Method and device for monitoring plasma state and plasma process state

Abstract

The present invention relates to a method and apparatus for monitoring a plasma state and a plasma process state, an electronic device collecting sensing data corresponding to a preset sensing data collection condition from sensing data obtained by a plurality of sensors, the electronic device simultaneously storing and preprocessing the sensing data, the electronic device screens input data for learning and performs learning, and the electronic device predicts a plasma state and a plasma process state based on learning data derived from the learning, and may also be applied to other embodiments.

Description

Technical Field

[0001] This invention relates to a method and apparatus for monitoring plasma state and plasma process state.

[0002] This invention was derived from research conducted as part of a national science and technology research operation funding support project of the National Association for Science and Technology.

[0003] Name of the project management (professional) organization: China Association for Science and Technology

[0004] Research Project Title: Research Operation Funding Support from the National Association for Science and Technology (Main Project Funding) Research Topic Title: Development and Empirical Demonstration of Intelligent Technology for Semiconductor Plasma Process Equipment

[0005] Name of the implementing organization: Korea Fusion Energy Research Institute

[0006] Research period: January 1, 2022 to December 31, 2022 Background Technology

[0007] Generally, in semiconductor device manufacturing processes, plasma equipment that utilizes plasma generated by high-frequency electricity to perform processes such as etching and deposition on semiconductor substrates is widely used. Various sensors are installed on this plasma equipment, and the operation and status of the plasma chamber or auxiliary components are confirmed based on the sensor data obtained from the sensors.

[0008] Currently, by analyzing sensor data acquired from sensors attached to the plasma chamber, the system determines the start and end points of the plasma process, identifies functional anomalies such as whether plasma is generated, determines the end point of wafer etching, and checks whether process byproducts generated inside the plasma chamber have been removed. However, the accuracy of these methods is very low. Therefore, situations frequently occur where the plasma process stops midway, is judged as abnormal during normal operation, or is judged as normal during abnormal operation, resulting in decreased productivity.

[0009] Therefore, when it is necessary to change the method of analyzing sensor data to improve the accuracy of monitoring plasma state or plasma process state, the resulting problems include the need to re-execute the sensor data analysis over a long period of time and to re-execute the verification process of the analysis results. Furthermore, it leads to the problem of needing to shut down the system to apply the analysis results. Summary of the Invention

[0010] Technical issues

[0011] To address the problems of the prior art, embodiments of the present invention provide a method and apparatus for monitoring plasma state and plasma process state, which trains the plasma state and plasma process state by inputting sensing data collected from multiple sensors attached to the plasma chamber into machine learning or deep learning, and predicts the plasma state and plasma process state based on the training results.

[0012] Furthermore, embodiments of the present invention provide a method and apparatus for monitoring plasma state and plasma process state, which compares the actual collected sensor data with the predicted data based on training results during the plasma process, thereby improving the accuracy of the predicted data.

[0013] Technical solution

[0014] The method for monitoring plasma state and plasma process state according to an embodiment of the present invention is characterized by comprising: an electronic device collecting sensor data corresponding to preset sensor data collection conditions from sensor data obtained from multiple sensors; the electronic device simultaneously performing storage and preprocessing of the sensor data; the electronic device filtering input data for learning and performing learning; and the electronic device predicting plasma state and plasma process state based on the learning data obtained through the learning.

[0015] Furthermore, the feature is that after predicting the plasma state and the plasma process state, the method further includes: the electronic device performing verification on the prediction result, and the electronic device confirming, based on the verification result, whether to stop the plasma process or whether to relearn the sensing data.

[0016] Furthermore, it is characterized in that the electronic device displays at least one of the prediction result and the verification result.

[0017] Furthermore, the feature is that the collection of sensor data includes: confirming the collection conditions, the collection conditions including the name, unique address, number of times sensor data collection failed, and message name set in the sensor data for at least one of the plurality of sensors.

[0018] Furthermore, the feature is that the collection of sensing data comprises: transmitting a request signal for the sensing data to the at least one sensor at a certain period or in real time, and collecting the sensing data from the at least one sensor based on the request signal.

[0019] Furthermore, the feature is that the collection of sensor data is characterized by: if the collection of sensor data is completed, then loading the sensor data into at least one message queue corresponding to the at least one sensor; and if it is confirmed that the loading of the sensor data is completed, then confirming that the collection of sensor data is completed.

[0020] Furthermore, the simultaneous preprocessing includes: confirming preprocessing setting information corresponding to the at least one sensor, and performing the sensor data preprocessing based on the preprocessing setting information.

[0021] Furthermore, the learning process is characterized by: setting a prediction interval for learning; selecting sensor data that falls within the prediction interval from the sensor data loaded in the message queue; performing statistical processing on the selected sensor data; and filtering the input data for learning through regression analysis of the sensor data.

