TW201830334A - Diagnostic methods for the classifiers and the defects captured by optical tools - Google Patents

Abstract

Wafer inspection with stable nuisance rates and defect of interest capture rates are disclosed. This technique can be used for discovery of newly appearing defects that occur during the manufacturing process. Based on a first wafer, defects of interest are identified based on the classified filtered inspection results. For each remaining wafer, the defect classifier is updated and defects of interest in the next wafer are identified based on the classified filtered inspection results.

Description

Diagnostic methods for classifiers and defects captured by optical tools

The invention relates to defect detection.

The evolution of the semiconductor manufacturing industry places higher demands on yield management and, in particular, measurement and inspection systems. Critical dimensions are shrinking and wafer sizes are increasing. Economy is driving the industry to shorten the time to achieve high yield and high value production. Therefore, minimizing the total time from detecting a yield problem to resolving it determines the return on investment for a semiconductor manufacturer. Manufacturing semiconductor devices such as logic and memory devices typically involves processing a semiconductor wafer using a larger number of manufacturing processes to form the various features and multiple levels of a semiconductor device. For example, lithography involves transferring a pattern from a reticle to a semiconductor manufacturing process that is a photoresist disposed on a semiconductor wafer. Additional examples of semiconductor manufacturing processes include, but are not limited to, chemical mechanical polishing (CMP), etching, deposition, and ion implantation. Multiple semiconductor devices can be manufactured in one configuration on a single semiconductor wafer and then separated into individual semiconductor devices. Algorithms can be used to detect defects on a wafer. When machine learning algorithms are used to generate defect classifiers and nuisance filters, algorithms tend to be viewed as black box solutions that have not been tuned or diagnosed. The evaluation of an inspection formula usually waits until it is observed that a new set of labeled data is received for the evaluation or instead some parts of the labeled data are not used and retained for verification. Both of these technologies are a waste of resources. When setting a test recipe, the overall performance evaluation can be achieved based on the quality of the data used to train the classifier and the ability of the classifier to learn and extract information from the data. If the quality of the data is poor and the actual defects and obstructions do not have a clear separation boundary, any classifier may fail. Two measures are used: discriminative power and reliability to evaluate the effectiveness of each formulation. There are many discriminative powers. A confusion matrix for training data, which consists of a set of conditional error rates. Based on these conditions, the error rate, the rate of return, and the rate of interference can be important to the semiconductor manufacturer. The rate of return is the ratio of the number of properly classified defects of interest (DOI) to the total number of DOIs in the wafer. The obstruction rate is the ratio of the number of obstructions classified as DOI to the total number of defects classified as DOI. A higher return rate and a lower nuisance rate mean a better formula. However, the obstruction rate and return rate can only be evaluated on training data sets containing actual data markers before. Reliability is a measure of how deterministic the classifier is in making decisions. It is a function of the posterior estimation implemented by the classifier. Previously, the reliability of the classifier was evaluated through the credibility calculation of each defect. Although discrimination and reliability can be important measures, discrimination and reliability can obscure reality when DOI and the underlying distribution have specific characteristics. This can be called a shadowing effect. The methods commonly used for classifier evaluation on Broadband Plasma (BBP) and Laser Scanning (LS) tools are based on the confusion matrix of the training set used to measure discriminative power and calculate the reliability of measurement Histogram. As can be seen in Figure 1 (a), the discriminative power is measured based on the training data only. If any special sampling method is used, the training data is biased. Credibility histograms have been used to measure reliability. There may be no information about the reliability of a DOI or a nuisance classification using this technology for a user. Using the confusion matrix of the training set is often insufficient to understand the nature of the recipe on the entire wafer. If defects in the training set have been selected in a specific way (usually achieved to reduce the number of defects used for scanning electron microscope (SEM) inspection and manual classification), the confusion matrix of the training set is biased toward those defects and not on the entire wafer One of the good estimators of classifier performance. Previous solutions retrained a binary classifier based on manual classification obtained during program monitoring (production sampling) (eg, obstruction vs. DOI). These previous solutions use updated classifiers to generate new DOI / obstruction separations for subsequent wafers and use new cells to generate production samples, which are then used to tune the next classifier. 50% of the previous solution samples were randomly sampled from the DOI bins of the nearest classifier, and the other 50% were randomly sampled from the entire cluster. Two samples are used to compare the statistical procedure control (SPC) of the two tests, and the second sample also provides a "subthreshold" defect for retraining the classifier. Another previous method for handling program / wafer changes relied on building a classifier from scratch and iteratively building a training set with the help of SEM automatic defect classification (ADC), and then producing production samples from the newly generated DOI bin. However, the need to build a classifier from scratch on each wafer has a high cost in terms of SEM tool time. In addition, the asserted facts used to train the BBP model are based on SEM ADCs without human verification, which potentially makes the asserted facts less reliable. Finally, this method does not take advantage of defects from previous wafers, and therefore increases the risk of insufficient data and instability during the training process. Without additional sampling, prior art could not find estimates of returns and obstructions (and for unlabeled data) across the wafer. As a result, users do not know how to tune the recipe can affect overall performance. The prior art also does not recognize shadowing effects. Therefore, a new defect detection technology and system are needed.

