Semiconductor manufacturing equipment particle pollution tracing method and system
By constructing a graph neural network model and Bayesian inference, the problem of misjudgment of interfering substances in the source tracing of particle contamination in semiconductor manufacturing equipment was solved, achieving accurate source location and tracing of contamination and improving the accuracy of source tracing of particle contamination in equipment.
Patent Information
- Application Number
- CN202511181642.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-22
- Publication Date
- 2025-10-31
AI Technical Summary
In existing technologies, the accuracy of particle contamination tracing in semiconductor manufacturing equipment is difficult to guarantee, especially due to misjudgment of interfering objects such as bubbles or water droplets, which increases the difficulty of identifying particle contamination sources.
A particle pollution tracing system using semiconductor manufacturing equipment includes a monitoring module, a data analysis module, and a graph neural network module. By constructing first and second graph neural network models, it handles the cases with and without interfering substances, respectively. Combined with Bayesian inference, it locates the pollution source equipment and corrects the true value of particle concentration.
It improves the accuracy of particle contamination tracing in semiconductor manufacturing equipment, shortens the contamination tracing time in wafer fabs, and enables accurate tracing of contamination source equipment.
Smart Images
Figure CN120869907A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor contamination tracing technology, specifically to a method and system for tracing particle contamination in semiconductor manufacturing equipment. Background Technology
[0002] Equipment particle contamination in semiconductor manufacturing refers to the phenomenon where particles generated by equipment during the semiconductor manufacturing process contaminate semiconductor wafers. As the feature size of semiconductor devices becomes smaller and smaller, when the particle size approaches or exceeds the feature size of the device, it will seriously affect the performance and reliability of the device. Therefore, it is necessary to monitor and analyze the particles generated during equipment operation in real time, promptly detect abnormalities in particle contamination, and take corresponding measures to deal with them. By analyzing the characteristics of the particles, such as particle size, composition, and shape, the source of particle contamination can be determined, thus narrowing down the scope to specific equipment or process steps.
[0003] However, extracting and identifying particle characteristics is not easy. The shape, size, and composition of particles are affected by various factors and can change, leading to unstable or unclear characteristics. In addition, different types of particles may have overlapping or similar characteristics, increasing the difficulty of accurate identification. Finally, in semiconductor cleanrooms, bubbles or water droplets may be misidentified as particles, leading to incorrect source tracing. Therefore, how to improve the accuracy of particle contamination source tracing in equipment has become an urgent problem to be solved. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for tracing the source of particulate contamination in semiconductor manufacturing equipment, in order to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a particle contamination tracing system for semiconductor manufacturing equipment, comprising a monitoring module, a data analysis module, a graph neural network module, and a data storage module; the monitoring module is used to collect monitoring data from the cleanroom and send it to the data storage module; the graph neural network module is used to generate a first graph neural network model when there are no interfering substances and a second graph neural network model when there are interfering substances, obtain the true value of particle concentration based on the graph neural network model, and send the true value of particle concentration to the data analysis module; the data analysis module locates the contamination source equipment through Bayesian inference; the data storage module is used to store historical monitoring data from the cleanroom.
[0006] Specifically, the graph neural network module further includes: a model generation unit, a similarity calculation unit, and a ground truth generation unit; the model generation unit is used to construct a first graph neural network model and a second graph neural network model; the similarity calculation unit is used to determine the similarity between feature matrices; and the ground truth generation unit is used to obtain the ground truth value of particle concentration.
[0007] Specifically, the data analysis module further includes: a judgment unit, an analysis unit, and a classification unit; the judgment is used to determine the presence of interfering substances in the clean room; the analysis unit is used to analyze the particle concentration change rate data to determine whether it is necessary to conduct source tracing analysis of particle pollution; the classification unit is used to obtain the probability of the presence of interfering substances from the particle concentration change rate.
[0008] Specifically, the real value generation unit acquires historical monitoring data of the cleanroom when interfering substances are present, extracts the features of the equipment nodes from the monitoring data, inputs the features of the equipment nodes into the first graph neural network model, uses the features of the particle counter nodes as the prediction target, and uses time series prediction to obtain the features of the particle counter nodes; obtains the predicted particle concentration value of the particle counter based on the features of the particle counter nodes, and obtains the error curve based on the predicted particle concentration value and the observed particle concentration value of historical monitoring data;
[0009] The system acquires monitoring data from the current cleanroom. If there are no interfering substances, the true value of the current particle concentration is obtained through the first neural network model. If there are interfering substances, the predicted value of the current particle concentration is obtained through the second neural network model. The system calculates the similarity between the features of the current equipment node and the features of historical equipment nodes in the cleanroom when interfering substances are present. The system determines the error curve corresponding to the features of the closest historical equipment node. The system calculates the true value of the current particle concentration by performing operations on the error curve and the predicted value of the current particle concentration. The closest refers to the greatest similarity between the features of the historical equipment node and the current equipment node.
