JP5693573B2 - How to build an optimal endpoint algorithm - Google Patents

Description

   For ease of discussion, some terms are defined as follows:

  Data group: A record of measurements taken as a function of time for one parameter relating to the processing tool.

  Change point: A point where some change occurs in time series.

  Endpoint: when a process (eg, etching a silicon layer) is completed or nearly completed.

  Endpoint domain: An interval in a set of data where an endpoint is considered to occur. Endpoint domains are usually relatively wide and based on user judgment.

  Partial Least Squares Discriminant Analysis (PLS-DA): A method for finding a relationship between two data groups. PLS-DA can be used when there are a plurality of independent variables (input matrix X) and possibly a plurality of dependent variables (input matrix Y). In PLS-DA, the Y variable is not continuous, but is composed of independent discrete values or discrete class groups. PLS-DA can be used to find a linear combination of X variables that can be used to classify input data into one of the discrete classes.

  Pre-endpoint domain: The portion of data that precedes the end-point domain.

  Back endpoint domain: The part of the data group that follows the endpoint domain.

  Sign: A characteristic change point (or combination of change points) in the development of a parameter or a combination of parameters indicating an end point in processing. The combination of parameters and the nature of the change form part of the signature.

  Stepwise regression: means fitting a straight line to data values in a finite time interval of data from individual sensor channels using a least square fitting algorithm.

  Advances in plasma processing have led to the development of the semiconductor industry. In order to gain a competitive advantage, semiconductor device manufacturers need to maintain strict control of the processing environment to minimize waste and produce high quality semiconductor devices.

  One way to maintain strict control is to identify processing endpoints. As described above, the term “endpoint” refers to the point in time when a process (eg, etching of a silicon layer) is completed or nearly completed. The process of identifying the endpoint may seem as simple as identifying the signal with the greatest change. However, signal changes do not always match the endpoint. Other factors such as channel noise can also cause signal pattern changes.

  For ease of discussion, FIG. 1 shows a simple method for building an endpoint algorithm. The method described in FIG. 1 is usually performed manually, for example by an experienced user.

  For example, consider a situation where a test substrate is being processed. Since there are various types of substrates, the same type of substrate that may be used in the production environment is often used as the test substrate. For example, when a specific patterned substrate is used during production, a similar patterned substrate may be used as the test substrate.

  In the first step 102, data is acquired for one substrate. For example, while processing a substrate, data is acquired with a sensor (pressure gauge, optical emission spectrometer (OES), temperature sensor, etc.). If not thousands, data for hundreds of sensor channels is collected.

  After processing the substrate, the collected data can be analyzed. If a large amount of data is available, finding endpoints in thousands of signal streams can be a difficult task requiring deep knowledge of tools and recipes. For this reason, it is usually a skilled user who is engaged in the analysis work.

  In the next step 104, the expert user examines the change in the signal pattern with one or more signals. An experienced user may perform analysis using one or more software programs. For example, the software program may be a simple analysis tool that performs simple calculations and analysis. As another example, the software program may be a simple data visualization program that illustrates signal history, for example.

  However, even with the skills and experience of skilled users, the amount of data that can be acquired by sensors and used for analysis is enormous. Therefore, the task of identifying the endpoint sign is a distracting task. For example, there are over 2000 wavelength measurements in just one OES sensor channel. (Sensor channels that provide data on temperature, pressure, voltage, etc.) Endpoint data is also available on other sensor channels, so if you need to analyze all signals and signal combinations, experienced users cannot reach them. Will be faced with serious work.

  Obviously, depending on the application, some signals may give better quality endpoint data than others. For example, if both signal A and signal B have endpoint data, signal B can provide a better quality endpoint sign if signal B is less noisy than signal A. Given that there are dozens or hundreds of signals, the task of analyzing a set of data to find one endpoint signature, and even the best endpoint signature, is a very tedious and time consuming process.

  When analyzing data, an expert user would be looking for signal changes (eg, signal pattern changes) as an indication of an endpoint. For example, if the signal has a downward slope, the peak of the signal slope may be considered to indicate a change. The task of manually identifying a signal change has conventionally been a troublesome task, but in recent years, this task has become more difficult as the signal change becomes unclear. Particularly difficult with recipes used to process small open areas on a substrate. For example, because the area of the opening that is processed (eg, etched) is very small (eg, less than 1% of the substrate area), the signal change is negligible and almost invisible to the human eye.

  To facilitate analysis, expert users may remove data values that they believe are not relevant to endpoint identification. One way to reduce the set of data is to identify and remove areas in the signal stream that an experienced user does not think the endpoint will occur. That is, an experienced user may limit the search for endpoints to a target area in the signal stream that is typically between the previous and subsequent endpoint domains. Finding and narrowing down endpoint signatures is very costly (experts' time), so it is possible to have a front endpoint domain and a back endpoint domain, so it is left as a search target for endpoints It is desirable to limit the area.

  Since experienced users are usually familiar with processing, it may be possible to further reduce the data group by analyzing only carefully selected signals. Carefully selected signals include, for example, signals and combinations of signals that may contain endpoint data based on the experience of a skilled user. When analyzing combinations of signals together, the signal combinations are usually from one and the same sensor source. In general, data from different sensor sources are not combined because differences between sensors make it difficult, if not impossible, to perform a correlation analysis manually.

  Obviously, if only the filtered data group is processed, the risk of unintentionally removing the optimal endpoint signature will increase. That is, by filtering the data, an expert user assumes that one endpoint signature, and even the best endpoint signature, is present in one of the remaining signals after filtering. As a result, the endpoint signature identified in the remaining signal is not necessarily the optimal endpoint signature.

  After identifying the signal change, an expert user can perform verification analysis to determine the robustness of the signal change as an endpoint candidate. For example, the skilled user may analyze the signal history to determine the uniqueness of the signal change. If the signal change is not unique (ie if it occurs more than once in the signal history), the signal is removed from the data set. The skilled user then resumes the cumbersome task of identifying “endless” endpoints with other signals.

  In the next step 106, a filter group (eg, a digital filter group) is applied to the data group to remove noise and perform data smoothing. Examples of applicable filters include, but are not limited to, time series filters and frequency based filters. Applying a filter to a data group can reduce the noise of the data group, but it is usually done with caution because the filter can also increase the real-time delay in the signal.