In a first embodiment, a system for detecting a defect of interest in a plurality of wafers is provided. The system includes: a central storage medium configured to store a plurality of classified inspection results and an initial defect classifier; a wafer inspection tool; an image data acquisition system; and a processor, which is stored with the central storage The media, the wafer inspection tool, and the image data acquisition system are in electronic communication. The processor is configured to execute instructions of a inspection engine, a sampling engine, and a tuning engine. The inspection engine instructs the processor to receive an inspection result of a first wafer from the wafer inspection tool. The sampling engine instructs the processor: retrieve the initial defect classifier from the central storage medium; filter the inspection results based on the initial defect classifier; and review the first from the image data acquisition system based on the filtered inspection results The location of interest on the wafer; classifying the filtered inspection results based on the initial defect classifier; storing the classified filtered inspection results in the central storage medium; and based on the classified experiences The filtering inspection results identify the defects of interest in the first wafer. The tuning engine instructs the processor to update the initial defect classifier based on the classified filtered inspection results stored in the central storage medium. For each remaining wafer, the inspection engine instructs the processor to receive the next wafer inspection result from the wafer inspection tool. For each remaining wafer, the sampling engine instructs the processor: to filter the inspection results of the next wafer based on the initial defect classifier; based on the filtered inspection results and historical analysis sampling of the next wafer to use The image data acquisition system inspects the position of interest on the next wafer; classifies the filtered inspection results of the next wafer based on the inspected positions of interest on the next wafer; The classified filtered inspection results of the next wafer are stored in the central storage medium; based on the classified filtered inspection results of the next wafer stored in the central storage medium, the processor is used Update the defect classifier; and identify the defect of interest in the next wafer based on the classified filtered inspection results of the next wafer. For each of the remaining wafers, the tuning engine may instruct the processor to update the defect with the processor based on the classified filtered inspection results of the next wafer stored in the central storage medium Classifier. The sampling engine may instruct the processor to perform the filtering step based on the updated defect classifier. The image data acquisition system can be a SEM inspection tool. The wafer inspection tool can perform a thermal scanner to capture inspection results. For example, the wafer inspection tool may be a wideband plasma inspection tool. The defect classifier can send the concerned defect data and obstruction data for retraining the defect classifier. The step of identifying a defect of interest may include: approaching classification boundary sampling of one of the nearest defect classifiers; obtaining information about classifier stability based on fluctuations in the defect classifier; observing movement of one of the classification boundaries; The predicted movement in the classification boundary identifies the defects of interest. The inspection results or inspected locations of interest may be stored in the central storage medium. In a second embodiment, a method for identifying a defect of interest in a plurality of wafers is provided. The method includes receiving a first wafer inspection result from a wafer inspection tool at a processor. The processor is used to filter the inspection results based on an initial defect classifier. Based on the filtered inspection results, an image data acquisition system is used to view the location of interest on the first wafer. The processor is used to classify the filtered inspection results based on the inspected locations of interest on the first wafer. The classified filtered inspection results are stored in a central storage medium. A defect of interest in the first wafer is identified based on the classified filtered inspection results. For each remaining wafer, the method includes receiving the next wafer inspection result from the wafer inspection tool at the processor. The processor is used to filter the inspection results based on the initial defect classifier. Based on the filtered inspection results and historical analysis samples of the next wafer, the image data acquisition system is used to view the location of interest on the next wafer. The processor is used to classify the filtered inspection results of the next wafer based on the reviewed locations of interest on the next wafer. The sorted filtered inspection results of the next wafer are stored in the central storage medium. The processor is used to update the defect classifier based on the classified filtered inspection results of the next wafer stored in the central storage medium. A defect of interest in the next wafer is identified based on the classified filtered inspection results of the next wafer. The image data acquisition system can be a SEM inspection tool. The wafer inspection tool can perform a thermal scanner to capture inspection results. For example, the wafer inspection tool may be a wideband plasma inspection tool. The defect classifier can send the concerned defect data and obstruction data for retraining the defect classifier. The step of identifying a defect of interest may include: approaching classification boundary sampling of one of the nearest defect classifiers; obtaining information about classifier stability based on fluctuations in the defect classifier; observing movement of one of the classification boundaries; and based on The predicted movement in the classification boundary identifies the defects of interest. For each of the remaining wafers, the method may include updating the defect classifier using the processor based on the classified filtered inspection results of the next wafer stored in the central storage medium. The filtering step may be performed based on the updated defect classifier. The inspection results or inspected locations of interest may be stored in the central storage medium. The step of updating the defect classifier based on the classified filtered inspection result devices stored in the central storage medium may include: estimating a rate of return based on a calculated training confusion matrix and based on the defect in the central storage medium The classifier, the classified filtered inspection results of the next wafer, and the estimated rate of return estimate an obstruction rate. The calculated training confusion matrix is based on the classified filtered inspection results of the next wafer stored in the central storage medium. The filtered inspection results may have at least two threshold values associated with the filtered inspection results. One of the at least two thresholds is directed to an inspection procedure and an inspection of defects. A second one of the at least two thresholds is less than the first threshold and is configured to capture secondary threshold defects during inspection.

