KR100679813B1 - Statistical method of screening potential problem process of semiconductor fabrication - Google Patents

Abstract

A statistical detection method of potential problem process in a semiconductor fabrication is provided to accurately derive a process to be improved by analyzing physical in-line data systematically. In-line data, PCM(Pulse Code Modulation) data, and probe test data are collected(10). A probability distribution of the probe test data is obtained, and is stratified into groups of different slopes to divide a good product and a bad product(20). The in-line data, the PCM data, and the probe test data are aligned as data for a chip corresponding to shot(30). A value P indicative of a correlation coefficient and significance level by analyzing the data, and a significance factor is selected on the basis of the value P(40). The significance factor is derived with reference to the value P(50).

Description

Statistical Method of Screening Potential Problem Process of Semiconductor Fabrication

1 is a flow chart illustrating a statistical detection method of a semiconductor manufacturing process according to an embodiment of the present invention.

2A and 2B are exemplary diagrams showing a normal distribution and stratification of probe test data.

3A and 3B are exemplary diagrams illustrating a table in which shot unit data is arranged in chip units.

4A and 4B are exemplary diagrams showing a result of correlation analysis on data.

5A to 5C are exemplary views showing the results of multiple linear regression analysis.

FIG. 5D is an illustration showing the results of a residual plot for the regression equation of FIG. 5C. FIG.

6A and 6B are exemplary diagrams showing the results of binary logistic regression.

7A-7C illustrate exemplary results of a reverse sample test.

8A and 8B are exemplary diagrams showing verification results through experiments.

The present invention relates to a method for statistical analysis of semiconductor products, and more particularly, to a method for statistical detection of potential problem processes in semiconductor manufacturing.

Systematic and statistical analytical methods are very necessary in the semiconductor industry where technology competition is intense. In particular, quality and yield are directly related to a company's profits, so it is important to quickly and accurately identify potential causes for it.

In order to improve the characteristics of semiconductor products manufactured by numerous processes, physical in-line data is generally managed in a unit process. In practice, however, in-line data is only used to manage unit processes, and its utilization is very low because systematic and statistical methods are not established to improve the characteristics of semiconductor products or perform defect analysis.

Accordingly, an object of the present invention is to provide a detection method capable of quickly analyzing in-line data through a systematic and statistical technique and making a clear judgment on a causative process.

In order to achieve the above object, the statistical detection method of the semiconductor manufacturing process according to the present invention comprises the steps of (a) collecting in-line data, PCM data, probe test data in units of shots, and (b) probe test data Stratified into different groups in different probability distributions to obtain a probability distribution of the good and the defective, and (c) in-line data, PCM data, and probe test data for each shot. Sorting the data on the chip, (d) correlating the data on the chip to obtain a p value indicating a correlation coefficient and a significance level, and selecting a significant factor primarily based on the p value as a reference scale; (e) conducting multiple linear regression analysis on the selected significant factors, taking into account the p-values for the entire regression equations, and looking at the p-values for each factor; It is configured by.

Statistical detection method of the semiconductor manufacturing process according to the present invention, after the step (d), (f) deriving the significant factors by binomial logistic regression analysis with the result of the classification of the good / defective product and the selected significant factors in the correlation analysis The method may further include a step, and after step (e) or step (f), further comprising: (g) conducting a reverse sample test on the significant factors to separate the input factors into two groups from the output factors. Can be.

Hereinafter, with reference to the accompanying drawings will be described in detail an embodiment of the present invention. In describing the embodiments, descriptions of technical contents which are well known in the technical field to which the present invention belongs and are not directly related to the present invention will be omitted. This is to more clearly communicate without obscure the subject matter of the present invention by omitting unnecessary description.

Definition of Terms

In-line data: Data measured after a unit process in a semiconductor manufacturing line. The width | variety of the photosensitive film measured after advancing a photo process is such an example.

PCM data (process control monitoring data): Data that can be used to measure test patterns to analyze whether a process is bad. For example, the threshold voltage of a transistor to be measured to determine whether the length of the transistor is properly formed.

Probe test data: data that measures the electrical characteristics of a chip For example, standby current measured to determine the battery life of a mobile phone.

Example

1 is a flowchart of a statistical detection method of a semiconductor manufacturing process according to an embodiment of the present invention. Referring to FIG. 1, the statistical detection method includes a data collection step 10, a distribution test and stratification step 20 of the probe test data, a chip unit sorting step 30 of shot data, and a significant factor through data correlation analysis. It consists of a screening step 40, a multilinear regression step 50, a binary logistic regression step 60, an inverse sample test step 70, an experiment verification step 80. Hereinafter, each step will be described in detail.

Data collection

First, the data for the in-line process in the data collection step collects measurements for each shot of the left, the lower, the middle, the upper, and the right for the same point using the same run wafer. . Here, the shot refers to a unit for patterning a plurality of chips at the same time in the mask manufacturing process, and usually consists of 8 to 30 chips.

PCM data is also collected for the same spot on the wafer where in-line data was measured. Probe test data also collects information about the same point, and it is important to gather data from log files as well as simple pass / fail information.