  In some cases, multivariate analysis (such as principal component analysis or partial least squares) may be performed when analyzing data. By performing multivariate analysis, the data group can be further reduced. In order to use multivariate analysis, an experienced user may need to define the shape (eg, curve) of the endpoint characteristics. That is, it is expected that the skilled user predicts the shape of the endpoint in a state where the endpoint candidate has not yet been specified. By predefining the shape of the endpoint, signals that do not show the desired shape are removed in multivariate analysis. For example, if the endpoint shape is defined as a peak, signals that do not indicate this shape are removed. Thus, if the optimal endpoint signature is not in the “expected” shape, the optimal endpoint signature will be missed.

  As can be seen from the above description, the work of identifying one endpoint sign from a large amount of data is a distracting work and takes a long time if it does not take several weeks. Furthermore, even if one endpoint sign is identified, little or no quantitative analysis has been made as to whether a signal or combination of signals is suitable as an endpoint sign. For example, to verify a signal change as an endpoint sign, an expert user analyzes other signals and searches for similar signal changes around the same time frame. If an experienced user has already spent a significant amount of time to identify the initial endpoint signature, there may not be time, resources, and / or passion left to validate the results.

  In the next step 108, the expert user selects an endpoint algorithm type based on the nature of the transition. The type of endpoint algorithm may be based on the shape of a spectral line that can represent the endpoint. For example, the endpoint may be represented by a change in slope. Thus, an expert user is considered to propose a gradient dependent algorithm.

  Further, the endpoint algorithm may be based on a derivative that can provide an optimal endpoint signature. However, the first derivative of the endpoint sign (eg, change in slope) does not necessarily give the optimal endpoint algorithm. Alternatively, a second derivative of the gradient (such as an inflection point) may give a more desirable endpoint algorithm. The ability to identify the best endpoint algorithm associated with an endpoint signature, not just the endpoint signature, requires less skill (even for experienced users).

  In the next step 110, the algorithm settings are optimized and / or tested. After identifying the endpoint algorithm, the endpoint algorithm may be converted to a production endpoint algorithm. Because the test environment is different from the production environment, it is necessary to adjust the endpoint algorithm before introducing the endpoint algorithm into the production process. Examples of the setting to be adjusted include, but are not limited to, a smoothing filter, a delay time, a setting specific to the type of algorithm, and the like.

  For example, a filter used to smooth data in a test environment can cause real-time delays that are unacceptable in a production environment. As used herein, real-time delay refers to the time difference between an unfiltered signal change and a filtered signal change. For example, if a signal peak occurs 40 seconds after the start of processing, after the filter is applied, the peak generation time is delayed by 5 seconds. If the endpoint algorithm is applied with a filter setting, the substrate may be over-etched before the endpoint algorithm identifies the endpoint. To minimize real-time delay, the filter needs to be adjusted.

  Before introducing the endpoint algorithm into the production process, a test may be performed to determine whether the settings have been optimized. For example, the endpoint algorithm may be applied to the data group used to create the endpoint algorithm. If the endpoint algorithm can correctly identify the endpoint using the adjusted settings, then the settings are considered optimized. On the other hand, if the endpoint algorithm cannot correctly identify the endpoint, the settings need to be adjusted. It may be necessary to perform this test multiple times (in a trial and error manner) until the settings are optimized.

  In the next step 112, it is determined whether to perform a robustness test on the endpoint algorithm. When performing a robustness test (step 114), the endpoint algorithm may be applied to a group of data related to other boards. For example, a second test board is processed to collect data. An endpoint algorithm is applied to this second data group. If the endpoint algorithm can identify the endpoint, the endpoint algorithm is considered robust and the endpoint algorithm is introduced into the production process (step 116). On the other hand, if the endpoint algorithm cannot identify the endpoint, the endpoint algorithm is considered not sufficiently robust and the expert user returns to step 104 to identify other endpoint candidates. , Resume work to build other endpoint algorithms.

  Considering that it takes time to perform and analyze the robustness test, many endpoint algorithms may be introduced into the production environment without performing the robustness test. That is, step 112 is considered an optional step in creating the endpoint algorithm.

  As can be seen from FIG. 1, the method of creating an endpoint algorithm is largely a manual process and is typically performed by a skilled person with the skills and experience to perform complex analyses. Given the limited resources, endpoint algorithms introduced into the production process may lack quantitative support. Furthermore, since it is impossible for a single person to analyze all signals and / or combinations of signals in a reasonable period of time, the resulting endpoint algorithm is not necessarily the optimal endpoint algorithm for processing. Not always.

Therefore, there is a need for a simple method for building a robust endpoint algorithm.
(1) A first aspect of the present invention is a method of automatically specifying an optimal endpoint algorithm for verifying a processing endpoint during substrate processing in a plasma processing system,
In the plasma processing system, receiving sensor data from a plurality of sensors while processing at least one substrate, the sensor data including a plurality of signal streams from a plurality of sensor channels, A sensor comprising an emission spectroscopy analyzer, wherein the plurality of signal streams include a wavelength measurement signal stream from the emission spectroscopy analyzer;
Identifying an endpoint domain, wherein the endpoint domain is a time period in which the processing endpoint is expected to occur;
Dividing each signal stream of the plurality of signal streams in the endpoint domain into a plurality of segments based on a time interval;
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the first gradient group and identifying a first candidate signal group from the plurality of signal streams, wherein the first candidate signal group is the other in the plurality of signal streams; A process having a greater slope change than the signal; and
Combining adjacent wavelength measurements of the sensor data having similar gradient changes to form a first signal wavelength region group;
By applying a class feature group to at least a part of the first candidate signal group and the first signal wavelength band group, a first endpoint / sign candidate group is identified, and the first candidate signal Determining whether at least a part of the group and the first signal wavelength band group have the class feature group, wherein the class feature group is at least one of a peak feature, a valley feature, and an inflection point feature. A process comprising:
Converting the first endpoint sign candidate group into an endpoint algorithm group;
Incorporating one endpoint algorithm of the endpoint algorithm group into the production environment as the optimal endpoint algorithm;
It is a method provided with.
(2) A second aspect of the present invention is a method for verifying an endpoint during substrate processing in a processing chamber,
Executing the recipe on the substrate;
Receiving sensor data from a group of sensors while processing the substrate, wherein the sensor data includes a plurality of signal streams from a plurality of sensor channels, the plurality of sensors including an emission spectroscopy analyzer; The plurality of signal streams includes a signal stream of wavelength measurements from the emission spectrometer;
Dividing each signal stream of the plurality of signal streams into a plurality of segments based on a time interval;
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the first gradient group and identifying a first candidate signal group from the plurality of signal streams, wherein the first candidate signal group is the other in the plurality of signal streams; A process having a greater slope change than the signal; and
Combining adjacent wavelength measurements of the sensor data having similar gradient changes to form a first signal wavelength region group;
By applying a class feature group to at least a part of the first candidate signal group and the first signal wavelength band group, a first endpoint / sign candidate group is identified, and the first candidate signal Determining whether at least a part of the group and the first signal wavelength band group have the class feature group, wherein the class feature group is at least one of a peak feature, a valley feature and an inflection point feature. A process comprising:
Identifying a processing endpoint using at least one endpoint signature of the first set of endpoint signatures;
Ending the substrate processing;
It is a method provided with.

