASIC hardware and executes diffraction theory algorithms to convert terahertz time-domain spectral data into three-dimensional tomographic images, showing the spatial distribution of defects inside the wafer.

[0082] The equipment docks with the wafer via a robotic arm or conveyor belt, ensuring non-contact scanning and avoiding surface contamination. The central control unit manages the inspection process, receiving raw data via Ethernet communication and running image processing software for preliminary filtering and enhancement. This step leverages the high penetration of terahertz waves and the high signal-to-noise ratio of superconducting detectors to achieve precise detection of micron-level defects.

[0083] Furthermore, from a technical principle perspective, terahertz wave scanning is based on the interaction principle between electromagnetic waves and semiconductor materials. When terahertz waves penetrate a wafer, they will be scattered or absorbed when they encounter defect regions. By analyzing the amplitude and phase changes of the time-domain spectrum, the defect characteristics can be deduced.

[0084] Image reconstruction algorithms are based on diffraction theory and convert signal data into a three-dimensional map by solving an inverse problem, such as using deconvolution methods to enhance resolution.

[0085] Its software logic includes a data calibration module that uses reference samples to correct system errors and ensure image accuracy. This step avoids physical contact using terahertz waves, reducing damage to the wafer, while the low-noise design of the superconducting detector improves signal quality and adapts to diverse wafer materials.

[0086] In summary, terahertz wave scanning provides high-resolution internal defect data, which, combined with image reconstruction, enables accurate identification of micron-level defects, reducing the false negative rate. The non-contact design ensures wafer integrity and is suitable for sensitive material inspection. Furthermore, high signal-to-noise ratio signal acquisition improves inspection reliability, overcomes the limitations of traditional optical inspection, enhances the equipment's compatibility with 8-12 inch wafers, and ultimately improves inspection efficiency and production line yield.

[0087] In some embodiments, to improve the signal-to-noise ratio and detection stability, and to ensure long-term accuracy maintenance, S3 further includes: The S300 adds a sealed cavity to the detection area to isolate the effects of temperature, humidity and airflow.

[0088] Specifically, in implementing this step, the sealed cavity adopts a multi-layer composite structure design, with an outer stainless steel metal shell, an inner electromagnetic shielding coating, and vacuum insulation material filling the middle.

[0089] The sealed chamber is connected to the testing equipment base using a precision packaging process, with fluororubber sealing rings at the interfaces to ensure airtightness. A high-transmittance terahertz wave window is installed on the front of the chamber, made of high-resistivity silicon material with an anti-reflective coating to reduce signal attenuation. An environmental monitoring sensor network, including high-precision temperature and humidity sensors and micro-differential pressure sensors, is arranged inside the chamber to monitor environmental parameters in real time. The chamber is completely isolated from the external environment, and a slight positive pressure is maintained internally by a mechanical pump system, effectively blocking interference from external temperature fluctuations, humidity changes, and airflow on the testing process.

[0090] The cavity is equipped with an automatic door mechanism on its side, driven by a stepper motor, which opens only briefly when the wafer is moving in or out. This step employs a multi-layered composite structure design combined with active pressure control, creating a stable testing environment through physical isolation.

[0091] Furthermore, the stainless steel outer shell provides mechanical strength and electromagnetic shielding, while the vacuum insulation layer stabilizes temperature by reducing heat conduction, and the slightly positive pressure prevents external air infiltration. An environmental monitoring sensor network, based on a multi-parameter fusion detection principle, collects real-time temperature, humidity, and pressure data within the chamber and dynamically adjusts the operating status of the mechanical pump unit via a PID controller. The chamber design eliminates external interference factors by creating a closed detection environment. Hardware support also includes an emergency pressure relief valve and safety interlock devices to ensure safe operation of the chamber under abnormal conditions. The entire system communicates with the main control unit via an industrial bus for centralized monitoring and management.

[0092] In summary, the sealed cavity effectively isolates external environmental fluctuations, providing a stable working environment for terahertz wave detection and reducing signal fluctuations caused by environmental factors. The multi-layer composite structure design ensures long-term sealing reliability and avoids performance degradation due to material aging. Combined with real-time environmental monitoring and active pressure control, the system can maintain optimal detection conditions under various operating conditions, improving the signal-to-noise ratio and detection stability, solving the environmental sensitivity problem of traditional open detection systems, and laying a solid foundation for maintaining long-term accuracy.

[0093] S301 integrates a high-precision temperature control module to maintain a constant temperature within the cavity.

[0094] Specifically, the high-precision temperature control module adopts a multi-stage temperature control system design, including a semiconductor cooling chip, a heating element, and a circulating liquid cooling device. The semiconductor cooling chip, based on the Peltier effect, is installed on the four walls of the cavity and achieves cooling or heating functions by changing the direction of the current; the heating element is a thin-film platinum resistance heater, evenly distributed on the inner surface of the cavity; the circulating liquid cooling device consists of a micro pump, a radiator, and heat conduction pipes, using an aqueous ethylene glycol solution as the heat conduction medium.

[0095] In addition, 12 monitoring points are distributed inside the cavity to collect temperature data in real time. The core of the temperature control module uses a 32-bit microcontroller that executes a fuzzy PID control algorithm to dynamically adjust the working state of each temperature control element according to the set temperature value. The system is also equipped with a temperature calibration unit to automatically calibrate the sensor accuracy periodically. This step adopts a multi-level composite temperature control architecture to achieve an ultra-stable thermal environment through precise temperature feedback control.

[0096] Furthermore, the thermoelectric cooler achieves directional heat transfer through carrier migration, while the heating element generates heat energy based on the Joule effect; the two work together to achieve rapid temperature regulation. The circulating liquid cooling device removes excess heat through forced convection heat transfer, maintaining the system's thermal balance.

[0097] Fuzzy PID control algorithm dynamically adjusts control parameters by calculating temperature deviation and its rate of change in real time, thus overcoming the shortcomings of traditional PID in nonlinear systems.

[0098] In summary, the high-precision temperature control system effectively controls temperature fluctuations within the cavity, greatly reducing the impact of thermal expansion and contraction on optical components. The multi-level temperature control architecture ensures rapid response and uniform distribution of temperature adjustment, avoiding the generation of local hot spots. Combined with intelligent control algorithms, the system can adaptively compensate for changes in ambient temperature, keeping the terahertz wave transmitter and detector at their optimal operating temperature, improving signal stability and detection repeatability, solving the accuracy attenuation problem caused by temperature drift, and ensuring that the system maintains detection accuracy during long-term operation.

[0099] S302. A reference calibration unit is embedded in the terahertz wave path to monitor the signal strength in real time and dynamically adjust the transmission parameters.