[0022] Furthermore, the apparatus for monitoring plasma state and plasma process state according to an embodiment of the present invention includes: a collection module for collecting sensor data corresponding to preset collection conditions from plasma process-related sensor data obtained from multiple sensors; a data storage module for receiving and storing the sensor data from the collection module; a preprocessing module for receiving the sensor data from the collection module and performing preprocessing; a learning execution module for filtering input data for learning from the preprocessed sensor data from the preprocessing module and performing learning; and a prediction module for predicting plasma state and plasma process state based on the learning data obtained through learning, wherein the collection module simultaneously transmits the sensor data to the data storage module and the preprocessing module.

[0023] Furthermore, it is characterized by including: a verification module, which verifies the prediction results of the prediction module and confirms whether the plasma process has been stopped or whether the sensing data has been relearned based on the verification results.

[0024] Furthermore, it is characterized by including: a visualization module that displays at least one of the prediction results and the verification results.

[0025] Furthermore, the feature is that the collection module confirms the collection conditions, which include the name, unique address, number of times sensor data collection failed, and message name set in the sensor data for at least one of the plurality of sensors.

[0026] Furthermore, the feature is that the collection module transmits a request signal for the sensing data to the at least one sensor at a certain period or in real time, and collects the sensing data from the at least one sensor based on the request signal.

[0027] Furthermore, the feature is that, once the collection of the sensing data is complete, the collection module loads the sensing data into at least one message queue corresponding to the at least one sensor.

[0028] Furthermore, the preprocessing module confirms preprocessing setting information corresponding to the at least one sensor and performs preprocessing of the sensing data based on the preprocessing setting information.

[0029] Furthermore, the learning execution module is characterized in that it sets a prediction interval for learning, selects sensor data that falls within the prediction interval from the sensor data loaded in the message queue, and filters the input data for learning by performing statistical processing and regression analysis.

[0030] The effects of the invention

[0031] As described above, the method and apparatus for monitoring plasma state and plasma process state according to the present invention input sensing data collected from multiple sensors attached to the plasma chamber into machine learning or deep learning to perform learning, and predict the plasma state and plasma process state based on the learning results, thereby having the effect of monitoring plasma state and plasma process state.

[0032] Furthermore, the method and apparatus for monitoring plasma state and plasma process state according to the present invention compare the actual sensor data collected during the plasma process with the predicted data based on the learning results, re-filter the input data used for learning or perform relearning, thereby improving the accuracy of the predicted data and thus improving the production efficiency of the plasma process. Attached Figure Description

[0033] Figure 1 A diagram illustrating a system for monitoring plasma state and plasma process state according to an embodiment of the present invention.

[0034] Figure 2 A flowchart illustrating a method for monitoring plasma state and plasma process state according to an embodiment of the present invention is provided.

[0035] Figure 3 A detailed flowchart illustrating a method for collecting sensor data according to an embodiment of the present invention is provided.

[0036] Figure 4 A detailed flowchart illustrating a method for filtering input data according to an embodiment of the present invention is provided.

[0037] Figure 5 A diagram illustrating the learning data table after learning is completed according to an embodiment of the present invention.

[0038] Figure 6 A diagram illustrating the prediction data table after the prediction is completed according to an embodiment of the present invention.

[0039] Figures 7a to 7fThis is an example diagram illustrating the visualization of operational results for monitoring plasma state and plasma process state according to an embodiment of the present invention. Detailed Implementation

[0040] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. The detailed description disclosed below, taken in conjunction with the drawings, is for illustrative purposes only and is not intended to represent the only possible embodiments in which the invention can be implemented. In the drawings, parts irrelevant to the description may be omitted for clarity, and the same reference numerals may be used for the same or similar constituent elements throughout the specification.

[0041] Figure 1 A diagram illustrating a system for monitoring plasma state and plasma process state according to an embodiment of the present invention.

[0042] Reference Figure 1 The system 10 according to the present invention may include a plurality of sensors 100 and electronic devices 200.

[0043] The multiple sensors 100 may include sensors such as OES sensors, VI sensors, and MFC sensors. The multiple sensors 100 acquire sensor data such as light-related spectral data and electrical characteristic data generated during the plasma process and provide this data to the collection module 210. Therefore, each sensor included in the multiple sensors 100 can communicate with the collection module 210 in a 1:1 manner, or be connected and communicate in an N:1 manner via LAN.

[0044] The electronic device 200 collects sensor data corresponding to preset sensor data collection conditions from sensor data obtained from multiple sensors 100, and simultaneously performs sensor data storage and preprocessing. To this end, the electronic device 200 may include a collection module 210, a preprocessing module 220, a prediction module 230, a data storage module 240, a learning execution module 250, a learning storage module 260, a verification module 270, a visualization module 280, and a connection module 290.

[0045] The collection module 210 collects sensor data in real time by communicating with multiple sensors 100 attached to the plasma chamber. For this purpose, the collection module 210 can communicate 1:1 with the multiple sensors 100 or 1:N via LAN. The collection module 210 performs sensor data collection according to preset collection conditions. More specifically, the collection conditions may include the names of the multiple sensors 100, unique addresses used to access the sensors (e.g., IP addresses), the number of sensor data collection failures and loading failures, and the message names set in the sensor data.