[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for tracing the source of particulate contamination in semiconductor manufacturing equipment, comprising the following steps:
[0011] Acquire historical monitoring data in the cleanroom; determine the method for identifying interfering substances and their presence based on the historical monitoring data; construct a first-graph neural network model based on historical monitoring data without interfering substances; acquire historical monitoring data with interfering substances, input the historical monitoring data into the first-graph neural network model to obtain the predicted particle concentration; obtain the error based on the predicted particle concentration and the observed values in the historical monitoring data, and form an error curve L based on the error.
[0012] A second-graph neural network model is constructed based on historical monitoring data of interfering substances. The current monitoring data of the cleanroom is then obtained and input into the second-graph neural network model to obtain the predicted value of the current particle concentration. The predicted value of the current particle concentration and the error curve L are calculated to obtain the true value of the current particle concentration. Based on the true value of the current particle concentration, Bayesian inference is performed to trace the source of particle contamination.
[0013] Specifically, the method for identifying interfering substances based on historical monitoring data in cleanrooms also includes the following steps:
[0014] In the historical monitoring data of the cleanroom, particle concentration data from the particle counter is obtained, and particle concentration change rate data is obtained from the particle concentration data. The particle concentration change rate data is arranged in ascending order, and the sequence value that is significantly different from k and greater than k is identified from the arrangement. The first threshold is determined based on the sequence value. If the particle concentration change rate data calculated from the particle concentration data of the particle counter is not less than the first threshold, it is assumed that there is an interfering substance. Then, the presence of the interfering substance is verified, where k is a constant.
[0015] Specifically, determining the presence of interfering substances based on historical monitoring data in the cleanroom also includes the following steps:
[0016] Obtain the particle concentration time series of the particle counter containing interference. Let the particle counter containing interference be the target particle counter. Record the time point when the target particle counter is identified as having an abnormal increase in particle concentration as t. Obtain the average rate of change of particle concentration of the target particle counter before time t-dt, the average rate of change of particle concentration of the target particle counter after time t+dt at2, and the average rate of change of particle concentration of the target particle counter between time [t-dt, t+dt] at; where dt is a preset time window.
[0017] Calculate the difference between at and at1 to obtain Δt, and perform unsupervised classification on [Δt, at2]:
[0018] Obtain historical data of particle counters with interfering objects, calculate the average change rate v1 of particle concentration before the appearance of interfering objects and the average change rate v2 of particle concentration after time dt of the appearance of interfering objects; obtain the average change rate v of particle concentration of particle counters within time dt of the appearance of interfering objects, calculate the difference between v and v1 to obtain Δv, and add [Δv, v2] to the unsupervised classification dataset.
[0019] Obtain historical particle concentration data of the particle counter when particle contamination occurs, and calculate the average change rate r1 of the particle concentration of the particle counter before particle contamination occurs and the average change rate r2 of the particle concentration of the particle counter after time dt of particle contamination occurs; obtain the change rate r of the particle concentration of the particle counter within time dt of particle contamination, calculate the difference between r and r1 to obtain Δr, and add [Δr, r2] to the unsupervised classification dataset.
[0020] Unsupervised classification is performed on [Δt, at2] using an unsupervised classification dataset. The probability of the presence of interference is obtained based on the proportion of data belonging to the cluster where particle contamination occurs. If the probability of the presence of interference is not less than a set threshold, interference is judged to exist; otherwise, particle contamination is judged to exist.
[0021] Specifically, constructing the first graph neural network model based on historical monitoring data in the absence of interfering objects also includes the following steps:
[0022] The process involves acquiring monitoring data from a cleanroom in the absence of interfering substances, extracting features of equipment nodes from the monitoring data, using the equipment and particle counters in the cleanroom as nodes in a graph neural network, and defining the connections between equipment and between particle counters and equipment as edges. A graph neural network model is then set up to process the graph data. The feature matrix of the equipment nodes is used as input, and the connection relationships between nodes are defined through the adjacency matrix of the graph. The graph neural network is then used to extract node embeddings, capture the characteristics of pollution source propagation to nodes, update node states, learn propagation patterns, and obtain the first graph neural network model.
[0023] Specifically, the process of calculating the true value of the current particle concentration from the predicted value and the error curve L also includes the following steps:
[0024] Historical monitoring data of the cleanroom in the presence of interfering substances is acquired. The characteristics of the equipment nodes are extracted from the monitoring data and input into the first graph neural network model. The characteristics of the particle counter nodes are used as the prediction target, and the characteristics of the particle counter nodes are obtained by time series prediction. The predicted particle concentration value of the particle counter is obtained based on the characteristics of the particle counter nodes. The error curve is obtained based on the predicted particle concentration value and the observed particle concentration value of historical monitoring data. The error curve and the characteristics of the equipment nodes are stored.