Determine the distribution of probe test data and Stratification

The second step of the statistical detection method is the identification and stratification of the probe test data. First, we need to look at the distribution of the output characteristics to be analyzed. The normality verification test is performed using Minitab, a kind of statistical software. Subsequently, the probability distribution distinguishes good and defective products by stratifying groups having different slopes.

The layering method includes a method based on an inflection point as shown in FIG. 2A and a method based on a predetermined distance from the inflection point as shown in FIG. 2B. The second method is to more clearly distinguish between good and bad products by using a point larger than the inflection point when the output characteristics are very sensitive as a criterion for the selection of the defective items. In some cases, it is possible to classify into three groups.

FIG. 2A shows a leakage current characteristic (IDDS) in an operation standby state, and FIG. 2B shows a black signal characteristic of a CMOS image sensor. In FIG. 2A, a product having an increased IDDS is a defective product that consumes a lot of power, and a product having a black signal in FIG. 2B is a product having an unclear image, and each has a problem somewhere.

Sort by chip of shot unit data

The third step of the statistical detection method is to arrange the data in shot units in chip units. That is, in-line data and PCM data obtained in units of shots are sorted by data about chips corresponding to each shot. This creates a data table corresponding to each lot ID, wafer number, shot, and chip. The table shown in FIG. 3A is an example. Finally, a table is created in which probe data, in-line data, and PCM data are all in one-to-one correspondence. The table shown in FIG. 3B is one example.

Through data correlation analysis Significance factor  Selection

The fourth step of the statistical detection method is a step of performing correlation analysis on the data using the table prepared above. This step uses the minitab described above. Minitab can be used to obtain p-values that represent significance levels as well as correlation coefficients. FIG. 4A illustrates an example of obtaining correlation coefficients and p values for 69 items of IDDS and in-line data, and FIG. 4B illustrates an example of obtaining correlation coefficients and p values for 10 items of black signal and in-line data.

Based on the p-values obtained as a reference scale, several items are selected from hundreds of significant factors. As a selection criterion, p value may be 0.05 or less in primary and p value may be 0.10 or less in secondary. Applying this criterion to FIG. 4A, firstly M1 FI2 (p value: 0.013) and LDD P-2 PEP (p value: 0.032) are captured, and secondly, M2C FI2 (p value: 0.076) and AA FICD2 (p). Value: 0.086). Similarly, applying the above criteria to FIG. 4B, SPACER SH having a p value of 0.000, NSAL WET, and GC Rox having a p value of 0.007 can be detected.

Multiple Linear Regression

The fifth step of the statistical detection method is a multilinear regression step. The correlation analysis simply proceeds to identify the linear characteristics between the two variables. However, since semiconductor manufacturing is a collection of processes, regression analysis is performed in consideration of mutual interference. As a result of the correlation analysis in the previous step, multiple regression analysis is performed on the selected significant factors to consider the p-values for the entire regression equations, and to derive the significant factors by looking at the p-values for each factor. Here, the meaning of multiple means that there are several factors, and firstly, all the factors detected in the correlation analysis are solved into one equation. The results of multiple linear regression analysis for IDDS consisting of primary selected factors are illustrated in FIG. 5A.

The significance of the regression equation in the regression results is determined by looking at the p-value for the analysis of variance. In the example of FIG. Regression analysis is repeated except for large AA FICD2). 5B is an illustration of the results of multiple linear regression analysis on IDDS composed of secondary screened factors. In the example of FIG. 5B, since the p value is 0.043, the reliability may be higher.

Figure 5c is an illustration of the regression analysis results between the primary screening factor for the Black Signal. The p value in the illustrated regression analysis shows that it is very significant as 0.000. If you want to use a mathematical model of the regression equation, check the residual plot to verify that it is uniform throughout. The residual plot represents the square of the difference between the regression equation and the actual difference, which can be checked for normality from the resulting graph. FIG. 5D shows the result of the residual plot for the regression equation of FIG. 5C. 5c and 5d, the p value is 0.000, which is smaller than the significance level of 0.05, and thus the result is significant because the normality is shown on the residual plot and the distribution is symmetrical from left to right.

Since the purpose of the detection method is to find a significant factor, looking at individual p values for the significant factor, it can be seen that GC Rox is 0.315, SPACER SH is 0.000, and NSAL WET is also 0.000. Therefore, in case of Black Signal, SPACER SH and NSAL WET can be summarized as significant factors.

transposition Logistic  Regression analysis

The sixth step of the statistical detection method is a binary logistic regression step. This step is very useful for identifying relationships with one or more factors, such that the output factor is an attribute, such as a pass / fail, not a continuous type. To make binary logistic regression more efficient, we used the conversion from the distribution of probe data, which was previously shown as a metric type, to the good and defective items. That is, binomial logistic regression analysis may be performed with the factors detected primarily in the good / bad product classification result and the correlation analysis of FIG. 3B.

