JP2008250910A - Data mining method and process management method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To mine an event having time dependence in a short period of time and with high accuracy by mechanical work, with respect to a data mining method and a process management method. <P>SOLUTION: Processing times including a waiting time in each process are mechanically classified by using a time classifying method, and a classified time group is used as an explanatory variable to perform data mining. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a data mining method and a process management method, and more particularly to a data mining method and a process management method characterized by a configuration for mining a time-dependent event in a short time and with high accuracy. is there.

Data mining can extract and analyze unique data from a large amount of data, and understand product sales trends and characteristic features related to manufacturing.
In the semiconductor manufacturing process, it is effective to extract elements that control the product yield.

  Many methods such as classification, grouping, association rules, pattern extraction, and visualization have been developed for this data mining, but the method called “classification” is mainly used in the process management of semiconductor manufacturing lines. (For example, refer to Patent Document 1).

In this “classification” method, as in the case of a decision tree analysis, an object that divides a process or device into two parts and explains the result is used as an explanatory variable, and an object variable that indicates a result such as yield is used as an objective variable It is a technique.
At this time, when the objective variable is a qualitative variable, the decision tree is called a classification tree, and when the objective variable is a quantitative variable such as a yield, it is called a regression tree.

Until a semiconductor product is produced, processes such as implants, film generation, and etching are performed in a number of manufacturing processes.
At this time, when process management is performed by data mining, when there is a decrease in yield in a specific LOT (lot) group, the yield of the lot that has passed through the combination of passing equipment for each process By looking at the degree of separation and arranging them from the higher degree, we grasp the equipment with a strong low-yield tendency and identify it as a malfunctioning equipment. Here, referring to FIGS. 13 to 17, A process management method based on the data mining method will be described.

See FIG.
FIG. 13 is a conceptual configuration diagram of the manufacturing process, which conceptually shows various processes consisting of several hundred processes along the flow of lots. In each process, a plurality of lots are indicated by a circle. In this example, it is assumed that an abnormal device is present in the process A and has an adverse effect on the manufacturing yield.

In this case, assuming that the apparatus ₃ is abnormal, the variation in yield is expressed as a “box-and-whisker chart” shown below.
Note that the “box-and-whisker diagram” is a statistical graph for easily expressing data with variations, and generally, the minimum value, the first quartile, the median, and the third quartile. , And five data of maximum values are displayed, and a “box” is formed by the first quartile, the median, and the third quartile.

See FIG.
FIG. 14 is an explanatory diagram of a method for analyzing a problematic process. For example, a plurality of used devices processed in the above-described step A are classified into two, and a lot group processed by each classified device. Each average value (X _ave1 , X _ave2 ) and the sum of squares (S ₁ , S ₂ ) obtained by adding the squares for each section are _obtained from the frequency distribution relating to the yield, which is the objective variable for each, and the entire lot From the frequency distribution, the average value (X _ave0 ) and the sum of squares (S ₀ ) in each section are obtained,
ΔS = S ₀ − (S ₁ + S ₂ )
Are divided into two so that ΔS shown in FIG.

Here, for the devices _{1 to 3} used in step A, [device ₁ ]: [device ₂ + device ₃ ], [device ₁ + device ₃ ]: [device ₂ ], and [device ₁ + device _2] ]: [Delta] S is obtained for the three combinations of [apparatus ₃ ], and the combination that maximizes [Delta] S is searched.
In the above case, since the device ₃ is an abnormal device, the ΔS in the combination of [device ₁ + device ₂ ]: [device ₃ ] is maximized, and therefore is divided into [device ₁ + device ₂ ] and [device ₃ ]. The

See FIG.
FIG. 15 is a conceptual explanatory diagram of the average yield for each device used. Here, attention is paid to three processes, step a, step b, and step c. In step a, the device group a ₁ and the device group a ₂ is used. In the process b, the lot processed in the apparatus group a ₂ in the process a is processed in the apparatus group b ₂ and the apparatus group b _{3. In} the process c, the lot is processed in the process a. Assume that the lot processed in a ₁ is processed in the device group c ₁ and the device group c ₂ .
The group division of each device is such that the above ΔS is maximized.

See FIG.
FIG. 16 is a regression tree created by performing data mining on the process shown in FIG. 15, and the average yield of 65 lots is 89%. The process a is divided into a group consisting of the apparatus a ₂ having an average yield of 82% and a group consisting of the apparatus a ₁ having an average yield of 95%, and the analysis is further performed by dividing the respective data.

