KR101094598B1 - Appartus and method for mornitoring plasma chamber - Google Patents

Abstract

The present invention relates to a plasma chamber monitoring apparatus and method thereof.
The plasma chamber monitoring apparatus of the present invention collects reflected light information of plasma corresponding to a light source from a plasma chamber performing a semiconductor process, and uses intensity according to wavelengths of radicals included in the collected reflected light information. To generate an optimal neural network model. Then, it is determined whether the plasma state included in the plasma chamber is normal by using the predicted value calculated by the optimal neural network model and the measured value which is reflected light information.
According to the present invention, after collecting the intensity information according to the wavelength (Wavelength) using the OES diagnostic apparatus and modeling it using a neural network, by monitoring the chamber in real time using a model prediction value, the plasma chamber You can expect the effect to accurately identify the leak state of the in real time.

Description

Plasma chamber monitoring device and its method {APPARTUS AND METHOD FOR MORNITORING PLASMA CHAMBER}

The present invention relates to a plasma chamber monitoring apparatus and method thereof, and more particularly, to an in-situ leak of a plasma chamber using an optical emission spectroscopy and a neural network model. It is to provide a plasma chamber monitoring apparatus and method that can accurately monitor the occurrence.

As a process which can be performed from the semiconductor device using a plasma, a dry etching process, a chemical vapor deposition process, etc. are mentioned. These processes utilize a plasma chamber capable of generating plasma to proceed with the process. Therefore, the semiconductor substrate is placed on a substrate support in the plasma chamber, the inside of the plasma chamber is formed under predetermined reaction conditions, and then plasma is generated to perform an etching process and a chemical vapor deposition process.

On the other hand, abnormalities may occur in the plasma process conditions, or components constituting the plasma equipment, for example, a mass flow controller, an RF source, a bias power, or anomaly. In this case, the plasma state changes, and thus, deposition or etching characteristics are changed. In this case, the quality of the process and the reliability of the device are deteriorated. In order to prevent this, a technology for strictly monitoring the semiconductor plasma state is required.

Currently, leak monitoring of plasma chambers is mainly checked using a helium detector. In the case of the leak monitoring using such a helium detector, since the equipment is stopped and leak detection is performed after the leak occurs, the equipment productivity is greatly reduced. In addition, due to the plasma state due to the occurrence of the leak, as the process proceeds in a changed state, there is a problem in that the quality of the manufactured thin film process is degraded and thus the yield of the device is significantly reduced. In order to overcome these limitations of conventional helium detectors, in-situ diagnostic tools such as optical emission spectroscopy (OES) have been developed. However, it was difficult to detect leaks by analyzing OES data in real time during the process when measuring the leakage state by measuring and analyzing the distribution of chamber radicals using OES. This is because OES data itself is multivariate and contains huge amounts of information, so there is no technology to process it in real time. Of course, it is possible to monitor a small number of radical variables, but in this case there is a problem that the leak monitoring performance is deteriorated because it can not use other useful radical information. Therefore, it is necessary to develop a technique for detecting leaks in real time using the entire OES information.

The problem to be solved by the present invention is to accurately monitor the in-situ leakage of the plasma chamber in real time using an optical emission spectroscopy and neural network model. It is to provide a plasma chamber monitoring apparatus and method.

According to an aspect of the present invention, there is provided a plasma chamber monitoring method comprising: collecting reflected light information of plasma corresponding to a light source from a plasma chamber performing a semiconductor process; Generating an optimal neural network model using an intensity according to wavelengths of radicals included in the collected reflected light information; And determining whether the plasma state included in the plasma chamber is normal by using the predicted value calculated by the optimal neural network model and the measured value which is reflected light information.

Here, the step of determining whether the plasma state is normal,

Calculating a root mean squared error (RMS) using the measured value included in the reflected light information of the plasma and the predicted value of the optimal neural network model; Calculating a Belief value using the calculated RMSE value; Determining whether the calculated Belief value exceeds a preset error range; And determining that the error range is exceeded, informing that the plasma state of the plasma chamber is abnormal when the error range is exceeded.

Here, as a result of determining whether the error range is exceeded, if the error range is not exceeded, the plasma state of the plasma chamber is determined to be normal, and then the reflected light information of the newly loaded wafer into the plasma chamber is collected. It further comprises the step of repeating after the collecting step.

Here, the step of generating the optimal neural network model,

Classifying all radicals included in the reflected light information of the plasma into all learning or test data based on the intensity according to the wavelength; Generating a first neural network model using the classified training data; And generating the optimal neural network model by applying a predetermined learning factor to optimize the first neural network model to the first neural network model.