Cross Reference to Related Applications This application claims U.S. Provisional Application No. 62 / 444,694 filed on January 10, 2017, U.S. Provisional Application No. 62 / 475,030 filed on March 22, 2017, and November 2017 The priority of U.S. Provisional Application No. 62 / 581,378, filed on the 3rd, is incorporated herein by reference. Although the claimed subject matter will be described in terms of specific embodiments, other embodiments (including those that do not provide all of the advantages and features set forth herein) are also within the scope of the invention. Various structural, logical, procedural, and electronic changes may be made without departing from the scope of the invention. Therefore, only the scope of the invention is defined with reference to the scope of the appended patents. The embodiments disclosed herein address new systems and methods for handling procedures and wafer instability in the early stages of an integrated circuit manufacturing process. An embodiment of the present invention is based on generating a small sample on a production batch in addition to production sampling, summing up the samples above several wafers and also establishing a new classifier and using the classifier to generate a new update on the next wafer. The idea of the sample. The embodiments disclosed herein may be particularly advantageous over existing methods for at least the following reasons. The presently disclosed systems and methods utilize supplemental (extended) samples produced using recently known program conditions and well adapted to return one of a superior classifier. Recently known procedural conditions and defects are far more useful for this purpose than random samples currently in use. The recently known procedural status is also a superior indication of a procedural change, and any new defects or defects exhibiting the greatest change will be effectively revealed in the sample. In other words, a valid incremental discovery with a small sample size results in smaller additional SEM inspection and classification costs. In addition, with additional monitoring of procedural instability and the ability of currently disclosed systems and methods to quantify the instability of the procedure, supplementary samples can be automatically tuned to match the status of such procedures. In general, the disclosed systems and methods allow more relevant wideband plasma inspection with more stable interference rates and DOI capture rates. The disclosed system and method allow for faster detection of new defects that occur during the manufacturing process and allows analysis of one of the stability of the manufacturing process. There are several ways to implement the currently disclosed systems and methods. One embodiment relies only on data from a central storage medium, and the system and method then use the classifier performance to manually classify defects in the remainder of the inspection. These embodiments make the classifier one wafer behind the wafer currently being inspected. Another embodiment adds the ability to update the classifier for the current wafer by performing sampling on a wafer defect inspection tool and then generating supplemental samples on a central storage medium. One advantage of this embodiment is that recent wafer conditions are also included in the classifier. Rates of return and obstructions can be estimated for data where actual signs are not available. Therefore, the expected value of the return rate and the obstruction rate can be provided. All estimates for technology display returns, obstructions, posteriors, and credibility are accurate or data have a shadow distribution. Algorithm-generated data (other than classification) can provide diagnostic information that cannot be obtained using manually generated classifiers such as the online defect combiner (iDO). iDO is an example of an algorithm that can instantly classify defects during inspection. A formula can be evaluated. These methods include an estimate of the rate of return; an estimate of the rate of nuisance; an assessment of the receiver operating curve (ROC) that can show the rate of return versus the rate of nuisance for fine-tuned formulations; and the detection of shadowing effects Are the estimates of reliability, rate of return, and nuisances trustworthy? ROC can be used to draw a curve between true positive rate and false positive rate. Instead of or in cooperation with ROC, you can use the DOI return (true positive rate) versus the obstruction rate (its non-false positive rate). Two outputs from the classifier can be used to build a diagnostic tool. First, the decision of the classification results provided by the classifier can be used. Second, each defect posterior can be used. There are different ways in which a classifier can find the posterior. Probability measurements of the distance or accuracy from the centroids of each category are two examples. To estimate the rate of return, the ratio of the number of correctly classified DOIs to the total number of DOIs in the training set can be used. This can be applied to the test set to find an estimate of the number of DOIs that may be lost in the test data. Assuming that there are two categories (DOI and obstruction), the confusion matrix appears as shown in Table 1. Table 1 S _nm is a collection of all defects that originally belonged to category m and were classified as category n. S _DD is a defect set that is classified as DOI and is actually a DOI. S _ND is a defect set that is classified as obstructive and is actually a DOI. S _DN is a defect set that is classified as DOI and is actually a nuisance. S _NN is a defect set that is classified as obstruction and is actually an obstruction. The return on the entire wafer is estimated in Equation 1. Equation 1 In Equation 1, | S | represents the size (cardinality) of the set S. To estimate the nuisance rate of unlabeled defects, the ratio of the cumulative posterior of the defects associated with the nuisances of defects classified as DOI to the total number of defects in the DOI bin can be used. Assuming two categories (DOI and nuisance), they are classified after a cluster of data showing test data (or any unlabeled data) in Table 2. S _D lines classified as a defective set of DOI. S _N is a defect set classified as an obstruction. Assuming that the hindrance class associated with the defect i is posterior p _i , the hindrance rate will be calculated as shown in Equation 2. Equation 2 S _D is a defect set classified as DOI. p _i is the posterior probability of the nuisance category associated with defect i. | S _D | Represents the size of the set SD. The rate of return can increase with greater nuisance. This can be achieved, for example, by moving the cutting line in the confidence histogram and changing the category code for defects with lower confidence. The rate of return and obstruction can be evaluated for all possible values