[0025] Historical monitoring data of the cleanroom when interference was present was acquired, and the characteristics of the equipment nodes were extracted from the monitoring data. A second neural network model was trained based on the characteristics of the equipment nodes. The current monitoring data of the cleanroom was acquired. If there was no interference, the true value of the current particle concentration was obtained through the first neural network model. If there was interference, the predicted value of the current particle concentration was obtained through the second neural network model. The similarity between the characteristics of the current equipment node and the characteristics of the historical equipment nodes in the cleanroom when interference was present was calculated. The error curve corresponding to the closest historical equipment node characteristics was determined. The error curve and the predicted value of the current particle concentration were calculated to obtain the true value of the current particle concentration. The closest means that the similarity between the historical equipment node characteristics and the current equipment node is the greatest.
[0026] Specifically, tracing the source of particle contamination through Bayesian inference based on the current true value of particle concentration also includes the following steps:
[0027] The posterior probability P{i|D} is calculated using Bayes' theorem: P{i|D} = (P{D|i} × P{i}) / P{D}, where P{i} is the prior probability, representing the probability that the i-th device is a pollution source; P{D} is the evidence factor, representing the probability of the observed data D occurring, used for normalization; and P{D|i} is the likelihood function, representing the probability of the observed data occurring when the i-th device is a pollution source.
[0028] The likelihood function P{D|i} is determined through the following steps:
[0029] Assuming device i is the pollution source, based on the current presence of interfering substances, the true value of the particle counter node is obtained through the first and second graph neural network models. A Gaussian distribution is obtained where the true value of the counter node is the average value and the noise of the counter node is the variance. Based on the Gaussian distribution, the probability P{D|i} of the observed data when the i-th device is the pollution source is obtained.
[0030] Based on the posterior probabilities of all devices, the device node with the highest posterior probability is selected as the candidate pollution source.
[0031] Compared with existing technologies, the beneficial effects of this invention are: developing a multi-sensor data fusion platform for cleanrooms, constructing a pollution propagation model based on graph neural networks, locating pollution source equipment through Bayesian inference, and shortening the time for tracing pollution sources in wafer fabs; and utilizing the rapid decay characteristics of bubbles or water droplets to correct the true value of particle concentration and accurately trace pollution source equipment. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the particle contamination tracing system for semiconductor manufacturing equipment of the present invention;
[0033] Figure 2 This is a flowchart of the method for tracing the source of particle contamination in semiconductor manufacturing equipment according to the present invention. Detailed Implementation
[0034] 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.
[0035] Example: Figure 1As shown, this invention provides a particle contamination tracing system for semiconductor manufacturing equipment, including a monitoring module, a data analysis module, a graph neural network module, and a data storage module. The monitoring module collects monitoring data from the cleanroom and sends it to the data storage module. The graph neural network module generates a first graph neural network model when there are no water droplets or bubbles and a second graph neural network model when there are water droplets or bubbles. Based on the graph neural network models, it obtains the true value of the particle concentration and sends the true value of the particle concentration to the data analysis module. The data analysis module locates the contamination source equipment through Bayesian inference. The data storage module stores historical monitoring data from the cleanroom.
[0036] The graph neural network module further includes: a model generation unit, a similarity calculation unit, and a ground truth generation unit; the model generation unit is used to construct a first graph neural network model and a second graph neural network model; the similarity calculation unit is used to determine the similarity between feature matrices; and the ground truth generation unit is used to obtain the ground truth value of particle concentration. The data analysis module further includes: a judgment unit, an analysis unit, and a classification unit; the judgment unit is used to determine the presence of water droplets or bubbles in the cleanroom; the analysis unit is used to analyze the particle concentration change rate data to determine whether source tracing analysis of particle contamination is necessary; and the classification unit is used to obtain the probability of the presence of water droplets or bubbles from the particle concentration change rate.
[0037] The real value generation unit acquires historical monitoring data of the cleanroom when water droplets or bubbles are present, extracts the features of the equipment nodes from the monitoring data, inputs the features of the equipment nodes into the first graph neural network model, uses the features of the particle counter nodes as the prediction target, and uses time series prediction to obtain the features of the particle counter nodes; obtains the predicted particle concentration value of the particle counter based on the features of the particle counter nodes, and obtains the error curve based on the predicted particle concentration value and the observed particle concentration value of historical monitoring data;
[0038] The system acquires monitoring data from the current cleanroom. If there are no water droplets or bubbles, the system obtains the true value of the current particle concentration using the first neural network model. If there are water droplets or bubbles, the system obtains the predicted value of the current particle concentration using the second neural network model. The system calculates the similarity between the features of the current equipment node and the features of historical equipment nodes in the cleanroom when water droplets or bubbles were present. It then determines the error curve corresponding to the closest historical equipment node features and calculates the true value of the current particle concentration using the error curve and the predicted value of the current particle concentration.
