KR20230055609A - Real time monitoring method and system for plasma appatus - Google Patents

Abstract

The present invention relates to a monitoring method of plasma equipment which comprises the steps of: collecting state data representing a state value of a plasma process performed on a substrate by the plasma equipment and test result data representing a process result after the plasma process is completed; creating a machine learning model using, as learning data, the state data as input data and the test result data as output data; collecting state data in real time representing a state of a plasma process performed on a new substrate by the plasma equipment; and determining whether there is an abnormality in the plasma equipment according to a test result value predicted from the state data using the machine learning model. Accordingly, a state of the plasma equipment can be diagnosed in real time.

Description

Real-time monitoring method and monitoring system of plasma equipment {REAL TIME MONITORING METHOD AND SYSTEM FOR PLASMA APPATUS}

The present invention relates to a real-time monitoring method and monitoring system of a plasma facility, and more particularly, to a method for monitoring a plasma facility in real time using a facility diagnosis algorithm and a monitoring system for performing the same.

Plasma equipment such as plasma deposition equipment and plasma etching equipment may be used to manufacture semiconductor devices. In order to diagnose the plasma facility, a method of interlocking the facility in real time when the sensor value of the facility part has a singularity (FDC (False Detection and Classification) Interlock) may be proposed, but the sensor value of the facility part is determined by the plasma process. There is a problem in that a diagnosis result completely different from a process result after completion may be provided. In addition, when equipment abnormalities are determined by inspection processes such as CD (Critical Dimension) measurement and destructive inspection, diagnosis results related to process results can be obtained, but real-time diagnosis is impossible.

One object of the present invention is to provide a method for monitoring a plasma facility capable of performing real-time diagnosis associated with product yield.

Another object of the present invention is to provide a monitoring system for performing the monitoring method described above.

In the method for monitoring a plasma facility according to exemplary embodiments for achieving one object of the present invention, state data indicating a state value of a plasma process performed on a substrate by the plasma facility and after the plasma process is completed Collect test result data representing the process results of A machine learning model is created using the state data as input data and the test result data as output data as training data. State data representing the state of a plasma process performed on a new substrate by the plasma facility is collected in real time. Using the machine learning model, it is determined whether or not there is an abnormality in the plasma facility according to a test result value predicted from the state data.

Plasma equipment monitoring system according to exemplary embodiments for achieving another object of the present invention, state data representing the state value of the plasma process performed by the plasma equipment and process results after the plasma process is completed A data storage unit for storing test result data, a learning model generation unit for generating a machine learning model using the state data as input data and the test result data as output data as learning data, and a new substrate by the plasma equipment. A data collection unit for collecting state data representing the state of the plasma process performed in real time, and a test result value predicted from the current state data collected by the data collection unit using the machine learning model. and a determination unit for determining whether or not the plasma equipment is abnormal.

According to exemplary embodiments, state data representing state values of a plasma process including OES data is collected at the plasma equipment level, and algorithm-based data preprocessing specialized for outlier detection is performed on the OES data to obtain input data. can be obtained Test result data representing process results after the plasma process is completed, such as optical measurement data, yield data, and destructive inspection data, may be obtained as output data. A machine learning model is generated using the input data and the output data as learning data, and a process result is predicted from the current state data using the machine learning model, thereby diagnosing the state of the plasma facility in real time.

Accordingly, before performing a measurement process or a test process (EDS process) after the plasma process, the plasma facility is monitored in real time and the facility is interlocked according to the monitoring result, thereby preventing the yield from being lowered. .

However, the effects of the present invention are not limited to the above-mentioned effects, and may be variously extended without departing from the spirit and scope of the present invention.

1 is a block diagram illustrating a semiconductor manufacturing system according to example embodiments.
FIG. 2 is a block diagram illustrating the plasma process monitoring system of FIG. 1 .
FIG. 3 is a block diagram illustrating the plasma processing apparatus of FIG. 1 .
4 is a graph showing an example of data pre-processing performed by the data pre-processing unit of the machine learning processing unit of FIG. 2 .
5 is a graph showing a correlation between data derived by the machine learning control unit of FIG. 2 .
FIG. 6 is a graph showing a determination of whether a facility is abnormal according to a facility monitoring index predicted by the machine learning controller of FIG. 2 .
7 is a flowchart illustrating a real-time monitoring method of a plasma facility according to exemplary embodiments.