of the cutting line. Three plots can then be shown, three examples of which are shown in FIG. 2. Graph (a) in Figure 2 shows the rate of return versus the cut line value. Graph (b) in Fig. 2 shows the nuisance ratio versus the cut line value. Panel (c) in Figure 2 shows the ROC. A ROC can be a useful representation of the performance of a classifier on a given data set. The user can find something that will be a nuisance to a desired capture rate, and vice versa. With these curves, a user can decide whether the value of the cutting line is trustworthy. In classification, well-separated distributions can be distributions with short overlap, as shown in Figures 3 (a), 3 (b), and 3 (c). Data can be well separated like drawing a clear boundary between two distributions, as shown in Figures 3 (a) and 3 (b). The distribution can be well separated and have multiple regions in space and separated using multiple boundaries, as shown in Figure 3 (c). Most classifiers can learn this situation. In this scenario, the performance of the classifier is average. This probability density function (PDF) is what they normally appear on a wafer, but this is not always the case. A large part of a distribution may have been obscured. The shadowing effect is a situation when a large part of the distribution of one category is under the PDF of another category. This situation can occur as an error during manual or automatic labeling or because it does not have good attributes to distinguish shaded parts from other categories. Figures (a) and (b) in Figure 4 are two examples of this situation. The detection of the first case ((a) in Fig. 4) is relatively simple, because the accuracy of determining a category can be determined only by observing the training confusion matrix. Detecting the second case ((b) in Figure 4) is more difficult. This situation can mislead a user about the data on the wafer, and no matter which classifier is used, most of a class will not be detected. The misclassification here is not due to the poor performance of the classifier, but can be due to the poor quality of the feature or logo. To detect this, a training set can be used to train a classifier. Then, the training can be sorted to set the defects obtained from the classifier in ascending order from the confidence value. An empty pool can be generated and defects can be added to the pool one by one from the lowest confidence to the highest confidence. After each defect is added, the defects in the pool can be calculated to the confusion matrix and the accuracy of the category and the number of defects in the pool can be saved. The accuracy of each category can be defined as the number of correctly classified defects in that category versus the total number of defects from that category. After using all the defects in the training set, the accuracy can be compared to the number of defects in the pool. An example of this algorithm is shown in Figure 5. For a general defect distribution on a wafer, the accuracy of all categories in the pool is expected to increase or remain constant as the number of defects increases. Although other basic principles are possible, a new defect in the pool may have greater or equal credibility than a previous defect in the pool. The plots in (a) and (b) of Figure 6 show this for two different wafers. The plot in (a) of FIG. 6 is from a wafer without a shadowed DOI, and both the DOI and the obstruction accuracy improve with the number of defects. However, the plot (b) in FIG. 6 shows that one of the wafers with a shadowing effect was observed for a DOI category. The DOI bin does not improve with the number of defects. Its indication adds a high-confidence defect, but these are incorrectly classified, which is an indication of shadowing effects. Details of one embodiment of a method are shown in the flowchart of FIG. FIG. 7 shows a flowchart of an algorithm for estimating the obstruction rate and capture rate and the detection of shadow effects. The training set is used to generate the classifier. The classifier is applied to the defects in the test set. The classifier is then used to evaluate the credibility and posterior of all defects (both in the training and test sets). Use a posteriori to achieve an estimate of the obstruction rate. The confusion matrix obtained from the training set is used to estimate the capture rate. Finally, a check is implemented to find out if the data is under shadow effects. If it is not under the shadow effect, the estimation is reliable. FIG. 8 is a flowchart of a method 100 for identifying a defect of interest in a plurality of wafers. At 101, such as receiving a first wafer inspection result from a wafer inspection tool at a processor, the tool may be a BBP tool or another inspection device. At 102, such as using a processor to filter inspection results based on an initial defect classifier. At 103, such as using an image data acquisition system to review the location of interest on the first wafer based on the filtered inspection results. The image data acquisition system may be an SEM inspection tool or another measurement, inspection or metrology tool. At 104, such as using a processor to classify the filtered inspection results based on the inspected locations of interest on the first wafer. At 105, the classified test results are stored in a central storage medium. At 106, such as using a processor to identify defects of interest based on the classified filtered inspection results. Filtered inspection results, such as for each sampled wafer, can be kept separate. At 107, for each remaining wafer, such as receiving the inspection results of the next wafer from the wafer inspection tool at a procedure. At 108, such as using a processor to filter inspection results based on the initial defect classifier. At 109, such as using an image data acquisition system to review the location of interest on the next wafer based on filtered inspection results and historical analysis samples. At 110, such as using a processor to sort the filtered inspection results based on the inspected location of interest on the next wafer. At 111, the classified filtering results are stored in a central storage medium. At 112, such as using a processor to update the defect classifier based on the classified results stored in the central storage medium. At 113, such as using a processor to identify defects of interest on the next wafer based on the sorted filtered inspection results of the next wafer. The next wafer may refer to the next sequential wafer, but may also mean a second, third, fourth, fifth, or later wafer. In the method 100, identifying a defect of interest may include classifying a boundary sample close to one of a nearest defect classifier. Information on classifier stability can be obtained based on fluctuations in the defect classifier. The movement of classification boundaries can be predicted. The defect of interest can be identified based on