[0046] The collection module 210 periodically or in real-time transmits request signals for sensing data to each sensor 100 based on collection conditions, and collects sensing data obtained from each sensor 100 according to the request signals. If the number of failed sensing data collection attempts by the collection module 210 reaches or exceeds the number of failures set in the collection conditions, it can confirm that a sensor error has occurred. Furthermore, if sensing data collection is confirmed to be complete, the collection module 210 loads the sensing data into a message queue. For example, the collection module 210 can load sensing data collected from the OES sensor into a message queue named M_OES, sensing data collected from the VI sensor into a message queue named M_VI, and sensing data collected from the MFC sensor into a message queue named M_MFC. If the number of failed loading attempts in a message queue reaches or exceeds the number of failures set in the collection conditions, it can confirm that a message queue error has occurred.

[0047] When the collection module 210 confirms that the sensor data loading is complete, it can set up a message queue program such as RabbitMQ to simultaneously transmit the sensor data to the preprocessing module 220 and the data storage module 240. Furthermore, the collection module 210 can generate threads or processes to perform sensor data collection that match the number of sensors, allowing it to operate independently for each sensor.

[0048] The preprocessing module 220 performs preprocessing such as correction, change, combination and normalization on the sensor data received from the collection module 210 through the message queue, and transmits the preprocessed sensor data to the prediction module 230 and the data storage module 240 through the message queue.

[0049] More specifically, the preprocessing module 220 confirms the preprocessing settings information for sensor data preprocessing and performs sensor data preprocessing based on the preprocessing settings information. The preprocessing settings information may include items such as message name, missing value handling method, minimum threshold, filter columns, normalization processing method, and data reduction method, and other items may also be added.

[0050] Based on preprocessing settings, the preprocessing module 220 retrieves sensor data with the same name as M_OES from the sensor data loaded in the message queue and processes missing values ​​based on the preprocessing settings. For example, if the missing value processing method for sensor data with message name M_OES is set to average, the preprocessing module 220 can process the missing values ​​identified in the sensor data using the average value of sensor data corresponding to the same plasma state and plasma process state. Furthermore, if the missing value processing method for sensor data with message name M_MFC is deletion, the preprocessing module 220 can delete the identified missing values ​​in the sensor data.

[0051] The preprocessing module 220 uses the minimum threshold value set in the preprocessing settings as a reference, and replaces all sensor data below the minimum threshold value with that minimum threshold value. At this time, the minimum threshold value is set to remove outliers from the sensor data.

[0052] The preprocessing module 220 removes data that is irrelevant to the plasma state and plasma process state by filtering only the sensor data that corresponds to the filter columns (items) set in the preprocessing settings information.

[0053] Subsequently, the preprocessing module 220 performs normalization of the sensor data according to the normalization processing method set in the preprocessing settings information. For example, if the normalization processing method of the sensor data with message name M_VI is set to Min / Max (minimum / maximum) mode, then the sensor data of M_VI is normalized using the Min / Max mode.

[0054] The preprocessing module 220 performs data reduction on the sensor data based on the data reduction method set in the preprocessing settings information. Furthermore, the preprocessing module 220 transmits the preprocessed sensor data to the prediction module 230 and the data storage module 240 via a message queue.

[0055] The prediction module 230 performs predictions of plasma state and plasma process state based on the learning results stored in the learning storage module 260 and the sensing data transmitted from the preprocessing module 220, and transmits the prediction data to the data storage module 240 and the verification module 270 through a message queue.

[0056] More specifically, the prediction module 230 can select at least one learning result from the learning results based on user-defined conditions, such as learning accuracy or prediction execution time. Furthermore, the prediction module 230 inputs the sensor data transmitted from the preprocessing module 220 into the learning information to predict the plasma state and plasma process state. At this time, the prediction module 230 can select multiple learning information, such as electron density, electron temperature, and deposition thickness. For example, to predict electron density, the prediction module 230 can select the MLR_OES_1 algorithm; to predict electron temperature, it can select the MLR_VI_1 algorithm; and to predict deposition thickness, it can select the XB_OES_VI_MFC_1 algorithm. In this case, the prediction module 230 performs predictions using sensor data from OES, VI, and MFC.

[0057] If the predictions for electron density, electron temperature, and deposition thickness are completed, the prediction module 230 stores the prediction data table in the data storage module 240. At this time, the prediction data table may include items such as prediction time, whether the process has been performed, the name of the algorithm used for prediction, state value distinction, prediction result, and actual measurement value.

[0058] The data storage module 240 stores in real time the data transmitted by the collection module 210, the preprocessing module (220), the prediction module (230), and the verification module 270 via a message queue. To store various data formats and forms, the data storage module 240 can use a NoSQL system for big data storage, and because the data is stored in a state that includes time information, it can be used as a benchmark for data synchronization and visualization.

[0059] Data storage module 240 stores sensor data. Since the form and format of the sensor data collected by collection module 210 differ depending on the sensor data obtained from multiple sensors 100, a NoSQL database system (such as MongoDB) can be used to store the sensor data. In this invention, to store data in real-time without loss, the database is not directly accessed in the detailed functions; instead, a message queue is used for storage, as illustrated below. For example, OES sensor data obtained from the OES sensor, among the sensor data collected by collection module 210, can be loaded into the message queue under the name M_OES.

