TWI480917B - Methods for constructing an optimal endpoint algorithm - Google Patents

Description

Construction method of optimal endpoint algorithm

The present invention relates to a method of constructing an optimal endpoint algorithm.

The present application claims the priority of the co-owned provisional patent application "Method and System for Advanced Equipment Control/Advanced Process Control of Plasma Processing Equipment", U.S. Patent Application No. 61/222,102, and Agent No. P2012P /LMRX-P183P1, filed on June 30, 2009 by the inventor of the s.

This section of the continuation claims the priority and interest of the commonly assigned patent application "identification of the arrangement and method of uncontrolled incidents at the process module stage". Application No. 12/555,674, Agent Case No. P2002 /LMRX-P179, filed by the inventor Huang et al. on September 8, 2009, which is related to and claims a co-transfer of a provisional patent application "identifying an uncontrolled event at the process module stage The priority of the arrangement and its method, the application number No. 61/222,024, the agent's case number P2002P/LMRX-P179P1, was filed by the inventor Huang et al. on June 30, 2009, the full text of which is incorporated herein. for reference.

For ease of discussion, define a few words as follows.

Data set - measurement record, which is a time function for processing implement parameters.

Change point - A point in the time series when several changes occur.

The end point - a process (such as ruthenium etching) has reached or near the point in time when it is completed.

Endpoint field - The interval during which the expected end point is expected to occur in the data set. The endpoint domain is usually quite wide and based on user estimates.

Partial Least Squares Discriminant Analysis (PLS-DA) - A method for finding the relationship between two sets of data. PLS-DA can be used when there are multiple independent variables (in the input matrix X) and multiple possible strain numbers (in the output matrix Y). In PLS-DA, the Y variable is not continuous, but consists of a set of independent discrete values or sets. PLS-DA will attempt to find a linear combination of X variables that can be used to classify the input data into one of the discrete sets.

Pre-terminating field - the part of the data set before the end point field.

Back end field - the part of the data set after the end point field.

Feature Notation - A special point of change (or combination of points of change) in the progression of a parameter or combination of parameters that indicates the end of a process. This combination of parameters and the nature of the change typically form part of a feature mark.

Stepwise regression - means that in the interval of limited temporary data (from a different sensor pipeline), the least squares fit algorithm is used to match the data values.

Advances in plasma processing have provided growth in the semiconductor industry. To achieve competitive advantages, semiconductor component manufacturers need to maintain tight control of the processing environment, minimizing consumption and producing high quality semiconductor components.

One way to maintain tight control is by identifying a process endpoint. As used herein, a terminology term refers to the point in time at which a process (such as a germanium layer etch) has reached or is near completion. The procedure for identifying the endpoint can be as simple as identifying the signal with the greatest change. However, signal changes do not always coincide with the end point. Other factors, such as pipe noise, can cause signal patterns to change.

To facilitate discussion, Figure 1 presents a simple method of constructing an endpoint algorithm. For example, the method as described in Figure 1 is often performed manually by an expert user.

For example, consider the situation in which the test substrate is being processed. Since there are different types of substrates, the type of test substrate tends to be the same as the type of substrate used in a production environment. For example, if a substrate of a particular pattern is used during production, a substrate of similar pattern is used as the test substrate.

In a first step 102, data is acquired for the substrate. In one example, a sensor (such as a pressure gauge, an optical emission spectrometer (OES), a temperature sensor, etc.) acquires data when processing the substrate. Data can be collected from hundreds (if not thousands) of sensor pipelines.

After the substrate has been processed, the collected data can be analyzed. Because there is too much data available, finding the end point in thousands of signal streams is a challenging task, often requiring in-depth knowledge of the tool and recipe. Therefore, the expert user is usually responsible for performing the analysis task.

In the next step 104, the expert user will review the changes in the signal pattern for more than one signal. Expert users use more than one software program to assist with analysis. In one example, the software program is a simple analysis tool that performs simple calculations and analysis. In another example, the software program is a simple data visualization program, for example, to graphically depict a signal history.

However, even with the expertise and experience of expert users, the amount of data that can be obtained by the sensor and that can be analyzed can be overwhelming. Therefore, the task of identifying the endpoint feature tokens can be a daunting task. In one example, there are more than 2,000 wavelength measurements in the OES sensor pipeline. Because the endpoint data can also be found in other sensor pipelines (such as sensor pipelines that provide information on temperature, pressure, voltage, etc.), if you need to analyze each signal and signal combination, expert users will face insurmountable task.

As can be expected, depending on the application, several signals can provide endpoint data that is superior to other signals. For example, both signals A and B have endpoint data. However, since signal B has less noise than signal A, signal B provides a better end point signature. Assuming tens or hundreds of signals, the task of analyzing the data set for the end point feature notation (not to mention the best end point feature mark) can become extremely tedious and time consuming.

Expert users will look for signal changes (such as changes in signal patterns) as a sign of the endpoint when analyzing the data. For example, if the signal is tilted downward, the peak value of the signal slope represents a change. Although manual identification of signal changes has been a tedious task in the past, in recent years, as signal changes have become less noticeable, this task has become more difficult. This is especially true for formulations that process small open areas on the substrate. In one example, the open area being processed (e.g., etched) is so small (e.g., <1% of the substrate area) that the signal changes are so subtle that the human eye is barely noticeable.

To facilitate analysis, expert users delete data values that they believe are not related to the endpoint of the identification. One way to narrow down a data set involves identifying and deleting areas of the signal flow that the expert user would not expect to end. In other words, the expert user limits its endpoint to the target area in the signal stream, which is often located between the pre- and post-end fields. Since the cost of finding and honing the end point feature is high (in expert time), the goal is to make the pre- and post-end fields as large as possible to limit the area in which the remaining end points are sought.

