Gaseous molecular pollutant on-line monitoring system for semiconductor manufacturing

Abstract

The invention belongs to the technical field of semiconductor manufacturing management, and particularly relates to a semiconductor manufacturing-oriented gaseous molecular pollutant online monitoring system, which comprises a small-size traceability dynamic historical influence analysis module, a multi-source real-time monitoring and early warning module, a prospective early warning analysis module, an active control detection module and an intelligent feedback interface, the small-size traceability dynamic historical influence analysis module is based on a hybrid model, historical data and real-time data are combined, the future influence trend of pollutants can be predicted and graded, and therefore enterprises can make coping strategies in advance, reasonably arrange production plans and reduce the influence of the pollutants on the product yield. According to the method, pollutant diffusion paths can be simulated, high-risk stations can be positioned, pollutant sources and propagation paths are analyzed through the Bayesian network, and a root cause probability sorting list is output, so that a basis is conveniently provided for prevention, control and treatment of pollutants, and targeted measures can be quickly and accurately taken.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing management technology, specifically an online monitoring system for gaseous molecular contaminants in semiconductor manufacturing. Background Technology

[0002] In the semiconductor manufacturing process, the presence of gaseous molecular contaminants has a significant impact on product quality and yield. As the feature size of semiconductor devices continues to shrink, the requirements for the cleanliness of the manufacturing environment are becoming increasingly stringent. Even extremely low concentrations of gaseous molecular contaminants can lead to a decline in device performance or even failure.

[0003] The invention patent publication number CN117129636A discloses an online monitoring system for gaseous molecular contaminants in semiconductor manufacturing. This system includes analyzing the impact of historical semiconductor processing within a clean area to determine the degree of influence of different gaseous molecular contaminants, facilitating the classification of these contaminants into different levels. This allows for a more intuitive analysis of the impact of gaseous molecular contaminants and enables targeted online monitoring based on different levels. Furthermore, by collecting real-time concentration data and combining it with the analysis of the control efficiency of gaseous contaminants, and by determining the composition of gaseous molecular contaminants through real-time concentration, the system further evaluates the control efficiency and provides real-time early warnings based on the efficiency assessment.

[0004] In summary, traditional environmental monitoring methods in semiconductor manufacturing are relatively limited, typically focusing on only a few parameters. This makes it difficult to comprehensively and accurately reflect the actual impact of pollutants. In other words, relying solely on concentration monitoring cannot fully consider the combined influence of environmental parameters and process stages on pollutant behavior. Furthermore, they lack the ability to predict future pollutant trends, only identifying problems after pollutants have already caused some impact, making it impossible to take preventative measures in advance. In addition, traditional methods are inefficient in locating pollution sources and conducting root cause analysis, making it difficult to quickly and accurately determine the source and transmission path of pollutants, resulting in untimely and ineffective response measures.

[0005] Therefore, the present invention provides an online monitoring system for gaseous molecular contaminants in semiconductor manufacturing. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: an online monitoring system for gaseous molecular pollutants in semiconductor manufacturing, comprising: A small-sized dynamic historical impact analysis module is used to dynamically adjust the historical impact coefficient threshold by combining real-time environmental parameters and process stages, and based on... The hybrid model takes historical and real-time data as input and predicts the future impact trend of pollutants and classifies them into levels. Historical data includes ; Real-time data is acquired through sensors, including concentration and temperature / humidity data. The multi-source real-time monitoring and early warning module integrates temperature, humidity, pressure, and equipment status parameters to build a multivariate early warning model, and dynamically allocates the weights of each dimension based on the process stage. The forward-looking early warning analysis module includes a digital twin sub-module, which is used to simulate pollutant diffusion paths and locate high-risk workstations, and a root cause reasoning sub-module, which analyzes pollutant sources and transmission paths based on Bayesian networks; The active control detection module uses the DQN algorithm to dynamically adjust the sampling frequency and sensor sensitivity with the optimization goals of detection accuracy and system load, and implements a risk-level sampling strategy. The intelligent feedback interface integrates visualization tools such as 3D heat maps and trend comparison charts, as well as an intelligent suggestion system based on a historical case library, which automatically generates response measures.