An apparatus for monitoring a plasma state of a plasma chamber performing a semiconductor process according to an embodiment of the present invention,

An optical spectroscopic analysis unit configured to collect reflected light information of plasma for radicals in all or specific wavelength sections in the plasma chamber; A neural network modeling construction module for generating an optimal neural network model using an intensity according to wavelengths of radicals included in the collected reflected light information; And calculating an RMSE value based on the prediction value obtained through the optimal neural network model and the measured value of the reflected light information, calculating a Belief value using the calculated RMSE value, and whether the calculated Belief value exceeds a preset error range. And a model application diagnostic module for determining whether the plasma chamber is normal.

Here, the model application diagnostic module collects reflected light information on the wafer newly introduced into the plasma chamber when the calculated Belief value does not exceed the error range.

Here, the model application diagnostic module notifies the manager that the plasma chamber is abnormal when the calculated Belief value exceeds the error range.

Here, the plasma chamber monitoring device,

The apparatus may further include a learning pattern configuration module classifying the collected reflected light information of the plasma into learning data and tester data.

Here, the neural network modeling construction model,

Generating a first neural network model using the training data, calculating an error value between the generated neural network model and the test data, and optimizing the first neural network model using the calculated error value and a learning factor Create a neural network model.

As described above, according to the present invention, after collecting the radical intensity information according to the wavelength of the plasma using the OES diagnostic apparatus and modeling it using a neural network, by monitoring the chamber in real time using a model prediction value In addition, the effect of accurately identifying the leak state of the plasma chamber in real time can be expected.

1 is a view showing a plasma chamber monitoring apparatus according to an embodiment of the present invention.
2 is a graph showing the intensity according to the wavelength of the plasma radicals measured by the plasma chamber monitoring apparatus according to an embodiment of the present invention.
3 is a graph showing a neural network model structure applied to the plasma chamber monitoring apparatus according to an embodiment of the present invention.
4 is a graph showing the learning data error of each plasma wavelength of the plasma chamber monitoring apparatus according to an embodiment of the present invention.
FIG. 5 is a graph illustrating test data errors of a plasma chamber monitoring apparatus according to an embodiment of the present invention for each wavelength.
FIG. 6 is a graph illustrating RMSE values measured by a test file of model prediction performance of a plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.
FIG. 7 is a graph illustrating Belief values measured by matching a plasma chamber monitoring apparatus according to an embodiment of the present invention with the number of test files.
8 is a flowchart illustrating a plasma chamber monitoring method of the plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is said to "include" a certain component, it means that it can further include other components, without excluding other components unless specifically stated otherwise.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention.

First, the configuration of a plasma chamber monitoring apparatus according to an embodiment of the present invention will be described.

1 is a view showing a plasma chamber monitoring apparatus according to an embodiment of the present invention.

As shown in FIG. 1, the plasma chamber monitoring apparatus 200 according to an embodiment of the present invention includes an optical emission spectroscopy 210 and a neural network 220.

The light reflection spectrometer 210 collects and measures the intensity according to the wavelength of all radicals of the plasma in the plasma chamber 100 by dispersing light (light) into a single light spectrum. The intensity information according to the wavelength is provided to the neural network unit 220. In this case, the light reflection spectrometer 210 may measure and use the intensity according to the wavelength of all radicals in a specific wavelength section or a plurality of wavelength sections arbitrarily set.

Here, the light reflection spectrometer 210 may include deposition and etching equipment using plasma to collect the intensity according to the wavelength of the radicals of the normal plasma in the plasma chamber 100.

Here, the light reflection spectrometer 210 may use a back-propagation neural network to measure the intensity according to the wavelength of all the radicals of the plasma. In addition to other back propagation neural networks, other pattern recognition systems may be used, such as fuzzy logic with learning.

On the other hand, the plasma collected by the light reflection spectrometer 210 according to the embodiment of the present invention is a plasma in a state where a failure does not occur, and means a plasma in a steady state.

The neural network unit 220 includes a learning pattern configuration module 222, a neural network modeling construction module 224, and a model application diagnostic module 226.

The training pattern configuration module 222 generates training data for creating a neural network model by using reflected light information received from the light reflection spectrometer 210, and test data for determining suitability of the trained model. ) The calculated training data and test data are provided to the neural network modeling construction module 224. Specifically, the learning pattern configuration module 222 expresses information according to the wavelength included in the reflected light information received from the light reflection spectrometer 210 as n (n is a positive integer) vectors, and this corresponds to an odd number of times. The data is divided into n / 2 odd vectors and n / 2 even vectors corresponding to even times. The divided odd vectors are used as training data, and the divided even vectors are used as test data. It can consist of different ratios of training and test data n.