Because expert users are often familiar with the process, expert users can further narrow down the data set by analyzing only the selected signals. The featured signal includes a signal or combination of signals that will contain the endpoint data based on the experience of the expert user. Typically, when signal combinations are analyzed as a group, the combination of signals often comes from a single source of signal sensors. In general, because differences between sensors can make correlation analysis difficult, if not impossible, to perform manually, data from different sensor sources is not combined.

As can be expected, the operation of only one filter data set will increase the risk of inadvertently deleting the best endpoint feature mark. In other words, by filtering the data, the expert user assumes that the endpoint feature signature (not to mention the best endpoint signature) is one of the remaining signals after filtering. Therefore, the endpoint feature signature identified in the remaining signal is not necessarily the best endpoint signature.

After the signal change has been identified, the expert user performs a verification analysis to determine the robustness of the signal change as the endpoint candidate. For example, an expert user analyzes the signal history and determines the uniqueness of the signal change. If the signal change is not unique (ie, occurs more than once in the signal history), the signal is removed from the data set. Expert users then return to their tedious tasks and identify "hard-to-find" endpoints in other signals.

In the next step 106, a set of filters (such as a set of digital filters) is used in the data set to remove noise and smooth the data. For example, examples of filters that may be used include, without limitation, a timing filter, a frequency filter. Although the use of filters for data sets reduces the noise of the data set, filters are often used with caution because they also increase the immediate delay of the signal.

In several cases, multivariate analysis (such as principal component analysis or partial least squares methods) is performed to analyze the data. Perform multivariate analysis to further narrow down the data set. In order to use multivariate analysis, expert users need to define the shape of the endpoint feature (such as a curve). In other words, even if the end point candidate value has not been recognized, the expert user is expected to predetermine the shape of the end point. By predetermining the endpoint shape, multivariate analysis essentially removes signals that do not exhibit the desired shape. In one example, if the end shape is defined as a peak, the signal that does not exhibit this shape is deleted. Therefore, if the best end point feature mark does not have the "expected" shape, the best end point feature mark will be missed.

As can be seen from the foregoing, the task of identifying a single endpoint feature token from too much data can be a prohibitive task and can be time consuming to perform (if not weekly). In addition, once the endpoint feature signature is identified, only a minimal or no execution signal or combination of signals is performed as a quantitative analysis of the suitability of the endpoint signature. In one example, to verify signal changes as endpoint feature signatures, the expert user analyzes other signals to find similar signal changes in approximately the same time zone. However, considering that the expert user has spent a significant amount of time identifying the first endpoint feature token, the expert user may not always have time, resources, and/or willingness to verify the outcome.

In the next step 108, the expert user selects the type of the endpoint algorithm based on the nature of the transition. Typically, the type of endpoint algorithm is based on the shape of the line, which for example represents the endpoint. In one example, the endpoint can be represented by a change in slope. Therefore, expert users will propose algorithms that depend on the slope.

In addition, the endpoint algorithm is based on the derivative that provides the best endpoint signature. However, the first derivative of the endpoint signature (such as a change in slope) may not provide an optimal endpoint algorithm. For example, the second derivative of the slope (such as the inflection point) may provide a better endpoint algorithm. The ability to identify not only the endpoint signature but also the best endpoint algorithm associated with the endpoint signature requires the expertise of a small number of users, even expert users.

In the next step 110, the settings of the algorithm are optimized and/or tested. Once the endpoint algorithm has been identified, the endpoint algorithm is transformed into a production endpoint algorithm. Because there is a difference between the test environment and the production environment, the setpoint of the endpoint algorithm needs to be adjusted before moving the endpoint algorithm to production. For example, the set values to be adjusted include, without limitation, a smoothing filter, a delay time, a specific set value of an algorithm type, and the like.

In one example, filters used to smooth data in a test environment can cause unacceptable immediate delays in a production environment. As discussed herein, immediate delay refers to the time difference between an unfiltered signal change and a filtered signal change. For example, the peak of the signal may occur in the 40th second of the process. However, after using the filter, peaks are not generated until after 5 seconds. If the endpoint algorithm is used with the filter setpoint, the substrate may be over-etched before the endpoint algorithm recognizes the endpoint. In order to minimize the immediate delay, the filter must be adjusted.

Before moving the endpoint algorithm to production, a test is performed to determine if the setpoint has been optimized. In one example, an endpoint algorithm is used on the data set that has been used to construct the endpoint algorithm. If the endpoint algorithm correctly recognizes the endpoint using the adjusted setpoint, the setpoint is considered to be optimal. However, if the endpoint algorithm does not correctly identify the endpoint, the setpoint must be adjusted. Multiple tests must be performed before the set value is optimal (through trial and error).

At next step 112, a determination is made for performing a robustness test at the endpoint algorithm. If the robustness test has been performed (step 114), the endpoint algorithm is used for the data sets associated with the other substrates. In one example, the second test substrate can be processed and data collected. The endpoint algorithm is then used in the second data set. If the endpoint algorithm is able to recognize the endpoint, then the endpoint algorithm is considered robust and proceeds to production using the endpoint algorithm (step 116). However, if the endpoint algorithm is unable to recognize the endpoint, then the endpoint algorithm is considered to be less robust, and the expert user returns to step 104 to resume recognizing the other endpoint candidate and constructing another endpoint algorithm.

Considering the robustness test takes time to perform and analyze, many end-point algorithms are used in the production environment without undergoing robustness testing. In other words, step 112 is often viewed as an optional step in constructing an endpoint algorithm.