[0008] Furthermore, the small-size source tracing dynamic historical impact analysis module The hybrid model employs a multi-head attention mechanism to process historical pollutant data. Temporal features are extracted and dynamic weights are assigned based on real-time environmental parameters such as temperature, humidity, and pressure. This represents the frequency difference in the concentration fluctuations of gaseous molecular contaminants between qualified and unqualified semiconductor processing periods in the clean area during historical operation. It reflects the correlation between contaminant concentration fluctuations and process anomalies; the frequency difference is positively correlated with the impact of contaminants on yield. The frequency difference between qualified and unqualified periods is calculated by statistically analyzing the number of times the pollutant concentration peaks or the rate of change per unit time. This represents the maximum fluctuation range of the concentration of the same gaseous molecular contaminant during different types of non-conforming semiconductor processing within a historical operating period. By extracting the range of contaminant concentration under etching defects or thin film contamination, the average or weighted value is obtained; During historical operating periods, the average concentration fluctuation of gaseous molecular pollutants in the same type of qualified semiconductor processing is calculated by measuring the standard deviation or moving average fluctuation of pollutant concentrations during the qualified period. The pollutant future impact trend and level information predicted by the small-sized source tracing dynamic historical impact analysis module is one of the reference factors when the multi-source real-time monitoring and early warning module dynamically allocates the weights of each dimension.

[0009] Furthermore, in the multi-source real-time monitoring and early warning module, the process of constructing a multivariate early warning model includes the following steps: Data collection and integration, variable selection and weight allocation, model building and training, model validation and optimization, and real-time early warning and application; The dynamic allocation of weights for each dimension is based on a process stage sensitivity matrix, which is obtained through training with historical data and maps the sensitivity coefficients of each process stage to the pollutant dimension.

[0010] Furthermore, the digital twin submodule uses the finite element analysis method to construct a cleanroom flow field model. After inputting real-time monitoring data, it simulates the pollutant diffusion path through CFD simulation and outputs the coordinates of high-risk areas and concentration distribution cloud maps. The construction of a cleanroom flow field model includes the following steps: collecting basic data, geometric modeling and mesh generation, physical model selection and parameter setting, model solving, result analysis and model verification, and model integration and application. The coordinates of high-risk areas simulated by the digital twin submodule in the forward-looking early warning analysis module and the root cause probability ranking list output by the root cause reasoning submodule will be transmitted to the active control detection module to guide it in adjusting the sampling strategy and sensor sensitivity, focusing on high-risk areas and possible pollution sources.

[0011] Furthermore, the process of analyzing pollutant sources and propagation pathways based on Bayesian networks includes: Various factors related to the source and spread of pollutants are identified as nodes in a Bayesian network, and the structure of the Bayesian network is constructed based on the causal and conditional dependencies between the factors. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is then used to learn and estimate the conditional probability table in the Bayesian network, determining the probability distribution of each node under different conditions. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is used to learn and estimate the conditional probability table in the Bayesian network to determine the probability distribution of each node under different conditions. When new monitoring data or information is input, the inference algorithm of Bayesian network is used to perform inference and calculate the posterior probability of each pollution source as the source of pollutants to determine the most likely pollution source. At the same time, the propagation path and diffusion range of pollutants under the influence of different factors are analyzed. The reasoning results are interpreted and analyzed, transforming complex probability calculations into understandable information, and based on the analysis results, providing decision support for the prevention and control of pollutants; In the root cause reasoning submodule, the Bayesian network nodes include equipment status, personnel operation, and environmental parameters. The contribution of each node to the pollutant event is quantified through a conditional probability table, and a root cause probability ranking list is output.

[0012] Furthermore, in the active control detection module, the reward function of the DQN algorithm is designed as follows: ; in, These are the weighting coefficients, which are optimized through reinforcement learning training.

[0013] Furthermore, in the risk-based sampling strategy, the initial sampling frequency for low-level pollutants is set to once every 10 minutes, and the initial sensor sensitivity is a signal-to-noise ratio ≥ 20 dB; the sampling frequency for high-level pollutants is set to once every 3.3 minutes, and the sensor sensitivity is dynamically increased to a signal-to-noise ratio ≥ 30 dB.