Hereinafter, with reference to the accompanying drawings, preferred embodiments of the present invention will be described in more detail.

1 is a block diagram illustrating a semiconductor manufacturing system according to example embodiments. FIG. 2 is a block diagram illustrating the plasma process monitoring system of FIG. 1 . FIG. 3 is a block diagram illustrating the plasma processing apparatus of FIG. 1 . 4 is a graph showing an example of data pre-processing performed by the data pre-processing unit of the machine learning processing unit of FIG. 2 . 5 is a graph showing a correlation between data derived by the machine learning control unit of FIG. 2 . FIG. 6 is a graph showing a determination of whether a facility is abnormal according to a facility monitoring index predicted by the machine learning controller of FIG. 2 .

Referring to FIGS. 1 to 6 , the semiconductor manufacturing system includes first to third processes for sequentially performing first, second, and third processes P1, P2, and P3 on a substrate such as a wafer W. The monitoring system 100 may include a monitoring system 100 for diagnosing and controlling abnormalities of the process facilities E1 , E2 , and E3 and the first processing facility E1 in real time.

In example embodiments, the first process equipment E1 is a plasma equipment for performing the plasma process P1 on the wafer W, and the second process equipment E2 is a plasma equipment where the plasma etching process is performed. It is equipment for performing the subsequent process (P2) on the wafer (W), and the third process equipment (E3) is to perform the plasma process and the test process (P3) for the wafer (W) on which the subsequent process has been performed. It may be a facility for

For example, the first process facility E1 may be a plasma etching facility that performs a plasma etching process for forming a through hole in a target film on the wafer W. The third process facility E3 may be a measurement facility for measuring the size of the through hole formed in the target film. Alternatively, the third process facility E3 includes an inspection facility (EDS (Electrical Die Sorting) facility) for inspecting the electrical characteristics of the semiconductor device formed on the wafer W or a destructive inspection facility for performing a destructive test. can do.

Examples of the first process facility E1 include a plasma dry etcher, a plasma dry cleaner, a plasma enhanced chemical vapor deposition, a sputtering device, and the like. .

As shown in FIG. 3 , the first process facility E1 may include at least one plasma processing device 10 . The plasma processing apparatus 10 may include a chamber 20 , a substrate support unit 30 having a lower electrode, an upper electrode 40 , a first power supply unit 50 and a second power supply unit 60 . In addition, the plasma processing apparatus 10 may further include a gas supply unit 70 and an exhaust unit 24 .

In example embodiments, the plasma processing apparatus 10 may be a device for etching an etch target film on a substrate such as a semiconductor wafer W disposed in a capacitively coupled plasma (CCP) chamber. However, the plasma generated by the plasma processing apparatus is not limited to capacitive coupled plasma, and may be, for example, inductively coupled plasma or microwave plasma. In addition, the plasma processing device is not necessarily limited to an etching device, and may be used as, for example, a deposition device, a cleaning device, and the like. Here, the substrate may include a semiconductor substrate, a glass substrate, and the like.

A substrate support 30 for supporting the substrate may be disposed inside the chamber 20 . For example, the substrate support 30 may serve as a susceptor for supporting the wafer W. The substrate support 30 may include an electrostatic chuck having an electrostatic electrode thereon to hold the wafer W by electrostatic attraction.

A gate (not shown) may be installed on a sidewall of the chamber 20 to allow entry and exit of the wafer W. A wafer W may be loaded and unloaded onto the substrate support through the gate.

The exhaust unit 24 may be connected to an exhaust port installed in the lower portion of the chamber 20 through an exhaust pipe. The exhaust unit 24 includes a vacuum pump such as a turbo molecular pump to adjust the processing space inside the chamber 20 to a desired vacuum level. In addition, process by-products and residual process gases generated in the chamber 20 may be discharged through the exhaust port.

The upper electrode 40 may be disposed above the substrate support 30 to face the lower electrode in the substrate support 30 . A chamber space between the upper electrode 40 and the lower electrode may be used as a plasma generating region. The upper electrode 40 may have a surface facing the wafer W on the substrate support 30 .