As can be seen from Figure 1, most of the methods for constructing the endpoint algorithm are manual procedures, which often perform complex analysis by experts with expertise and experience. Considering the limitations of resources, the endpoint algorithm that moves to production may lack quantitative support. In addition, because within a reasonable time, a single person cannot have a way to analyze all signals and/or signal combinations, the constructed end-point algorithm may not always be the best end-point algorithm for the process.

Therefore, a simplified approach to constructing robust endpoint algorithms is highly anticipated.

An embodiment of the invention is directed to a method of automatically identifying an optimal endpoint algorithm for identifying a process endpoint during processing of a substrate in a plasma processing system. The method includes during processing of at least one substrate in the plasma processing system, The sensor data is received from a plurality of sensors, wherein the sensor data includes a plurality of signal streams from a plurality of sensor tubes. The method also includes identifying a endpoint domain, wherein the endpoint domain is a summary interval at which an end of the process is expected to occur. The method further includes analyzing the sensor data to generate a set of possible endpoint feature signatures. The method also includes converting the set of possible endpoint feature tokens into a set of optimal endpoint algorithms. The method additionally includes introducing an optimal endpoint algorithm of the set of optimal endpoint algorithms into the production environment.

The above summary is only one of the many embodiments of the invention disclosed herein, and is not intended to limit the scope of the invention. These and other features of the present invention will be described in detail in the embodiments of the invention described below.

The invention will now be described in detail with reference to a number of embodiments as illustrated in the accompanying drawings. In the following description, numerous specific details are set forth However, it will be apparent to those skilled in the art that the present invention may be practiced without some or all of the specific details. In other instances, well known process steps and/or structures have not been described in detail to avoid undesired interference with the present invention.

The following description includes various embodiments of the methods and techniques. It should be borne in mind that the present invention may also cover manufactured products, including computer readable media stored by computer readable instructions for performing embodiments of the present technology. By way of example, computer readable media includes semiconductor, magnetic, photomagnetic, optical, or other forms of computer readable media for storing computer readable code. In addition, the present invention also encompasses an apparatus for carrying out embodiments of the present invention. Such devices include circuitry (dedicated and/or programmable) to perform the tasks associated with embodiments of the present invention. Examples of such devices include general purpose computers and/or dedicated computing devices that are suitably programmed and include a combination of computer/computing devices and special/programmable circuits suitable for the various tasks associated with embodiments of the present invention.

In accordance with an embodiment of the present invention, a method for automatically detecting and optimizing an endpoint algorithm is provided. Embodiments of the invention include methods of constructing an endpoint algorithm that determines the optimal endpoint of a process. Embodiments of the invention also include the use of endpoint calculus in a production environment Local method of law.

In this document, various embodiments are discussed using an endpoint as an example. However, the invention is not limited to the end point, but includes any point of change that can occur in the process. Accordingly, the discussion is intended as an example, and the invention is not limited to the illustrated examples.

In an embodiment of the invention, a method of constructing an end point algorithm is provided. The method includes a simple, affinitive, automated method that can be used by both expert and non-expert users. The method includes obtaining sensor data, automatically defining a rough endpoint, automatically analyzing data, automatically determining a set of possible endpoint feature tokens, and automatically introducing an optimal endpoint algorithm to production.

In the prior art, a single person was unable to analyze all signals within a reasonable period of time simply because of the amount of data. Unlike prior art, the analysis in the examples involved little or no human intervention. Instead, in one embodiment, an algorithmic engine can be employed to perform the analysis. Because the data is analyzed automatically, not manually, more (if not all) of the data can be analyzed. In one embodiment, all possible signals are analyzed and each signal is characterized by its correlation with possible endpoint feature signatures. In addition, because the analysis is now performed by the algorithmic engine, the analysis is no longer limited to data files from a single substrate. Therefore, more information can be analyzed to construct a robust set of optimal endpoint algorithms.

The algorithm engine is a software program based on a time function associated with the region of the endpoint (such as the endpoint field). Once the user has defined a rough end point area (such as the end point field), the algorithm engine can be used to analyze the data and find a set of best end point feature marks.

In an embodiment, the algorithm engine identifies a set of possible shapes that represent possible endpoint feature tokens in the multivariate analysis. Unlike prior art, the user does not need to have prior knowledge (such as peaks, troughs, spacing, etc.) for the shape of each possible endpoint feature token. Instead, once the algorithm engine has identified the possible endpoint feature tokens, the algorithm engine produces a list of possible shapes. Therefore, the possible endpoint feature signatures recognized by the algorithm engine are not limited to a single shape (such as a curve). In one embodiment, the algorithm engine is configured to perform data conditioning and testing of known endpoint candidate values to identify the best endpoint feature token for a process. Parameters as a function of time The variability can be derived by performing a stepwise regression to determine the slope of each data input parameter over a finite time interval throughout the manufacturing process. In one embodiment, the time interval used to calculate the slope can be set to remove noise in the splicing data and to remove the slow drift of the data unrelated to the endpoint.

In one embodiment, the OES signals are grouped according to the degree of change (i.e., slope) of the variability that is visible as the process progresses. In one example, successive wavelengths with similar slope variations are grouped together. By grouping the OES signals by the slope, the amount of signal to be analyzed and the noise in the signal are greatly reduced. This result is presented as a list of signal and signal groups, most likely containing information related to the endpoint.

In an embodiment, picking is performed to reduce the amount of possible end point feature tokens. In an embodiment, the robust endpoint feature symbol is a signal present in all processed substrates. In one example, if an endpoint feature signature is not a feature in all or a very large portion of the test substrate, the endpoint feature signature is not robust and can be deleted. However, if an end point feature mark appears on the control substrate, the end point feature mark can also be deleted because the control substrate is a substrate that does not receive etching and therefore should not produce an end point feature mark.