The present invention can also be realized as the following application examples.
[Application Example 1]
A method of automatically identifying an optimal endpoint algorithm for verifying a processing endpoint during substrate processing in a plasma processing system,
Receiving sensor data from a plurality of sensors during substrate processing of at least one substrate in the plasma processing system, the sensor data comprising a plurality of signal streams from a plurality of sensor channels; ,
Identifying an endpoint domain, wherein the endpoint domain is an estimated time period in which the processing endpoint is expected to occur;
Analyzing the sensor data to generate an endpoint / signature candidate group;
Converting the endpoint sign candidate group into an optimal endpoint algorithm group;
Incorporating one optimal endpoint algorithm from the optimal endpoint algorithm group into the production environment;
A method comprising:
[Application Example 2]
A method described in application example 1,
The method, wherein the sensor data is collected from two or more substrates.
[Application Example 3]
A method described in application example 1,
Analyzing the sensor data includes performing linear fitting on the sensor data to divide each signal stream from the plurality of signal streams into a plurality of segments based on time intervals. .
[Application Example 4]
A method described in application example 3,
The method wherein each segment of the plurality of segments is uniform.
[Application Example 5]
A method described in application example 3,
Analyzing the sensor data to generate a first endpoint signature group of the endpoint signature candidate groups,
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the gradient group and identifying a high contrast signal group having a large gradient change from the plurality of signal streams;
Combining adjacent wavelengths having similar gradient changes to form signal wavelength ranges;
Ranking the high contrast signal;
Ranking the signal wavelength ranges; and
Identifying the first endpoint / signature candidate group by applying a class feature group to at least a part of the high contrast signal and the signal wavelength range group, wherein the class feature group is a peak, A step comprising at least one of a valley and an inflection point;
Including a method.
[Application Example 6]
A method described in application example 5,
Analyzing the sensor data to generate a second endpoint signature group of the endpoint signature candidate group,
Multivariate analysis was performed by combining the gradient scaled by the corresponding gradient noise value of the first gradient group and the first corresponding gradient noise value group, and normalized with the normalized gradient group Generating a corresponding gradient noise value group;
Identifying a normalized signal having a large gradient and a small change from the plurality of signal streams by calculating a gradient change in the normalized gradient group;
Combining adjacent wavelengths having similar gradient changes to form a normalized signal wavelength range;
Ranking the normalized signal;
Ranking the normalized signal wavelength groups; and
Generating the second endpoint / sign candidate group by applying a class feature group to a ratio of the high-contrast signal group and the signal wavelength band group to the normalized signal and the normalized signal wavelength band group When,
Including a method.
[Application Example 7]
A method described in application example 5,
Converting the endpoint sign candidate group into the optimal endpoint algorithm group,
The first endpoint / signature candidate when the first endpoint / signature candidate and the second endpoint / signature candidate in the endpoint / signature candidate group have the same shape and the same period And combining the second endpoint sign candidate;
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing multivariate correlation analysis to identify an optimal endpoint signature group among the endpoint signature candidate groups;
Converting the optimal endpoint signatures into endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Generating the optimal endpoint algorithms, comprising:
Excluding a real-time endpoint algorithm with a corresponding real-time delay greater than a predetermined threshold;
Excluding the real-time endpoint algorithm if the real-time endpoint algorithm does not pass the robustness test;
Performing at least one of:
Ranking each endpoint algorithm of the optimal endpoint algorithm group based on at least one of fidelity ratio and real-time delay;
Including a method.
[Application Example 8]
The method according to Application Example 6,
Converting the endpoint / signature candidate group into an optimal endpoint algorithm group,
The first endpoint / signature candidate when the first endpoint / signature candidate and the second endpoint / signature candidate in the endpoint / signature candidate group have the same shape and the same period And combining the second endpoint sign candidate;
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing multivariate correlation analysis to identify an optimal endpoint signature group among the endpoint signature candidate groups;
Converting the optimal endpoint signatures into real-time endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Generating the optimal endpoint algorithms, comprising:
Excluding a real-time endpoint algorithm with a corresponding real-time delay greater than a predetermined threshold;
Excluding the real-time endpoint algorithm if the endpoint algorithm does not pass the robustness test;
Performing at least one of:
Ranking each endpoint algorithm of the optimal endpoint algorithm group based on at least one of fidelity ratio and real-time delay;
Including a method.
[Application Example 9]
A method described in application example 1,
The method wherein the capturing step is based on at least one of a ranking of the endpoint algorithm group and a user defined condition group.
[Application Example 10]
A method for verifying endpoints during substrate processing in a processing chamber, comprising:
Executing the recipe on the substrate;
Receiving process data from the sensor group during substrate processing;
Applying the optimal endpoint algorithm to analyze the processing data;
Identifying a processing endpoint;
Ending the substrate processing;
A method comprising:
[Application Example 11]
The method according to application example 10,
The method, wherein the analysis is performed by a calculation engine, the calculation engine being a high speed processing module configured to process a large amount of data.
[Application Example 12]
The method according to application example 10,
The optimal endpoint algorithm is
Receiving sensor data from a plurality of sensors during processing of at least one substrate in the plasma processing system, wherein the sensor data includes a plurality of signal streams from a plurality of sensor channels;
Identifying an endpoint domain, wherein the endpoint domain is an estimated time period in which the processing endpoint is expected to occur;
Analyzing the sensor data to generate an endpoint / signature candidate group;
Converting the endpoint sign candidate group into an optimal endpoint algorithm group;
Incorporating one optimal endpoint algorithm from the optimal endpoint algorithm group into the production environment;
Consists of, the method.
[Application Example 13]
The method according to application example 12,
The method, wherein the sensor data is collected from two or more substrates.
[Application Example 14]
A method described in application example 13,
The method of analyzing the sensor data includes performing linear fitting on the sensor data to divide each signal stream from the plurality of signal streams into a plurality of segments based on time intervals.
[Application Example 15]
The method according to application example 14,
Analyzing the sensor data to generate a first endpoint signature group of the endpoint signature candidate groups,
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the gradient group and identifying a high contrast signal group having a large gradient change from the plurality of signal streams;
Combining adjacent wavelengths having similar gradient changes to form signal wavelength ranges;
Ranking the high contrast signal;
Ranking the signal wavelength ranges; and
Identifying the first endpoint / signature candidate group by applying a class feature group to at least a part of the high contrast signal and the signal wavelength range group, wherein the class feature group is a peak, A step comprising at least one of a valley and an inflection point;
Including a method.
[Application Example 16]
The method according to application example 15,
Analyzing the sensor data to generate a second endpoint signature group of the endpoint signature candidate group,
A multivariate analysis is performed by combining the gradient scaled by the corresponding gradient noise value of the first gradient group and the first corresponding gradient noise value group, and the normalized gradient group is normalized. Generating a corresponding gradient noise value group;
Identifying a normalized signal having a large gradient and a small change from the plurality of signal streams by calculating a gradient change in the normalized gradient group;
Combining adjacent wavelengths having similar gradient changes to form a normalized signal wavelength range;
Ranking the normalized signal;
Ranking the normalized signal wavelength groups; and
Generating the second endpoint / sign candidate group by applying a class feature group to a ratio of the high-contrast signal group and the signal wavelength band group to the normalized signal and the normalized signal wavelength band group When,
Including a method.
[Application Example 17]
The method according to application example 15,
Converting the endpoint sign candidate group into the optimal endpoint algorithm group,
The first endpoint / signature candidate when the first endpoint / signature candidate and the second endpoint / signature candidate in the endpoint / signature candidate group have the same shape and the same period And combining the second endpoint sign candidate;
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing multivariate correlation analysis to identify an optimal endpoint signature group among the endpoint signature candidate groups;
Converting the optimal endpoint signatures into endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Generating the optimal endpoint algorithms, comprising:
Excluding a real-time endpoint algorithm with a corresponding real-time delay greater than a predetermined threshold;
Excluding the real-time endpoint algorithm if the real-time endpoint algorithm does not pass the robustness test;
Performing at least one of:
Ranking each endpoint algorithm of the optimal endpoint algorithm group based on at least one of fidelity ratio and real-time delay;
Including a method.
[Application Example 18]
The method according to application example 16,
Converting the endpoint / signature candidate group into an optimal endpoint algorithm group,
The first endpoint when the first endpoint signature candidate and the second endpoint signature candidate of the first endpoint signature candidate group have the same shape and the same period Combining the sign candidate and the second endpoint sign candidate;
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing a multivariate correlation analysis to identify an optimal endpoint signature group among the first endpoint signature candidate groups;
Converting the optimal endpoint signatures into real-time endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Generating the optimal endpoint algorithms, comprising:
Excluding a real-time endpoint algorithm with a corresponding real-time delay greater than a predetermined threshold;
Excluding the real-time endpoint algorithm if the endpoint algorithm does not pass the robustness test;
Performing at least one of:
Ranking each endpoint algorithm of the optimal endpoint algorithm group based on at least one of fidelity ratio and real-time delay,
Method.
[Application Example 19]
The method according to application example 12,
The method of capturing the optimal endpoint algorithm is based on at least one of a ranking and a user-defined set of conditions.
[Application Example 20]
The method according to application example 14,
The method wherein each segment of the plurality of segments is uniform.
The present invention will be described below with reference to the accompanying drawings, which are merely illustrative and do not limit the present invention. In the following description, the same components are denoted by the same reference numerals.