predicted movements in the classification boundary. The wafer inspection tool may execute a thermal scanner to capture inspection results using the method 100. The defect classifier can send the concerned defect data and obstruction data for retraining the defect classifier. For each remaining wafer, a defect classifier may be updated, such as using a processor, based on the classified results stored in a central storage medium. The filtering step may be performed based on the updated defect classifier. Inspection results or inspected locations of interest may be stored in a central storage medium. Updating the defect classifier based on the classified results stored in the central storage medium may include estimating a rate of return based on a calculated training confusion matrix. The calculated training confusion matrix may be based on a filtered filtered inspection result of the next wafer stored in the central storage medium. An obstruction rate can be estimated based on the defect classifier in the central storage medium, the filtered inspection results of the classification of the next wafer, and the estimated rate of return. These steps may be performed by a processor. A confidence value can also be calculated based on the initial defect classifier. In this example, updating the defect classifier based on the classified results stored in the central storage medium may further include detecting a shadowing effect based on the defect classifier and the calculated confidence value. The filtered inspection results may have at least two threshold values associated with the filtered inspection results. One of the at least two thresholds is for one of the inspection procedures and inspections that can be used for defects. One of the at least two thresholds is less than the first threshold and can be configured to capture secondary threshold defects during inspection. This enables sampling on both sides of the threshold to allow the classification boundary to be changed in both directions. This technology offers several advantages. It provides a fast rate of return estimator. Generally, the estimation of the rate of return is an expensive and / or imprecise task. A user must sample a large number of defects from an obstructing cell, review them using a tool (eg, a SEM tool), classify them, and try to provide an estimate of the number of DOIs in the obstructing cell. This method is not feasible most of the time because of the large number of defects in the DOI bin. The embodiments disclosed herein do not require any samples, which makes them extremely fast. A faster estimate of the nuisance is also provided. Usually to estimate the nuisance rate, the user randomly samples from the DOI bin and then SEM inspects them and classifies them. The techniques disclosed herein can be used to remove this extra time for sampling, SEM review and classification. The estimation of the ROC curve across the wafer can be a useful tool for semiconductor manufacturers to tune recipes and identify the best conditions for inspection to give the desired output. The disclosed technique also provides one of the detection methods for shadowing effects. Identifies the inseparable part of the distribution in the data. This phenomenon usually occurs due to errors during manual labeling, poor SEM image quality, or lack of strong features. FIG. 9 is a block diagram of a system 200 for detecting defects of interest in a plurality of wafers. The system 200 includes a wafer inspection tool 201, an image data acquisition system 204, a central storage medium 203, and a processor 202. The image data acquisition system 204 may be a SEM viewing tool. The wafer inspection tool 201 may be a BBP inspection tool, which may be configured to perform a thermal scan to capture inspection results. The wafer inspection tool 201 may also be an LS tool or an unpatterned wafer surface inspection system, such as Surfscan SPx manufactured by KLA-Tencor Corporation. The central storage medium 203 is configured to store a plurality of classified inspection results and an initial defect classifier. The processor 202 is in electronic communication with the central storage medium 203, the wafer inspection tool 201, and the image data acquisition system 204. The processor 202 is configured to execute instructions of a inspection engine, a sampling engine, and a tuning engine. The inspection engine instructs the processor to receive an inspection result of a first wafer from the wafer inspection tool. The sampling engine instructs the processor: retrieve the initial defect classifier from the central storage medium; filter the inspection results based on the initial defect classifier; and review the first from the image data acquisition system based on the filtered inspection results The location of interest on the wafer; classifying the filtered inspection results based on the initial defect classifier; storing the classified filtered inspection results in the central storage medium; and based on the classified experiences The filtering inspection results identify the defects of interest in the first wafer. The tuning engine instructs the processor to update the initial defect classifier based on the classified results stored in the central storage medium. For each remaining wafer, the inspection engine instructs the processor to receive the next wafer inspection result from the wafer inspection tool. The sampling engine instructs the processor: to filter the inspection results based on the initial defect classifier; to use the image data acquisition system to sample and view the location of interest on the next wafer based on the filtered inspection results and historical analysis; based on the inspected location on the next wafer Classify the filtered inspection results by attention; store the classified results in a central storage medium; use a processor to update the defect classifier based on the classified results stored in the central storage medium; and classify based on the next wafer The filtered inspection results identify defects of interest on the next wafer. For each remaining wafer, the tuning engine may instruct the processor to use the processor to update the defect classifier based on the classified results stored in the central storage medium. The sampling engine may instruct the processor to perform a filtering step based on the updated defect classifier. The number of results and wafers used to update the defect classifier can be determined by the algorithm and can be controlled by settings. These quantities may depend on usage and inspection. For research and development applications, only a few recent wafers can be used. In a more mature high-volume manufacturing process, training data can come from more wafers. It can be limited by time and information adequacy. The defect classifier can send the concerned defect data and obstruction data for retraining the defect classifier. The steps of