[0060] If no predictive data is available for predicting the plasma state and plasma process state, the learning execution module 250 retrieves preprocessed sensor data from the data storage module 240. The learning execution module 250 inputs the retrieved sensor data into the learning algorithm to accurately predict the plasma state and plasma process state, and obtains the result value through repeated learning.

[0061] More specifically, the learning execution module 250 can be set with prediction ranges for predicting plasma states and plasma process states. For example, if the application range of the semiconductor / display device is a plasma electron density of 3×0.8 cm⁻³ to 5×10⁹ cm⁻³ and a plasma electron temperature of 2 eV to 7 eV, then an electron density of 8×10⁸ cm⁻³ to 3×10⁹ cm⁻³ and an electron temperature of 3 eV to 6 eV can be set as the prediction range.

[0062] The learning execution module 250 selects sensor data that falls within a set prediction interval and performs statistical processing on the selected sensor data. The learning execution module 250 obtains statistical information such as the minimum value, maximum value, median value, average value, standard deviation, kurtosis, and skewness of each item in the selected sensor data.

[0063] The learning execution module 250 performs correlation analysis between all or selected sensor data items and the plasma state or plasma process state to be predicted. The learning execution module 250 can input all or selected sensor data items into a multiple linear regression algorithm to predict the plasma state and plasma process state. Through regression analysis, the learning execution module 250 can obtain information such as significance probability (p-value), regression coefficient (weight) values, and F-test values.

[0064] The learning execution module 250 can filter sensor data used as input data for the learning algorithm. In this case, the input data can be data filtered after its importance is determined by the basic regression analysis process, or data filtered by analyzing logs and information generated during the learning process.

[0065] The learning execution module 250 applies the selected input data to the learning algorithm and performs the learning. More specifically, the learning execution module 250 can determine the form of the plasma state and plasma process state and select a learning algorithm. For example, the learning execution module 250 can select multiple learning algorithms based on the number of result values ​​to be predicted (e.g., predicting electron density and electron temperature separately, and predicting electron density and electron temperature simultaneously) or the form of the result values ​​(e.g., normalized numerical form, normal / abnormal classification form, etc.). In this case, the learning execution module 250 can select at least one learning algorithm from supervised learning, unsupervised learning, and reinforcement learning algorithms; if multiple learning algorithms are selected, they can be combined.

[0066] The learning execution module 250 applies the input data and result values ​​to the selected algorithm and performs learning. At this time, the learning execution module 250 can adjust the number of learning iterations, sensor data loading size, number of layers, etc., to improve learning accuracy. Furthermore, the learning execution module 250 inputs test data into the learning algorithm to confirm the prediction accuracy and prediction time, and can use only the learning results that conform to the accuracy and response time settings of the electronic device 200 user. At this time, the learning execution module 250 can set one or more learning execution results that conform to the user-set accuracy and response time. Through this, as learning is repeated, the predicted values ​​are continuously recorded, allowing the selection of the learning algorithm with the highest accuracy.

[0067] The learning execution module 250 stores the learning results in the learning storage module 260. The learning results stored in the learning storage module 260 may include the prediction item, algorithm name, input data name, algorithm used for learning, number of learning iterations, number of learning layers, number of combined learning models, size of the learning algorithm, learning accuracy, and prediction execution time, etc.

[0068] The learning storage module 260 stores the result values ​​received from the learning execution module 250. At this point, the result values ​​can be stored using storage methods provided by machine learning or deep learning frameworks such as Tensorflow, PyTorch, and Keras, or they can simply store the extracted information.

[0069] After the actual plasma process is completed, the verification module 270 compares the actual measurement results obtained during the inspection process with the predicted plasma state and plasma process state data predicted by the prediction module 230. To this end, the verification module 270 verifies the pre-stored verification settings information. If the error between the predicted data and the actual measurement results exceeds the allowable error range contained in the verification settings information, and the number of times the error exceeds the allowable error count contained in the verification settings information, then it confirms whether the process is in at least one state requiring process stoppage or relearning.

[0070] The visualization module 280 displays the sensor data stored in the data storage module 240, the preprocessed sensor data, the plasma state and plasma process state predicted from the sensor data, the learning algorithm used for prediction, and the learning results of the sensor data. The visualization module 280 can select graphics, tables, or other visualization formats and can display data changes over time.

[0071] The connection module 290 performs connections with external systems, namely a central server (not shown) and multiple electronic devices (not shown) located outside the electronic device 200. Therefore, the present invention can implement a central server and multiple electronic devices as an edge computing environment. Furthermore, the connection module 290 can transmit measurement information measured in the electronic device 200 (e.g., wafer thickness information, etching results, normal / defect judgment results, etc.) and the plasma state and plasma process state predicted by the electronic device 200 to the central server or multiple electronic devices.

[0072] Figure 2 A flowchart illustrating a method for monitoring plasma state and plasma process state according to an embodiment of the present invention is provided.