[0100] Specifically, the reference calibration unit is installed in the optical path between the terahertz wave transmitter and the detector, and consists of three main parts: a reference sample wheel, a signal monitoring module, and a parameter adjuster.

[0101] Its reference sample wheel, driven by a stepper motor, carries standard wafer samples and empty sample positions, periodically intervening in the optical path for benchmark calibration. Its signal monitoring module, based on a Schottky diode detector array, monitors the intensity, frequency, and phase characteristics of the terahertz wave in real time. Its parameter regulator, composed of a voltage-controlled attenuator and a phase shifter, dynamically adjusts the output power (adjustment range 10-100mW) and frequency (adjustment range 0.1-3THz) of the terahertz transmitter based on the monitoring results.

[0102] Its reference calibration unit is connected to the main control system via a high-speed data interface and uses an adaptive control algorithm to optimize transmission parameters in real time. This step employs an embedded real-time calibration mechanism, maintaining optimal signal quality through closed-loop feedback.

[0103] Furthermore, the reference calibration unit operates based on the principles of comparative measurement and adaptive control. The standard wafer sample provides a reference for known reflection and transmission characteristics. By comparing the difference between the measured signal and the reference signal, the system can accurately assess the current signal state.

[0104] The signal monitoring module is based on the interaction principle between terahertz waves and semiconductor materials, deriving signal quality by detecting amplitude and phase changes in the transmitted signal. The parameter regulator uses voltage-controlled elements to precisely adjust the intensity and spectral characteristics of the terahertz wave by changing the bias voltage.

[0105] The adaptive control algorithm is based on the least squares principle and finds the optimal combination of launch parameters through iterative calculation. This step establishes a complete detection-comparison-adjustment closed-loop control system.

[0106] In summary, the embedded reference calibration unit enables real-time monitoring and compensation of terahertz signals, effectively controlling signal fluctuation amplitude. The dynamic parameter adjustment function allows the system to adaptively compensate for the effects of device aging and environmental changes, maintaining the optimal signal-to-noise ratio. The periodic automatic calibration mechanism ensures the accuracy of the detection benchmark, avoiding accuracy drift during long-term use. These improvements solve the signal attenuation and drift problems in traditional systems, significantly improving detection stability and reliability, and ensuring that the system maintains the specified detection accuracy even after continuous operation.

[0107] S303 uses multi-band terahertz technology to switch frequency bands and offset environmental effects.

[0108] Specifically, the multi-band terahertz system consists of a tunable terahertz source, a broadband detector, and a frequency band control unit.

[0109] The tunable terahertz source, based on a photoconductive antenna array design, achieves rapid switching across 16 discrete frequency bands within the 0.1-3 THz range by adjusting the pump parameters of the femtosecond laser. The broadband detector employs a pyroelectric detector array, coupled with a bandpass filter bank, to simultaneously receive echo signals from multiple frequency bands. The frequency band control unit, based on an FPGA chip design, incorporates a built-in frequency band selection algorithm that automatically selects the optimal operating frequency band based on environmental monitoring data.

[0110] The system is also equipped with a frequency band calibration database, which stores the performance characteristics of each frequency band under different environmental conditions.

[0111] The frequency band switching process is achieved through an RF switch matrix, with a switching time of less than 100ms. This step employs a multi-band collaborative working mechanism, effectively overcoming environmental interference through intelligent frequency band selection.

[0112] Furthermore, multi-band terahertz technology is based on the frequency dependence principle of the interaction between terahertz waves and matter. Terahertz waves in different frequency bands have varying sensitivities to environmental factors. For example, the low-frequency band of 0.1-1 THz is less sensitive to changes in humidity, while the high-frequency band of 1-3 THz provides higher spatial resolution.

[0113] The frequency band selection algorithm is based on fuzzy logic decision-making, comprehensively considering multiple factors such as temperature, humidity, and wafer material to calculate the suitability index of each frequency band. Frequency band switching is achieved by changing optical pump conditions and filter settings, ensuring fast and reliable frequency band conversion.

[0114] This system employs time-division multiplexing technology, allowing multiple frequency bands to be used sequentially for scanning within a single detection cycle, thereby improving detection reliability through data fusion. This step optimizes system performance by dynamically adjusting the operating frequency band.

[0115] In summary, multi-band terahertz technology enables the system to intelligently select the optimal operating frequency band based on environmental conditions, effectively reducing signal attenuation caused by environmental factors; multi-band data fusion provides more comprehensive defect information, improving the ability to identify minute defects; and the fast frequency band switching function ensures continuous detection and avoids detection interruptions caused by sudden environmental changes. These improvements address the inherent sensitivity of single-band systems to the environment, significantly enhancing the system's adaptability and reliability in changing environments, ensuring stable long-term detection accuracy, and expanding the applicability of the equipment in diverse production lines.

[0116] Based on this, the problem of accuracy fluctuation caused by interference of terahertz waves is solved by environmental adaptive shielding and compensation, achieving sealed cavity and multi-band technology; significantly improving the signal-to-noise ratio and detection stability, and ensuring long-term accuracy maintenance.

[0117] In some embodiments, to extend component lifespan, enable seamless inspection of wafers of different sizes, and improve equipment durability, S3 further includes: S305. Detection is performed using a reconfigurable optical array.

[0118] The array consists of multiple independently controlled microlens units.

[0119] Specifically, the reconfigurable optical array is deployed in the optical path of the detection system and consists of 36 independent microlens units. Each unit is designed based on microelectromechanical systems technology and includes a microlens and a drive interface.

[0120] Its microlens unit is managed by an array controller, which uses a multi-channel digital signal processor. Each channel independently outputs a control voltage to adjust the tilt angle and position of the microlens to meet the inspection needs of wafers of different sizes, such as 8 inches or 12 inches.

[0121] Its hardware includes an array substrate, wiring layers, and a heat dissipation module. The software logic is integrated into the central control unit, which runs a configuration algorithm to automatically calculate the optimal parameters for each microlens based on the wafer size and sends commands via a serial communication protocol. This step leverages the independent controllability of the microlens units and the reconfigurability of the array to achieve flexible optimization of the optical path through distributed adjustment.

[0122] Furthermore, the reconfigurable optical array operates based on the principles of wave optics and adaptive optics. Each microlens unit acts as an independent optical element, adjusting the beam focusing characteristics by changing its curvature or position; the array as a whole compensates for optical aberrations caused by changes in wafer surface morphology through wavefront sensing and correction algorithms.