[0039] In another embodiment of the present invention, the present invention provides a method for tracing the source of particle contamination in semiconductor manufacturing equipment, which involves acquiring historical monitoring data in a cleanroom; determining the judgment method and the presence of water droplets or bubbles based on the historical monitoring data in the cleanroom; constructing a first graph neural network model based on historical monitoring data in which no water droplets or bubbles are present; acquiring historical monitoring data in which water droplets or bubbles are present, and inputting the historical monitoring data as input into the first graph neural network model to obtain a predicted value of particle concentration; obtaining the error based on the predicted value of particle concentration and the observed value in the historical monitoring data, and forming an error curve L based on the error.
[0040] A second-graph neural network model is constructed based on historical monitoring data of water droplets or bubbles. The current monitoring data of the cleanroom is then obtained and input into the second-graph neural network model to obtain the predicted value of the current particle concentration. The predicted value of the current particle concentration and the error curve L are calculated to obtain the true value of the current particle concentration. Based on the true value of the current particle concentration, Bayesian inference is used to trace the source of particle contamination.
[0041] Historical monitoring data includes particle concentration data from particle counters and environmental data, including but not limited to temperature, humidity, and wafer fab process parameters; for different process parameters, non-numerical data is converted into numerical data by assignment.
[0042] The network structures of the first and second neural networks differ. The second neural network is trained on historical monitoring data showing the presence of water droplets or bubbles. Since it includes the influence of water droplets or bubbles on particle concentration, misidentifying them as particles can interfere with the results and cause errors. This error needs to be captured, and then the input to the second neural network is used to calculate the true value. The first neural network, however, does not include the influence of water droplets or bubbles on particle concentration. Therefore, when historical monitoring data showing the presence of water droplets or bubbles is input into the first neural network, the output also does not include this influence, leading to errors compared to the observed values. The observed values, which do include the influence of water droplets or bubbles, can be used to calculate the error by comparing the observed values with the output of the first neural network. By analyzing the continuous historical monitoring data over time, the errors between the continuous observations and the output of the first neural network are obtained, forming an error sequence. From this error sequence, an error curve can be derived. The current monitoring data of the cleanroom is input into the second neural network model to obtain the predicted value of the current particle concentration. Based on the time point of water droplet appearance obtained from the bubble, and corresponding to the starting point of the error curve, the predicted value of particle concentration output by the second neural network is calculated with the corresponding position of the error curve to obtain the true value of particle concentration.
[0043] Let L1 represent the particle concentration data in the historical monitoring data where water droplets or bubbles exist, and L2 represent the predicted value obtained by inputting the historical monitoring data into the first graph neural network. Then we get the error L1-L2, and construct the error curve based on the continuous error over time.
[0044] Let L(T) represent the error value after a time T when a water droplet or bubble appears on the error curve. Let L be the predicted value of the current particle concentration output by the neural network in the second graph after a time T when a water droplet or bubble appears. L corresponds to L(T) in time. Subtract L(T) from L to obtain the true value.
[0045] The number of error curves that can be obtained is not unique, so it is necessary to select an appropriate error curve for calculation.
[0046] The method for determining whether a water droplet or an air bubble is present based on historical monitoring data in the cleanroom also includes the following steps:
[0047] In the historical monitoring data of the cleanroom, the particle concentration data of the particle counter is obtained, and the particle concentration change rate data is obtained from the particle concentration data. The particle concentration change rate data is arranged in ascending order, and the sequence value that is significantly different from k and greater than k is identified from the arrangement. The first threshold is determined based on the sequence value. If the particle concentration change rate data calculated from the particle concentration data of the particle counter is not less than the first threshold, it is assumed that water droplets or bubbles are present. Then, the presence of water droplets or bubbles is verified, where k is a constant.
[0048] k can be set to 0. Significance can be determined statistically by obtaining the average and variance of the particle concentration change rate. The average and variance are then used to determine if the particle concentration is significantly different from 0. When particle contamination occurs, the particle counter detects an increase in particle concentration, thus increasing the particle concentration change rate. Besides particle contamination, water droplets or bubbles can also increase the particle concentration detected by the particle counter. However, water droplets or bubbles can evaporate or burst, significantly reducing the detected particle concentration. Therefore, by using sequence values greater than 0 that are significantly different from 0, such as setting the confidence threshold to 0.9, the confidence level of the sequence value differing from 0 is obtained statistically. When the confidence level is greater than 0.9, the sequence value is considered significantly different from 0, and the corresponding particle concentration data is found. This particle concentration data is the first threshold. First, the first threshold is used to determine whether pollution source tracing is necessary. Then, it is determined whether the misjudgment is due to water droplets or bubbles. If not, pollution source tracing begins.