As a result, the average yield of the five lots processed by the combination of the apparatus a ₂ and the apparatus b ₂ is the lowest at 65%, and the average yield of the 20 lots processed by the combination of the apparatus a ₁ and the apparatus c ₁ is 98%. The highest yield.

See FIG.
FIG. 17 is a box-and-whisker plot of the yield of the four combinations shown in FIG. 16. The combination of the device a ₁ and the device c ₁ not only has a high average yield but also has the smallest variation, while the device a ₂ The combination of device b ₂ not only has the lowest average yield, but also has a very large variation, suggesting that the process is not stable.

In this way, when looking at the regression tree in order, the relationship between multiple steps such as the relationship between the stepper and the etcher can be seen. By analyzing this correlation, it is possible to investigate the cause of the decrease in manufacturing yield. Become.
JP 2006-041081 A

As mentioned above, data mining has been performed so far in consideration of classification by processing device number, but it is clear from experience that the manufacturing yield depends not only on the processing device number but also on the processing time. Therefore, it is necessary to analyze the effect of processing time on yield.
However, when trying to analyze the processing time of each lot in each process, the time is not a number but a quantity, so it cannot be analyzed immediately by the data mining method.

Therefore, the applicant of the present application attempts to classify the effect of the processing time on the manufacturing yield by observing the processing time of each process and see the relationship with the yield.
First,
a. The time from loading to unloading including waiting time for each lot is defined as processing time, and data is extracted for each process and each lot. Then
b. A person looks at the extracted time data for each process and divides it into a long time group and a short time group.
This is because automation is possible if it can be classified on the basis of a fixed time, but since it is necessary to classify by looking at the entire processing time result, it is not possible to automate because the time to be divided points cannot be set in advance. Then
c. The yield for long-time group lots is totaled, and the yield for short-time group lots is also totaled. Then
d. Search for a process with a large yield difference and pick it up as a process candidate to be improved.

  Alternatively, a process management method is adopted in which the processing time is divided into a long time group and a short time group, and the relationship between the processing time and the yield is made into a correlation graph for all the processes and referenced.

However, the time division method described above has various problems. First,
A. Time classification is sensuous because it is performed by humans, and it is easy to make mistakes and takes time.
That is, there are hundreds of processes for manufacturing one semiconductor device, and the number of analysis cases is one or two. However, three or more cases do not continue, and as a result, analysis cannot be performed.
B. Even after the time classification is completed, there is a yield summary for each lot, which is a daunting task.
C. There is a problem that the yield hits due to the time group are not so many and it is serious, so the analysis tends to be postponed and the turnaround time becomes long.
D. There is a problem that it takes a lot of man-hours to check the contents of the correlation diagram when there are hundreds of references. In addition, when the low yield is concentrated in a specific time group (time B, time C...) In the middle, there is a problem that it is easy to overlook because it is a human work.

  Therefore, an object of the present invention is to perform mining of time-dependent events in a short time and with high accuracy by mechanical work.

Now, with reference to FIG. 1, means for solving the problems in the present invention will be described.
See FIG. 1 In order to solve the above-mentioned problem, the present invention is a data mining method in which processing times including waiting time in each step are mechanically classified using a time classification method, and the classified time group is used as an explanatory variable. It is characterized by performing data mining.

  In this way, by classifying time, which is not a number, into a time group by using a time classification method, it becomes possible to handle the number of groups of time, and thus the conventional data mining method can be used. The process management can be performed by applying it as it is.

  In particular, when classifying by time classification method, since it is mechanically classified, human sensory elements are not included, and even if the number of data is enormous, only data is input to the microcomputer. Since the time group can be instantly classified, the analysis time can be greatly shortened.

  Further, as a time group classification method in this case, the processing time of each processing unit in each process is automatically arranged in order of size by mechanical processing (sorting), and the difference before and after the processing time arranged in the order is large. Points (change points) may be taken out in descending order by a predetermined number and separated to divide the time group.

  The change point determination method shown in FIG. 1 is a case where the number of change point selections is 2, and the difference between before and after the processing time arranged in order is large. In the figure, the difference is the largest ΔT = 65. The boundary with ΔT = 8 is the change point that is the second largest difference.