[0123] The control logic employs a time-division multiplexing strategy, activating the microlens array in turn to avoid signal interference, and utilizes a feedback loop to optimize lens parameters in real time. This step decomposes the complex optical task into multiple micro-units for parallel processing, improving the system's adaptability and accuracy. Hardware support also includes a power management chip to ensure stable power supply to each unit and reduce the impact of voltage fluctuations on performance.

[0124] In summary, firstly, the independently controlled microlens array enables the system to quickly reconstruct the optical path and seamlessly switch between detection modes for wafers of different sizes, reducing equipment changeover time to less than 30 seconds; secondly, the reconfigurable design reduces the need for mechanical adjustments, lowers component wear, and extends the lifespan of the optical system; furthermore, adaptive optimization improves the consistency of the detection field of view and focal length, ensuring the accuracy of defect identification, solving the compatibility issues of traditional fixed optical systems, and ultimately improving the durability and reliability of the equipment in diverse production lines.

[0125] S306: Adjusts lens position using non-contact drive technology, reducing mechanical wear.

[0126] Specifically, non-contact drive technology employs electromagnetic levitation or electrostatic drive mechanisms and is applied to the displacement adjustment of microlens units.

[0127] Its electromagnetic levitation drive is based on the Lorentz force principle. Each microlens unit is embedded with a micro copper wire wound coil and an array of permanent magnets is arranged on the outside. By controlling the magnitude and direction of the current, magnetic force is generated to push the lens to move in three-dimensional space without contact.

[0128] Its electrostatic drive utilizes the capacitance effect, applying voltage between the lens and the electrode plate to generate electrostatic force and achieve nanoscale positioning.

[0129] The drive hardware includes a power amplifier, position sensor, and feedback circuitry to ensure precise control. The software logic integrates a PID control algorithm to calculate the deviation between the target position and the actual position in real time and adjust the drive signal accordingly. This step employs a non-mechanical contact drive method, fundamentally eliminating wear caused by friction.

[0130] Furthermore, non-contact drive technology, based on electromagnetic or electrostatic principles, uses magnetic or electrostatic forces to replace traditional mechanical linkages or gears, achieving smooth adjustment of the lens position. Its electromagnetic levitation utilizes Ampere's and Faraday's laws, where the interaction between coil current and magnetic field generates thrust; electrostatic drive is based on Coulomb's law, where voltage changes adjust the magnitude of the attractive force.

[0131] In the control system, position feedback is achieved through Hall effect sensors or capacitive sensors, forming a closed-loop regulation to ensure stability and repeatability. This step improves the long-term reliability of the system by eliminating physical contact points, reducing wear and heat generation. Hardware support also includes overvoltage protection circuitry to prevent damage to the drive components.

[0132] In summary, non-contact drive technology significantly reduces mechanical wear and fatigue failure, extending the service life of lens components; smooth displacement adjustment avoids jitter and impact, improving detection stability and accuracy; in addition, the low-power design reduces heat load, reduces reliance on cooling systems, and solves durability issues under high-frequency adjustments, ultimately improving the overall economy and sustainability of the equipment by extending component life.

[0133] The S307 features a built-in optical encoder that monitors lens displacement in real time and performs a self-calibration cycle to compare with a standard template.

[0134] Specifically, its optical encoder is integrated near each microlens unit and is designed based on the principle of grating ruler or interferometer, including light sources such as LEDs or laser diodes, grating disks and photodetectors.

[0135] The encoder monitors lens displacement in real time with a resolution of 10 nanometers, and the data is transmitted to the processing unit via a high-speed serial interface.

[0136] Its self-calibration cycle is performed by a calibration module that stores image data of a standard template, such as a reference wafer with known defects. For example, it automatically compares the current lens position with the template reference in a periodic manner after inspecting 20 wafers, calculates the deviation, and generates a correction signal.

[0137] The software logic includes a calibration algorithm that uses the least squares method to fit the error curve and update the control parameters. The hardware includes an encoder circuit board, a template library robotic arm, and a communication bus. This step integrates real-time monitoring and automatic calibration, maintaining system accuracy through closed-loop feedback.

[0138] Furthermore, the optical encoder operates based on the principle of moiré fringes or interference fringes. When the lens moves, the grating generates a periodic signal, which is then decoded by a counter to determine the displacement value. Self-calibration identifies systematic errors such as drift or deviation by comparing the measured data with known values ​​of a standard template.

[0139] This calibration cycle employs an event-triggered mechanism, automatically initiating when the cumulative usage time or number of checks reaches a threshold, ensuring timely correction. This step enables the system to autonomously optimize performance, reducing the need for external intervention. Hardware support includes a redundant encoder array, enhancing monitoring reliability.

[0140] In summary, real-time displacement monitoring ensures the accuracy of lens position and effectively controls long-term accuracy drift; self-calibration cycle automatically corrects system errors, reducing the frequency of manual calibration and improving testing efficiency; in addition, comparison with standard samples provides a traceable benchmark, enhancing the reliability of test results, solving the problem of accuracy decay caused by mechanical wear, and ultimately extending the effective service life of the equipment by maintaining high accuracy.

[0141] The S308 employs a micro-motor for nanometer-level compensation to ensure accuracy in focal length and field of view.

[0142] Specifically, its micro motors, based on piezoelectric ceramic or voice coil motor technology, are installed in the drive mechanism of the microlens unit to provide nanometer-level displacement compensation.

[0143] Piezoelectric micromotors utilize the inverse piezoelectric effect, where ceramic deformation drives a lens when voltage is applied; voice coil motors are based on electromagnetic principles, where the coil moves in a magnetic field to achieve precise adjustment.

[0144] The compensation hardware includes a motor driver, position feedback sensors such as strain gauges or laser rangefinders, and a control unit. The software logic integrates an adaptive compensation algorithm that dynamically calculates the compensation amount based on optical encoder data and calibration results, and outputs pulse signals to drive the micromotor. The hardware also includes heat sinks and vibration damping mounts to ensure stable operation. This step employs a nanometer-precision active compensation mechanism to counteract the effects of environmental or mechanical changes through fine adjustments.

[0145] Furthermore, this micro-motor compensation is based on the principles of precision drive and closed-loop control. The piezoelectric motor achieves high-resolution movement through the voltage-displacement linear relationship, while the voice coil motor utilizes the current-force proportional characteristics for rapid response.

[0146] Its compensation algorithm is based on a predictive model, taking into account external factors such as temperature and vibration, and pre-calculates compensation values ​​to reduce latency. This step improves macroscopic accuracy through minor adjustments, enhancing the system's robustness. Hardware support includes a temperature compensation circuit to counteract the effects of thermal expansion.