Diagram showing a simple way to build an endpoint algorithm.

5 is a flowchart outlining a method for detaining an endpoint algorithm in an embodiment of the present invention.

4 is a flowchart outlining the steps performed by an endpoint algorithm to find the optimal endpoint algorithm in one embodiment of the invention. 4 is a flowchart outlining the steps performed by an endpoint algorithm to find the optimal endpoint algorithm in one embodiment of the invention.

6 is a flowchart illustrating an overview of executing an optimal endpoint algorithm in a production environment in an embodiment of the present invention.

The block diagram which shows the example which expand | deploys a data group into the list | wrist of the optimal endpoint algorithm in one Embodiment of this invention.

  Hereinafter, the present invention will be described in detail with reference to embodiments shown in the accompanying drawings. Numerous details described below are for the purpose of fully understanding the present invention. It will be apparent to those skilled in the art that the present invention may be practiced without some or all of these details. In other instances, well known process steps and / or structures have not been described in detail in order not to unnecessarily obscure the features of the present invention.

  Various embodiments including methods and techniques are described below. In addition, the present invention can be configured as a product including a computer-readable medium storing computer-readable instructions for implementing the technique of the present invention. Computer-readable media include, for example, semiconductor media, magnetic media, magneto-optical media, optical media, and other computer-readable media that can store computer-readable code. Furthermore, the present invention can be configured as an apparatus for executing an embodiment of the present invention. Such an apparatus may comprise dedicated and / or programmable circuitry that performs the tasks associated with embodiments of the present invention. Examples of such devices include general purpose computers and / or dedicated computing devices that are programmed as needed, with dedicated / programmable circuits and computers suitable for various tasks associated with embodiments of the present invention. / A combination of computing devices may be used.

  In one embodiment of the present invention, a method is provided for automatically finding and optimizing endpoint algorithms. Embodiments of the present invention include a method for constructing an endpoint algorithm for determining the optimal endpoint in a process. Embodiments of the present invention also include in situ methods that apply the endpoint algorithm to the production environment.

  In the present specification, various embodiments and examples will be described using an endpoint as an example, but the present invention is not limited to the endpoint, and the same applies to any change point that may occur during processing. It is applicable to. Accordingly, the following description is merely an example, and the present invention is not limited to these examples.

  In one embodiment of the present invention, a method for constructing an endpoint algorithm is provided. The method may be a simple and easy-to-use automated method that can be used by both skilled and unskilled users. The method acquires sensor data, automatically defines an approximate endpoint period, automatically analyzes the data, automatically determines endpoint signature candidates, and uses the optimal endpoint algorithm in the production process. You may capture it.