[0049] Determining the presence of water droplets or air bubbles based on historical monitoring data in the cleanroom also includes the following steps:
[0050] Obtain the particle concentration time series of the particle counter containing water droplets or bubbles. Let the particle counter containing water droplets or bubbles be the target particle counter. Record the time point when the target particle counter is identified as having an abnormal increase in particle concentration as t. Obtain the average rate of change of particle concentration of the target particle counter before time t-dt, the average rate of change of particle concentration of the target particle counter after time t+dt at2, and the average rate of change of particle concentration of the target particle counter between time [t-dt, t+dt] at; where dt is a preset time window.
[0051] Calculate the difference between at and at1 to obtain Δt, and perform unsupervised classification on [Δt, at2]:
[0052] Obtain historical data of particle counters containing water droplets or bubbles, and calculate the average rate of change of particle concentration v1 before the appearance of water droplets or bubbles and the average rate of change of particle concentration v2 after time dt. Obtain the average rate of change v of particle concentration v of particle counters within time dt when water droplets or bubbles appear, calculate the difference between v and v1 to obtain Δv, and add [Δv, v2] to the unsupervised classification dataset.
[0053] Obtain historical particle concentration data of the particle counter when particle contamination occurs, and calculate the average change rate r1 of the particle concentration of the particle counter before particle contamination occurs and the average change rate r2 of the particle concentration of the particle counter after time dt of particle contamination occurs; obtain the change rate r of the particle concentration of the particle counter within time dt of particle contamination, calculate the difference between r and r1 to obtain Δr, and add [Δr, r2] to the unsupervised classification dataset.
[0054] Unsupervised classification is performed on [Δt, at2] using an unsupervised classification dataset. The probability of the presence of water droplets or bubbles is obtained based on the proportion of data belonging to the cluster where particle contamination occurs. If the probability of the presence of water droplets or bubbles is not less than a set threshold, it is determined that water droplets or bubbles exist; otherwise, it is determined that particle contamination exists.
[0055] To determine whether there is interference from water droplets or bubbles, unsupervised classification is performed on [Δt, at2]. For this purpose, unsupervised classification is performed on the data [Δr, r2] obtained when there is particle contamination, the data [Δv, v2] obtained when there are water droplets or bubbles, and the current data [Δt, at2] to determine the classification cluster to which the current data [Δt, at2] belongs. The presence of water droplet or bubble interference is determined based on the ratio of the two classes in the classification cluster. [Δt, at2], [Δv, v2], and [Δr, r2] are all vectors and serve as features for unsupervised classification.
[0056] Particle counters may misidentify particles, even mistaking bubbles or water droplets for particles, leading to inaccurate detection results and affecting the reliability of semiconductor monitoring. Therefore, it is necessary to identify bubbles or water droplets. From a physical perspective, bubbles or water droplets may disappear during propagation due to mechanisms such as evaporation, bursting, and condensation, while solid particles typically persist in a solid state—this is the most significant difference between the two. When particle concentration begins to increase, it could be due to solid particle contamination or interference from bubbles or water droplets. However, bubbles or water droplets disappear quickly, reducing particle concentration, while solid particle contamination can maintain particle concentration for a longer period. Therefore, two sets of data are used for cluster analysis. First, the difference is used to distinguish between particle contamination (which could be interference from bubbles or water droplets) and normal conditions. Then, the average rate of change of particle concentration from the particle counter after time dt is used to distinguish between bubbles or water droplets and solid particles. Historical data where the presence of bubbles or water droplets is known, and historical data where their absence is known, are used for classification to determine the presence of water droplets or bubbles.
[0057] Building the first graph neural network model based on historical monitoring data where no water droplets or bubbles are present also includes the following steps:
[0058] The process involves acquiring monitoring data from a cleanroom when there are no water droplets or bubbles, extracting features of equipment nodes from the monitoring data, using the equipment and particle counters in the cleanroom as nodes in a graph neural network, and defining the connections between equipment and between particle counters and equipment as edges. A graph neural network model is then set up to process the graph data. The feature matrix of the equipment nodes is used as input, and the connection relationships between nodes are defined through the adjacency matrix of the graph. The graph neural network is then used to extract node embeddings, capture the characteristics of pollution source propagation to nodes, update node states, learn propagation patterns, and obtain the first graph neural network model.
[0059] Construct a graph neural network model. The graph is defined as follows: Nodes: Representing devices or particle counters. Device and particle counter node features include static and dynamic characteristics. Edges: The connections between devices and between particle counters and devices, such as the flow path and diffusion direction of groundwater. Edge features: Including static and dynamic edge features. Static edge features include distance, direction, and airflow path; dynamic edge features include airflow velocity and direction. Graph neural network type: Select an appropriate graph neural network model (such as GCN, GraphSAGE, GAT) to process the graph data.