[0147] In summary, nanometer-level compensation ensures long-term accuracy of focal length and field of view, reducing detection errors to below the micrometer level; the active adjustment mechanism responds quickly to environmental changes, maintaining detection stability; in addition, low-power micromotors reduce energy consumption, extend component life, and solve drift problems in high-precision systems, ultimately supporting the need for seamless inspection of wafers of different sizes by improving device durability.

[0148] Based on this, the problems of mechanical wear and precision drift are solved by using a low-wear intelligent optical mechanism, enabling non-contact drive and self-calibration; significantly extending component life, achieving seamless inspection of wafers of different sizes, and improving equipment durability.

[0149] In this embodiment, the method further includes, S4. Compare the actual defect detection data with the predicted defect value, and calculate the deviation value.

[0150] The comparison operation includes data normalization and feature extraction.

[0151] Specifically, during this step, the comparison operation is performed by a dedicated processing module on the central control host, which is designed based on a multi-core processor and high-speed memory.

[0152] Data normalization first converts the actual detection data and predicted values ​​into a uniform scale, such as through min-max scaling or Z-score normalization, to eliminate dimensional differences. Feature extraction uses digital signal processing technology to extract key features from terahertz images and predicted values, such as defect area, contour shape, and grayscale distribution, and uses principal component analysis or wavelet transform to reduce data dimensionality.

[0153] The deviation value is calculated using Euclidean distance or correlation coefficient methods to quantify the difference between the actual and the predicted values, and the results are stored in a temporary buffer.

[0154] Its software logic includes a parallel computing framework that utilizes OpenMP or CUDA to accelerate the alignment process and ensure real-time performance. The hardware also includes a data bus for high-speed data transmission between modules. This step leverages multi-feature fusion and real-time alignment to improve the accuracy of deviation analysis through refined data processing.

[0155] Furthermore, from a technical principle perspective, the comparison process is based on pattern recognition and statistical learning principles. Data normalization ensures comparability of data from different sources, for example, mapping terahertz signal amplitude and predicted size values ​​to the [0,1] range; feature extraction highlights key defect attributes and reduces redundant information through convolution operations or filter banks.

[0156] The deviation calculation combines a weighted average, assigning different weights based on defect type to more accurately reflect differences. This step generates control signals through real-time comparison, providing input for model updates. Software support also includes anomaly logging for subsequent analysis and optimization.

[0157] In summary, accurate deviation calculation enables the system to promptly identify detection errors and trigger subsequent update procedures, thereby reducing the false positive rate. Data normalization and feature extraction improve the robustness of the comparison, adapting to wafers of different sizes and materials. Furthermore, real-time processing capabilities ensure the efficiency of high-throughput production lines, solving the lag problem of traditional static inspection and enhancing the system's adaptability and reliability.

[0158] S5. Determine whether the deviation value exceeds the preset threshold of 5%.

[0159] Specifically, in implementing this step, the judgment operation is performed by the threshold comparison module on the central control host, which is based on digital comparator hardware or software logic.

[0160] Its preset threshold is set to 5% and stored in non-volatile memory, which can be dynamically adjusted through the configuration interface according to production line needs. After the deviation value is obtained from step S4, the comparison module performs arithmetic operations to calculate the relative error percentage, i.e., (|actual value - predicted value| / predicted value) × 100%, and compares it with the threshold.

[0161] Its software logic includes multi-condition judgments, such as setting differentiated thresholds for different defect types to improve flexibility. The hardware also includes a real-time clock to add timestamps to the judgment operations for historical traceability. This step utilizes dynamically adjustable thresholds and multi-condition judgments to enhance system adaptability through intelligent decision-making.

[0162] Furthermore, from a technical principle perspective, the judgment process is based on the threshold detection principle in control theory, which uses a comparator to output a binary result "yes / no" to trigger subsequent actions.

[0163] Its software logic integrates state machine management to ensure that judgment operations are executed under strict timing, avoiding race conditions. This step decomposes complex decisions into simple comparisons, improving reliability and response speed. Hardware support includes a watchdog timer to prevent system deadlock and ensure continuous operation.

[0164] In summary, automated threshold judgment reduces human intervention, improves detection efficiency, and ensures timely triggering of re-inspection when deviations exceed limits; dynamic threshold adjustment allows the system to adapt to process fluctuations, reducing the risk of false detections. Furthermore, it solves the rigidity problem of traditional fixed threshold methods, enhances the applicability of the equipment in diverse production lines, and ultimately contributes to improved yield stability.

[0165] S6. When the deviation value exceeds the preset threshold, the model update program is automatically triggered.

[0166] The model update procedure includes dynamically adjusting the parameters of the digital twin model based on real-time process data and actual defect detection data. This adjustment is achieved through an incremental learning mechanism, eliminating the need for full retraining.

[0167] Specifically, during this step, the model update procedure is executed by a dedicated update module integrated into the model server or central control host. The hardware includes an incremental learning accelerator for efficient parameter tuning.

[0168] Its incremental learning mechanism is based on online machine learning algorithms, such as stochastic gradient descent or adaptive moment estimation. It updates the model weights using only newly acquired real-time process data and actual detection data, without retraining the entire model, thus saving computational resources and time.

[0169] Its dynamic adjustment process includes first mixing new data with historical data and calculating the gradient through the loss function; then updating the model parameters while retaining the original knowledge.

[0170] Its software logic includes version control to ensure model consistency before and after updates. Data transmission is completed via a high-speed network, and the central control host monitors the update status and rolls back to the backup model in case of failure. This step leverages the efficiency and real-time nature of incremental learning, avoiding system interruption through partial updates.

[0171] Furthermore, from a technical principle perspective, incremental learning is based on optimization theory. It adjusts parameters by iteratively minimizing the error function, such as using mini-batch processing of new data to reduce memory usage.

[0172] This model update procedure enables the system to adapt to environmental changes and avoid model aging. Hardware support includes a distributed computing cluster for parallel processing of update tasks, improving speed; software logic also includes health checks to verify data quality before updating and prevent error propagation.

[0173] In summary, the incremental learning mechanism significantly reduces model update time and computational overhead, ensuring continuous system operation on high-speed production lines without requiring downtime for retraining. Dynamic parameter adjustment improves the accuracy of the digital twin model, making defect predictions more realistic and further reducing the false positive rate. Furthermore, it solves the identification bias problem caused by the lag in traditional model updates, enhances the system's adaptive detection capabilities, and ultimately improves the yield and production efficiency of semiconductor mass production.

[0174] In some embodiments, to reduce accidental re-inspection, improve the system's adaptability to diverse production lines, and reduce operational risks, S6 also includes: S600: Construct a process-defect correlation model and dynamically optimize the parameters of the digital twin model based on real-time data streams.