[0014] Furthermore, in the intelligent feedback interface, the 3D heat map supports multi-level drilling, which is used to display the distribution of pollutant concentrations in cleanrooms, workstations, and equipment layers by layer, and to mark the locations of historical warning events.

[0015] Furthermore, the intelligent suggestion system includes a dual-mode system of rule engine and case reasoning: the rule engine generates standardized suggestions based on an expert knowledge base, while case reasoning provides customized solutions by matching historical cases with similarity.

[0016] Furthermore, it includes an adaptive traceability verification and calibration module, which is used to periodically compare the sensor's measured values ​​with the digital twin's simulated values, automatically correct the model parameters and threshold settings, and ensure the long-term stability of the system.

[0017] The beneficial effects of this invention are as follows: The small-sized source-tracing dynamic historical impact analysis module of this invention is based on a hybrid model and combines historical and real-time data. It can predict the future impact trend of pollutants and classify them into levels, thereby facilitating enterprises to formulate response strategies in advance, rationally arrange production plans, and reduce the impact of pollutants on product yield. The forward-looking early warning analysis module uses a digital twin sub-module and a root cause reasoning sub-module to simulate the pollutant diffusion path, locate high-risk workstations, and analyze the source and propagation path of pollutants through Bayesian network analysis, outputting a root cause probability ranking list. This provides a scientific basis for the prevention and control of pollutants and enables rapid and accurate implementation of targeted measures. The active control detection module adopts the DQN algorithm to achieve high detection accuracy and system... With system load as the optimization goal, the sampling frequency and sensor sensitivity are dynamically adjusted, and a risk-based sampling strategy is implemented to reduce system load and improve detection efficiency while ensuring detection accuracy. The intelligent feedback interface integrates 3D heat maps and trend comparison chart visualization tools, as well as an intelligent suggestion system based on a historical case library. It can intuitively display the distribution of pollutant concentrations and the location of historical warning events, and automatically generate response measures. This helps operators quickly understand the monitoring results and make reasonable decisions. The adaptive traceability verification and calibration module regularly compares the sensor's measured values ​​with the digital twin simulation values, automatically corrects the model parameters and threshold settings, ensures long-term stable operation of the system, and improves the system's reliability and accuracy. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a monitoring flowchart of the present invention; Figure 2 This is a diagram illustrating the internal interfaces of the system of this invention; Figure 3 This is a system deployment diagram of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1-3 , This embodiment provides an online monitoring system for gaseous molecular contaminants in semiconductor manufacturing, comprising: a small-size source tracing dynamic historical impact analysis module, a multi-source real-time monitoring and early warning module, a forward-looking early warning analysis module, an active control and detection module, an intelligent feedback interface, and an adaptive source tracing verification and calibration module. The modules exchange data and interact through data interfaces to ensure timely information sharing and collaborative work. These data interfaces are mainly divided into API interfaces, message interfaces, and file sharing interfaces. The correspondence between the interface name, interface content, and implementation method is as follows: Figure 2 As shown; At the same time, such as Figure 3 The system platform shown provides unified access and interaction functions for users through user permission management.