  In the prior art, since the amount of data is enormous, it is almost impossible for a single person to analyze all signals well in an appropriate period. Unlike the prior art, the analysis method of the embodiment requires little or no human involvement. In an embodiment of the present invention, the analysis is performed using an algorithm engine instead. Since data is analyzed automatically instead of manually, more data can be analyzed even if not all data. In one embodiment, all signals can be analyzed and each signal characterized in terms of validity as an endpoint sign candidate. Furthermore, since the analysis is performed by the algorithm engine, the analysis is not limited to the data file from one substrate. Therefore, it is possible to analyze a larger amount of data and construct an optimal endpoint algorithm group having high robustness.

  The algorithm engine may be a software program based on a function of time against a target range (eg, endpoint domain) for the endpoint. Once the user has defined an approximate endpoint range (eg, endpoint domain), an algorithm engine can be used to analyze the data and find the optimal endpoint signature set.

  In one embodiment, the algorithm engine may identify a set of shape candidates that may indicate endpoint signature candidates in multivariate analysis. Unlike the prior art, it is not necessary for the user to know in advance the shapes (eg, peaks, valleys, steps, etc.) related to each endpoint / signature candidate. Instead, the algorithm engine creates a list of shape candidates after identifying endpoint sign candidates. Thus, the endpoint algorithm candidates identified by the algorithm engine are not limited to one shape (eg, a curve). In one embodiment, the algorithm engine is configured to perform data adjustments and test known endpoint candidates to identify the optimal endpoint signature for processing. By performing stepwise regression to determine the slope of each data input parameter over a series of finite time intervals over the processing history, the variation of each parameter can be derived as a function of time. In one embodiment, the time interval used for the slope calculation may be set so as not to accept noise in the input data and to accept slow drift in data not related to the endpoint.

  In one embodiment, OES signals may be grouped according to a degree of change (ie, slope) that varies as processing proceeds. For example, adjacent wavelengths having similar gradient changes may be grouped. By grouping the OES signals based on the gradient, the number of signals that need to be analyzed and the noise contained in the signals can be significantly reduced. The resulting list of signals and signal groups is very likely to contain information about the endpoint.

  In one embodiment, selective removal may be performed to reduce the number of endpoint sign candidates. In one embodiment, the robust endpoint signature is an endpoint signature that is present on all processed substrates. For example, if an endpoint sign is not a property common to all test boards or substantially most test boards, the endpoint sign is not robust and is excluded. If an endpoint signature is seen on the control substrate, the control substrate is an unetched substrate and the endpoint signature should not occur, so that endpoint signature is excluded.

  In one embodiment, multivariate analysis may be performed. For example, the analysis result may be used as an input for partial least square discriminant analysis (PLS-DA) to optimize the weight of individual signals in each group based on the gradient. In one embodiment, instead of prompting the user to enter the expected shape of the endpoint curve (as in the prior art), the PLS-DA uses the endpoint target range and shape provided by the algorithm engine. It may be performed based on.

  In one embodiment, the PLS-DA results of the OES signal may be combined and combined with other sensor signals. In one embodiment, PLS-DA may be repeated for a new group of signals to generate a compact and optimal combination of endpoint signs with low computational load and high contrast for real-time endpoint calculations. .

  In one embodiment, the endpoint sign candidate is converted to an endpoint algorithm with a minimum delay time. Endpoint sign candidates that cannot be converted to a real-time endpoint algorithm with a minimum real-time delay may be excluded. That is, if the real time delay associated with the algorithm exceeds the maximum allowable real time delay, the real time endpoint algorithm may be discarded.

  In one embodiment, endpoint algorithm candidates may be ranked based on a ratio of useful information to irrelevant information and / or information about real-time delay, hereinafter referred to as fidelity ratio. Good. For example, an algorithm with a high fidelity ratio and a small real-time delay is considered as a more robust algorithm. After ranking, one of the real-time endpoint algorithms may be selected and introduced into the production process.

  The features and advantages of the present invention will be further understood from the following description with reference to the drawings.

  The flowchart of FIG. 2 outlines a method for constructing an endpoint algorithm in one embodiment of the present invention.

  In an initial step 202, data is acquired by a group of sensors within the processing chamber. The situation considered here is, for example, a situation where a test substrate is being processed. While the substrate is being processed, data (optical emission, electrical signals, pressure data, plasma data, etc.) are collected by the sensor group.

  In one embodiment, the data used to form the optimal endpoint algorithm may be obtained from multiple test boards. By integrating data from different test substrates, noise due to material differences between the substrates and process variability can be eliminated. In one embodiment, data may be obtained from test substrates processed in different chambers. By integrating data from different chambers, the noise associated with the differences between the chambers can be removed.

  In the next step 204, the approximate duration that the processing endpoint is expected to occur is identified. That is, it defines an endpoint domain. Unlike the prior art, the endpoint domain means a relatively broad estimation period in which the algorithm engine searches for valid endpoint signatures. For example, for fast searching, the endpoint domain can be extended to include a portion of what was previously the endpoint domain in the prior art. This allows the algorithm engine to identify endpoint signatures that may occur early in the process. Such early endpoints can reduce the risk of damaging the underlying semiconductor layer.

  In the next step 206, the algorithm engine is activated to perform data analysis and generate an optimal set of endpoint algorithms. In one embodiment, data analysis is not done manually, so data files from multiple substrates can be analyzed. As will be appreciated by those skilled in the art, an endpoint algorithm built from data files from multiple boards can be applied to multiple boards being analyzed, despite the greater amount of data involved. Since endpoint characteristics that are not commonly found can be excluded, the robustness tends to increase.

  The flowchart of FIGS. 3A and 3B outlines the steps performed by the algorithm engine to analyze a collection of data and generate a list of optimal endpoint algorithms in one embodiment of the invention. For ease of discussion, FIGS. 3A and 3B will be described in conjunction with FIG. FIG. 5 is a block diagram illustrating an example of expanding a data group into a list of optimal endpoint algorithms in one embodiment.

  In an initial step 302, the algorithm engine performs a linear fitting on the available data group (first data group 502). That is, each signal is divided into fixed segments based on the time interval (data group 504). The segment length is important to minimize the noise while maximizing the possibility of identifying endpoint characteristics. If the length of the segment is too long, the endpoints can be averaged and lost. On the other hand, if the length of the segment is too short, the gradient (described later in step 304) may be affected by noise. In one embodiment, a minimum value and a maximum value of the segment length may be defined in advance. In one embodiment, the minimum segment length is greater than 1 / 10th of a second. In another embodiment, the maximum segment length is less than 2 seconds for data collection at 10 Hz.