[0175] Specifically, in implementing this step, the process-defect correlation model is deployed on a correlation model server, which is equipped with a multi-core processor and large-capacity memory and is connected to the digital twin system via a high-speed data bus.

[0176] The model is built on a real-time data stream processing architecture. The hardware includes a stream processing engine and a real-time database, such as Apache Kafka or similar industrial-grade data stream platforms, to continuously receive process parameters and defect detection data from sensors, such as etching rate and film thickness.

[0177] Its dynamic optimization process is achieved through a parameter adjustment module, which is based on an incremental learning algorithm. It updates the model weights using only new data within the sliding time window, such as the most recent 1000 data points, without involving full retraining.

[0178] Its correlation model employs a graph neural network structure, using process parameters as nodes and defect types as edges, learning their nonlinear relationships through a message passing mechanism. The software logic includes a version control interface, allowing model parameters to be snapshotted and rolled back before and after optimization.

[0179] Its model updates are triggered by physical signal thresholds; for example, when the terahertz wave signal intensity deviation exceeds a set range, the optimization process is automatically initiated, avoiding complex mathematical calculations. This step employs a real-time data stream-driven dynamic optimization mechanism, continuously learning the process-defect correlation to maintain high model accuracy.

[0180] Furthermore, from a technical principle perspective, this process-defect correlation model operates based on causal reasoning and graph computation principles. After receiving real-time data streams, the model maps process parameters into high-dimensional vectors through a feature extraction layer, and then uses an attention mechanism to calculate the correlation weights between parameters and defects, dynamically adjusting the internal parameters of the digital twin model, such as the weights and biases of neural network layers.

[0181] The optimization process employs a variant of gradient descent, minimizing the prediction error through a loss function while using regularization techniques to prevent overfitting.

[0182] Hardware support includes a GPU cluster to accelerate graph neural network computation; software logic integrates a data verification module to ensure the integrity and consistency of input data. This step enables the system to adapt to process fluctuations and improve model robustness through real-time optimization. Physical signal threshold triggering, based on a hardware comparator circuit, generates an interrupt signal when the sensor signal exceeds a preset range, directly initiating the optimization process and reducing software latency.

[0183] In summary, firstly, dynamic optimization based on real-time data streams enables the digital twin model to quickly adapt to process changes, reducing prediction errors caused by model lag and thus lowering the probability of accidental re-inspection. Secondly, the process-defect correlation model improves the accuracy of defect identification by exploring the deep relationship between parameters and defects, enabling the system to maintain stable performance in diverse production lines. Furthermore, the physical signal threshold triggering mechanism avoids complex calculations, reduces system resource consumption, and solves the problem of slow response in traditional optimization methods. Ultimately, by improving model adaptability, it reduces operational risks and contributes to improving production yield.

[0184] S601. When the model health index indicates that the prediction error has accumulated, automatically switch to the backup model.

[0185] Specifically, when implementing this step, the model health monitoring is performed by a health assessment module, which is integrated into the model server or a standalone embedded controller. The hardware includes a performance monitoring unit such as a hardware performance counter and an error detection circuit.

[0186] Its health index is calculated based on the cumulative prediction error, such as by using a sliding window to statistically calculate the average absolute error or root mean square error. When the index exceeds a preset threshold, a switching mechanism is triggered, for example, when the error exceeds 5% for 5 consecutive times.

[0187] Its backup model is stored in a redundant storage array, including multiple versions of the training model, each optimized for different process conditions.

[0188] Its automatic switching process is achieved through a switching controller, which is based on FPGA or microprocessor design. When a health alarm signal is received, it automatically switches from the main model to the backup model to ensure that the detection process is uninterrupted.

[0189] Its software logic includes a heartbeat detection mechanism to periodically verify the status of the master model and log switching events for auditing. This step employs a health-driven automatic fault-tolerance mechanism, ensuring continuous system operation through redundant backups.

[0190] Furthermore, from a technical principle perspective, the health assessment of this model is based on the principle of statistical process control, and identifies the cumulative trend of error by calculating the moving average and standard deviation of the prediction error in real time.

[0191] The switching mechanism is based on a redundant system design, with the primary model and backup model deployed in parallel, but only the primary model is active. When the health indicators are abnormal, the switching controller selects the backup model for output through a multiplexer, while simultaneously initiating the repair process of the primary model.

[0192] The hardware support includes a watchdog timer and a power management unit to ensure the switching process can be completed even in abnormal situations such as power outages; the software logic integrates a rollback strategy, allowing recovery to the previous stable state in the event of a switching failure. This step improves system reliability through redundant design and avoids detection interruptions caused by single points of failure.

[0193] In summary, the automatic switching mechanism ensures timely activation of backup models when model performance degrades, reducing false positives and false negatives caused by the accumulation of prediction errors, thereby lowering the frequency of unintended re-tests. Health monitoring provides early warning capabilities, enabling the system to take corrective measures before problems escalate, improving adaptability to diverse production lines. Furthermore, it addresses the operational risks associated with model aging in traditional systems, improves equipment availability and production stability through fault-tolerant design, and ultimately contributes to the continuous optimization of semiconductor mass production efficiency and yield.

[0194] S602. Utilize digital thread technology to connect the entire manufacturing process, ensuring that the virtual model and the physical entity are updated synchronously.

[0195] The model update is triggered by a physical signal threshold that is not a mathematical algorithm, thus avoiding complex calculations.

[0196] Specifically, in implementing this step, the digital thread technology is implemented through a digital thread platform, which is based on an industrial IoT architecture and includes hardware such as edge gateways, cloud servers, and distributed database systems.

[0197] Its digital thread connects various units throughout the manufacturing process, such as etching equipment, deposition equipment, and testing stations, and collects process parameters, equipment status, and defect data in real time through standardized protocols.

[0198] The process of synchronizing the virtual model with the physical entity is executed through a data synchronization engine. This engine is based on a publish-subscribe model. When the physical entity data changes, the virtual model is automatically updated accordingly. For example, when the etching rate is updated in real time, the digital twin model immediately adjusts its simulation parameters.

[0199] Its model update is triggered by physical signal thresholds. For example, when the temperature sensor reading exceeds the set range or the terahertz wave signal strength deviation reaches the threshold, an update command is directly generated, bypassing complex algorithm processing.

[0200] Its software logic includes an event-driven architecture and version management to ensure data consistency and traceability. This step employs end-to-end digital thread integration and direct physical signal triggering to improve system response speed through end-to-end data synchronization.