Multivariate analysis is performed in an embodiment. In one example, the results obtained from the analysis are used as input to Partial Least Squares Discriminant Analysis (PLS-DA), so that the weight of each individual signal in each group of slopes is the highest. good. In one embodiment, the user is not required to enter the desired shape of the endpoint curve (as required by the prior art), but rather the area under which the PLS-DA will follow the endpoint and the shape provided by the algorithm engine.

In one embodiment, the results of the PLS-DA from the OES signal can be combined and combined with other sensor signals. In one embodiment, the PLS-DA can be repeated to the new joint signal set to produce a refined best possible endpoint feature signature combination with high computational load for high contrast and immediate endpoint calculations.

In an embodiment, the possible endpoint feature token is converted to an endpoint algorithm with the smallest possible delay time. Possible end feature signatures that cannot be converted to an immediate endpoint algorithm with minimal immediate delay are deleted. In other words, if the immediate delay associated with the algorithm exceeds the maximum allowable immediate delay, then the immediate endpoint algorithm Will be abandoned.

In one embodiment, the possible endpoint algorithms are ranked based on the ratio of useful information to irrelevant information (hereinafter referred to as fidelity) and/or immediate delay. In one example, algorithms with high fidelity and low real-time latency are considered to be more robust algorithms. Once the grading has been performed, one of the immediate endpoint algorithms will be selected and moved to production.

The features and advantages of the present invention will become more apparent from the understanding of the appended claims.

2 presents a simplified flow diagram depicting a method of constructing an end point algorithm in an embodiment of the present invention.

In a first step 202, data is retrieved by a set of sensors in the processing chamber. For example, consider the case in which the test substrate is being processed. When the substrate is being processed, data (eg, light emissions, electrical signals, pressure data, plasma data, etc.) are collected by a set of sensors.

In one embodiment, the data used to construct the optimal endpoint algorithm may come from more than one test substrate. By incorporating data from different test substrates, noise associated with material differences or process variability between the substrates can be eliminated. In an embodiment, the data may be from test substrates processed in different chambers. By incorporating data from different chambers, noise associated with differences between the chambers can also be eliminated.

In the next step 204, a summary time period in which the end of the process is expected to occur is identified. In other words, define a destination field. Unlike the prior art, the endpoint field is a rough and relatively wide time interval in which the algorithm engine will search for valid endpoint feature signatures. For example, due to the high speed of the search, the user can extend the endpoint field to incorporate portions of the prior art that are pre-finished domains. In this way, the algorithm engine can identify the end point signatures that occur earlier in the process. These early endpoints reduce the risk that the process will damage the underlying semiconductor layer.

In the next step 206, the algorithm engine is launched to perform data analysis and generate a set of optimal endpoint algorithms. In one embodiment, data files from more than one substrate can be analyzed because data analysis is not performed manually. Those skilled in the art are aware that even if a relatively large amount of data is involved, the endpoint algorithm constructed by the plurality of substrates tends to be more robust because the endpoint features that are not commonly found in the analyzed substrate are deleted.

3A and 3B present a simplified flow diagram of an embodiment of the present invention depicting the steps and algorithm engine for performing a data set analysis and generating a list of best end point algorithms. To facilitate discussion, Figures 3A and 3B will be discussed in conjunction with Figure 5. Figure 5 presents a block diagram depicting an example of a data set progression as a list of best end-point algorithms in an embodiment.

In a first step 302, the algorithm engine performs a linear adaptation of the available data sets (initial data group 502). In other words, each signal is divided into uniform segments (data group 504) based on time intervals. In order to minimize noise and maximize the likelihood of identifying endpoint characteristics, the length of the segment is quite important. If the segment length is too long, the end point will be averaged out and the end point will be missed. If the segment is too short, the slope (as described later in step 304) will be affected by noise. In an embodiment, the minimum and maximum amounts of segment length may be predetermined. In an embodiment, the minimum segment length is longer than 1/10 seconds. In another embodiment, the maximum segment length is shorter than 2 seconds for data collected at 10 Hz.

In the next step 304, the algorithm engine calculates the slope and its corresponding slope noise value for each segment (adjusting the uncertainty of the slope). In one example, if signal A has been divided into ten segments, the ten slopes and slope noise values of signal A are determined (data group 506A). In an embodiment, the slope noise value can be used to normalize the slope (data group 506B).

Additionally or alternatively, the algorithm engine uses the slope scaled by the slope noise value as an input to perform multivariate analysis (eg, partial least squares analysis), based on data from the sensor pipeline combination, to generate slope and slope noise values. An additional list (also included in data group 506A). In an embodiment, the slope noise value can be used to normalize the slope (also included in data group 506B).

Once the slope and slope noise value lists for each segment have been constructed (data group 506A), in the next step 306, the algorithm engine will identify the signal candidate values with the endpoint data. In one example, the algorithm engine analyzes each signal (with its fragments) and quantifies the amount of variation in the slope of each signal. One method of quantifying the amount of slope variation involves calculating the standard deviation of the normalized slope. In one example, the high standard deviation represents a signal that the slope changes. In this example, the high standard deviation represents a signal with possible endpoint data. Therefore, a signal with a high slope variation (relative to slope noise) is recognized as a signal candidate (data group 508).

Since the OES data includes a large number of wavelength measurements (at least 2,000 signals), in the next step 308, the algorithm engine reduces OES semaphores by combining successive wavelengths with similar slope variations to the signal wavelength band (data group 510). . In one example, if there are 100 wavelength measurements between 255 nm and 208 nm and the wavelength measurements have similar slope variations, the 100 wavelength measurements are combined into a single signal wavelength band. And is considered a single unit in the analysis. For example, if there are 2,000 wavelength measurements, then only 10 signal wavelength bands may need to be analyzed. By grouping the wavelength measurement results, the amount of objects that need to be analyzed has been greatly reduced, so that the computational load can be reduced.