  In the next step 304, the algorithm engine calculates, for each segment, a gradient and a corresponding gradient noise value (gradient fitting uncertainty). For example, when the signal A is divided into 10 segments, 10 gradients and gradient noise values are obtained for the signal A (data group 506A). The gradient may be normalized using the gradient noise value (data group 506B).

  In addition or alternatively, the algorithm engine performs a multivariate analysis (such as a partial least squares analysis) using the gradient scaled by the gradient noise value as input, and from the sensor channel combination An additional list of gradients and gradient noise values based on these signals may be generated (included in data group 506A). In one embodiment, gradient noise values may be used to normalize the gradient (included in data group 506B).

  After creating a list of gradients and gradient noise values for each segment (data group 506A), at the next step 306, the algorithm engine identifies candidate signals that may have endpoint data. For example, the algorithm engine may analyze each signal (and its segments) and digitize the amount of change in slope for each signal. One way of quantifying the slope change is to calculate the standard deviation of the normalized slope. For example, a large standard deviation represents a signal with a change in slope. In this example, a large standard deviation is considered to mean a signal that may have endpoint data candidates. Therefore, a signal having a large gradient change may be specified as a signal candidate (data group 508).

  Since the OES data contains a large amount of wavelength measurements (at least 2000 signals), the algorithm engine combines the adjacent wavelengths with similar gradient changes in the next step 308 to form a signal wavelength range ( The number of OES signals may be reduced by the data group 510). For example, if there are 100 wavelength measurements between 255 and 280 nanometers, and these wavelength measurements have similar slope changes, the 100 wavelength measurements can be combined into one signal. A wavelength range can be formed and handled as one unit during analysis. For example, even if there are 2000 wavelength measurement values, only 10 signal wavelength ranges need to be analyzed. By grouping the wavelength measurement values, the number of items to be analyzed is greatly reduced, so that the calculation load can be reduced.

  In the next step 310, the algorithm engine identifies a list of normalized signals (data group 506B) that can capture drift and noise in the basic processing. That is, the algorithm engine identifies signals that are suitable for normalization, with large gradients (relative to gradient noise) and small changes. The normalized signal (data group 512) may indicate possible candidates for removing common mode changes (eg, drift, noise, etc.) in the sensor signal.

  In the next step 312, the algorithm engine attempts to reduce the number of normalized OES signals by combining adjacent wavelengths with similar gradient changes to form a normalized signal wavelength band (data group 514). Also good. This step 312 is similar to step 308 except that it is applied to the normalized OES signal.

  In the next step 314, the algorithm performs high contrast sensor signal (data group 508), high contrast sensor signal wavelength range (data group 510), normalized signal (data group 512), and normal for all sensor channels. A list of the converted signal wavelength range (data group 514) is generated. In one embodiment, the signals in each data group are ranked. Since the possibility of the endpoint data in each signal is quantified, the signals of each data group can be ranked. For example, a signal with a large slope change can be ranked higher than a signal with a small slope change.

  In the next step 316, the algorithm engine searches for high-contrast sensor signals and / or wavelength ranges that are endpoint signature candidates (data group 516) within the endpoint domain. In one embodiment, the endpoint signature may be specified through a class feature group (peak, valley, inflection point, etc.). In one embodiment, class feature groups may be defined in advance. The class feature group may also be searched for in various derivatives of the signal.

  In one embodiment, a filter may be applied to the data groups 508 and 510 to remove noise and smooth the data. In one embodiment, the filter applied to the data group may be a time symmetric filter. A time symmetric filter uses the same number of points before and after a particular point to calculate an average value. Such a filter may be applied only in the post-processing mode instead of executing the processing in real time. Unlike time asymmetric filters, time symmetric filters are considered to introduce negligible time and / or amplitude distortion. As a result, it is believed that the real time delay of the filtered data is minimized.

  As can be seen from the above description, each data group may include a large amount of signals. In one embodiment, since each data group is ranked, data analysis time can be significantly reduced by reducing the number of searches. For example, instead of searching all items included in the data group 508, only the top 10 high-contrast sensor signals may be analyzed. The number of items to be searched can be changed. You may make it calculate | require optimal number by performing a decreasing regression analysis.

  In the next step 318, the algorithm engine performs a high contrast sensor signal / wavelength range for normalized signals / wavelength ranges (data groups 512 and 514) that are endpoint signature candidates (data groups 518) within the endpoint domain. The ratio of (data groups 508 and 510) is searched. By taking the ratio of the high contrast sensor signal / wavelength range to each normalized signal / wavelength range, endpoint sign candidates with a higher fidelity ratio can be identified.

  In the next step 320, the algorithm engine searches the resulting data (data groups 516 and 518) and ranks the combinations (data group 520). That is, matching for combining endpoint signs having similar shapes and periods is performed to improve contrast and SNR. In one embodiment, a linear combination is performed on the same derivative. That is, even if it occurs within the same time interval, the peak produced by the first derivative is not combined with the peak produced by the second derivative.

  In the next step 322, the algorithm engine performs a robustness test to remove endpoint signatures that are not considered reproducible. In one embodiment, the robustness test may verify consistency across multiple substrates. For example, if an endpoint sign candidate is not consistent across multiple substrates, the endpoint sign candidate is discarded because it may be due to noise / drift.

  As another example, the robustness test may verify the similarity between the test substrate and the control substrate (or group of control substrates). For example, assume that the test substrate is a substrate with a resist mask having a portion exposing a silicon region. The control substrate has the same characteristics as the test substrate except that it is completely covered with a resist mask. The same substrate treatment is applied to both the test substrate and the control substrate. However, since the entire surface of the control substrate is covered with a resist mask, no sign of etching is observed. That is, the control board has no endpoint. If the change on the control board matches any of the endpoint sign candidates, the matching endpoint sign candidate is discarded.

  As another example, the robustness test may verify uniqueness. For example, when the endpoint sign candidate to be tested has a peak characteristic, the remaining signal may be analyzed to determine whether another peak characteristic occurs before or after the endpoint sign candidate is generated. . If another peak is identified, the endpoint sign candidate is deleted.

  An example of various robustness criteria that can be applied to exclude signatures that appear not to be correct endpoint signatures has been described. By performing a robustness test on the endpoint sign candidates, a list of endpoint sign candidates that may be actual endpoints can be further solidified.