[0201] Furthermore, from a technical principle perspective, this digital thread technology is based on the principles of data integration and real-time communication. Through a unified data model, it maps physical entities such as equipment and sensors in the manufacturing process with virtual models such as digital twins as digital objects, thereby achieving bidirectional data flow.

[0202] This synchronous update process is achieved through a change data capture mechanism, such as using database triggers or message queues to listen for data change events and automatically push them to the virtual model.

[0203] The physical signal threshold triggering process is based on a hardware interrupt mechanism. When the sensor output signal exceeds the threshold detected by the comparator circuit, an interrupt request is sent directly to the digital thread platform to initiate the model update process. This step accurately replicates the physical world state through the digital thread, ensuring that the virtual model is always synchronized with the entity. Hardware support also includes a time synchronization module, which uses GPS or network time protocols to align the clocks of each node, avoiding data timing errors.

[0204] In summary, digital threading technology achieves data integration across the entire manufacturing process, ensuring real-time synchronization between the virtual model and the physical entity. This reduces model deviations caused by data lag, thereby mitigating the risk of accidental re-inspection. The physical signal threshold triggering mechanism simplifies the update process, improves response speed, and enables the system to quickly adapt to production line changes. Furthermore, it solves the problems of data silos and update delays in traditional systems, enhances adaptability to diverse production lines, and reduces operational risks by improving synchronization accuracy and reducing computational complexity. Ultimately, it supports the efficient and stable operation of semiconductor mass production.

[0205] Based on this, the accuracy of digital twins is enhanced by linking processes and defects and monitoring model health, enabling dynamic optimization and backup switching mechanisms; it significantly reduces accidental re-inspection, improves the system's adaptability to diverse production lines, and reduces operational risks.

[0206] In this embodiment, the method further includes: S7. Based on the updated digital twin model, perform defect verification on subsequent wafers, output defect identification results, and control the detection equipment to continue running when the verification is successful; otherwise, trigger the re-inspection process.

[0207] Specifically, in implementing this step, the defect verification process is executed by the verification module, which is based on a central control host or an independent embedded system. The hardware includes a result output interface and a device control unit.

[0208] The verification process involves first inputting the real-time process data of the subsequent wafers into the updated digital twin model to generate new defect prediction values; then comparing them with the actual inspection data. If the deviation is within the allowable range of less than 5%, a "pass" result is output, and a control signal is sent to make the inspection equipment continue to run; otherwise, a re-inspection process is triggered, for example, by moving the wafer back to the inspection position and re-scanning it using a robotic arm.

[0209] Its software logic includes result logging and alarm generation for quality traceability. Control signals are transmitted via industrial protocols to ensure equipment synchronization. This step utilizes closed-loop control and real-time verification, with a feedback mechanism to ensure detection accuracy.

[0210] Furthermore, from a technical principle perspective, defect verification is based on the feedback loop principle of the control system, with model output as a reference and actual data as feedback, and a comparator is used to make decisions on actions.

[0211] The software logic integrates an event-driven architecture; for example, when verification fails, the normal process is immediately interrupted and a re-examination subroutine is initiated. This step achieves self-optimizing detection, reducing external intervention. Hardware support includes redundant control units to prevent single points of failure and improve reliability.

[0212] In summary, real-time verification and closed-loop control ensure the accuracy of defect identification, minimizing false positives and false negatives. The automated re-inspection process improves processing efficiency, adapting to the high-throughput demands of large-scale production lines. Furthermore, it solves the problem of verification lag in traditional testing, enhancing the system's stability and compatibility in dynamic production lines, ultimately significantly improving the consistency of yield in semiconductor mass production.

[0213] Based on this, by combining real-time data acquisition with dynamic updates of the digital twin model, the problem of inaccurate defect identification caused by lagging model updates is solved, achieving adaptive defect detection and significantly improving detection accuracy and the system's adaptability in dynamic production lines. Its integrated real-time feedback and incremental learning mechanisms non-obviously improve model response speed; significantly reduce the false positive rate, and enhance the stability of yield in semiconductor mass production.

[0214] In some embodiments, to ensure that the equipment maintains detection accuracy and improves yield stability during continuous operation, S7 further includes: The S700 integrates adaptive beamforming technology, which automatically optimizes detection parameters based on wafer size and material.

[0215] Specifically, the adaptive beamforming system is deployed at the terahertz wave transmitter, and the hardware includes a phased array antenna unit composed of multiple microstrip patch antennas and a digital beamforming controller based on FPGA design.

[0216] The controller receives wafer dimensions and material data obtained through material sensors, and dynamically adjusts the phase and amplitude of the antenna elements through a weighted algorithm to optimize the focus and coverage of the terahertz beam.

[0217] The software logic integrates a parameter configuration module, which generates optimal beam parameters using lookup tables or real-time calculations and sends control commands via a serial interface. The system is equipped with a temperature compensation circuit to ensure beam stability. This step employs intelligent adaptive beam adjustment, achieving automatic optimization of detection parameters through hardware and software collaboration.

[0218] Furthermore, adaptive beamforming technology, based on array signal processing principles, synthesizes a directional beam to match wafer characteristics by adjusting the excitation parameters of multiple antenna elements. This technology enables the system to adjust the beam pattern and gain in real time based on input data, reducing signal scattering and energy loss. Hardware support includes a feedback sensor network to monitor beam quality and perform closed-loop correction.

[0219] In summary, this system can automatically adapt to different wafers, improve testing compatibility and accuracy, overcome the limitations of traditional fixed parameter systems, ensure accuracy during continuous operation, and significantly enhance yield stability.

[0220] S701: The detection threshold is dynamically updated through the particle swarm optimization algorithm, and automatic calibration is performed after every 20 wafers are inspected.

[0221] The calibration process uses a standard defect template library for benchmark comparison.

[0222] Specifically, the particle swarm optimization module runs on the central processing unit, with hardware including high-speed memory and a computing accelerator. The algorithm uses detection parameters, such as signal thresholds, as particles, iteratively updating position and velocity to find the optimal value. After inspecting every 20 wafers, the system automatically calls a standard defect sample library for benchmark comparison and error calculation. The software logic integrates an event-triggered mechanism and calibration protocol to ensure seamless execution. This step employs an integration of dynamic optimization and periodic calibration, maintaining threshold accuracy through intelligent algorithms.

[0223] Furthermore, this particle swarm optimization algorithm, based on the principle of swarm intelligence, simulates social behavior to search for the global optimum; calibration identifies deviations and adjusts thresholds by comparing measured data with a benchmark. This technology enables continuous learning of the system.