In the next step 310, the algorithm engine recognizes the normalized signal list (data group 506B), which captures the offset and noise in the following process. In other words, the algorithm engine identifies signals that are suitable for normalization because of its high slope but low variation (relative to slope noise). The normalized signal (data group 512) represents a possible candidate value for removing general mode variations (eg, offset, noise, etc.) in the sensor signal.

In the next step 312, the algorithm engine reduces the normalized OES semaphore by combining successive wavelengths with similar slope variations to the normalized signal wavelength band (data group 514). Step 312 is somewhat similar to step 308 except that step 312 is used to normalize the OES signal.

In the next step 314, the algorithm generates a high contrast sensor signal (data group 508), a high contrast sensor signal wavelength band (data group 510), a normalized signal (data group 512), and A list of standardized wavelength bands (data group 514). In one embodiment, the signals in each data set are graded. Since the possibility of endpoint data in each signal has been quantified, the signals in each data set can be graded. In one example, signals with high slope variability have higher grading than signals with low slope variability.

In the next step 316, the algorithm engine searches for high contrast sensor signals and/or frequency bands to find possible endpoint feature signatures in the endpoint domain (data group 516). In an embodiment, the endpoint feature symbol can be identified by a set of classification features (peaks, troughs, inflection points, etc.). The set of classification features can be predetermined in an embodiment. The set of classification features can be searched for in different signal derivatives.

In an embodiment, filters may be applied to data groups 508 and 510 to remove noise and smooth data. In an embodiment, the filter applied to the data group is a time symmetric filter. The time symmetric filter uses an equal number of points before and after a particular point to calculate an average. This filter can only be used in post-processing mode, not during immediate process execution. Unlike time-asymmetric filters, time-symmetric filters tend to cause minimal time warping and/or amplitude distortion. Therefore, filtering data will experience minimal immediate delay.

As can be seen from the foregoing, each data group includes an excess of signals. In an embodiment, since each data group has been graded, the data analysis time can be greatly reduced by reducing the search value. In one example, instead of searching for all of the items in the data group 508, only the top 10 high contrast sensor signals are analyzed. The amount of objects that can be searched may vary. Perform a recycling reduction analysis to determine the optimal amount.

In the next step 318, the algorithm engine searches for a ratio of high contrast sensor signals/bands (data groups 508 and 510) to normalized sensor signals/bands (data groups 512 and 514) to find possible in the endpoint domain. Endpoint feature symbol (data group 518). By using the ratio of each high contrast sensor signal/band to each normalized sensor signal/band, the identifiable possible end feature signature will have a higher fidelity.

In the next step 320, the algorithm engine searches for data results (data groups 516 and 518) to rank the combination (data group 520). In other words, matching is performed to combine endpoint signatures with similar shapes and time periods to enhance the signal-to-noise ratio (SNR). In an embodiment, linear combining is performed on the same derivative. In other words, even if the peak occurring in the first derivative occurs in the same time interval as the peak occurring in the second derivative, the two are not combined.

In the next step 322, the algorithm engine performs a robustness test to remove possible non-repetitive endpoint feature signatures. In an embodiment, the robustness test will confirm consistency between multiple substrates. In an example, if the possible end point feature mark does not coincide between the plurality of substrates, for example, because the end point feature mark may be the result of the noise/offset, the possible end point feature mark is discarded.

In another example, the robustness test confirms the similarity between the test substrate and the control substrate (or a set of control substrates). For example, consider the test substrate as having one The substrate is divided into a photoresist mask of the exposed area. The control substrate has the same characteristics as the test substrate except that the control substrate is entirely covered by the photoresist mask. Both the test substrate and the control substrate undergo the same substrate processing. However, since the entire surface of the control substrate is covered by the photoresist mask, the control substrate should not exhibit signs of etching. Therefore, the control substrate should have no end point. Therefore, if the change of the control substrate matches one of the possible end point signatures, the matching possible end feature signature will be discarded.

In another example, the robustness test includes test uniqueness. In an example, the possible endpoint feature signature being tested has a peak feature. The remainder of the signal is analyzed to determine if other peak features occur before or after the possible endpoint feature signature occurs. If other peaks are identified, the possible end feature signature is deleted.

The above are examples of different robustness criteria that can be used to delete feature signatures for non-true endpoint feature tokens. By applying a robustness test to the possible endpoint feature tokens, a list of possible endpoint feature tokens for the true endpoint can be further determined.

In an embodiment, the algorithmic engine performs a multivariate correlation analysis, such as a partial least squares discriminant analysis (PLS-DA) based on correlation, to optimize the list of possible endpoint feature tokens. As previously mentioned, multivariate analysis (such as correlation-based PLS analysis) typically requires the shape of the endpoint feature signature to be defined first. In other words, multivariate analysis requires knowledge of the expected shape of the signature curve. In the prior art, the user is typically a person who must provide an endpoint signature shape (such as a crest, trough, slope, etc.). Considering that the shape of the endpoint candidate value (in the prior art) is time consuming (if not calculated in weeks), the user typically can only provide one shape feature as an input to the multivariate analysis. Unlike prior art, the possible endpoint feature signatures recognized by the algorithm engine have different shape characteristics. Thus, the input that can be input to the multivariate correlation analysis is in accordance with the shape of the possible endpoint feature signature that has been identified.