The upper electrode 40 may be supported by an insulating shield member (not shown) above the chamber 22 . The upper electrode 40 may be provided as a part of a shower head for supplying gas into the chamber 20 . The upper electrode 40 may include a circular electrode plate. The upper electrode 40 may have a plurality of injection holes through which gas is supplied into the chamber 20 .

The gas supply unit 70 is a gas supply element, and may include a gas supply source, a flow controller, and a gas supply pipe. The gas supply pipe may be connected to a gas diffusion chamber in the shower assembly, and the flow controller may control a supply flow rate of gas introduced into the chamber 20 through the gas supply pipe. For example, the gas supply source may include a plurality of gas tanks, and the flow controller may include a plurality of mass flow controllers (MFCs) respectively corresponding to the gas tanks. The mass flow controllers may independently control supply flow rates of the gases.

The first power supply unit 50 and the second power supply unit 60 may apply high-frequency power to the lower electrode and the upper electrode 40 to form plasma in the chamber 20 .

The second power supply 60 may include an upper RF power source 62 and an upper RF matching unit 64 . The upper RF power source 62 may generate a radio frequency (RF) signal. The upper RF matcher 64 may match the impedance of the RF signal generated from the upper RF power supply 62 to control plasma to be generated.

The first power supply 50 may include a lower RF power source 52 and a lower RF matching unit 54 . The lower RF power supply 52 may generate a radio frequency (RF) signal. The lower RF matcher 54 may match the impedance of the bias RF. The lower RF power source 52 and the upper RF power source 62 may be synchronized or asynchronous with each other through the tuner of the system controller 90 .

The system controller 90 may be connected to the first power supply unit 50 and the second power supply unit 60 to control their operations. The system controller may include a microcomputer and various interfaces, and may control the operation of the plasma processing device according to program and recipe information stored in an external memory or an internal memory.

In example embodiments, the plasma processing apparatus 10 may include a temperature controller for adjusting the temperature of the substrate support 30 . The temperature controller may include a heater and/or a cooler.

For example, the temperature controller may include a heater 32 disposed inside the electrostatic chuck to adjust the temperature of the electrostatic chuck and a heater power supply 34 to supply electric power to the heater 32 .

Further, the temperature controller may include a coolant assembly 38 for regulating coolant flow (pressure and flow rate) through the substrate channels as the cooler for cooling the substrate support 30 . For example, coolant assembly 38 may include a coolant pump and reservoir. Coolant assembly 38 is operable by system controller 90 to selectively flow coolant through the channels. The system controller 90 may control the flow rate and temperature of the coolant.

In addition, the plasma processing apparatus 10 includes a chamber temperature sensor for detecting the temperature inside the chamber 20, a chamber pressure sensor for detecting the pressure inside the chamber 20, and a temperature sensor for detecting the temperature of the substrate support 30. An electrostatic chuck temperature sensor 36, temperature sensors for detecting the temperature of the gas and the coolant, a voltage sensor for detecting the supplied RF voltage, a pressure sensor for detecting the pressure of the gas and the coolant, and the like can do. Sensor detection signals detected from the sensors may be provided to the system controller 90 through interfaces. As will be described later, the system controller 90 may output the sensor detection data to the first data collection unit 112 of the input data collection unit 110 as facility parts monitoring data.

Also, the plasma processing apparatus 10 may include an optical detector 80 for analyzing the plasma generated in the chamber 20 . For example, the photodetector 80 may include an optical emission spectrometer (OES).

An optical window 22 for optical access to the inside of the chamber 20 may be provided on one side wall of the chamber 20 . The plasma may emit light having a unique wavelength according to a plasma gas or an etching gas and a type of a thin film to be etched that reacts thereto. Light emitted from the plasma may be emitted to the outside through the optical window 22 .

The emission spectrometer may be installed adjacent to the optical window 22 of the chamber 20 . The emission spectrometer may monitor light generated from the plasma P transmitted through the optical window 22 of the chamber 20 . The light incident through the optical window 22 is transmitted to the emission spectrometer through an optical fiber, and the emission spectrometer splits the light according to the wavelength, and then measures the intensity for each wavelength with a CCD sensor. It can be confirmed and the intensity can be amplified through a photo multiplier tube (PMT).