[0022] A small-sized dynamic historical impact analysis module is used to dynamically adjust the historical impact coefficient threshold by combining real-time environmental parameters and process stages, and based on... The hybrid model takes historical and real-time data as input and predicts the future impact trend of pollutants and classifies them into levels. Historical data includes ; Real-time data is acquired through sensors, including concentration and temperature / humidity data. To enable users to gain a deeper understanding of the core technologies of the small-scale source tracing dynamic historical impact analysis module and to clarify how the model utilizes various data to accurately predict the future impact trends and levels of pollutants, this invention further discloses the small-scale source tracing dynamic historical impact analysis module. The hybrid model employs a multi-head attention mechanism to process historical pollutant data. Temporal features are extracted and dynamic weights are assigned based on real-time environmental parameters such as temperature, humidity, and pressure. This represents the frequency difference in the concentration fluctuations of gaseous molecular contaminants between qualified and unqualified semiconductor processing periods in the clean area during historical operation. It reflects the correlation between contaminant concentration fluctuations and process anomalies; the frequency difference is positively correlated with the impact of contaminants on yield. The frequency difference between qualified and unqualified periods is calculated by statistically analyzing the number of peak occurrences or the rate of change of contaminant concentration per unit time. For example, in a certain month's production, during qualified semiconductor processing periods, the number of peak occurrences of contaminant concentration per unit time (e.g., per hour) is 3; during unqualified semiconductor processing periods (e.g., the production stage with etching defects), the number of peak occurrences of contaminant concentration per unit time is 8. Therefore, the difference in concentration fluctuation frequency is 8-3=5 times. This indicates that in the unqualified production stage, there is a strong correlation between contaminant concentration fluctuations and process abnormalities (etching defects), and the significant frequency difference may have a serious impact on yield. This represents the maximum fluctuation range of the concentration of the same gaseous molecular contaminant during different types of non-conforming semiconductor processing within a historical operating period. By extracting the range of contaminant concentration under etching defects or thin film contamination, the average or weighted value is taken; for example, for different types of non-conforming semiconductor processing, such as another product production stage where thin film contamination occurs, the range of contaminant concentration under this stage is extracted. Assuming that the ranges obtained from multiple measurements are 10ppm, 12ppm, and 8ppm respectively, the average value (10+12+8) / 3≈10ppm is taken as the maximum fluctuation range of the concentration of the same gaseous molecular contaminant in this type of non-conforming processing. During historical operating periods, the average concentration fluctuation of gaseous molecular contaminants in the same type of qualified semiconductor processing is calculated by taking the standard deviation or moving average fluctuation of the contaminant concentration within the qualified period. For example, during the same type of qualified semiconductor processing, the contaminant concentration is continuously measured for a week, with multiple measurements taken each day. The standard deviation of the daily concentration is calculated, and the average of the weekly standard deviations is taken as the average concentration fluctuation. For example, if the daily standard deviations within a week are 1.5ppm, 1.2ppm, 1.8ppm, 1.3ppm, 1.6ppm, 1.4ppm, and 1.710ppm, the average is (1.5+1.2+1.8+1.3+1.6+1.4+1.7) / 7≈1.5ppm; The pollutant future impact trend and level information predicted by the small-sized source tracing dynamic historical impact analysis module is one of the reference factors when the multi-source real-time monitoring and early warning module dynamically allocates the weights of each dimension.

[0023] For example, suppose historical data on semiconductor manufacturing includes data on gaseous molecular pollutants from different quarters and production stages of the past year. A multi-head attention mechanism can simultaneously consider the impact of seasonal variations (such as higher temperatures and humidity in summer potentially affecting pollutant dispersion) and process differences during the production of different products on pollutant concentrations. For instance, in the production of a certain high-precision chip, a specific lithography process may cause fluctuations in the concentration of a certain gaseous molecular pollutant over a specific time period. The multi-head attention mechanism can capture this temporal characteristic and, combined with current real-time temperature, humidity, and pressure data, dynamically adjust the weights of each factor to more accurately predict future pollutant impact trends and levels. The multi-source real-time monitoring and early warning module integrates temperature, humidity, pressure, and equipment status parameters to build a multivariate early warning model, and dynamically allocates the weights of each dimension based on the process stage. The detailed description of the construction process helps technicians develop models according to standard procedures, while the introduction of the sensitivity matrix for each process stage explains the basis for weight allocation. To make the technical solution of the entire multi-source real-time monitoring and early warning module more complete and reliable, this invention further discloses the process for constructing a multivariate early warning model in the multi-source real-time monitoring and early warning module, which includes the following steps: Data collection and integration: Integrate multi-dimensional parameter data such as temperature, humidity, pressure, and equipment status. These data come from different sensors and monitoring devices and need to be uniformly time-aligned and cleaned to ensure the accuracy and consistency of the data; collect dynamic information of process stages and clarify the impact weight of different process stages on each dimension parameter. For example, in some key production links, changes in certain parameters may have a greater impact on the generation and diffusion of pollutants. Variable selection and weight allocation: Analyze the correlation between each variable and gaseous molecular pollutants to determine which variables have a significant impact on the generation, accumulation and diffusion of pollutants; dynamically allocate the weight of each dimension based on the process stage. For example, in the high temperature and high pressure process stage, temperature and pressure parameters may have a greater weight in the early warning model. Model building and training: Select appropriate modeling methods, such as multiple linear regression, support vector machine (SVM) or neural network, to build a multivariate early warning model; use historical data to train the model and adjust the model parameters so that the model can accurately capture the complex relationship between each variable and pollutants.