  In one embodiment, the algorithm engine may perform multivariate correlation analysis, such as correlation based partial least square discriminant analysis (PLS-DA), to optimize the list of endpoint sign candidates. . As described above, multivariate analysis (such as correlation-based PLS analysis) typically requires the definition of the endpoint sign shape. That is, in multivariate analysis, it is necessary to know the desired shape of the sine curve. In the prior art, the shape of the endpoint sign (peak, valley, slope, etc.) is usually defined by the user. Given that it takes a long time to determine the shape of an endpoint candidate (in the prior art), if it takes several weeks, the user can finally give one shape characteristic as input for multivariate analysis. Met. Unlike the prior art, endpoint sign candidates identified by an algorithm engine may have different shape characteristics. As a result, the number of inputs that can be included in the multivariate correlation analysis depends on the shape of the identified endpoint sign candidate.

  In one embodiment, shapes / shapes (as determined by the endpoint sign candidate list) are correlated to each signal to generate a correlation matrix between the endpoint sign candidates and the signal in each sensor channel. You may make it do. The correlation matrix may be applied to each signal with optimal weight and / or optimal load that can maximize the contrast of each endpoint sign candidate. While multivariate analysis helps optimize the list of endpoint sign candidates (data group 522), multivariate correlation analysis is not always necessary to identify the list of optimal endpoint algorithms. Moreover, although the PLS analysis based on correlation was used in the above-described example, the present invention is not limited to the PLS analysis based on correlation, and multivariate analysis based on any kind of correlation can be used.

  In the next step 324, the algorithm engine converts the remaining endpoint sign candidates (data group 522) into a real-time endpoint algorithm (data group 524) with minimal real-time delay. That is, the algorithm engine is configured to convert endpoint sign candidates into an endpoint algorithm that can be executed at production with a small real-time delay. In one embodiment, the settings required by each endpoint algorithm may be automatically calculated. For example, real-time filter settings may be automatically optimized to determine the endpoint with the minimum filter delay for each processed test board. The initial transient generated in the infinite impulse response filter may be minimized by initializing the cascade memory section with the real-time filter as a cascade structure. This is particularly important for endpoint algorithms that may have an endpoint near the beginning of the data history.

  For each endpoint sign candidate, the algorithm engine may provide a real-time endpoint algorithm. In one embodiment, if the algorithm engine cannot build a real-time endpoint algorithm, it does not provide any endpoint algorithm. For example, if the algorithm engine cannot build a real-time endpoint algorithm that can determine / identify the endpoint for each processing test board, it may not provide any endpoint algorithm.

  In the next step 326, the algorithm engine excludes endpoint algorithms that may exceed the maximum allowable real-time delay. For example, if the time required to identify an endpoint exceeds a predetermined threshold, the endpoint algorithm is eliminated because real-time delays can lead to substrate over-etching during production.

  In the next step 328, the algorithm engine excludes real-time endpoint algorithms that do not meet the robustness criteria. An example of a robustness criterion is to identify endpoints for all test boards with minimal real-time delay. That is, each endpoint algorithm is required to identify endpoints for all test boards. Another example of a robustness criterion is not to identify an endpoint for the control board. That is, if an endpoint algorithm can identify an endpoint for a control board, the endpoint algorithm is not robust and may be discarded.

  In the next step 330, the algorithm engine ranks the real-time endpoint algorithm. In one embodiment, the ranking may be based on fidelity ratio and / or real time delay. For example, if two real-time endpoint algorithms have the same fidelity ratio, the endpoint algorithm with the smaller real-time delay is ranked higher. As another example, if two endpoint algorithms have the same real-time delay, the higher fidelity endpoint algorithm is ranked higher.

  Returning to FIG. 2, in the next step 208, the real-time endpoint algorithm is introduced into the production process. In one embodiment, the highest ranking real-time endpoint algorithm may be automatically introduced into the production process. In another embodiment, the real-time endpoint algorithm introduced into the production process may be controlled by the user and may select an endpoint algorithm that best matches the user's requirements. For example, real-time delay is an important issue for device manufacturers. For this reason, device manufacturers may choose an endpoint algorithm with less robustness if the delay time is short.

  Empirical evidence has shown that by automating the process, the task of creating an optimal real-time endpoint algorithm can be completed in a few minutes. Furthermore, since the algorithm engine is configured to perform analysis with minimal human input, even an unskilled user can implement the process of building an endpoint algorithm. Thus, if a list of acceptable endpoint algorithms cannot be generated for a given endpoint domain, the user can immediately re-define the endpoint domain and send the algorithm engine a new endpoint name in a few minutes. An algorithm list can be generated.

  The flowchart of FIG. 4 shows an overview of executing a real-time endpoint algorithm in a production environment in one embodiment of the present invention.

  In the first step 402, the recipe is executed.

  In the next step 404, data is acquired by the sensor group during substrate processing.

  In the next step 406, the endpoint algorithm is used in-situ to analyze the data and identify the processing endpoint. In one embodiment, the data may be analyzed using a calculation engine. Since a large amount of data is collected, the calculation engine may be a high-speed processing module configured to process a large amount of data. Instead of going through the production equipment host controller or the processing module controller, the data may be sent directly from the sensor. In U.S. Application No. 12 / 555,674 filed September 8, 2009 by Huang et al., An example of an analysis computer suitable for performing an analysis is described.

  In the next step 408, the system determines whether an endpoint has been identified.

  If the endpoint has not been identified, the system returns to step 404.

  On the other hand, when the end point is specified, the recipe is ended in the next step 410.

  As can be appreciated from the foregoing description, one or more embodiments of the present invention provide a method for identifying an optimal real-time endpoint algorithm. These methods do not require skilled users by automating the analysis. By using the method described herein, a more robust endpoint algorithm can be introduced into the production process. Also, the time required to create an endpoint algorithm is greatly reduced, so updating an endpoint algorithm or creating a new endpoint algorithm is no longer resource intensive and time consuming. It's not a lot of work.

  Although the present invention has been described according to some preferred embodiments, various changes, modifications, substitutions, and the like are possible within the scope of the gist of the present invention. The various embodiments and examples described above are merely examples and do not limit the present invention. Also, throughout this specification, endpoints have been used as examples, but the present invention is also applicable to change points that are signal change events that may occur during processing.

  The names and summary of the invention are for convenience only and are not intended to interpret the claims. Further, the summary of the invention describes a very simplified form and is for convenience, and does not interpret or limit the entire invention described in the claims. As used herein, the term “group” is used in its commonly understood mathematical sense and includes 0, 1 or 2 or more. The method and apparatus of the present invention can be implemented with various modifications. The scope of the claims set forth below includes various modifications, substitutions, and equivalents within the scope of the present invention.