[0073] Reference Figure 2 In step 201, the electronic device 200 confirms whether it has received a monitoring start signal for monitoring the plasma state and plasma process state. The monitoring start signal can be received by the user of the electronic device 200 after the plasma process has started. If the monitoring start signal is received, the electronic device 200 proceeds to step 203; otherwise, it waits for the monitoring start signal to be received.

[0074] In step 203, the collection module 210 collects sensing data obtained from the multiple sensors 100 and executes step 205. At this time, the method for collecting the sensing data will utilize the following... Figure 3 To provide a more specific explanation. Furthermore, the multiple sensors 100 may include OES sensors, VI sensors, and MFC sensors. Figure 3 A detailed flowchart illustrating a method for collecting sensor data according to an embodiment of the present invention is provided.

[0075] Reference Figure 3 In step 301, if the collection conditions of the collection module 210 are in a preset state, then step 303 is executed; otherwise, step 305 is executed. In step 305, the collection module 210 can be configured by the user of the electronic device 200 to collect the sensing data obtained from the sensor 100, and step 307 is executed. Furthermore, in step 303, the collection module 210 invokes the preset collection conditions and executes step 307.

[0076] In stage 307, the collection module 210 performs sensor data collection based on the set collection conditions. More specifically, the collection conditions may include the names of multiple sensors 100, unique addresses used to connect to the sensors (e.g., IP addresses), the number of sensor data collection failures and loading failures, and the message names set in the sensor data, as shown in Table 1 below. Furthermore, in addition to sensor names, unique addresses, number of failures, and message names, other items may be added to the collection conditions.

[0077] Table 1

[0078] Sensor name Unique address Number of failures Message name OES 192.168.0.2 3 M_OES VI 192.168.0.3 3 M_VI MFC 192.168.0.4 5 M_MFC

[0079] In step 307, the collection module 210 transmits request signals for sensing data to each sensor 100 at certain time intervals or in real time, and collects sensing data from each sensor 100 according to the request signals. In step 309, if it is confirmed that the sensing data collection is complete, the collection module 210 executes step 311; if it is confirmed that the sensing data collection is incomplete, it executes step 317. At this time, if the number of failed sensing data collections reaches or exceeds the number of failures set in the collection conditions, the collection module 210 executes step 321; if the number of failures has not been reached, it executes step 307 to continue collecting sensing data. For example, if the number of failed sensing data collections of the OES sensor reaches or exceeds 3 times, the collection module 210 executes step 321, thereby confirming that an error has occurred in the OES sensor.

[0080] Conversely, if the number of data collection failures from the OES sensor is less than 3, the collection module 210 returns to step 307 and can continue collecting sensor data. In step 311, the collection module 210 loads the sensor data into message queues. For example, the collection module 210 can load the sensor data collected from the OES sensor into a message queue named M_OES, the sensor data collected from the VI sensor into a message queue named M_VI, and the sensor data collected from the MFC sensor into a message queue named M_MFC.

[0081] In step 313, if it is confirmed that the sensor data has not been fully loaded, the collection module 210 executes step 319. In step 319, if the number of loading failures in the message queue reaches or exceeds the failure count set in the collection conditions, the collection module 210 executes step 321; if the failure count has not been reached, step 311 is executed to continue loading the sensor data. For example, if the number of loading failures of the sensor data obtained from the OES sensor reaches or exceeds 3, the collection module 210 executes step 321, thereby confirming that an error has occurred in the corresponding message queue M_OES.

[0082] In step 313, if it is confirmed that the sensor data loading is complete, the collection module 210 executes step 315, thereby confirming that the sensor data collection is complete. Thus, if the sensor data collection is confirmed to be complete, the electronic device 200 can install a message queue program such as RabbitMQ to return to... Figure 2 In step 205, the collected sensor data is simultaneously transmitted to the preprocessing module 220 and the data storage module 240.

[0083] More specifically, the preprocessing module 220 confirms the preprocessing settings information for sensor data preprocessing and performs sensor data preprocessing based on the preprocessing settings information. The preprocessing settings information is shown in Table 2 below. Furthermore, the items in the preprocessing settings information may include message name, missing value handling method, minimum threshold value, filter column, normalization processing method, and data reduction method, and other items can also be added.

[0084] At this point, the learning execution module 250, based on the statistical processing results of the sensor data, sorts the sensor data items according to their standard deviation values ​​and uses regression analysis to select 10 sensor data items with high transformation coefficients (weights) and significance probabilities (p-values) not exceeding 0.05 as the selection column (items). If the correlation value obtained from the analysis reaches or exceeds 80%, the learning execution module 250 can set the data reduction method to "PCA". This allows for the setting of preprocessing settings. Furthermore, the preprocessing settings generated in the above manner can be reset if learning fails in the learning execution module 250, the learning accuracy is low, or if errors persist in the verification module 270 requiring relearning. That is, the preprocessing settings can be reset during learning execution or verification execution.