Accordingly, the photodetector 80 may detect the intensity of light for each wavelength by splitting the light from the plasma generated in the chamber 20 . Signals (OES data) detected from the photodetector 80 may be provided to the system controller 90 through an interface. As will be described later, the system controller 90 may output the OES data as plasma monitoring data to the second data collection unit 114 of the input data collection unit 110 . The OES data may be data representing emission intensity of each wavelength measured at each time. The OES data may include intensity data according to a wavelength range of 200 nm to 850 nm for each process time.

In example embodiments, the monitoring system 100 collects, in real time, state data representing state values of the plasma processing process performed by the plasma processing apparatus 10 of the first process facility E1, and performs machine learning. It is possible to predict a test result value indicating a product yield after the plasma processing process is completed from the state data using a model, and determine whether or not the plasma processing apparatus 10 has an abnormality based on the prediction.

If there is an abnormality in the plasma processing apparatus 10 according to the predicted test result value, the monitoring system 100 outputs an interlock control signal to the system controller 90 of the plasma processing apparatus 10, and the system controller ( 90) may control the plasma processing apparatus 10 to be interlocked according to the interlock control signal.

As shown in FIG. 2 , the monitoring system 100 may include a learning data collection unit and a machine learning processing unit 200 . The learning data collection unit may include an input data collection unit 110 and an output data collection unit 150 . The machine learning processor 200 may include a data preprocessor 210 and a machine learning control unit 220 . In addition, the monitoring system 100 may further include a display unit 300 .

The input data collection unit 110 may collect state data representing state values of a plasma treatment process performed by the plasma processing apparatus 10 . The state data may include at least one of equipment part monitoring data indicating a state of parts of the plasma processing device 10 and plasma monitoring data indicating a state of plasma generated in the plasma processing device 10 . The input data collector 110 may include a first data collector 112 and a second data collector 114 .

The first data collection unit 112 may collect facility part monitoring data indicating the state of parts of the plasma processing apparatus 10 . Sensor detection data detected from component sensors installed in the plasma processing device 10 are provided to the system controller 90, and the first data collection unit 112 receives the sensor as the facility component monitoring data from the system controller 90. Detection data may be received. Examples of the equipment parts monitoring data include RF (Radio Frequency) power, RF frequency, electrostatic chuck temperature, coolant flow rate, gas flow rate, backside gas (He) pressure of the electrostatic chuck, and upper electrode temperature.

The second data collector 114 may collect plasma monitoring data indicating a state of plasma generated in the chamber 20 of the plasma processing apparatus 10 . The OES data detected from the photodetector 80 installed in the plasma processing device 10 is provided to the system controller 90, and the second data collector 112 receives the OES as the plasma monitoring data from the system controller 90. data can be received.

In example embodiments, the input data collector 110 may further include a third data collector 116 . The third data collection unit 116 may collect virtual metrology data. The virtual measurement data may refer to processing data defined by a manager or defined and generated by a specific processing algorithm to easily determine an abnormality of a plasma facility, rather than data directly obtained from the sensors.

For example, a specific point in time at which the wavelength of a specific gas (eg, CN gas) detected during a plasma etching process for channel hole formation in a V-NAND product decreases or is not detected is set as the virtual measurement data you can collect this. Examples of the virtual measurement data include an etch rate prediction value for each wafer film quality, a gas flow rate difference value for each process step, and the like. As will be described later, the machine learning processing unit 200 may discover or update the virtual measurement data having a correlation with process results (product yield) through a specific processing algorithm.

The output data collection unit 150 may collect test result data indicating a process result (product yield) after the plasma processing process (or a subsequent process thereafter) performed by the plasma processing apparatus 10 is completed. The test result data may include at least one of measurement data measured by a measurement facility, test data measured by an electrical inspection facility (EDS facility), and destructive test data measured by a destructive test facility. The output data collector 150 may include a fourth data collector 152 , a fifth data collector 154 and a sixth data collector 156 .

The fourth data collection unit 152 may collect measurement data measured by measurement facilities. For example, the fourth data collection unit 152 may receive OCD (Optical Critical Dimension) data indicating the size of a through hole formed in the target film on the wafer W from the measurement equipment.