[0024] Model validation and optimization: The trained model is validated using an independent validation dataset to evaluate performance metrics such as alert accuracy and recall. The model is then optimized based on the validation results, which may include adjusting variable weights, improving model structure, or adding new variables.

[0025] Real-time early warning and application: The constructed multivariate early warning model is applied to real-time monitoring data. When the model's prediction results exceed the set early warning threshold, an early warning signal is issued in a timely manner. Based on the early warning information, combined with the visualization tools and intelligent suggestion system provided by the intelligent feedback interface, corresponding countermeasures are taken.

[0026] The dynamic allocation of weights for each dimension is based on the process stage sensitivity matrix, which is obtained by training with historical data and maps the sensitivity coefficient of each process stage to the pollutant dimension.

[0027] For example, data collection and integration: In a semiconductor manufacturing workshop, temperature and humidity sensors collect data every minute, pressure sensors collect data every 5 minutes, and equipment status parameters (such as equipment running time, fault alarm information, etc.) are obtained from the equipment control system. These data from different sensors and devices are time-aligned, for example, to a time granularity of one minute, and data cleaning is performed to remove obvious outliers (such as extreme and unreasonable values ​​suddenly appearing from temperature and humidity sensors). At the same time, process stage information is collected, such as whether photolithography, etching, or thin film deposition is currently underway, to clarify the sensitivity of different processes to parameters such as temperature, humidity, and pressure. Variable selection and weight allocation: Data analysis revealed that during the etching process, pressure parameters were highly correlated with the concentration of gaseous molecular contaminants. This is because pressure changes during etching may affect the distribution of reactive gases and the generation of contaminants. Based on the sensitivity matrix of the process stage, the weight of pressure parameters in the early warning model may be set to 0.4, while the weight of temperature and humidity parameters may be set to 0.3, and the weight of equipment status parameters may be set to 0.3. Model building and training: A neural network was selected to build a multivariate early warning model. The model was trained by using historical data on temperature, humidity, pressure, equipment status parameters and corresponding pollutant concentrations as inputs and outputs. After multiple training sessions, the parameters were adjusted so that the model could fit the relationship between each variable and the pollutant concentration well. Model Validation and Optimization: The trained model was validated using an independent validation dataset. It was found that the model had a low early warning accuracy in certain special process transition stages. Therefore, a process transition flag was added as a new variable, and the model was retrained. After optimization, the early warning accuracy of the model was improved. Real-time early warning and application: When real-time monitoring data is input into the model, the model predicts that the pollutant concentration will exceed the early warning threshold within 10 minutes. The system will issue an early warning signal in a timely manner. Operators can view the 3D thermal map and intelligent suggestions through the intelligent feedback interface and find that it may be caused by abnormal pressure of a certain etching equipment. They can then check and adjust the equipment according to the suggestions.

[0028] The forward-looking early warning analysis module includes a digital twin sub-module, which is used to simulate pollutant diffusion paths and locate high-risk workstations, and a root cause reasoning sub-module, which analyzes pollutant sources and transmission paths based on Bayesian networks; The description of the construction process helps ensure the accuracy and reliability of the model, while the description of information transmission clarifies the collaborative working method between modules, making the connection between the entire forward-looking early warning analysis module and the active control and detection module closer, and improving the overall performance of the system. Therefore, a further improvement of the present invention is to describe that the digital twin sub-module uses the finite element analysis method to construct a cleanroom flow field model. After inputting real-time monitoring data, it simulates the pollutant diffusion path through CFD simulation and outputs the coordinates of high-risk areas and the concentration distribution cloud map. The construction of a cleanroom flow field model includes the following steps: collecting basic data, geometric modeling and mesh generation, physical model selection and parameter setting, model solving, result analysis and model verification, and model integration and application. The coordinates of high-risk areas simulated by the digital twin submodule in the forward-looking early warning analysis module and the root cause probability ranking list output by the root cause reasoning submodule will be transmitted to the active control detection module to guide it in adjusting the sampling strategy and sensor sensitivity, focusing on high-risk areas and possible pollution sources.