[0085] Table 2

[0086]

[0087] The preprocessing module 220, based on preprocessing settings, retrieves sensor data with the same name as M_OES from the sensor data loaded in the message queue and processes missing values ​​based on the preprocessing settings. Specifically, since the missing value processing method for M_OES sensor data is averaging, when a missing value is confirmed, it can be processed using the average value of sensor data corresponding to the same plasma state and plasma process state. Missing values ​​confirmed in M_MFC sensor data can be deleted. The preprocessing module 220 uses the minimum threshold value set in the preprocessing settings as a benchmark, replacing all sensor data below the minimum threshold value with the minimum threshold value. Here, the minimum threshold value is set to remove outliers from the sensor data. To remove data unrelated to or irrelevant to the plasma state and plasma process state, the preprocessing module 220 only filters sensor data corresponding to the filter columns (items) set in the preprocessing settings.

[0088] Next, the preprocessing module 220 performs normalization of the sensor data according to the normalization method set in the preprocessing settings. For example, if the normalization method for sensor data with message name M_VI is Min / Max, then the Min / Max method is used to normalize the sensor data of M_VI. That is, if the values ​​of the first item in the sensor data with message name M_VI are distributed between 900 and 1000, and the values ​​of the second item are distributed between 0 and 10, then the value of the second item has a greater impact on the plasma state and plasma process state. However, during the learning process, since the change in the value of the first item is greater than the change in the value of the second item, increasing the influence of the first item during learning will result in problems such as poor learning performance and reduced accuracy. Therefore, in machine learning or deep learning, in order to ensure that the first and second items have the same influence, the above normalization is used to change the size of all sensor data item values ​​to the same size.

[0089] The preprocessing module 220 performs data reduction on the sensor data based on the data reduction method set in the preprocessing settings. Generally, using all sensor data will cause problems such as learning failing to execute properly and inaccurate predictions based on the learning results. Furthermore, the large volume of sensor data will incur overhead during learning. Therefore, to address this issue, the preprocessing module 220 performs data reduction. Then, the preprocessing module 220 transmits the preprocessed sensor data to the prediction module 230 and the data storage module 240 via a message queue.

[0090] Next, the sensor data transmitted to the data storage module 240 in step 207 is transmitted to the learning execution module 250, which filters out the input data to be used for learning. This will be achieved using the following... Figure 4 To provide a more specific explanation. Figure 4 A detailed flowchart illustrating a method for filtering input data according to an embodiment of the present invention is provided.

[0091] Reference Figure 4 In step 401, the learning execution module 250 can set a prediction range to predict the plasma state and plasma process state. For example, if the application range of the semiconductor / display device is a plasma electron density of 3×10^8 cm^-3 to 5×10^9 cm^-3 and a plasma electron temperature of 2 eV to 7 eV, then the range of electron density of 8×10^8 cm^-3 to 3×10^9 cm^-3 and electron temperature of 3 eV to 6 eV can be set as the prediction range.

[0092] In step 403, the learning execution module 250 can select sensor data included in the prediction interval set in step 401. For example, when confirming the change results of the sensor data stored in the data storage module 240, if the change in the sensor data of M_OES and M_MFC is greater than the change in the sensor data of M_VI, the learning execution module 250 can select the sensor data of M_OES and M_MFC. For example, if the plasma electron density changes between 3×10^8 cm^-3 and 5×10^9 cm^-3, the sensor data of M_OES, M_MFC, and M_VI can be confirmed. The confirmation results show that the change in the sensor data of M_OES is 1000 to 1200, the change in the sensor data of M_MFC is 10 to 15, and the change in the sensor data of M_VI is 1.5 to 1.7. Therefore, the learning execution module 250 can select the sensor data of M_OES and M_MFC.

[0093] In step 405, the learning execution module 250 performs statistical processing on the selected sensor data. More specifically, the learning execution module 250 obtains statistical information for each item of the selected sensor data, such as minimum value, maximum value, median value, average value, standard deviation, kurtosis, and skewness. At this point, the numerical values ​​of the sensor data, for example, in the case of sensor data obtained from an OES sensor that typically has 3600 wavelengths and their corresponding intensity values, can refer to the intensity value corresponding to each wavelength. Furthermore, in the case of sensor data obtained from a VI sensor, it can refer to the data for 15 harmonic components and their corresponding voltage, current, and phase items.

[0094] In step 407, the learning execution module 250 performs a correlation analysis by analyzing the correlation between all or selected sensor data items and the plasma state or plasma process state to be predicted.

[0095] In step 409, the learning execution module 250 inputs all or selected sensor data items into a multiple linear regression algorithm to predict the plasma state and plasma process state. Through regression analysis, the learning execution module 250 can obtain information such as significance probability (p-value), transformation coefficient (weight) values, and F-test values.

[0096] In step 411, the learning execution module 250 filters the sensor data to be applied to the learning algorithm as input data and returns it to... Figure 2 Step 209.

[0097] In step 209, the learning execution module 250 will Figure 4 The selected input data is applied to the learning algorithm, and learning is performed. More specifically, the learning execution module 250 can determine the state of the plasma and the plasma process state, and select a learning algorithm. For example, if the plasma state or plasma process state value is a fixed value, the learning algorithm used will differ depending on whether it is normal or a failure. Furthermore, if more than two values ​​need to be predicted simultaneously, the learning algorithm used will also be different.