[0060] Data acquisition and integration include historical data for equipment nodes: This includes both static and dynamic characteristics of the equipment. Static characteristics include equipment type (e.g., etching machine, lithography machine, which can be encoded using one-hot encoding), equipment coordinates, installation time, and maintenance status. Dynamic characteristics include current operating status (running, standby, maintenance, which can be encoded using one-hot encoding), current power, energy consumption, temperature, and vibration amplitude. Historical data for particle counter nodes includes static characteristics such as location, type, and detection range; dynamic characteristics include particle concentration. Real-time data is obtained through sensors on the equipment and the particle counter itself. Data fusion is used to form a dataset, constructing a graph structure containing node features and edge information.
[0061] Model Input and Feature Extraction: Input features include: Node feature matrix: features of device nodes and particle counter nodes; Graph adjacency matrix: defining the connections between nodes. Feature processing: Utilizing a graph neural network to extract node embeddings to capture the characteristics of pollution propagation. Updating node states to learn the dynamic trends of pollution diffusion.
[0062] Diffusion trend calculation, output target: particle concentration data of particle counter nodes. Propagation simulation: using time series prediction (such as methods based on time-series graphical neural networks) to simulate changes in particle contamination.
[0063] Model training and optimization outputs node-level tasks: regression values of particle concentration for particle counter nodes; training data: historical observation data as the training set; pollution diffusion information at different time points; loss function: using root mean square error (RMSE) or other loss functions to minimize the difference between predicted and observed values; optimization method: using Adam or SGD optimizers; selecting optimal model parameters through cross-validation.
[0064] The process of calculating the true value of the current particle concentration from the predicted value and the error curve L also includes the following steps:
[0065] Historical monitoring data of the cleanroom in the presence of water droplets or bubbles is acquired. Features of the equipment nodes are extracted from this data and input into the first graph neural network model. The features of the particle counter nodes are used as the prediction target, and time series prediction is employed to obtain the features of the particle counter nodes. Based on the features of the particle counter nodes, the predicted particle concentration value of the particle counter is obtained. An error curve is generated based on the predicted particle concentration value and the observed particle concentration values from historical monitoring data. The error curve and the features of the equipment nodes are then stored.
[0066] Historical monitoring data of the cleanroom when water droplets or bubbles were present was acquired. Features of the equipment nodes were extracted from this data, and a second neural network model was trained based on these features. Current monitoring data of the cleanroom was then acquired. If no water droplets or bubbles were present, the true particle concentration was obtained using the first neural network model. If water droplets or bubbles were present, the predicted particle concentration was obtained using the second neural network model. The similarity between the features of the current equipment node and the historical features of equipment nodes in the cleanroom when water droplets or bubbles were present was calculated. The error curve corresponding to the closest historical equipment node features was determined. The error curve and the predicted particle concentration were then used to calculate the true particle concentration. Since multiple error curves could exist, the closest error curve was selected for calculation.
[0067] The features of a device node are in the form of a feature matrix. Therefore, the closest historical device node features can be determined through matrix similarity calculations, including but not limited to co-similarity, Euclidean distance, and singular values. After determining the corresponding error curve, the values on the error curve and the predicted values are calculated based on the time points to obtain the true values. The term "closest" refers to the historical device node features having the highest similarity to the current device node. The features of a device node include, but are not limited to, the device's industrial parameters, temperature, humidity, and other environmental data.
[0068] The observed values here refer to the measurement results of the particle counter, which are the actual values; while the predicted values and the true values are both predicted values obtained through the graph neural network model. The difference is that the predicted values include the interference of water droplets or bubbles, while the true values do not have the interference of water droplets or bubbles. Therefore, Bayesian inference is performed based on the true values.
[0069] Tracing the source of particulate contamination using Bayesian inference based on the current true value of particle concentration also includes the following steps:
[0070] The posterior probability P{i|D} is calculated using Bayes' theorem: P{i|D} = (P{D|i} × P{i}) / P{D}, where P{i} is the prior probability, representing the probability that the i-th device is a pollution source; P{D} is the evidence factor, representing the probability of the observed data D occurring, used for normalization; and P{D|i} is the likelihood function, representing the probability of the observed data occurring when the i-th device is a pollution source.
[0071] The likelihood function P{D|i} is determined through the following steps:
[0072] Assuming device i is the pollution source, based on the current presence of water droplets or bubbles, the true value of the particle counter node is obtained through the first and second graph neural network models. A Gaussian distribution is obtained where the true value of the counter node is the average value and the noise of the counter node is the variance. Based on the Gaussian distribution, the probability P{D|i} of the observed data when the i-th device is the pollution source is obtained.
[0073] Based on the posterior probabilities of all devices, the device node with the highest posterior probability is selected as the candidate pollution source.