[0029] For example, when constructing a cleanroom flow field model, the basic data such as the length, width, and height of the cleanroom, equipment layout, ventilation outlet location, and wind speed are first collected. Geometric modeling and mesh generation are then performed, dividing the cleanroom into multiple small unit meshes. A suitable physical model, such as the k-ε turbulence model, is selected, and parameters such as air density and viscosity are set. After inputting real-time monitored data such as pollutant concentration, temperature, and humidity, CFD simulation is used to output and display that the pollutant concentration is high in an area near a certain workstation, with coordinates (x=5m, y=3m, z=2m), and the gradient change of pollutant concentration in that area is shown in the form of a concentration distribution cloud map.

[0030] A detailed description of the process helps technicians to correctly implement Bayesian network analysis; therefore, a further description of the process for analyzing pollutant sources and propagation paths based on Bayesian network analysis in this invention includes: Various factors related to the source and spread of pollutants are identified as nodes in the Bayesian network, such as the type of pollution source, meteorological conditions (wind direction and speed), topography, and equipment operating status. Based on the causal and conditional relationships between these factors, the structure of the Bayesian network is constructed. For example, the type of pollution source affects the initial concentration and composition of pollutants, and meteorological conditions affect the direction and spread of pollutants. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is then used to learn and estimate the conditional probability table in the Bayesian network, determining the probability distribution of each node under different conditions. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is used to learn and estimate the conditional probability table in the Bayesian network to determine the probability distribution of each node under different conditions. When new monitoring data or information is input, inference is performed using Bayesian network inference algorithms (such as belief propagation algorithms).

[0031] Calculate the posterior probability of each pollution source as a source of pollutants to determine the most likely pollution source. Simultaneously, analyze the transmission paths and diffusion range of pollutants under the influence of different factors. The reasoning results are interpreted and analyzed, transforming complex probability calculations into understandable information. For example, the possible sources and transmission paths of pollutants are visualized. Based on the analysis results, decision support is provided for the prevention and control of pollutants, such as adjusting the layout of monitoring points and taking targeted pollution source control measures. In the root cause reasoning submodule, the Bayesian network nodes include equipment status, personnel operation, and environmental parameters. The contribution of each node to the pollutant event is quantified through a conditional probability table, and a root cause probability ranking list is output.

[0032] For example, the Bayesian network nodes are defined as including equipment status (e.g., whether the equipment is operating normally, whether the equipment temperature is normal), personnel operation (whether the operation is standardized, whether the operation time is reasonable), and environmental parameters (temperature, humidity, pressure). By collecting historical data, a conditional probability table is learned. When an abnormal increase in pollutant concentration occurs, the belief propagation algorithm is used for reasoning. It is found that the posterior probability of an abnormal temperature of a certain lithography device in the equipment status node is relatively high. At the same time, the posterior probability of non-standard operation near that device is also relatively high in the personnel operation node. The root cause probability ranking list output shows that abnormal equipment temperature and non-standard personnel operation are the main reasons that may lead to an increase in pollutant concentration.

[0033] The active control detection module uses the DQN algorithm to dynamically adjust the sampling frequency and sensor sensitivity with the optimization goals of detection accuracy and system load, and implements a risk-level sampling strategy. The design of the reward function is the core of the DQN algorithm. Defining its form helps in the training and optimization of the algorithm. Therefore, a further improvement of this invention lies in describing the reward function design of the DQN algorithm in the active control detection module as follows: ; in, These are the weighting coefficients, which are optimized through reinforcement learning training.

[0034] The detailed description of the risk-based sampling strategy enables technicians to reasonably adjust the detection parameters according to the level of pollutants, thereby improving detection efficiency and resource utilization. Therefore, a further improvement of this invention is that in the risk-based sampling strategy, the initial sampling frequency for low-level pollutants is set to 1 time / 10 minutes, and the initial sensor sensitivity is a signal-to-noise ratio ≥20 dB; the sampling frequency for high-level pollutants is set to 1 time / 3.3 minutes, and the sensor sensitivity is dynamically increased to a signal-to-noise ratio ≥30 dB.