Claims (14)

A method of automatically identifying an optimal endpoint algorithm for verifying a processing endpoint during substrate processing in a plasma processing system,
In the plasma processing system, while processing at least one substrate, comprising the steps of receiving sensor data from a plurality of sensors, looking contains a plurality of signal streams from the sensor data a plurality of sensors channels, said plurality The plurality of signal streams includes a wavelength measurement signal stream from the emission spectroscopy analyzer; and
Identifying an endpoint domain, wherein the endpoint domain is a time period in which the processing endpoint is expected to occur;
Dividing each signal stream of the plurality of signal streams in the endpoint domain into a plurality of segments based on a time interval;
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the first gradient group and identifying a first candidate signal group from the plurality of signal streams, wherein the first candidate signal group is the other in the plurality of signal streams; A process having a greater slope change than the signal; and
Combining adjacent wavelength measurements of the sensor data having similar gradient changes to form a first signal wavelength region group;
By applying a class feature group to at least a part of the first candidate signal group and the first signal wavelength band group, a first endpoint / sign candidate group is identified , and the first candidate signal Determining whether at least a part of the group and the first signal wavelength band group have the class feature group, wherein the class feature group is at least one of a peak feature, a valley feature, and an inflection point feature. A process comprising:
Converting the first endpoint sign candidate group into an endpoint algorithm group;
Incorporating one endpoint algorithm of the endpoint algorithm group into the production environment as the optimal endpoint algorithm;
A method comprising: The method of claim 1, comprising:
The method, wherein the sensor data is collected from a plurality of substrates. The method of claim 1, comprising:
The method wherein each segment of the plurality of segments has a uniform length. The method of claim 1, further comprising:
Identifying a second candidate signal group from the plurality of signal streams, wherein the second candidate signal group has a larger gradient and a smaller gradient change than other signals of the plurality of signal streams. A process,
Combining adjacent wavelength measurements of the second candidate signal group having similar gradient changes to form a second signal wavelength band group;
The ratio of the signal amount between the first candidate signal group and the first signal wavelength range group, and the ratio of the signal amount between the second candidate signal group and the second signal wavelength range group are set to a second class. Generating a second endpoint / signature candidate group by applying the feature group;
A method comprising: The method of claim 1, comprising:
The step of converting comprises
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the first endpoint signature candidate group;
Performing a multivariate correlation analysis to identify an optimal endpoint signature group among the first endpoint signature candidate groups;
Converting the optimal endpoint signatures into optimal endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Excluding a subset of optimal endpoint algorithms with a corresponding real-time delay greater than a predetermined threshold;
Ranking the optimal endpoint algorithms based on at least one of fidelity ratio and real-time delay;
Including a method. The method of claim 1, comprising:
The step of converting comprises
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing multivariate correlation analysis to identify an optimal endpoint signature group among the endpoint signature candidate groups;
Converting the optimal endpoint signatures into real-time endpoint algorithms that minimize real-time delay, wherein the real-time delay is based on filter delays;
Excluding a subset of said real-time endpoint algorithm group that did not pass the robustness test of the endpoint algorithm;
Including a method. A method for verifying endpoints during substrate processing in a processing chamber, comprising:
Executing the recipe on the substrate;
Comprising the steps of receiving sensor data from the sensor group during the processing the substrate, wherein the sensor data is seen containing a plurality of signal streams from a plurality of sensors channels, the plurality of sensors includes optical emission spectrometer The plurality of signal streams includes a signal stream of wavelength measurements from the emission spectrometer ; and
Dividing each signal stream of the plurality of signal streams into a plurality of segments based on a time interval;
Calculating a first gradient group and a first corresponding gradient noise value group for the sensor data, wherein a gradient and a corresponding gradient noise value are calculated for each segment of the plurality of segments;
Calculating a gradient change in the first gradient group and identifying a first candidate signal group from the plurality of signal streams, wherein the first candidate signal group is the other in the plurality of signal streams; A process having a greater slope change than the signal; and
Combining adjacent wavelength measurements of the sensor data having similar gradient changes to form a first signal wavelength region group;
By applying a class feature group to at least a part of the first candidate signal group and the first signal wavelength band group, a first endpoint / sign candidate group is identified , and the first candidate signal Determining whether at least a part of the group and the first signal wavelength band group have the class feature group, wherein the class feature group is at least one of a peak feature, a valley feature, and an inflection point feature. A process comprising:
Identifying a processing endpoint using at least one endpoint signature of the first set of endpoint signatures;
Ending the substrate processing;
A method comprising: The method of claim 7 , comprising:
The method of calculating the gradient change is performed by a calculation engine. The method of claim 7 , further comprising:
A method comprising: identifying an endpoint domain, wherein the endpoint domain is an estimated time period in which the processing endpoint is expected to occur. The method of claim 7 , comprising:
The method, wherein the sensor data is collected from a plurality of substrates. The method of claim 7 , further comprising:
Identifying a second candidate signal group from the plurality of signal streams, wherein the second candidate signal group has a larger gradient and a smaller gradient change than other signals of the plurality of signal streams. A process,
Combining adjacent wavelength measurements of the second candidate signal group having similar gradient changes to form a second signal wavelength band group;
The ratio of the signal amount between the first candidate signal group and the first signal wavelength range group, and the ratio of the signal amount between the second candidate signal group and the second signal wavelength range group are set to a second class. Generating a second endpoint / signature candidate group by applying the feature group;
A method comprising: The method of claim 7 , further comprising:
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the first endpoint signature candidate group;
Performing a multivariate correlation analysis to identify an optimal endpoint signature group among the first endpoint signature candidate groups;
Converting the optimal endpoint signatures into optimal endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Excluding a subset of real-time endpoint algorithms with a corresponding real-time delay greater than a predetermined threshold;
A method comprising: The method of claim 7 , further comprising:
Performing a robustness test and excluding endpoint signatures that are considered not reproducible from the endpoint signature candidate group; and
Performing a multivariate correlation analysis to identify an optimal endpoint signature group among the first endpoint signature candidate groups;
Converting the optimal endpoint signatures into real-time endpoint algorithms with minimal real-time delay, wherein the real-time delay is based on a filter delay;
Excluding a subset of said real-time endpoint algorithm group that did not pass the robustness test of the endpoint algorithm;
A method comprising: The method of claim 7 , comprising:
The method wherein each segment of the plurality of segments has a uniform length.

Images