[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for tracing the source of particulate contamination in semiconductor manufacturing equipment, characterized in that, Includes the following steps: Obtain historical monitoring data in the cleanroom; The method for identifying interfering substances and their presence is determined based on historical monitoring data in the cleanroom; a first-graph neural network model is constructed based on historical monitoring data where no interfering substances are present; historical monitoring data where interfering substances are present is obtained and used as input into the first-graph neural network model to obtain the predicted value of particle concentration; the error is obtained based on the predicted value of particle concentration and the observed value in the historical monitoring data, and an error curve L is formed based on the error. A second-graph neural network model is constructed based on historical monitoring data of interfering substances. The current monitoring data of the cleanroom is then obtained and input into the second-graph neural network model to obtain the predicted value of the current particle concentration. The predicted value of the current particle concentration and the error curve L are calculated to obtain the true value of the current particle concentration. Based on the true value of the current particle concentration, Bayesian inference is performed to trace the source of particle contamination.
2. The method for tracing the source of particulate contamination in semiconductor manufacturing equipment according to claim 1, characterized in that, The method for determining interfering substances based on historical monitoring data in the cleanroom also includes the following steps: In the historical monitoring data of the cleanroom, particle concentration data from the particle counter is obtained, and particle concentration change rate data is obtained from the particle concentration data. The particle concentration change rate data is arranged in ascending order, and the sequence value that is significantly different from k and greater than k is identified from the arrangement. The first threshold is determined based on the sequence value. If the particle concentration change rate data calculated from the particle concentration data of the particle counter is not less than the first threshold, it is assumed that there is an interfering substance. Then, the presence of the interfering substance is verified, where k is a constant.
3. The method for tracing the source of particulate contamination in semiconductor manufacturing equipment according to claim 2, characterized in that, Determining the presence of interfering substances based on historical monitoring data in the cleanroom also includes the following steps: Obtain the particle concentration time series of the particle counter containing interference. Let the particle counter containing interference be the target particle counter. Record the time point when the target particle counter is identified as having an abnormal increase in particle concentration as t. Obtain the average rate of change of particle concentration of the target particle counter before time t-dt, the average rate of change of particle concentration of the target particle counter after time t+dt at2, and the average rate of change of particle concentration of the target particle counter between time [t-dt, t+dt] at; where dt is a preset time window. Calculate the difference between at and at1 to obtain Δt, and perform unsupervised classification on [Δt, at2]: Obtain historical data of particle counters with interfering objects, calculate the average change rate v1 of particle concentration before the appearance of interfering objects and the average change rate v2 of particle concentration after time dt of the appearance of interfering objects; obtain the average change rate v of particle concentration of particle counters within time dt of the appearance of interfering objects, calculate the difference between v and v1 to obtain Δv, and add [Δv, v2] to the unsupervised classification dataset. Obtain historical particle concentration data of the particle counter when particle contamination occurs, and calculate the average change rate r1 of the particle concentration of the particle counter before particle contamination occurs and the average change rate r2 of the particle concentration of the particle counter after time dt of particle contamination occurs; obtain the change rate r of the particle concentration of the particle counter within time dt of particle contamination, calculate the difference between r and r1 to obtain Δr, and add [Δr, r2] to the unsupervised classification dataset. Unsupervised classification is performed on [Δt, at2] using an unsupervised classification dataset. The probability of the presence of interference is obtained based on the proportion of data belonging to the cluster where particle contamination occurs. If the probability of the presence of interference is not less than a set threshold, interference is judged to exist; otherwise, particle contamination is judged to exist.
4. The method for tracing the source of particulate contamination in semiconductor manufacturing equipment according to claim 3, characterized in that, The process of constructing the first graph neural network model based on historical monitoring data in the absence of interfering objects also includes the following steps: The process involves acquiring monitoring data from a cleanroom in the absence of interfering substances, extracting features of equipment nodes from the monitoring data, using the equipment and particle counters in the cleanroom as nodes in a graph neural network, and defining the connections between equipment and between particle counters and equipment as edges. A graph neural network model is then set up to process the graph data. The feature matrix of the equipment nodes is used as input, and the connection relationships between nodes are defined through the adjacency matrix of the graph. The graph neural network is then used to extract node embeddings, capture the characteristics of pollution source propagation to nodes, update node states, learn propagation patterns, and obtain the first graph neural network model.