The fifth data collection unit 154 may collect test data measured by an electrical test facility. For example, the fifth data collection unit 154 may receive EDS data of the semiconductor chip formed on the wafer W from the electrical inspection facility.

The sixth data collection unit 156 may collect destructive test data measured by a destructive test facility. For example, the sixth data collection unit 154 may receive destructive inspection data of the wafer W from the destructive inspection facility.

The data preprocessing unit 210 of the machine learning processing unit 200 may perform data preprocessing to convert the collected data into data suitable for a machine learning algorithm. The data pre-processing unit 210 may provide one data value (representative statistical value) by statistically processing time-series data obtained while one wafer is processed in the plasma processing apparatus 10 . Time-series data for one wafer is pre-processed into one statistical value, and the statistical values collected for a predetermined period of time may indicate a deterioration tendency of the plasma processing apparatus 10 moving over time. The data preprocessor 210 may include a first preprocessor 212 and a second preprocessor 214 .

The first pre-processing unit 212 may statistically process the equipment parts monitoring data collected from the first data collection unit 112 to provide one statistical data value. For example, the statistical value may include an average value, a median value, a minimum value, a maximum value, a sum, a standard deviation, and the like.

The second preprocessor 214 may perform preprocessing on the plasma state data collected from the second data collector 114 . Since the OES data includes intensity data according to a wavelength range of 200 nm to 850 nm for each process time over time, the second pre-processor 214 performs interpolation and peak detection for use in the machine learning algorithm. ), shift calibration, normalization, and additional normalization. In addition, errors between equipment can be reduced in the same plasma process through preprocessing of equipment matching (TTTM).

Specifically, as shown in FIG. 4 , the collected OES data may be interpolated and preprocessed at regular wavelength intervals. When the wavelength interval of the OES data is not constant, linear correction may be performed at regular intervals to compensate for this.

Subsequently, peak detection preprocessing may be performed in order to use only wavelengths having peak intensity values over time as data. In this case, only wavelengths having a peak value at a specific time point may be extracted by applying a peak detection algorithm.

Thereafter, the peak value of the OES wavelength may be shifted due to an interpolation effect or a preventive maintenance (PM) effect of the plasma chamber, and shift calibration preprocessing may be performed to correct this. For example, in the case of a plasma etching process for forming channel holes in V-NAND products, the oxygen plasma peak of the in-situ dry clean process can be used for shift calibration. there is. After fitting with an N-order polynomial function how much the oxygen peaks are shifted in the ISD process using the already known peak of the oxygen component, the function is interpolated to a wavelength of 200 nm to 850 nm to correct the wavelength. can

Subsequently, the intensity value in the OES data tends to decrease due to contamination of the optical window as the radio frequency (RF) time of the chamber elapses. To correct for this effect, the signal intensity is divided by the ambient wavelength intensity to normalize ) preprocessing can be performed.

In addition, a signal of a specific wavelength may be additionally normalized by utilizing the known characteristics of the process. For example, in the case of analyzing the CN wavelength (387 nm), in a process where NF ₃ gas is not supplied, the minimum intensity value of the CN wavelength (387 nm) is subtracted from all process steps to obtain the effect of removing the base value of CN gas. can

The machine learning control unit 220 may generate a machine learning model by using the preprocessed data as training data and predict a test result value from current state data collected in real time using the machine learning model. The machine learning control unit 220 may include a data storage unit 222 , a learning model generation unit 224 and a determination unit 226 .

The data storage unit 222 stores the data preprocessed by the data pre-processing unit 210 and the virtual measurement data provided from the third data collection unit 116 as input data, and outputs data from the data collection unit 150. Provided test result data can be stored as output data.

The learning model generation unit 224 may generate a machine learning model by using the input data and the output data as training data. The learning model generation unit 224 may use a machine learning algorithm having the highest consistency between the input data and the output data. Examples of the machine learning algorithm include Multi-Variable Linear Regression, Lasso, Decision Tree, Random Forest, and Gradient Boosting.

A linear regression model may be an algorithm that predicts a result from a linear correlation between data, and a multi-variable linear regression model may be an algorithm that predicts a result from a plurality of features. Lasso is an algorithm that can reduce the number of features in data by minimizing the size of coefficients in a linear regression model. The decision tree algorithm is an algorithm that classifies and predicts by tabulating decision-making rules, and the random forest is an algorithm that reduces the instability of the prediction model by adding randomness to the decision tree. Gradient Boosting is an algorithm that learns by adjusting sample weights in classifiers, and can update sample weights through gradient descent.