  In this case, at the time of classification by the time group classification method, a median processing time may be obtained, and a processing time larger than a predetermined value with respect to the median may be deleted as noise, thereby eliminating an abnormal value. This enables analysis with high accuracy.

  Further, in such a data mining method, the manufacturing process can be managed easily and accurately by using the manufacturing yield as an objective variable.

According to the present invention,
α. Since the processing time classification of hundreds of processes is automatically performed mechanically, the time classification, which conventionally took several days, is completed in a few minutes and is extremely efficient, and the classification accuracy is also high. Become .
β. Since the time classification can be performed efficiently, the relationship between processing time and yield can be easily performed using data mining technology, so the classification accuracy is improved and analysis results can be obtained in a short time.
γ. From the results of β above, processing time analysis can be applied to a large number of projects, and the turnaround time for low yield analysis can be reduced.
δ. By classifying time and applying it to mining, it becomes possible to obtain knowledge resulting from complex conditions, and the quality of analysis is dramatically improved.
The effect is obtained.

Here, with reference to FIG.2 and FIG.3, the outline | summary of embodiment of this invention is demonstrated.
FIG. 2 is a conceptual block diagram showing the outline of the embodiment of the present invention. In the present invention, the waiting time in each process is determined instead of using the processing device number as an explanatory variable for analysis in each process. The processing time included is automatically mechanically classified using a time classification method, data mining is performed using the classified time group as an explanatory variable, and the yield as an objective variable, and process management in the production line is performed.

See Figure 3
FIG. 3 is a flowchart of the data mining method using the time group classification method according to the embodiment of the present invention.
(1) The processing time including the standby time is automatically converted from quantitative handling to numerical handling.
This is because the explanatory variables for data mining by the regression tree are the equipment numbers DD85A01, DD85A02. The DD85A05 must be counted such as three units (numerical treatment), but the processing performance time is different in each process and numerically different for each lot.

For this reason, in a conventional attempt, classification is performed by human observation, and data mining is performed after dividing the classification into, for example, 50-minute group / 120-minute group.
However, in the first embodiment of the present invention, this classification is executed by the following algorithm, the processing time is automatically converted from quantitative treatment to numerical treatment, and the result is assigned to an explanatory variable for data mining. Mining analysis of the relationship between process time group and yield.

That is,
1) The processing time of each process is arranged for each lot and sorted in order of size. Then
2) A processing time difference ΔT between adjacent lots arranged in order of size is obtained. Then
3) With respect to ΔT, a predetermined number is taken out in descending order, and the values are taken as change points, and divided into time groups such as time group A, time group B, time group C,.
Note that the number of change points is appropriately determined according to the analysis target.
Also, when dividing the time group, as a noise countermeasure, the median of the processing time is obtained in advance, and N times or more of the median is cut as an abnormal value.
Note that the value of N is set according to the contents of the time data, and is set to 10 times or more, for example.

Next, data mining is performed for the time groups (time group A, time group B...) Classified in this way.
(2) Using the explanatory variables as the time group and the objective variable as the yield, for each step, find the division by the combination that maximizes the above ΔS. Then
(3) The OMP (significant difference judgment index) and the “grouping degree” obtained from ΔS in each process are sorted in descending order and arranged in the order of processes having high processing time dependency. Then
(4) A regression tree is prepared in the order of processes having a high processing time dependency, and a factor that decreases the yield is specified.

Note that OMP is a numerical value (%) that estimates the probability that a set that is divided into two by discriminant analysis (here, a set of time groups) is not significantly different, and is an estimated value of the probability of making an incorrect decision. (%), And thus, the smaller the OMP, the more significant the two splits are.
On the other hand, the “degree of unity” is an index indicating how much the sum of squares has changed after being divided into two parts. In this case, it indicates that “the variation does not change even when divided”, and in the case of 100%, it indicates that “there is no variation due to the division”.

Next, a data mining method using the time group classification method according to the first embodiment of the present invention will be described with reference to FIGS.
See Figure 4
FIG. 4 is data showing the processing time and yield in each process to which the data mining method using the time group classification method of Example 1 of the present invention is applied. Here, the data of the first polycrystalline silicon layer is shown. An example of the film forming process is shown for 10 lots, the upper figure is raw data, and the lower figure is a graph showing the relationship between processing time and yield for each lot.