5. The method for tracing the source of particulate contamination in semiconductor manufacturing equipment according to claim 4, characterized in that, The process of calculating the predicted value of the current particle concentration and the error curve L to obtain the true value of the current particle concentration also includes the following steps: Historical monitoring data of the cleanroom in the presence of interfering substances is acquired. The characteristics of the equipment nodes are extracted from the monitoring data and input into the first graph neural network model. The characteristics of the particle counter nodes are used as the prediction target, and the characteristics of the particle counter nodes are obtained by time series prediction. The predicted particle concentration value of the particle counter is obtained based on the characteristics of the particle counter nodes. The error curve is obtained based on the predicted particle concentration value and the observed particle concentration value of historical monitoring data. The error curve and the characteristics of the equipment nodes are stored. Historical monitoring data of the cleanroom when interference was present was acquired, and the characteristics of the equipment nodes were extracted from the monitoring data. A second neural network model was trained based on the characteristics of the equipment nodes. The current monitoring data of the cleanroom was acquired. If there was no interference, the true value of the current particle concentration was obtained through the first neural network model. If there was interference, the predicted value of the current particle concentration was obtained through the second neural network model. The similarity between the characteristics of the current equipment node and the characteristics of the historical equipment nodes in the cleanroom when interference was present was calculated. The error curve corresponding to the closest historical equipment node characteristics was determined. The error curve and the predicted value of the current particle concentration were calculated to obtain the true value of the current particle concentration. The closest refers to the greatest similarity between the historical equipment node characteristics and the current equipment node.
6. The method for tracing the source of particulate contamination in semiconductor manufacturing equipment according to claim 5, characterized in that, The method of tracing the source of particle contamination using Bayesian inference based on the current true value of particle concentration also includes the following steps: The posterior probability P{i|D} is calculated using Bayes' theorem: P{i|D} = (P{D|i} × P{i}) / P{D}, where P{i} is the prior probability, representing the probability that the i-th device is a pollution source; P{D} is the evidence factor, representing the probability of the observed data D occurring, used for normalization; and P{D|i} is the likelihood function, representing the probability of the observed data occurring when the i-th device is a pollution source. The likelihood function P{D|i} is determined through the following steps: Assuming device i is the pollution source, based on the current presence of interfering substances, the true value of the particle counter node is obtained through the first and second graph neural network models. A Gaussian distribution is obtained where the true value of the counter node is the average value and the noise of the counter node is the variance. Based on the Gaussian distribution, the probability P{D|i} of the observed data when the i-th device is the pollution source is obtained. Based on the posterior probabilities of all devices, the device node with the highest posterior probability is selected as the candidate pollution source.
7. A particle contamination tracing system for semiconductor manufacturing equipment, characterized in that, It includes a monitoring module, a data analysis module, a graph neural network module, and a data storage module. The monitoring module collects monitoring data from the cleanroom and sends it to the data storage module. The graph neural network module generates a first graph neural network model when there are no interfering substances and a second graph neural network model when there are interfering substances. Based on the graph neural network model, it obtains the true value of particle concentration and sends the true value of particle concentration to the data analysis module. The data analysis module locates the pollution source device through Bayesian inference. The data storage module stores historical monitoring data from the cleanroom.
8. The semiconductor manufacturing equipment particle contamination tracing system according to claim 7, characterized in that, The graph neural network module further includes: a model generation unit, a similarity calculation unit, and a ground truth generation unit; the model generation unit is used to construct a first graph neural network model and a second graph neural network model; the similarity calculation unit is used to determine the similarity between feature matrices; and the ground truth generation unit is used to obtain the ground truth value of particle concentration.
9. The semiconductor manufacturing equipment particle contamination tracing system according to claim 7, characterized in that, The data analysis module further includes: a judgment unit, an analysis unit, and a classification unit; the judgment unit is used to determine the presence of interfering substances in the clean room; the analysis unit is used to analyze the particle concentration change rate data to determine whether it is necessary to conduct source tracing analysis of particle pollution; the classification unit is used to obtain the probability of the presence of interfering substances based on the particle concentration change rate.
10. The semiconductor manufacturing equipment particle contamination tracing system according to claim 8, characterized in that, The real value generation unit acquires historical monitoring data of the cleanroom when there are interfering substances, obtains the features of the equipment nodes from the monitoring data, inputs the features of the equipment nodes into the first graph neural network model, uses the features of the particle counter nodes as the prediction target, and uses time series prediction to obtain the features of the particle counter nodes; obtains the predicted particle concentration value of the particle counter based on the features of the particle counter nodes, and obtains the error curve based on the predicted particle concentration value and the observed particle concentration value of historical monitoring data. The system acquires monitoring data from the current cleanroom. If there are no interfering substances, the true value of the current particle concentration is obtained through the first neural network model. If there are interfering substances, the predicted value of the current particle concentration is obtained through the second neural network model. The system calculates the similarity between the features of the current equipment node and the features of historical equipment nodes in the cleanroom when interfering substances are present. The system determines the error curve corresponding to the features of the closest historical equipment node. The system calculates the true value of the current particle concentration by performing operations on the error curve and the predicted value of the current particle concentration. The closest refers to the greatest similarity between the features of the historical equipment node and the current equipment node.