In one embodiment, the shape/shapes (as determined by the list of possible endpoint feature tokens) are associated with each signal to produce a correlation matrix between the possible endpoint signatures and the signals in each of the sensor conduits. The correlation matrix includes the optimal weights and/or units that can be applied to each signal to optimize the comparison of the possible endpoint feature signatures. Although multivariate analysis can help optimize the list of possible endpoint feature markers (data groups) 522), multivariate correlation analysis is not required to identify the best endpoint algorithm list. Moreover, even if a correlated PLS analysis is used in the foregoing examples, the invention is not limited to a correlated PLS analysis, but can be a multivariate analysis of the correlation of any type.

In the next step 324, the algorithm engine converts the remaining possible endpoint feature tokens (data group 522) into an immediate endpoint algorithm with a minimum immediate delay (data group 524). In other words, the algorithmic engine is configured to translate the possible endpoint feature tokens into an endpoint algorithm that is executed during production with minimal immediate delay. In one embodiment, the setpoints required for each endpoint algorithm are automatically calculated. In one example, the set value of the instant filter is automatically optimized to summon the endpoint on each processed test substrate with a minimum filter delay. The instant filters are serialized and use the initial values of the serial memory components to minimize the initial transient pulses that occur in the infinite impulse response filter. This is especially important for the end-point algorithm that ends at the beginning of the data history.

The algorithm engine provides an immediate endpoint algorithm for each possible endpoint feature token. In an embodiment, if the algorithm engine cannot construct an immediate endpoint algorithm, the endpoint algorithm will not be provided. In one example, if the algorithmic engine is unable to construct an immediate endpoint algorithm that can summon/identify the endpoint on each processing test substrate, then the endpoint algorithm will not be provided.

In the next step 326, the algorithm engine deletes the endpoint algorithm that exceeds the maximum allowable immediate delay. In one example, if the time required to identify the endpoint exceeds a predetermined threshold, the endpoint algorithm will be deleted because the immediate delay will result in over-etching of the substrate during production.

In the next step 328, the algorithm engine deletes the immediate endpoint algorithm that does not pass a set of robustness criteria. Examples of robustness criteria include identifying endpoints on all test substrates with minimal immediate delay. In other words, each endpoint algorithm needs to identify the endpoint on all test substrates. Another example of a robustness criteria includes not identifying an endpoint on a control substrate. In other words, if the endpoint algorithm can identify the endpoint on the control substrate, the endpoint algorithm is not robust enough and the endpoint algorithm will be discarded.

In the next step 330, the algorithm engine will classify the immediate endpoint algorithm. In an embodiment, the rating is based on fidelity and/or immediate delay. In an example, if two instant endpoint algorithms have the same fidelity, then there is a small delay. The late endpoint algorithm will be ranked higher. In another example, if the two endpoint algorithms have the same immediate delay, the endpoint algorithm with higher fidelity will have a higher rank.

Referring back to Figure 2, in the next step 208, the immediate endpoint algorithm is moved to production. In one embodiment, the instant score algorithm with the highest score is automatically moved to production. In another embodiment, the instant endpoint algorithm moved to production is controllable by the user, thereby allowing the user to select the endpoint algorithm that best meets their needs. In an example, immediate latency is a concern for component manufacturers. To this end, the component manufacturer would prefer a less robust endpoint algorithm if it provides a shorter delay.

Empirical evidence shows that by automating the process, the best real-time endpoint algorithm can be implemented in minutes. In addition, because the algorithmic engine is configured to perform analysis with minimal human input, the procedure for constructing the endpoint algorithm can now be performed by non-expert users. Thus, if the method fails to produce an acceptable endpoint algorithm list for the known endpoint domain, the user can quickly redefine the endpoint field and pass it back to the algorithm engine to generate a new endpoint algorithm list in minutes.

4 presents a simplified flow diagram of an embodiment of the present invention for implementing an instant endpoint algorithm in a production environment.

A recipe is executed in a first step 402.

In the next step 404, data is acquired during substrate processing by a set of sensors.

In the next step 406, the endpoint algorithm is used in situ to analyze the data to identify the end of the process. In one embodiment, the data is analyzed using a computing engine. Because it is possible to collect a large amount of data, the computing engine is configured as a high-speed processing module that processes large amounts of data. The data is sent directly from the sensor without going through the manufacturing master or even the process module controller. An example of an analytical computer suitable for performing this analysis is described in U.S. Patent Application Serial No. 12/555,674, filed on Jan.

At next step 408, the system makes a determination as to whether to identify the end point.

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

However, if the endpoint has been identified, the recipe is stopped at the next step 410.

As can be appreciated from the foregoing, one or more embodiments of the present invention provide a method of identifying an optimal instant endpoint algorithm. By automated analysis, this approach essentially removes the need for expert users. A more robust endpoint algorithm can be moved into production via the methods as described herein. And because the time required to build an endpoint algorithm has been greatly reduced, updating or constructing a new endpoint algorithm is no longer a resource-intensive and time-consuming task.

Although the invention has been described in terms of several preferred embodiments, modifications, alternatives, and equivalents are possible within the scope of the invention. While the examples are provided herein, the examples are intended to be illustrative and not restrictive. Moreover, even though the document uses the end point as an example throughout, the present invention can also be used for a change point, which is a signal change event that occurs during processing.

Moreover, for the sake of convenience, the headings and the inventive content are provided herein, but they are not intended to be used in the scope of the patent application scope herein. In addition, the abstract is written in a highly simplified form and is provided for convenience, and thus is not intended to constitute or limit the overall invention, which is set forth in the claims. If the term "group" is used herein, the vocabulary is intended to have a general understanding of the mathematical meaning, including zero, one, or more than one component. It should be noted that there are many alternative ways of implementing the methods and apparatus of the present invention. Therefore, the following claims are intended to cover all such modifications, alternatives, and equivalents, as long as they fall within the true spirit and scope of the invention.