The learning data and test data generation of the learning pattern configuration module 222 will be described later in detail with reference to FIG. 2.

The neural network modeling building module 224 constructs an optimal neural network model using the training data and the test data received from the learning pattern construction module 222.

In detail, the neural network modeling building module 224 generates a neural network model for the states of radicals in the plasma chamber 100 using the training data. At this time, the neural network modeling module 224 generates a neural network model using training data based on a generalized delta rule, and optimizes the neural network model generated using the test data. Here, the neural network modeling construction module 224 optimizes the neural network model by changing various learning factors involved in learning to improve the predictive performance of the test data of the generated neural network model. In this case, the learning factors include factors that can affect neural network models such as the number of neurons in the hidden layer, the distribution of neural network initial weights, the type and slope of the activation function of the neurons, and the learning allowance, and the neural network modeling construction module 224 ) Optimizes neural network models by individually adjusting learners or using learning factors optimized through genetic algorithms.

The model application diagnostic module 226 calculates an error value between the test data and the optimal neural network model by applying test data to the neural network model (hereinafter, also referred to as an "optimal neural network model") optimized by the neural network modeling construction module 224. . The error value is calculated using root mean squared error (RMS). At this time, RMSE is calculated through the following equation (1).

Where q is the total number of test data, d _j With out _j Are measured and predicted values, respectively. At this time, the measured value is actually measured value using the reflected light information, and the predicted value is a predicted value calculated through the optimized neural network model.

The model application diagnostic module 226 calculates the Belief value using the calculated RMSE value, and determines whether the plasma chamber 100 is normal according to whether the Belief value is included in the reference range. At this time, the Belief value is calculated using Equation 4 below.

In detail, the model application diagnostic module 226 determines whether the plasma chamber 100 is normal according to whether the calculated Belief value exceeds a preset error range. That is, the model application diagnostic module 226 determines that the plasma chamber 100 is abnormal (leak detection) and informs the manager when the increase rate or decrease rate exceeds the set error range (for example, 0.5). At this time, the model application diagnostic module 226 calculates the Belief value using the following equations (2), (3) and (4).

Here, Equation 2 is used when the cumulative sum (CUSUM) of the RMSE is increased, and Equation 3 is used when the cumulative average value of the RMSE is decreased. In this case, x is the RMSE value calculated from the model, μ is the average value of the RMSE values predicted from the model for the n test data determined to be normal. At this time, the number of n may vary depending on the process.

On the other hand, the model application diagnostic module 226 calculates the Belief value to determine the increase or decrease of the RMSE value over time, the Belief value is calculated through the following equation (4).

Where h / l is the value for h or l and set matches the value of μ.

The model application diagnostic module 226 determines that a leak has occurred when the Belief value exceeds a predetermined threshold (for example, 0.5) to determine whether the plasma chamber 100 is normal. In this case, the threshold value may vary depending on the process.

Such a plasma chamber monitoring apparatus 200 according to an embodiment of the present invention uses a neural network model by using an intensity according to a wavelength of a plasma, thereby using the state of plasma equipment more accurately and quickly. There is an advantage that can be grasped. In addition, there is a big advantage that it is possible to clearly know whether the equipment is normal through the state of the accurately measured plasma equipment.

FIG. 2 is a graph showing the intensity according to the wavelength of the plasma radicals of the plasma equipment measured by the plasma chamber monitoring apparatus according to an embodiment of the present invention.

As shown in FIG. 2, the plasma apparatus detecting apparatus 200 according to an embodiment of the present invention uses the reflected light reflected from the plasma of the plasma chamber 100 to generate a radical intensity in a preset wavelength range (178.2 to 887.77 nm). Sample at preset intervals (0.22 second intervals). In this case, the wavelength range may be arbitrarily set, and the sampling interval may also be adjusted according to the amount of samples to be secured. In FIG. 2, normal and abnormal mean normal and abnormal states of plasma, respectively.

3 is a diagram illustrating a neural network model structure applied to a plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As shown in FIG. 3, the plasma chamber monitoring apparatus 200 according to an embodiment of the present invention calculates a predicted value by using an auto-correlated neural time series model. The autocorrelation time series model predicts the intensity (I) information at λ + k using the intensity corresponding to the current wavelength λ and the past wavelength information λ-m.