[0035] The intelligent feedback interface integrates visualization tools such as 3D heat maps and trend comparison charts, as well as an intelligent suggestion system based on a historical case library, which automatically generates response measures.

[0036] The description of the 3D heat map function helps users better understand the distribution of pollutant concentrations, and the marking of historical warning event locations makes it easier for users to summarize lessons learned. Therefore, a further improvement of this invention is that in the intelligent feedback interface, the 3D heat map supports multi-level drilling to display the distribution of pollutant concentrations in cleanrooms, workstations, and equipment layers by layer, and to mark the locations of historical warning events.

[0037] The introduction of the dual-mode intelligent suggestion system allows users to choose the appropriate suggestion acquisition method according to the actual situation. Therefore, a further improvement of the present invention is that the intelligent suggestion system includes a dual-mode system of rule engine and case reasoning: the rule engine generates standardized suggestions based on the expert knowledge base, and the case reasoning provides customized solutions by matching historical cases with similarity.

[0038] For example, when abnormal pollutant concentrations occur, the rule engine generates standardized suggestions based on the expert knowledge base, such as "check whether the ventilation system is normal and check whether the operating parameters of the relevant equipment are within the normal range." Case reasoning, on the other hand, matches historical cases based on similarity. If similar abnormal pollutant concentrations have occurred before under similar process stages and equipment layouts, it provides customized solutions, such as "adjust the wind speed of a certain ventilation device and clean and maintain a certain device."

[0039] The long-term stability of the system is a key issue in practical applications. Clarifying the role and working method of the adaptive traceability verification and calibration module enables technicians to pay attention to and reasonably configure this module, ensuring that the system can operate continuously and effectively. Therefore, a further improvement of this invention is to set up an adaptive traceability verification and calibration module, which is used to periodically compare the sensor's measured values ​​with the digital twin simulation values, automatically correct the model parameters and threshold settings, and ensure the long-term stability of the system.

[0040] The terms "front," "back," "left," "right," "top," and "bottom" all refer to the figures in the accompanying drawings. Figure 1 Based on the perspective of the observer, the side of the device facing the observer is defined as the front, the left side of the observer is defined as the left, and so on.

[0041] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0042] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An online monitoring system for gaseous molecular contaminants in semiconductor manufacturing, characterized in that: include: A small-sized dynamic historical impact analysis module is used to dynamically adjust the historical impact coefficient threshold by combining real-time environmental parameters and process stages, and based on... The hybrid model takes historical and real-time data as input and predicts the future impact trend of pollutants and classifies them into levels. Historical data includes ; This represents the frequency difference in the concentration fluctuation of gaseous molecular contaminants between qualified and unqualified semiconductor processing periods in the clean area during historical operation. It reflects the correlation between contaminant concentration fluctuations and process anomalies; the frequency difference is positively correlated with the impact of contaminants on yield. The frequency difference between qualified and unqualified periods is calculated by statistically analyzing the number of times the pollutant concentration peaks or the rate of change per unit time. This represents the maximum fluctuation range of the concentration of the same gaseous molecular contaminant during different types of non-conforming semiconductor processing within a historical operating period. By extracting the range of contaminant concentration under etching defects or thin film contamination, the average or weighted value is obtained; During historical operating periods, the average concentration fluctuation of gaseous molecular pollutants in the same type of qualified semiconductor processing is calculated by measuring the standard deviation or moving average fluctuation of pollutant concentrations during the qualified period. Real-time data is acquired through sensors, including concentration and temperature / humidity data. The multi-source real-time monitoring and early warning module integrates temperature, humidity, pressure, and equipment status parameters to build a multivariate early warning model, and dynamically allocates the weights of each dimension based on the process stage. The forward-looking early warning analysis module includes a digital twin sub-module, which is used to simulate pollutant diffusion paths and locate high-risk workstations, and a root cause reasoning sub-module, which analyzes pollutant sources and transmission paths based on Bayesian networks; The active control detection module uses the DQN algorithm to dynamically adjust the sampling frequency and sensor sensitivity with the optimization goals of detection accuracy and system load, and implements a risk-level sampling strategy. The intelligent feedback interface integrates visualization tools such as 3D heat maps and trend comparison charts, as well as an intelligent suggestion system based on a historical case library, which automatically generates response measures.

2. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: The small-size source tracing dynamic historical impact analysis module The hybrid model employs a multi-head attention mechanism to process historical pollutant data. Temporal features are extracted and dynamic weights are assigned based on real-time environmental parameters such as temperature, humidity, and pressure. The pollutant future impact trend and level information predicted by the small-sized source tracing dynamic historical impact analysis module is one of the reference factors when the multi-source real-time monitoring and early warning module dynamically allocates the weights of each dimension.

3. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: The process of constructing a multivariate early warning model in the multi-source real-time monitoring and early warning module includes the following steps: Data collection and integration, variable selection and weight allocation, model building and training, model validation and optimization, and real-time early warning and application; The dynamic allocation of weights for each dimension is based on a process stage sensitivity matrix, which is obtained through training with historical data and maps the sensitivity coefficients of each process stage to the pollutant dimension.

4. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: The digital twin submodule uses the finite element analysis method to construct a cleanroom flow field model. After inputting real-time monitoring data, it simulates the pollutant diffusion path through CFD simulation and outputs the coordinates of high-risk areas and concentration distribution cloud map. The construction of a cleanroom flow field model includes the following steps: collecting basic data, geometric modeling and mesh generation, physical model selection and parameter setting, model solving, result analysis and model verification, and model integration and application. The coordinates of high-risk areas simulated by the digital twin submodule in the forward-looking early warning analysis module and the root cause probability ranking list output by the root cause reasoning submodule will be transmitted to the active control detection module to guide it in adjusting the sampling strategy and sensor sensitivity, focusing on high-risk areas and potential pollution sources.

5. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: The process of analyzing pollutant sources and transmission pathways based on Bayesian networks includes: Various factors related to the source and spread of pollutants are identified as nodes in a Bayesian network, and the structure of the Bayesian network is constructed based on the causal and conditional dependencies between the factors. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is then used to learn and estimate the conditional probability table in the Bayesian network, determining the probability distribution of each node under different conditions. Historical data is collected, including pollutant monitoring data, meteorological data, and equipment operation data. This historical data is used to learn and estimate the conditional probability table in the Bayesian network to determine the probability distribution of each node under different conditions. When new monitoring data or information is input, the inference algorithm of Bayesian network is used to perform inference and calculate the posterior probability of each pollution source as the source of pollutants to determine the most likely pollution source. At the same time, the propagation path and diffusion range of pollutants under the influence of different factors are analyzed. The reasoning results are interpreted and analyzed, transforming complex probability calculations into understandable information, and based on the analysis results, providing decision support for the prevention and control of pollutants; In the root cause reasoning submodule, the Bayesian network nodes include equipment status, personnel operation, and environmental parameters. The contribution of each node to the pollutant event is quantified through a conditional probability table, and a root cause probability ranking list is output.

6. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: In the active control detection module, the reward function of the DQN algorithm is designed as follows: ； in, These are the weighting coefficients, which are optimized through reinforcement learning training.

7. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: In the risk-based sampling strategy, the initial sampling frequency for low-level pollutants is set to 1 time / 10 minutes, and the initial sensor sensitivity is a signal-to-noise ratio ≥20 dB; the sampling frequency for high-level pollutants is set to 1 time / 3.3 minutes, and the sensor sensitivity is dynamically increased to a signal-to-noise ratio ≥30 dB.

8. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: In the intelligent feedback interface, the 3D heat map supports multi-level drilling, which is used to display the distribution of pollutant concentrations in cleanrooms, workstations, and equipment layers by layer, and to mark the locations of historical warning events.

9. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: The intelligent suggestion system includes a dual-mode system of rule engine and case reasoning: the rule engine generates standardized suggestions based on an expert knowledge base, while the case reasoning provides customized solutions by matching historical cases with similarity.

10. The online monitoring system for gaseous molecular contaminants in semiconductor manufacturing according to claim 1, characterized in that: It includes an adaptive traceability verification and calibration module, which is used to periodically compare the actual sensor values ​​with the digital twin simulation values, automatically correct the model parameters and threshold settings, and ensure the long-term stability of the system.

Images