The learning model generation unit 224 may use an evaluation index capable of evaluating the performance of the generated machine learning model. Examples of the evaluation index include R-squared, precision, and F-score. The R square is a coefficient of determination and may represent the goodness of fit of the regression model. As an index used when the value to be predicted is continuous, the R squared value is between 0 and 1, and the closer to 1, the closer the machine learning model is to predicting the result value. The precision may be defined as a ratio of what is actually 'True' among what the model predicts to be 'True' in the classification model. The precision value is between 0 and 1, and the closer to 1, the higher the precision of the model. The F-score represents the harmonic average of precision and recall (ratio of those predicted by the model to be 'True' among those that are actually 'True') in the classification model, and considers both precision and recall to determine the accuracy of the model. can indicate The closer the value is to 1, the higher the accuracy of the model.

The learning model generating unit 224 checks these evaluation indicators and selects a model capable of predicting the plasma process result with the highest consistency, or may combine a plurality of learning algorithms through an ensemble learning method.

The learning model generating unit 224 may derive a facility monitoring index through the generated machine learning model. The facility monitoring index may be a result value (output value) predicted by the machine learning model (OCD data, EDS data, destructive inspection data). The excavated facility monitoring index may be automatically collected through real-time data processing after generating a facility log. The index may be computerized (virtualized) through a Fab backbone network and collected as the virtual measurement data through the third data collection unit 116 .

As shown in FIG. 5, it is possible to predict VCD (Vertical CD), a major process parameter representing quality, through the facility monitoring index derived only from state data representing the state value of the plasma process performed by the plasma processing facility. can know Here, the model consistency is 0.72 for all chambers, the X axis is the equipment monitoring index (Index) derived by the machine learning model, and the Y axis is the CD data (vertical CD) of the word line in the V-NAND channel hole. .

The determination unit 226 predicts a test result value representing a process result (product yield) from the current state data collected by the data collection unit 110 using the machine learning model, and according to the predicted test result value It is possible to determine whether or not there is an abnormality in the plasma processing device 10 .

As shown in FIG. 6, the determination unit 226 derives a specification by linking the facility monitoring index and process test results, and when the facility monitoring index derived in real time is out of the tolerance range (A In case), it is possible to prevent the yield from deteriorating by applying the facility interlock.

For example, when the OCD value predicted using the machine learning model is in the defective area (too large or too small), it may be determined that the probability that the corresponding wafer is defective is high. Since the defective area of the OCD value can be known in advance through the correlation analysis between the actual OCD value and the actual EDS data, it can be determined that the equipment is abnormal when the predicted OCD value is in the defective area.

When it is determined that the plasma processing apparatus 10 has an abnormality according to the predicted test result value, the determination unit 226 outputs an interlock control signal to the system controller 90 of the plasma processing apparatus 10, and the system controller 90 may control the plasma processing apparatus 10 to be interlocked according to the interlock control signal.

The display unit 300 may include a display device for displaying the test result value determined by the determination unit 226 to a user. The display unit 300 may provide information about the state data and the preprocessed data used in the machine learning and information about the consistency of the machine learning model to the user.

As described above, the monitoring system 100 of the plasma facility provides state data indicating the state value of the plasma process performed by the plasma facility 12 and test result data indicating the process result (product yield) according to the plasma process. collected, and a machine learning model may be generated using the state data as input data and the test result data as output data as learning data.

The monitoring system 100 receives state data in real time when a plasma process is performed by the plasma facility 12, and diagnoses the state of the plasma facility from the current state data in real time using the machine learning model and processes the process. An index that can predict the result can be derived. When the derived index is out of the permissible range, it is determined that there is an abnormality in the plasma facility, and the yield reduction may be prevented by interlocking the facility.

Hereinafter, a method of monitoring a plasma facility in real time using the above-described monitoring system will be described.

7 is a flowchart illustrating a real-time monitoring method of a plasma facility according to exemplary embodiments.