6A is an exemplary diagram of binary logistic regression results for IDDS. Here, the p-value can be seen whether or not the regression equation is significant. The p-value is 0.089, indicating that it is significant when the significance level is 10%. This is the result of judging from the p value for each factor. That is, M1 FI2 indicated that the p-value is 0.019, which is a very significant factor.

6B is an exemplary diagram of binary logistic regression results for Black Signal. Here, we can see that the regression equation is significant by looking at the p value, but it is very significant because the p value is 0.000. Next, when the p value for each factor is judged, it can be seen that all of them are significant factors of 0.05 or less.

As mentioned above, the reason for the regression analysis and the binary logistic regression analysis is that, in general, when the normality is shown, the factor can be derived only by the regression analysis using the quantitative characteristic. This is because additional verification can lead to higher confidence factors.

Taken together, the factors most common in the regression and binomial logistic regression are the most significant factors. Especially, in this case, the factors detected in the binomial logistic regression due to the nature of the distribution (ie, GC Rox, SPACER). It can be concluded that SH, NSAL WET) are all significant factors.

Reverse sample test

A seventh step of the statistical detection method is a reverse sample test step. In order to analyze the input factor from the output factor into two groups, a sample test for each significant factor is performed using the inflection point of the probability distribution described above. A typical sample test would look at the average difference between the two groups for the output factor, but the reverse sample test is different in that it validates the input factor.

As an example, FIG. 7A shows the results of applying an inverse sample test on M1 FI2, which appeared as a significant factor. Again, the criterion is p value. Since p is 0.021, there is no significant difference between the two groups. The reverse sample test is also applied to M1 FI1 to determine the priority. 7B shows the result. Since the p value of M1 FI1 is 0.068 and the p value of M1 FI2 is 0.021, it can be seen that M1 FI2 is the first management factor.

As another example, FIG. 7C illustrates a result of applying a pass / fail test to SPACER SH by dividing a pass / fail based on the inflection point of FIG. 2B in the case of a black signal. In this example, the p value is 0.000, indicating a very high confidence level (1-p value). In addition, the 95% confidence interval shows that the difference between the two groups is 3.1Å.

Experimental Verification

The eighth step of the statistical detection method is a verification step through experiments. As a result of analysis up to now, M1 FI2 is a significant factor in the case of IDDS, and when analyzing PCM data in the same way, the VIA1 Chain resistance is significant as a factor, which is the first metal wire contacting the VIA1 contact. The result is indirect proof that width is a problem.

In the case of Black Signal, NSAL WET and SPACER SH were the most significant factors. Substantial verification of NSAL WET is based on the same stepwise procedure for PCM data. salicide resistance was found to be a significant factor.

Next, the verification experiment for SPACER SH was divided into two groups, and the general sample test was conducted to confirm the results. That is, after spacer etching, the average verification of the Black Signal of 14AT06 lot having a residual oxide thickness of 155Å and 152617 lot having 190Å was performed, and the result is shown in FIG. 8A.

The difference between the two groups can be said to be clearly different from the p value (0.000), and when the Spacer Rox is raised according to general experience, it can be seen that the progress is in a bad direction. This is again a box plot (FIG. 8B). From Figure 8b it was confirmed once again that SPACER SH is a significant factor.

As described so far, the statistical detection method of the semiconductor manufacturing process according to the present invention is to analyze the physical in-line data obtained in the semiconductor manufacturing unit process by systematic and statistical techniques, so that improvement measures must be taken among numerous semiconductor manufacturing processes. Several processes can be derived accurately and quickly. Therefore, the statistical detection method of the present invention can be effectively applied to defect analysis of semiconductor products, and can greatly contribute to quality improvement and yield improvement.

In the present specification and drawings, preferred embodiments of the present invention have been disclosed, and although specific terms have been used, these are merely used in a general sense to easily explain the technical contents of the present invention and to help the understanding of the present invention. It is not intended to limit the scope. It will be apparent to those skilled in the art that other modifications based on the technical idea of the present invention can be carried out in addition to the embodiments disclosed herein.

Claims (3)

(a) collecting in-line data, PCM data, and probe test data on a shot basis; (b) obtaining a probability distribution of the probe test data and stratifying it into groups having different slopes in the probability distribution to distinguish between good and defective products; (c) arranging the in-line data, the PCM data, and the probe test data with data for a chip corresponding to each shot; (d) correlating the data on the chip to obtain a p value indicating a correlation coefficient and a significance level, and selecting the significant factor primarily based on the p value as a reference scale; And (e) conducting multiple linear regression analysis on the selected significant factors to consider the p values for the entire regression equations and to derive the significant factors by looking at the p values for each factor; Statistical detection method of a semiconductor manufacturing process comprising a. The method of claim 1, After the step (d), (f) deriving a significant factor by performing a binary logistic regression analysis with the result of the classification of good or bad goods and the significant factors selected in the correlation analysis; Statistical detection method of the semiconductor manufacturing process further comprising. The method according to claim 1 or 2, After step (e) or step (f), (g) performing an inverse sample test on the significant factors to separate the input factors into two groups from the output factors; Statistical detection method of the semiconductor manufacturing process further comprising.

Images