In the plasma chamber monitoring apparatus 200 according to an embodiment of the present invention, the variables m and k determine the number of training and test data, and the activation functions used for the neurons of the output layer and the hidden layer are included in the activation function. There are a unipolar sigmoid function and a bipolar sigmoid function, and a linear function can be used for output layer neurons.

On the other hand, the composition of the training and test data is determined by (m, k), which is a basic factor of the time series model. Here, m refers to the amount of historical information used for model development, and k refers to the past time point at which the prediction occurs. The training data and test data constituted by the combination of (m, k) are different, and the developed model performance is also different. Therefore, the model performance can be evaluated and the optimized (m, k) can be determined while changing the (m, k) value to optimize the model performance.

4 is a graph illustrating a prediction error for each wavelength of a model for learning data of a plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As shown in FIG. 4, the neural network modeling construction module 224 of the plasma chamber monitoring apparatus 200 according to an embodiment of the present invention calculates an error with respect to training data through Equation 5 below.

According to an embodiment of the present invention, it can be seen that the error range of the training data is mostly distributed between -100 and 100. The error range may vary depending on the process.

FIG. 5 is a graph illustrating a prediction error for each wavelength of a test data of a plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As shown in FIG. 5, the plasma chamber monitoring apparatus 200 according to an exemplary embodiment of the present invention calculates a test data error in the same manner as the learning data error of FIG. 4 described above. It can be seen that the error range of the test data is mostly distributed between -50 and 50. The error range may vary depending on the process.

FIG. 6 is a graph illustrating RMSE values measured by a test file of model prediction performance of a plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As shown in FIG. 6, the plasma chamber monitoring apparatus 200 according to an exemplary embodiment of the present disclosure may determine a plasma that is not in a normal state through RMSE values of a total of 47 test files. That is, 1 to 42 times are included in the normal range (40 <normal <60) of the predetermined RMSE value determined to be normal, but 42 to 46 times RMSE value is determined to be out of the normal state.

In an embodiment of the present invention, in order to more accurately detect the occurrence of the leak, the RMSE value calculated from the model is transferred to the CUSUM control chart, and finally, whether the plasma chamber is normally confirmed through the Belief value obtained through the Belief function. The normality of the plasma chamber using the Belief value will be described later with reference to FIG. 7.

FIG. 7 is a graph illustrating Belief values measured by matching the number of test files of the plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As illustrated in FIG. 7, the plasma chamber monitoring apparatus 200 according to an exemplary embodiment of the present invention monitors an increase or decrease of the Belief value to determine whether the plasma equipment is normal. That is, if the Belief value increases or decreases beyond the set error range, it is determined that the plasma chamber 100 is not normal and informs the manager.

The Belief value calculated by the plasma chamber monitoring apparatus 200 according to the embodiment of the present invention using the RMSE is a Model-based CUSUM value, and the Belief value calculated using a statistical average value of the reflected light information is a CUSUM value. It can be seen from FIG. 7 that the results obtained through the Model-based CUSUM value according to the embodiment of the present invention can more clearly determine whether the plasma state is normal than the results obtained through the CUSUM.

Furthermore, since the difference can be more pronounced with the Belief value calculated using each radical, the Model-based CUSUM value according to the embodiment of the present invention can be used to quickly and accurately determine whether the plasma equipment is in a steady state. have.

Next, a plasma chamber monitoring method of the plasma chamber monitoring apparatus 200 for accurately determining whether a plasma state of the plasma chamber 100 is normal or not will be described in detail.

8 is a flowchart illustrating a plasma chamber monitoring method of the plasma chamber monitoring apparatus according to an exemplary embodiment of the present invention.

As shown in FIG. 8, the plasma chamber monitoring apparatus 200 according to an embodiment of the present invention collects reflected light information on radicals in the plasma chamber 100 (S100), and uses the collected reflected light information to provide a neural network. Create a model (S102). In this case, the plasma chamber monitoring apparatus 200 generates a neural network model using learning data among the reflected light information.

The test data is applied to the generated neural network model to calculate an error value, that is, the RMSE (S104), and the optimal neural network model is generated to minimize the RMSE while changing the learning factors (S106).

Thereafter, the plasma chamber monitoring apparatus 200 calculates a Belief value by using the RMSE value calculated for the test data (S110).

The plasma chamber monitoring apparatus 200 determines whether the calculated Belief value exceeds a preset error range (S112), and when the error range is exceeded, the plasma apparatus is abnormal (leak is detected) to the manager. Inform (S114).