[0224] In summary, dynamic threshold updates reduce false positives and false negatives, automatic calibration ensures long-term accuracy, improves the reliability of equipment during continuous operation, and ultimately supports a stable increase in yield.

[0225] Based on this, adaptive beamforming and dynamic calibration enhance detection compatibility and accuracy, enabling automatic parameter optimization and periodic calibration; significantly ensuring that the equipment maintains detection accuracy during continuous operation and improving yield stability.

[0226] In some embodiments, to reduce failure rates and operating costs, and improve device reliability and scalability, the system further includes: S8. Decompose the detection system into independent modules.

[0227] It includes a terahertz emission module, a detection module, and an optical array module, with each module using a standardized interface for quick plugging and unplugging.

[0228] Specifically, the detection system adopts a modular architecture design. Its terahertz emission module includes a femtosecond laser source and an array emitter; the detection module integrates a superconducting detector and a signal conditioning circuit; and the optical array module consists of a microlens array and its driving system.

[0229] Each module is electrically and optically connected via a unified multi-core connector, supporting hot-swapping. The module housing is made of lightweight alloy material and equipped with a guiding mechanism and quick-locking device to ensure installation accuracy. This step utilizes the system's highly modular design, enabling seamless interoperability and rapid replacement of modules through standardized interfaces.

[0230] Furthermore, the modular design is based on the principle of functional partitioning, with each module independently encapsulating core functional components and achieving data interaction and power supply through clearly defined interface protocols. The standardized interface includes high-speed data channels, power buses, and control signal lines, and employs a design to prevent mis-plugging. This implementation decouples the complex system into functionally independent sub-modules.

[0231] In summary, modular implementation significantly reduces system maintenance complexity while supporting flexible configuration of functional modules according to production line needs, thus improving the scalability and availability of the equipment.

[0232] S9 monitors the status of each module in real time through a built-in sensor network.

[0233] It includes temperature, vibration, and power consumption.

[0234] Specifically, each functional module embeds multiple types of sensor nodes. Temperature monitoring uses patch thermocouples, vibration monitoring uses MEMS accelerometers, and power consumption monitoring is achieved through Hall current sensors. The sensor network communicates with the main control unit via a CAN bus, with a sampling frequency of no less than 100Hz. Monitoring data is displayed in real time on the HMI interface and stored in a time-series database for trend analysis. This step employs a comprehensive condition monitoring system to provide a data foundation for predictive maintenance.

[0235] Furthermore, based on the principle of distributed data acquisition, the sensor network uses built-in signal conditioning circuits in each module to preprocess and filter the raw data before uploading it to the monitoring center via a unified communication protocol. The system employs a time synchronization mechanism to ensure the consistency of multi-parameter data.

[0236] In summary, real-time status monitoring enables the system to detect abnormal signs in a timely manner, advances the fault warning time, effectively prevents the occurrence of sudden failures, and significantly improves the reliability of equipment operation.

[0237] S10. When an anomaly is detected, the diagnostic program is automatically triggered or the system switches to a backup module.

[0238] Specifically, the system sets threshold ranges for each status parameter. When the monitored data exceeds the limits, a three-level response mechanism is automatically triggered: for primary anomalies, a self-diagnosis program is initiated; for intermediate anomalies, the system switches to a backup module; and for severe anomalies, a safe shutdown is executed.

[0239] Its diagnostic procedure is based on an expert system rule base, analyzing failure modes and their impacts. The backup module maintains a hot standby status, ensuring a seamless switchover process. This step enables intelligent self-handling of faults.

[0240] Furthermore, the self-diagnostic system uses fault tree analysis to locate fault sources through a pre-defined logical judgment process. The module switching mechanism adopts a redundancy design concept, with backup units configured for critical modules, and fast switching is achieved through solid-state relays.

[0241] In summary, automated anomaly handling mechanisms shorten the average system recovery time, significantly reduce production downtime, and improve overall equipment utilization.

[0242] S11. Predictive maintenance algorithms are used to warn of component degradation, and material optimization is combined to extend module life.

[0243] Specifically, the predictive maintenance algorithm, based on a machine learning model, analyzes historical sensor data to identify component degradation trends. When the predicted remaining lifespan is below a threshold, the system generates a maintenance warning in advance.

[0244] Key components utilize wear-resistant coatings and composite materials, such as diamond-coated terahertz windows and gold-plated connector contacts. The maintenance strategy is integrated with the spare parts management system, automatically generating procurement plans.

[0245] Furthermore, the predictive model employs time series analysis and degradation modeling techniques to estimate component lifespan through feature extraction and pattern recognition. Material optimization, based on surface engineering principles, improves the wear resistance and fatigue resistance of key components.

[0246] In summary, predictive maintenance effectively reduces unplanned downtime, material optimization extends module lifespan, significantly reduces total lifecycle operating costs, and improves return on investment for equipment.

[0247] Based on this, the modular design and self-maintenance mechanism solve the maintenance difficulties caused by the high system complexity, realize standardized interfaces and predictive maintenance, significantly reduce failure rate and operating costs, and improve equipment reliability and scalability.

[0248] The wafer defect detection system provided in the embodiments of this application is described below. The wafer defect detection system described below can be referred to in correspondence with the wafer defect detection method described above.

[0249] refer to Figure 2 The wafer defect detection system includes: Data acquisition module 1 is configured to acquire real-time process data during wafer manufacturing via a sensor network. The real-time process data includes etching rate and thin film thickness, and the sampling frequency is not less than 1 kHz.

[0250] Digital twin processing module 2 is configured to store and run digital twin models. The digital twin models simulate the entire wafer production process based on historical manufacturing data and generate defect prediction values.

[0251] The detection data acquisition module 3 is configured to acquire actual defect detection data of the wafer through non-contact inspection equipment. The actual defect detection data includes terahertz wave scanning signals and image reconstruction results.

[0252] The comparison and analysis module 4 is configured to receive actual defect detection data and defect prediction values, compare the data and calculate the deviation value to determine whether it exceeds the preset threshold of 5%.

[0253] Model update module 5 is configured to trigger a model update program when the deviation value exceeds a preset threshold, dynamically adjusting the parameters of the digital twin model. The adjustment is achieved through an incremental learning mechanism, eliminating the need for full retraining.

[0254] Defect verification module 6 is configured to perform defect verification based on the updated digital twin model, output defect identification results, and control the detection equipment to continue running when the verification is successful, otherwise trigger the re-inspection process.

[0255] The modules are connected via a bus to enable real-time data interaction.

[0256] Based on this, the system solves the problem of lagging updates in traditional models by tightly integrating digital twin models with real-time data feedback, realizing modular design and adaptive update mechanism; it significantly improves detection accuracy and efficiency, supports the high-throughput requirements of large-scale production lines, and reduces maintenance costs.