Referring to FIGS. 1 to 7 , first, a semiconductor manufacturing process may be performed on a wafer W, and a machine learning model may be generated using data collected during the semiconductor manufacturing process as learning data.

In example embodiments, first, second, and third processes P1, P2, and P3 are sequentially performed on a substrate such as a wafer W to form a structure or semiconductor device ( semiconductor chip). The first process (P1) is a plasma treatment process, the second process (P2) is a subsequent manufacturing process performed after the plasma treatment process, and the third process (P3) is the structure or the semiconductor device (semiconductor chip). It may be an inspection process for inspecting defects.

For example, the first process P1 may be a plasma etching process for forming through holes in the target film on the wafer W. The third process P3 may be a measurement process for measuring the size of the through hole formed in the target film. Alternatively, the third process P3 may be an inspection process or a destructive inspection process for inspecting the electrical characteristics of the semiconductor device formed on the wafer W.

Next, a test showing the state data collected from the plasma processing apparatus 10 performing the plasma processing process (P1) and the process result (product yield) after the plasma processing process (or subsequent manufacturing process (P2)) is completed. Result data may be collected, and a machine learning model may be generated using the state data and the test result data as learning data.

In exemplary embodiments, the state data is selected from among facility parts monitoring data indicating the state of parts of the plasma processing device 10 and plasma monitoring data indicating the state of plasma generated in the chamber 20 of the plasma processing device 10. At least one may be included. In addition, the state data may further include virtual measurement data having a correlation with the process result.

The facility part monitoring data may be collected from part sensors installed in the plasma processing device 10 . Examples of the equipment parts monitoring data include RF (Radio Frequency) power, RF frequency, electrostatic chuck temperature, coolant flow rate, gas flow rate, backside gas (He) pressure of the electrostatic chuck, and upper electrode temperature.

The plasma monitoring data may be collected from an optical detector 80 installed in the plasma processing device 10 . For example, the photodetector 80 may include an optical emission spectrometer (OES). The plasma monitoring data may include OES data. The OES data may be data representing emission intensity of each wavelength measured at each time.

The virtual measurement data may include processing data generated by a manager or generated by a specific processing algorithm to easily determine an abnormality of a plasma facility.

In exemplary embodiments, the test result data is at least one of measurement data measured by a measurement facility, test data measured by an electrical inspection facility (EDS facility), and destructive test data measured by a destructive test facility. can include

For example, the measurement data may include Optical Critical Dimension (OCD) data representing dimensions of a structure formed on a target film on the wafer W. The test data may include EDS data of a semiconductor chip formed on the wafer (W).

Subsequently, data preprocessing may be performed to convert the collected data into data suitable for a machine learning algorithm.

In example embodiments, the first preprocessor 212 may statistically process the facility part monitoring data collected from the first data collector 112 to provide one statistical data value. For example, the preprocessed statistical value may include an average value, a median value, a minimum value, a maximum value, a sum, a standard deviation, and the like.

The second preprocessor 214 may perform preprocessing on the plasma state data collected from the second data collector 114 . Since the OES data includes intensity data according to a wavelength range of 200 nm to 850 nm for each process time over time, the second pre-processor 214 performs interpolation and peak detection for use in the machine learning algorithm. ), shift calibration, normalization, and additional normalization. In addition, errors between equipment can be reduced in the same plasma process through preprocessing of equipment matching (TTTM).

Subsequently, the learning model generating unit 224 may generate a machine learning model by using the input data and the output data as training data. The learning model generation unit 224 may use a machine learning algorithm having the highest consistency between the input data and the output data. Examples of the machine learning algorithm include Multi-Variable Linear Regression, Lasso, Decision Tree, Random Forest, and Gradient Boosting.

As shown in FIG. 7, state data indicating the state of a plasma treatment process performed on a new wafer by the plasma processing apparatus 10 is collected in real time (S10), and the collected state data is preprocessed (S20). ), predicting a test result representing a process result (product yield) from the preprocessed state data using the machine learning model (S30), and according to the predicted test result, an abnormality in the plasma processing device 10 If it is determined that there is, the plasma processing device 10 may be interlocked (S40).

In example embodiments, the determination unit 226 of the machine learning control unit 220 may, when a facility monitoring index derived from the status data collected in real time using the machine learning model is out of a permissible range, the facility interface It is possible to prevent a decrease in yield by applying a lock.