identifying a defect of interest may include: approaching classification boundary sampling of one of the nearest defect classifiers; obtaining information about classifier stability based on fluctuations in the defect classifier; observing one of the classification boundaries moving; Defects of Predictive Motion Recognition are of interest. Observation-movement can be performed on some recent wafers. The inspection results or inspected locations of interest may be stored in the central storage medium 203, which may include a database. In a particular example, a central storage medium 203 may store the classified defects and the remainder of the inspection cluster. After adding each new data to the database, a tuning and analysis engine can operate on the stored data. A sampling engine can retrieve the nearest classifier from the central server to identify the most appropriate defect. This is achieved by one or more of the following techniques. First, the nearest classifier is used to sample the classification boundaries close to the model (as two sides of the boundary). Second, use classifier stability information obtained from recent classification fluctuations on the wafer. Third, guide most of the sample to the side of the classification boundary that is most likely to be in the direction of the boundary move. One embodiment relies only on a central storage medium and then uses the performance of the classifier and manual classification of defects for the remainder of the inspection. This configuration holds the classifier behind one wafer. Another embodiment adds the ability to update the model of the current wafer by performing sampling on the wafer defect inspection tool and then generating supplemental samples for central storage, which means that it also includes recent wafer conditions. Two examples are shown in FIGS. 10 and 11. In Figures 10 and 11, the use of one of the functions of a standard frustration filter is to hinder the DOI classifier to make the test run hotter. This reservation hinders defects on both sides of the DOI boundary for retraining. Stability information from historical analysis sampling settings and recent classifiers is used for sampling. Although the processor 202 and the central storage medium 203 are shown as separate, these may be part of the same control unit. Both the processor 202 and the central storage medium 203 may be part of the wafer inspection tool 201 or the image data acquisition system 204 or another device. In an example, the processor 202 may be an independent control unit or a centralized quality control unit. Multiple processors 202 and / or central storage media 203 may be used. For example, three processors 202 may be used for the inspection engine, the sampling engine, and the tuning engine. The processor 202 may be implemented in practice by any combination of hardware, software, and firmware. Likewise, its functions as described herein may be performed by a unit, or divided between different components, each of which may then be implemented by any combination of hardware, software, and firmware. The code or instructions for the processor 202 to implement various methods and functions may be stored in a controller-readable storage medium (such as one of the central storage medium 203 or other memory). The processor 202 and the central storage medium 203 may be coupled to the components of the system 200 in any suitable manner (eg, via one or more transmission media, which may include wired and / or wireless transmission media) such that the processor 202 and the central storage medium 203 may receive output generated by the system 200. The processor 202 may be configured to use the output to perform several functions. The processor 202 and the central storage medium 203, (some) other systems, or (some) other subsystems described herein may be part of various systems, including a personal computer system, an image computer, a host computer system, a workstation, a network device, the Internet Network equipment or other devices. The subsystem (s) or system (s) may also include any suitable processor known in the art, such as a parallel processor. In addition, the subsystem (s) or system (s) may include a platform with high-speed processing and software as a stand-alone or networked tool. If the system includes more than one subsystem, different subsystems can be coupled to each other so that images, data, information, instructions, etc. can be sent between the subsystems. For example, a subsystem may be coupled to additional subsystem (s) by any suitable transmission medium, which may include any suitable wired and / or wireless transmission medium known in the art. Two or more of these subsystems can also be effectively coupled via a shared computer-readable storage medium (not shown). An additional embodiment relates to a non-transitory computer-readable medium that stores program instructions executable on a controller for performing a computer-implemented method according to one of the embodiments disclosed herein. In particular, the processor 202 may be coupled to one of a central storage medium 203 or other electronic data storage medium having a non-transitory computer-readable medium containing program instructions executable on the processor 202. A computer-implemented method may include any step (s) of any method (s) described herein. For example, the processor 202 may be programmed to perform some or all of the steps of FIG. 8. The memory in the central storage medium 203 or other electronic data storage medium may be a storage medium, such as a magnetic or optical disk, a magnetic tape, or any other suitable non-transitory computer-readable medium known in the art. Program instructions may be implemented in any of a variety of ways, including, in particular, program-based technology, component-based technology, and / or object-oriented technology. For example, program instructions may be implemented using ActiveX controls, C ++ objects, JavaBeans, Microsoft Foundation Classes (MFC), SSE (Streaming SIMD Extension), or other technologies as needed. Each of the steps of the method may be performed as described herein. The method may also include any other step (s) that may be performed by the controller and / or computer subsystem (s) or system (s) described herein. The steps may be performed by one or more computer systems, which may be configured according to any of the embodiments described herein. In addition, the methods described above may be performed by any of the system embodiments described herein. Although the invention has been described with respect to one or more specific embodiments, it will be understood that other embodiments of the invention can be made without departing from the spirit and scope of the invention. Therefore, the present invention is deemed to be limited only by the scope of patent application of the accompanying invention and its reasonable interpretation.