102-116‧‧‧Steps

202-208‧‧‧Steps

302-330‧‧‧Steps

402-410‧‧‧Steps

502‧‧‧Initial data group

504‧‧‧Signal uniform fragment

506A, 506B, 508-524‧‧‧ data groups

The present invention is illustrated by way of non-limiting example in the accompanying drawings, in which like reference numerals refer to the like elements, and in which: FIG. 1 shows a simple method of constructing an end point algorithm.

2 presents a simplified flow diagram of an embodiment of the present invention depicting a method of constructing an end point algorithm.

3A and 3B present a simplified flow diagram of an embodiment of the present invention depicting the steps and algorithm engine for performing the discovery of the best endpoint algorithm.

4 presents a simplified flow diagram of an embodiment of the present invention for implementing the best endpoint algorithm in a production environment.

Figure 5 shows a block diagram of an embodiment of the present invention, depicting the data set to be the best An example of a list of endpoint algorithms.

202-208. . . step

Claims (14)

A method for automatically identifying an optimal endpoint algorithm for identifying a process endpoint during processing of a substrate in a plasma processing system, comprising: during processing of at least one substrate in the plasma processing system Sensors receive sensor data, wherein the sensor data includes a plurality of signal streams from a plurality of sensor tubes, the sensor data is collected from more than one substrate; a endpoint field, wherein the endpoint field is a rough period at which an end point of the process is expected to occur; analyzing the sensor data to generate a set of possible endpoint feature signatures; converting the set of possible endpoint feature tokens into a set of optimal endpoint algorithms; And introducing an optimal endpoint method of the set of optimal endpoint algorithms into the production environment, wherein the analyzing of the sensor data comprises performing linear adaptation on the sensor data to be from the plurality of signal streams Each of the incoming signal streams is divided into segments based on the time interval. For example, the method of automatically identifying an optimal endpoint algorithm in claim 1 of the patent scope, wherein each segment of the plurality of segments is uniform. A method of automatically identifying an optimal endpoint algorithm as in claim 1 wherein the analysis of the sensor data includes the sensing for generating the first set of possible endpoint feature tokens of the set of possible endpoint feature tokens Data, calculating a first set of slopes and a first set of corresponding slope noise values, wherein a slope and a corresponding slope noise value are calculated for each of the plurality of segments, and a slope variation of the slope is calculated to obtain from the plurality of segments A set of high contrast signals is identified in the signal stream, wherein the set of high contrast signals has a high slope variation, and a continuous wavelength having a similar slope variation becomes a set of signal wavelength bands, the high contrast signal is graded, and the set of signal wavelengths is set. Band grading, and by applying a set of classification features to at least a portion of the high contrast signal and the set of letters The first wavelength endpoint identifies the first set of possible endpoint feature signatures, wherein the set of classification features includes at least one of a peak, a trough, and an inflection point. A method for automatically identifying an optimal endpoint algorithm as in claim 3, wherein the analysis of the sensor data comprises combining by a second set of possible endpoint feature signatures for the set of possible endpoint feature signatures Performing a multivariate analysis to generate a set of normalized slopes and a set of normalized corresponding slope noise values, the slope of the first set of slopes corresponding to the slope of the slope and the first set of corresponding slope noise values, to calculate A slope variation of the normalized slope of the set to identify a normalized signal from the plurality of signal streams, wherein the normalized signal has a high slope and a low variation, and a continuous wavelength having a similar slope variation becomes a set of standardized signal wavelength bands, The normalized signal is categorized to classify the set of normalized signal wavelength bands and apply a set of classification features to the high contrast signal and a set of signal wavelength bands to the ratio of the normalized signal to the set of normalized signal wavelength bands to produce the The second set of possible endpoint feature tokens. For example, the method of automatically identifying an optimal end point algorithm in claim 3, wherein converting the set of possible end point feature tokens into the set of best end point algorithms includes the first possible end point feature mark of the set of possible end point feature marks During a shape and time period similar to the second possible endpoint feature token, in conjunction with the first possible endpoint feature token and the second possible endpoint feature token, a robustness test is performed to remove possible from the set of possible endpoint feature tokens Non-repetitive endpoint feature signatures, performing a multivariate correlation analysis to identify a set of optimal endpoint signatures for the set of possible endpoint feature tokens, transforming the set of optimal endpoint signatures into a set of instants with minimal immediate delay An end point algorithm, wherein the immediate delay is based on filter delay, by performing at least one of: generating the set of best endpoint algorithms to remove an immediate endpoint algorithm having a corresponding immediate delay greater than a predetermined threshold, and if so The immediate endpoint algorithm cannot pass a robustness test, deletes the immediate endpoint algorithm, and ranks the best endpoint algorithms of the set of best endpoint algorithms based on the fidelity rate and the immediate delay. At least one of them. For example, the method for automatically identifying an optimal end point algorithm in claim 4, wherein converting the set of possible end point feature tokens into the set of best end point algorithms includes the first possible end point feature mark of the set of possible end point feature marks During a shape and time period similar to the second possible endpoint feature token, in conjunction with the first possible endpoint feature token and the second possible endpoint feature token, a robustness test is performed to remove possible from the set of possible endpoint feature tokens Non-repetitive endpoint feature signatures, performing a multivariate correlation analysis to identify a set of optimal endpoint feature signatures for the set of possible endpoint feature tokens, transforming the set of optimal endpoint signatures into a set of immediate endpoint calculus with minimal immediate delay The method, wherein the immediate delay is based on a filter delay, by performing at least one of: generating the set of optimal endpoint algorithms to remove an immediate endpoint algorithm having a corresponding immediate delay greater than a predetermined threshold, and if the immediate endpoint The algorithm cannot pass a robustness test, delete the instant endpoint algorithm, and Each of the best endpoint algorithms of the set of best endpoint algorithms is