[0098] Therefore, the learning execution module 250 determines the shape of the result value to be predicted and selects a machine learning or deep learning algorithm (hereinafter referred to as the learning algorithm). The learning execution module 250 can select at least one learning algorithm from algorithms related to supervised learning, unsupervised learning, and reinforcement learning, and can combine multiple learning algorithms if multiple learning algorithms are selected.

[0099] The learning execution module 250 applies the input data and result values ​​to the selected algorithm and performs learning. At this time, the learning execution module 250 can adjust the number of learning iterations, the size of the sensor data loading, and the number of layers to improve learning accuracy. Furthermore, it can provide a function that allows repeated input at set intervals within a certain range to find the set value that yields the optimal result. Here, the set interval within a certain range can refer to the value set using the GridSearchCV technique to find the learning condition that most accurately predicts the learning result value.

[0100] Furthermore, the learning execution module 250 inputs test data into the learning algorithm to confirm the prediction accuracy and the time consumed in the prediction. Only learning results that meet the accuracy and response time settings set by the user of the electronic device 200 can be used. At this time, the learning execution module 250 can set one or more learning execution results that meet the accuracy and response time settings set by the user. In this way, as the prediction values ​​are continuously logged during repeated learning, the learning algorithm with the highest accuracy can be selected.

[0101] like Figure 5 As shown, the learning execution module 250 can store the learning results to the learning storage module 260. Figure 5 A diagram illustrating the learning data table after learning is completed according to an embodiment of the present invention. (Refer to...) Figure 5 The learning results stored in the learning storage module 260 may include the prediction item, algorithm name, input data name, algorithm used for learning, number of learning iterations, number of learning layers, number of combined learning models, learning algorithm capacity, learning accuracy, and prediction execution time.

[0102] In step 211, the prediction module 230 performs a prediction based on the learning results stored in the learning storage module 260 and the sensor data that has been preprocessed in step 205 and transmitted to the preprocessing module 220. More specifically, the prediction module 230 can select at least one learning result based on conditions set by the user in the learning results, such as learning accuracy or prediction execution time.

[0103] Furthermore, the prediction module 230 inputs the sensing data transmitted from the preprocessing module 220 into the learning information to perform predictions of the plasma state and plasma process state. At this time, the prediction module 230 can select multiple learning information parameters, such as electron density, electron temperature, and deposition thickness. For example, in... Figure 5 In the learning information shown, if the MLR_OES_1 algorithm is selected to predict electron density, the MLR_VI_1 algorithm is selected to predict electron temperature, and the XB_OES_VI_MFC_1 algorithm is selected to predict deposition thickness, then the prediction is performed using the sensor data of OES, VI, and MFC.

[0104] After predicting the electron density, electron temperature, and deposition thickness, the prediction module 230 will, as follows: Figure 6 The predicted data table shown is stored in the data storage module 240 after the prediction is completed. Figure 6 A diagram illustrating the prediction data table after the prediction is completed according to an embodiment of the present invention.

[0105] Reference Figure 6 The prediction data table can include items such as prediction time, whether the process has been carried out, the name of the algorithm used for prediction, the distinction of state values, prediction results and actual measurement values.

[0106] Subsequently, in step 213, the verification module 270 compares the actual measurement results obtained during the detection process after the actual plasma process is completed with the predicted data predicted by the prediction module 230, thereby performing verification of the predicted data. For example, as Figure 6 As shown, if the wafer inspection result for the process performed between 13:58:00 and 13:59:59 on July 1, 2022 is 2.38nm, then in Figure 6 Enter 2.38nm in the actual measured value of the project at 13:59:59 on July 1, 2022.

[0107] The verification module 270 can confirm the pre-stored verification settings information shown in Table 3 below. The verification settings information can be entered and stored by the user and can be changed by the user.

[0108] Table 3

[0109]

[0110] Verification module 270 periodically or in real-time compares the predicted data stored in data storage module 240 in step 211 with the actual measurement results. In step 215, if the comparison result shows that the error between the predicted data and the actual measurement results exceeds the allowable error range in the verification setting information, and the number of times this exceeds the allowable error count in the verification setting information, then verification module 270 confirms whether the process needs to be stopped. For example, if the error in electron density reaches or exceeds 10 times and is greater than 1 × 10⁹ cm⁻¹, then verification module 270 confirms whether the process needs to be stopped. Referring to Table 3, for electron density, since process stop is set to N, verification module 270 executes step 219. In step 219, since the relearning of electron density is set to Y, verification module 270 executes step 221 to perform relearning. Furthermore, if the allowable error in deposition thickness reaches or exceeds 5 times and is greater than 0.3 nm, then verification module 270 confirms whether the process needs to be stopped. Referring to Table 3, for deposition thickness, since its process stop is set to Y, verification module 270 executes step 217. In step 217, verification module 270 may stop the process and execute step 219. In step 219, for the deposition thickness, since whether to relearn is set to Y, verification module 270 executes step 221 to perform relearning. Conversely, in step 219, if whether to relearn is set to N, verification module 270 may execute step 223.