100‧‧‧ Method

101‧‧‧ steps

102‧‧‧step

103‧‧‧step

104‧‧‧step

105‧‧‧ steps

106‧‧‧ steps

107‧‧‧ steps

108‧‧‧ steps

109‧‧‧step

110‧‧‧step

111‧‧‧ steps

112‧‧‧step

113‧‧‧step

200‧‧‧ system

201‧‧‧ Wafer Inspection Tools

202‧‧‧Processor

203‧‧‧Center Storage Media

204‧‧‧Image data acquisition system

For a more complete understanding of the nature and purpose of the present invention, reference should be made to the following implementations in conjunction with the drawings, where: Figure 1 contains the prior art flowcharts (a) and (b); Figure 2 contains a return versus cutting line, respectively Graph (a) of the curve, graph (b) of a nuisance ratio versus cutting line curve, and (c) of a ROI curve (c); Figure 3 contains distributions (a), (b), and (c); 4 includes distributions (a) and (b); Figure 5 is a flowchart of an embodiment of a shadow detection algorithm according to the present invention; Figure 6 includes a general wafer (a) and a masked wafer ( b) a graph of accuracy versus the number of defects in the pool; Figure 7 is a flowchart of an embodiment of a diagnostic model according to the present invention; Figure 8 is a flowchart of an embodiment of the present invention; FIG. 9 is a block diagram of a system according to the present invention; FIG. 10 is a diagram of a dynamic classifier with dynamic sampling and stability analysis according to the present invention; and FIG. 11 is a diagram of dynamic sampling and stability according to the present invention. A graph of static classifiers for sexual analysis.

Claims (18)