categorized, wherein the score is based on at least one of a fidelity rate and the immediate delay. For example, the method for automatically identifying an optimal endpoint algorithm in claim 1 of the patent scope, wherein the introduction of the optimal endpoint algorithm is based on grading and a set of user definitions At least one of the conditions. A method of identifying an endpoint during processing of a substrate in a processing chamber, comprising: executing a recipe on a substrate; receiving processing data from a set of sensors during substrate processing; applying an optimal endpoint algorithm Analyzing the processing data; identifying a process endpoint; and stopping the substrate processing, wherein the optimal endpoint algorithm is constructed by the following description: during processing of the at least one substrate in the plasma processing system, from a plurality of sensors Receiving sensor data, wherein the sensor data includes a plurality of signal streams from a plurality of sensor tubes, the sensor data being collected from more than one substrate; identifying a destination region, wherein The endpoint is a rough period at which an end of the process is expected to occur; the sensor data is analyzed to generate a set of possible endpoint feature tokens; the set of possible endpoint feature tokens is converted to a set of optimal endpoint algorithms; and the group is introduced An optimal endpoint method of the best endpoint algorithm into the production environment, wherein the analysis of the sensor data includes performing linear adaptation of the sensor data to Each signal stream from the plurality of signal streams based on the time interval is divided into several segments. A method of identifying an endpoint during processing of a substrate in a processing chamber as in claim 8 wherein the analysis of the sensor data is performed to generate a first set of possible endpoint signatures for the set of possible endpoint signatures Included, for the sensor data, calculating a first set of slopes and a first set of corresponding slope noise values, wherein a slope and a corresponding slope noise value are calculated for each of the plurality of segments, and a slope variation of the slope is calculated, Identifying a set of high contrast signals from the plurality of signal streams, wherein the set of high contrast signals has a high slope variation, and a continuous wavelength having a similar slope variation becomes a set of signal wavelength bands, Classifying the high contrast signal, classifying the set of signal wavelength bands, and identifying the first set of possible end point signatures by applying a set of classification features to at least a portion of the high contrast signal and the set of signal wavelength bands, The group classification feature includes at least one of a peak, a trough, and an inflection point. A method of identifying an endpoint during processing of a substrate in a processing chamber as in claim 9 wherein the analysis of the sensor data is performed to generate a second set of possible endpoint signatures for the set of possible endpoint signatures The performing multivariate analysis is performed by combining a slope scaled by respective slope noise values of the first set of slopes and the first set of corresponding slope noise values to generate a set of normalized slopes and a set of normalized respective slopes a noise value, a slope variation of the normalized slope of the set is calculated to identify a normalized signal from the plurality of signal streams, wherein the normalized signal has a high slope and a low variation, and the continuous wavelengths with similar slope variations become a set of normalization a signal wavelength band, classifying the normalized signal, classifying the set of normalized signal wavelength bands, and applying a set of classification features to the high contrast signal and a set of signal wavelength bands to the normalized signal and the set of normalized signal wavelength bands Ratio to generate the second set of possible endpoint feature signatures. A method of identifying an end point during processing of a substrate in a processing chamber as in claim 9 wherein converting the set of possible endpoint feature tokens into the set of optimal endpoint algorithms includes the first possible feature of the set of endpoint characteristics A possible end point feature mark has a similar shape and time period with the second possible end point feature mark, in conjunction with the first possible end point feature mark and the second possible end point feature mark, performing a robustness test to mark from the set of possible end point features Remove possible non-repetitive endpoint feature tokens, Performing a multivariate correlation analysis to identify a set of optimal endpoint feature tokens of the set of possible endpoint feature tokens, converting the set of optimal endpoint feature tokens into a set of instant endpoint algorithms with minimal immediate delay, wherein the instant delay system Based on the filter delay, the set of best endpoint algorithms is performed by performing at least one of the following to remove an immediate endpoint algorithm having a corresponding immediate delay greater than a predetermined threshold, and if the instant endpoint algorithm fails to pass a robustness Testing, deleting the immediate endpoint algorithm, and classifying each of the best endpoint algorithms of the set of optimal endpoint algorithms, wherein the ranking is based on at least one of a fidelity rate and the immediate delay. A method of identifying an end point during processing of a substrate in a processing chamber, as in claim 10, wherein converting the set of possible endpoint feature signatures into the set of optimal endpoint algorithms includes the first possible feature of the set of endpoint characteristics A possible end point feature mark has a similar shape and time period with the second possible end point feature mark, in conjunction with the first possible end point feature mark and the second possible end point feature mark, performing a robustness test to mark from the set of possible end point features Removing a possible non-repetitive endpoint feature signature, performing a multivariate correlation analysis to identify a set of optimal endpoint feature tokens for the set of possible endpoint feature tokens, converting the set of optimal endpoint feature tokens into a group with minimal instant A delayed immediate endpoint algorithm, wherein the immediate delay is based on filter delay, by performing at least one of the following: generating the set of optimal endpoint algorithms to remove an immediate endpoint algorithm having a corresponding immediate delay greater than a predetermined threshold, And if the instant endpoint algorithm fails to pass a robustness test, delete the instant Algorithms, and the algorithms of the end of each set of the best best classification algorithms end, wherein the sub- The system is based on at least one of a fidelity rate and the immediate delay. A method of identifying an end point during processing of a substrate in a processing chamber as in claim 8 wherein the introduction of the optimal endpoint algorithm is based on at least one of a ranking and a set of user defined conditions . A method of identifying an end point during processing of a substrate in a processing chamber as in claim 8 of the patent application, wherein each of the plurality of fragments is uniform.