[0257] This application provides an electronic device, such as... Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 The illustrated electronic device 300 includes a processor 301 and a memory 303. The processor 301 and the memory 303 are connected, for example, via a bus 302. Optionally, the electronic device 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this electronic device 300 does not constitute a limitation on the embodiments of this application.

[0258] Processor 301 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in connection with the embodiments of this application. Processor 301 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0259] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0260] The memory 303 may be a ROM (Read-Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or it may be an EEPROM (Electrically Erasable Programmable Read-Only Memory), a CD-ROM (Compact Disc Read-Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0261] The memory 303 is used to store application code that executes the scheme of the embodiments of this application, and its execution is controlled by the processor 301. The processor 301 is used to execute the application code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0262] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments described in this application.

[0263] This application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wafer defect detection method described above.

[0264] Since the embodiments of the computer-readable storage medium portion correspond to the embodiments of the method portion, please refer to the description of the embodiments of the method portion for the embodiments of the computer-readable storage medium portion.

[0265] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0266] The above are only some embodiments of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for detecting wafer defects, characterized in that, include: Real-time process data during wafer manufacturing is acquired, including etching rate and thin film thickness, and is collected by sensors at a sampling frequency of not less than 1 kHz. The real-time process data is input into a pre-built digital twin model, which simulates the entire wafer production process based on historical manufacturing data to generate defect prediction values, wherein the defect prediction values ​​include defect type, location and size. The actual defect detection data of the wafer is obtained by non-contact inspection equipment, and the actual defect detection data includes terahertz wave scanning signals and image reconstruction results; The actual defect detection data is compared with the predicted defect value, and the deviation value is calculated. The comparison includes data normalization and feature extraction. Determine whether the deviation value exceeds a preset threshold of 5%; When the deviation value exceeds a preset threshold, the model update program is automatically triggered. The model update program includes dynamically adjusting the parameters of the digital twin model based on real-time process data and actual defect detection data. The adjustment is achieved through an incremental learning mechanism, without the need for full retraining. Based on the updated digital twin model, subsequent wafers are subjected to defect verification, and the defect identification results are output. If the verification is successful, the detection equipment is controlled to continue running; otherwise, a re-inspection process is triggered.

2. The method according to claim 1, characterized in that, The step of acquiring real-time process data also includes: The real-time process data is preprocessed by distributed edge nodes, redundant information is filtered and key features are extracted, and only deviation-related data is sent to the central processing unit. The preprocessing includes data compression and noise reduction, and uses multi-level caching technology to temporarily store data to ensure seamless data transmission. Dynamically allocate computing resources and automatically adjust task priorities based on the detected load to avoid processing delays; The edge nodes include dedicated hardware processors for locally performing data normalization and feature comparison.

3. The method according to claim 1, characterized in that, The step of triggering the model update procedure also includes: Construct a process-defect correlation model and dynamically optimize the parameters of the digital twin model based on real-time data stream; When the model health metric indicates that prediction errors have accumulated, the system will automatically switch to the backup model. Digital threading technology is used to connect the entire manufacturing process, ensuring that the virtual model and the physical entity are updated synchronously. The model update is triggered by a physical signal threshold without mathematical algorithms, thus avoiding complex calculations.

4. The method according to claim 1, characterized in that, The step of acquiring actual defect detection data also includes: A sealed cavity was installed in the testing area to isolate the effects of temperature, humidity, and airflow. An integrated high-precision temperature control module maintains a constant temperature within the cavity; A reference calibration unit is embedded in the terahertz wave path to monitor the signal strength in real time and dynamically adjust the transmission parameters. Multi-band terahertz technology is used to switch frequency bands and offset environmental effects.

5. The method according to claim 1, characterized in that, The step of acquiring actual defect detection data through non-contact detection equipment further includes: Detection is performed using a reconfigurable optical array, which consists of multiple independently controlled microlens units; The lens position is adjusted using non-contact drive technology, reducing mechanical wear; The built-in optical encoder monitors lens displacement in real time and performs a self-calibration cycle to compare with a standard sample; Micro-motors are used for nanometer-level compensation to ensure accuracy of focal length and field of view.

6. The method according to claim 1, characterized in that, The subsequent wafer defect verification step based on the updated digital twin model also includes: Integrated adaptive beamforming technology automatically optimizes detection parameters based on wafer size and material; The detection threshold is dynamically updated using a particle swarm optimization algorithm, and the system is automatically calibrated after every 20 wafers are inspected. The calibration process uses a standard defect template library for benchmark comparison.

7. The method according to claim 1, characterized in that, Also includes: The detection system is decomposed into independent modules, including a terahertz emission module, a detection module, and an optical array module. Each module uses a standardized interface to enable quick plugging and unplugging. The built-in sensor network monitors the status of each module in real time, including temperature, vibration, and power consumption. When an anomaly is detected, the diagnostic program is automatically triggered or the system switches to a backup module. Predictive maintenance algorithms are used to warn of component degradation, and material optimization is combined to extend module life.

8. A wafer defect detection system, characterized in that, Implementing the method as described in any one of claims 1 to 7, comprising: The data acquisition module is configured to acquire real-time process data during wafer manufacturing via a sensor network. The real-time process data includes etching rate and thin film thickness, and the sampling frequency is not less than 1 kHz. A digital twin processing module is configured to store and run a digital twin model, which simulates the entire wafer production process based on historical manufacturing data and generates defect prediction values. The detection data acquisition module is configured to acquire actual defect detection data of the wafer through a non-contact detection device. The actual defect detection data includes terahertz wave scanning signals and image reconstruction results. The comparison and analysis module is configured to receive actual defect detection data and defect prediction values, compare the data and calculate the deviation value to determine whether it exceeds a preset threshold of 5%. The model update module is configured to trigger a model update procedure when the deviation value exceeds a preset threshold, dynamically adjusting the parameters of the digital twin model. The adjustment is achieved through an incremental learning mechanism, eliminating the need for full retraining. The defect verification module is configured to perform defect verification based on the updated digital twin model, output defect identification results, and control the detection equipment to continue running when the verification is successful; otherwise, it triggers the re-inspection process. The modules are connected via a bus to enable real-time data interaction.

9. An electronic device, characterized in that, include: One or more processors; One or more memory units; And one or more computer programs, wherein the one or more computer programs are stored in the one or more memories, the one or more computer programs including instructions that, when executed by the one or more processors, cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program or instructions that, when executed, implement the method as described in any one of claims 1 to 7.

Images