For example, when the OCD value predicted using the machine learning model is in the defective area (too large or too small), it may be determined that the probability that the corresponding wafer is defective is high. Since the defective area of the OCD value can be known in advance through the correlation analysis between the actual OCD value and the actual ESD data, it can be determined that the equipment is abnormal when the predicted OCD value is in the defective area.

When it is determined that the plasma processing apparatus 10 has an abnormality according to the predicted test result value, the determination unit 226 outputs an interlock control signal to the system controller 90 of the plasma processing apparatus 10, and the system controller 90 may control the plasma processing apparatus 10 to be interlocked according to the interlock control signal.

The semiconductor manufacturing process described above may be used to manufacture a semiconductor package including a logic device or a memory device. The semiconductor package may be, for example, a logic device such as a central processing unit (CPU, MPU) or an application processor (AP), such as an SRAM (SRAM) device, a DRAM (DRAM), or a high-bandwidth memory (HBM) device device. It may include a volatile memory device such as the like, and a non-volatile memory device such as a flash memory device, a PRAM device, an MRAM device, and an RRAM device.

Although the above has been described with reference to the embodiments of the present invention, those skilled in the art can variously modify and change the present invention without departing from the spirit and scope of the present invention described in the claims below. You will understand that you can.

10: plasma processing device 20: chamber
22: optical window 24: exhaust unit
30: substrate support 32: heater
36: electrostatic chuck temperature sensor 38: coolant assembly
40: upper electrode 50: first power supply
52: lower RF power supply 54: lower RF matching unit
60: second power supply unit 62: upper RF power supply
64: upper RF matcher 70: gas supply unit
80: photodetector 90: system controller
100: monitoring system 110: input data collection unit
112: first data collection unit 114: second data collection unit
116: third data collection unit 150: output data collection unit
152: fourth data collector 154: fourth data collector
156: sixth data collection unit 200: machine learning processing unit
210: data pre-processing unit 212: first pre-processing unit
214: second pre-processing unit 220: machine learning control unit
222: data storage unit 224: learning model generation unit
226: determination unit 300: display device

Claims (10)

Collecting state data indicating a state value of a plasma process performed on a substrate by a plasma facility and test result data indicating a process result after the plasma process is completed;
generate a machine learning model using the state data as input data and the test result data as output data as training data;
collect in real time status data representing the status of a plasma process performed on a new substrate by the plasma facility; and
A method of monitoring a plasma facility comprising determining whether or not there is an abnormality in the plasma facility according to a test result value predicted from the state data using the machine learning model. The method of claim 1 , wherein the state data includes at least one of equipment part monitoring data representing a state of parts of the plasma equipment and plasma monitoring data representing a state of plasma generated in the plasma equipment. The method of claim 2 , wherein the plasma monitoring data includes OES data. 3. The method of claim 2, wherein generating the machine learning model further comprises defining virtual measurement data correlated with the process result as the state data. 5. The method of claim 4, wherein collecting the state data further comprises collecting the virtual measurement data. According to claim 1,
A method of monitoring a plasma facility further comprising pre-processing data to convert the state data into data suitable for a machine learning algorithm. 7. The method of claim 6, wherein preprocessing the state data includes statistically processing the state data. The plasma facility of claim 1, wherein the test result data includes at least one of measurement data measured by a measurement facility, test data measured by an electrical inspection facility, and destructive test data measured by a destructive test facility. monitoring method. The method of claim 1, wherein determining whether or not the plasma facility has an abnormality according to the predicted test result value
derive a facility monitoring index from the current condition data; and
A method of monitoring a plasma facility comprising interlocking the plasma facility when the facility monitoring index is out of the allowable range. a data storage unit for storing state data indicating a state value of a plasma process performed by a plasma facility and test result data indicating a process result after the plasma process is completed;
a learning model generating unit generating a machine learning model using the state data as input data and the test result data as output data as learning data;
a data collection unit for collecting, in real time, state data indicating a state of a plasma process performed on a new substrate by the plasma equipment; and
A plasma facility monitoring system comprising a determination unit for determining whether or not the plasma facility is abnormal according to a test result value predicted from the current state data collected by the data collection unit using the machine learning model.

Images