[0111] Next, in step 223, the visualization module 280 will perform visualization functions on the collected sensor data, prediction data, and learning results in the form of charts and graphs. This will utilize the following... Figures 7a to 7f Please provide an explanation. Figures 7a to 7f This is an example diagram illustrating the visualization of operational results for monitoring plasma state and plasma process state according to an embodiment of the present invention.

[0112] Figure 7a To show a graph of the OES sensing data obtained from the OES sensor at the current time, then Figure 7b To illustrate the time-cumulative curve of electron density data for a specific region of the plasma chamber predicted using OES sensing data, Figure 7c To show a graph of electron density in all chamber regions at the current time, Figure 7d A graph showing the difference between the electron density of all chamber regions predicted at the previous time and the electron density of all chamber regions predicted at the current time. Figure 7e To illustrate the set power value, predicted power value, and the difference between the set power value and the predicted power value within the plasma chamber, Figure 7f A graph showing the set pressure value, the predicted pressure value, and the difference between the set pressure value and the predicted pressure value within the plasma chamber.

[0113] The embodiments of the present invention disclosed in this specification and accompanying drawings are merely specific examples intended to facilitate the explanation of the technical content of the present invention and to aid in understanding it, and are not intended to limit the scope of protection of the present invention. Therefore, in addition to the embodiments disclosed in this specification, the scope of the present invention should also be understood to include all modifications or variations derived from the technical concept of the present invention.

Claims

1. A monitoring method, characterized in that, include: The electronic device collects sensor data corresponding to preset sensor data collection conditions from sensor data obtained from multiple sensors; The electronic device simultaneously performs the storage and preprocessing of the sensor data; The electronic device filters the input data for learning and performs the learning process; as well as The electronic device predicts the plasma state and plasma process state based on the learning data obtained through the learning process.

2. The monitoring method according to claim 1, characterized in that, After predicting the plasma state and plasma process state, the method further includes: The electronic device performs verification of the prediction results; and The electronic device determines whether to stop the plasma process or relearn the sensor data based on the verification results.

3. The monitoring method according to claim 2, characterized in that, Also includes: The electronic device displays at least one of the prediction result and the verification result.

4. The monitoring method according to claim 1, characterized in that, The collected sensor data includes: Confirm the collection conditions, which include the name, unique address, number of times sensor data collection failed, and message name set in the sensor data for at least one of the plurality of sensors.

5. The monitoring method according to claim 4, characterized in that, The collected sensor data is: Transmit a request signal for the sensing data to the at least one sensor at regular intervals or in real time; and The sensing data is collected from the at least one sensor based on the request signal.

6. The monitoring method according to claim 5, characterized in that, The collected sensor data includes: If the collection of the sensing data is completed, the sensing data is loaded into at least one message queue corresponding to the at least one sensor; and If it is confirmed that the loading of the sensor data is complete, then it is confirmed that the collection of the sensor data is complete.

7. The monitoring method according to claim 6, characterized in that, The simultaneous preprocessing includes: Confirm the preprocessing settings information corresponding to the at least one sensor; and The sensor data preprocessing is performed based on the preprocessing settings information.

8. The monitoring method according to claim 7, characterized in that, The execution learning includes: Set the prediction interval for learning; Select sensor data that falls within the prediction interval from the sensor data loaded in the message queue; Perform statistical processing on the selected sensor data; and The input data used for learning is selected through regression analysis of the sensor data.

9. A monitoring device, characterized in that, include: The collection module collects sensor data corresponding to preset collection conditions from multiple sensors related to plasma processes. A data storage module receives and stores the sensor data from the collection module; The preprocessing module receives the sensing data from the collection module and performs preprocessing. The learning execution module selects input data for learning from the preprocessed sensor data from the preprocessing module and performs learning. as well as The prediction module predicts the plasma state and plasma process state based on the learning data obtained through the learning process. The collection module simultaneously transmits the sensor data to the data storage module and the preprocessing module.

10. The monitoring device according to claim 9, characterized in that, Also includes: The verification module verifies the prediction results of the prediction module and confirms whether the plasma process should be stopped or whether the sensing data should be relearned based on the verification results.

11. The monitoring device according to claim 10, characterized in that, Also includes: The visualization module displays at least one of the prediction results and the verification results.

12. The monitoring device according to claim 11, characterized in that, The collection module confirms the collection conditions, which include the name, unique address, number of times sensor data collection failed, and message name set in the sensor data for at least one of the plurality of sensors.

13. The monitoring device according to claim 12, characterized in that, The collection module transmits a request signal for the sensing data to the at least one sensor at a certain period or in real time, and collects the sensing data from the at least one sensor based on the request signal.

14. The monitoring device according to claim 13, characterized in that, If the collection of the sensor data is completed, the collection module loads the sensor data into at least one message queue corresponding to the at least one sensor.

15. The monitoring device according to claim 14, characterized in that, The preprocessing module confirms the preprocessing setting information corresponding to the at least one sensor, and performs preprocessing of the sensor data based on the preprocessing setting information.

16. The monitoring device according to claim 15, characterized in that, The learning execution module sets a prediction interval for learning, selects sensor data that falls within the prediction interval from the sensor data loaded in the message queue, and filters the input data for learning by performing statistical processing and regression analysis.

Images