On the other hand, in step S112, if the error range does not exceed, the plasma chamber monitoring apparatus 200 determines that the plasma chamber 100 is normal, and from the step S100 for the wafer newly introduced into the plasma chamber 100 Start over. S102 to S106, which are involved in model construction and optimization for real-time monitoring, may be replaced by preconfigured models. When it is necessary to reconstruct the previously configured model, the process from S102 to S106 is performed.

As described above, according to the exemplary embodiment of the present invention, the plasma chamber monitoring apparatus checks whether the plasma chamber including the leak is abnormal through the Belief value calculated using the test data through the neural network model constructed using the intensity according to the wavelength of the reflected light. You can expect a big effect that can make an accurate and quick judgment.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

100: plasma chamber, 200: plasma chamber monitoring device,
210: light reflection spectrometer, 220: neural network part,
222: learning pattern construction module, 224: neural network modeling building module,
226: model application diagnostic module,

Claims (9)

Collecting reflected light information of a plasma corresponding to a light source from a plasma chamber performing a semiconductor process;
Generating an optimal neural network model using an intensity according to wavelengths of radicals included in the collected reflected light information; and
Determining whether the plasma state included in the plasma chamber is normal by using the predicted value calculated by the optimal neural network model and the measured value which is reflected light information.
Determining whether the plasma state is normal,
Calculating a root mean squared error (RMS) using the measured value included in the reflected light information and the predicted value of the optimal neural network model;
Calculating a Belief value using the calculated RMSE value,
Determining whether the calculated Belief value exceeds a preset error range; and
If exceeding the error range, informing that the plasma state included in the plasma chamber is abnormal;
Calculating the Belief value using the calculated RMSE value,
Method for monitoring the plasma chamber by dividing the Belief value as the cumulative sum (CUSUM; Cumulative Sum) of the Belief value increases or decreases as follows:

Where h / l is the value for h or l, set matches the value of μ,
When the cumulative average value (CUSUM) of the RMSE is increased, the following S _h is used.

When the cumulative average value (CUSUM) of the RMSE is decreased, the following S _i is used.

x is the RMSE value calculated from the model, and μ is the average value of the RMSE values predicted from the model for n test data judged to be normal. delete The method of claim 1,
As a result of determining whether the error range is exceeded, if the error range is not exceeded, the plasma state of the plasma chamber is determined to be normal, and the reflected light information of the newly loaded wafer into the plasma chamber is collected and the collection is performed. Steps to repeat after step
Plasma chamber monitoring method further comprising. The method of claim 1,
Generating the optimal neural network model,
Classifying all radicals included in the reflected light information of the plasma into all learning or test data based on the intensity according to the wavelength;
Generating a first neural network model using the classified training data; And
Generating the optimal neural network model by applying a predetermined learning factor to optimize the first neural network model to the first neural network model
Plasma chamber monitoring method comprising a. An apparatus for monitoring a plasma state of a plasma chamber performing a semiconductor process,
A light reflection spectrometer for collecting reflected light information for radicals in all or specific wavelength sections in the plasma chamber;
A neural network modeling construction module for generating an optimal neural network model using an intensity according to wavelengths of radicals included in the collected reflected light information; And
The RMSE value is calculated using the prediction value obtained from the optimal neural network model and the measured value of the reflected light information, the Belief value is calculated using the calculated RMSE value, and whether the calculated Belief value exceeds a preset error range. And a model application diagnostic module for determining whether the plasma chamber is normal.
The model application diagnostic module,
As the cumulative sum (CUSUM; Cumulative Sum) of the Belief value increases or decreases, the plasma chamber monitoring apparatus for dividing the Belief value as follows:

Where h / l is the value for h or l, set matches the value of μ,
When the cumulative average value (CUSUM) of the RMSE is increased, the following S _h is used.

When the cumulative average value (CUSUM) of the RMSE is decreased, the following S _i is used.

x is the RMSE value calculated from the model, and μ is the average value of the RMSE values predicted from the model for n test data judged to be normal. The method of claim 5,
The model application diagnostic module,
And if the calculated Belief value does not exceed the error range, reflecting light information is collected for the newly loaded wafer into the plasma chamber. The method of claim 6,
The model application diagnostic module,
And informing the manager that the plasma chamber is abnormal when the calculated Belief value exceeds the error range. The method of claim 5,
And a learning pattern configuration module for classifying the collected reflected light information into learning data and tester data. The method of claim 8,
The neural network modeling construction model,
Generating a first neural network model using the training data, calculating an error value between the generated neural network model and the test data, and optimizing the first neural network model using the calculated error value and a learning factor Plasma chamber monitoring device for generating neural network model.

Images