A system for detecting defects of interest in a plurality of wafers, comprising: a central storage medium configured to store a plurality of classified inspection results and an initial defect classifier; a wafer inspection tool; An image data acquisition system; and a processor in electronic communication with the central storage medium, the wafer inspection tool, and the image data acquisition system, the processor is configured to execute the following instructions: an inspection engine that instructs the A processor: receiving an inspection result of a first wafer from the wafer inspection tool; a sampling engine instructing the processor: retrieving the initial defect classifier from the central storage medium; filtering the initial defect classifier based on the initial defect classifier Waiting for inspection results; inspecting the location of interest on the first wafer from the image data acquisition system based on the filtered inspection results; classifying the filtered inspection results based on the initial defect classifier; classifying the filtered inspection results The filtered inspection results are stored in the central storage medium; and the first wafer is identified based on the classified filtered inspection results A defect of interest; a tuning engine instructing the processor: updating the initial defect classifier based on the classified filtered inspection results stored in the central storage medium; wherein for each remaining wafer: the inspection The engine instructs the processor: to receive the inspection result of the next wafer from the wafer inspection tool; the sampling engine instructs the processor: to filter the inspection results of the next wafer based on the initial defect classifier; The filtered inspection results and historical analysis samples of a wafer use the image data acquisition system to check the position of interest on the next wafer; based on the viewed positions of interest on the next wafer, Sort the filtered inspection results of the next wafer; store the sorted filtered inspection results of the next wafer in the central storage medium; based on the next stored in the central storage medium The classified filtered inspection results of the wafer using the processor to update the defect classifier; and the classified filtered based on the next wafer Test results to identify the focus of the next wafer defect by. If the system of claim 1, wherein for each of the remaining wafers: the tuning engine instructs the processor: based on the classified filtered inspection results of the next wafer stored in the central storage medium The processor is used to update the defect classifier; wherein the sampling engine instructs the processor to perform the filtering steps based on the updated defect classifier. The system of claim 1, wherein the image data acquisition system is a SEM inspection tool. The system of claim 1, wherein the wafer inspection tool performs a thermal scan to capture inspection results. The system of claim 1, wherein the defect classifier sends the defect information of interest and the obstruction data for retraining the defect classifier. If the system of claim 1, wherein the step of identifying the defect of interest includes: approaching classification boundary sampling of one of the nearest defect classifiers; obtaining information about classifier stability based on fluctuations in the defect classifier; observing the classification boundary One of the movements; and identifying the defects of interest based on predicted movements in the classification boundary. If the system of item 1 is requested, it further comprises storing the inspection results or the viewed positions of interest in the central storage medium. The system of claim 1, wherein the wafer inspection tool is a broadband plasma inspection tool. A method for identifying a defect of interest in a plurality of wafers, comprising: receiving an inspection result of a first wafer from a wafer inspection tool at a processor; using the processor based on an initial defect classifier Filtering the inspection results; using an image data acquisition system to inspect the location of interest on the first wafer based on the filtered inspection results; using the process based on the inspected locations of interest on the first wafer The device classifies the filtered inspection results; stores the classified filtered inspection results in a central storage medium; identifies the defect of interest in the first wafer based on the classified filtered inspection results And for each remaining wafer: receiving the next wafer inspection results from the wafer inspection tool at the processor; using the processor to filter the inspection results based on the initial defect classifier; based on the next wafer The filtered inspection results and historical analysis samples use the image data acquisition system to view the location of interest on the next wafer; based on the next wafer The inspected positions of interest above use the processor to classify the filtered inspection results of the next wafer; store the classified filtered inspection results of the next wafer in the center In the storage medium; using the processor to update the defect classifier based on the sorted filtered inspection results of the next wafer stored in the central storage medium; and the sorted based on the next wafer The filtered inspection results identify the defects of interest in the next wafer. The method of claim 9, wherein the image data acquisition system is a SEM inspection tool. The method of claim 9, wherein the wafer inspection tool performs a thermal scan to capture inspection results. The method of claim 9, wherein the defect classifier sends the defect information of interest and the obstruction data for retraining the defect classifier. The method of claim 9, wherein the step of identifying the defect of interest includes: approaching classification boundary sampling of one of the nearest defect classifiers; obtaining information about classifier stability based on fluctuations in the defect classifier; observing the classification boundary One of the movements; and identifying the defects of interest based on predicted movements in the classification boundary. The method of claim 9, further comprising, for each of the remaining wafers: using the processor to update based on the classified filtered inspection results of the next wafer stored in the central storage medium The defect classifier; wherein the filtering step is performed based on the updated defect classifier. If the method of item 9 is requested, it further comprises storing the inspection results or the viewed positions of interest in the central storage medium. The method of claim 9, wherein the wafer inspection tool is a broadband plasma inspection tool. The method of claim 9, wherein the step of updating the initial defect classifier based on the classified filtered inspection results stored in the central storage medium includes: estimating a rate of return based on a calculated training confusion matrix, wherein the The calculated training confusion matrix is based on the classified filtered inspection results of the next wafer stored in the central storage medium; and based on the defect classifier, the next wafer in the central storage medium The classified filtered inspection results and the estimated rate of return estimate an obstruction rate. The method of claim 9, wherein the filtered inspection results have at least two thresholds associated with the filtered inspection results, and one of the at least two thresholds is targeted for monitoring One of the procedures and defects is inspected, and one of the at least two thresholds is less than the first threshold and is configured to capture a sub-threshold defect during the inspection.