See Figure 5
FIG. 5 sorts the data shown in FIG. 4 in order of increasing processing time in order to handle the processing time numerically, obtains the processing time difference ΔT between adjacent lots, and calculates the processing time difference ΔT in order from the largest one. These are divided into three time groups, time group A, time group B, and time group C, with one point as a change point.
Here, ΔT = 65 is the change point ₂ and ΔT = 8 is the change point ₁ mechanically selected. This operation is performed by simply inputting the data shown in the upper diagram of FIG. Sorting and change point determination are automatically performed by the existing program.
In the following figure, the results of sorting are graphed and the change points are displayed.

See FIG.
FIG. 6 is a graph showing the results of the above-described data mining. As a result of data mining, [Time Group A]: [Time Group B + Time Group C], [Time Group A + Time Group C]: [Time Group B] , [Time Group A + Time Group B]: [Time Group C], and [Time Group A + Time Group C]: [Time Group B] has the largest ΔS, and [Time Group A + It is divided into time group C] and [time group B] and connected with a line, and the separation of yield is most clearly shown, and the processing process of [time group A + time group C] is reduced in yield. It is estimated that.

See FIG.
FIG. 7 is an explanatory diagram of the mining result of the etching process of the polycrystalline silicon layer, the upper figure is a data table after sorting and time group division, and the middle figure graphs the sorting result and displays the change point. The figure below is a graph of the data mining results.
As shown in the figure below, yield separation is not clearly seen.

See FIG.
FIG. 8 is an explanatory diagram of the mining result of the first layer via formation process, the upper diagram is a data table after sorting and time group division, the middle diagram is a graph of the sorting result, and the change point is shown. The figure below shows a graph of the data mining results.
As shown in the figure below, the separation of the yield appears more clearly than the etching process of FIG. 7, but the separation is worse than the film formation process of the polycrystalline silicon layer shown in FIG.

See FIG.
FIG. 9 is an explanatory diagram of the mining result of the etching process of the second layer Al wiring film forming process, the upper figure is a data table after sorting and time group division, and the middle figure is a graph of the sorting result. In addition, the change points are displayed, and the following figure is a graph of the data mining results.
As shown in the middle figure, the processing time dependence is hardly observed, and as shown in the following figure, the separation of the yield is not clearly shown.

See FIG.
FIG. 10 is a table in which the OMP, the degree of unity, the yield difference, and the like for the above four steps are arranged in the order in which the yield separation clearly appears in the data mining results.
1. 1. Film forming step of first-layer polycrystalline silicon layer 2. First layer via formation process 3. Etching step of the first polycrystalline silicon layer The second layer Al wiring film forming process is performed in this order.

See FIG.
FIG. 11 is a report in which candidates presumed to be the cause of the decrease in yield based on FIG. 10 are arranged in order. The first candidate is a process for forming a polycrystalline silicon layer, and the second candidate is the first layer. A process for forming an eye via is mentioned.

  The lower part shows data when the cause of the decrease in yield is analyzed under a complex condition. Here, a low yield is shown in the film formation process of the first candidate polycrystalline silicon layer [time group A + The seven lots of time group C] are extracted from the above-described data mining results for the first-layer via formation process, and the average yield and the yield difference are shown.

See FIG.
FIG. 12 is an explanatory diagram of the regression tree of the data mining result, [Time group A + Time group C] in the film formation process of the first polycrystalline silicon layer and [Formation of the first layer via] It can be seen that the combination of [time group A + time group B] shows a strong correlation with the decrease in yield.

  Therefore, the process analysis of the combination of [time group A + time group C] in the film forming process of the first polysilicon layer and [time group A + time group B] in the process of forming the first layer via. This makes it easier to determine the actual cause of yield loss.

  As described above, in the first embodiment of the present invention, since the processing time is divided into time groups using the mechanical time group classification method, it is possible to apply the data mining method using the time groups as explanatory variables. In addition, since there is no artificial factor in dividing the time group, objective and single-time data mining becomes possible.

  As an application example of the present invention, a data mining method for process management in a semiconductor manufacturing process having hundreds of manufacturing processes is typical, but not limited to a semiconductor manufacturing process. It is also applied to process management of processes, and is not limited to manufacturing processes.It is also applied to analysis of causal relationships between a series of events and processes that are considered to be time-dependent. Is.

It is explanatory drawing of the determination method of the change point in this invention. It is a notional block diagram which shows the outline | summary of embodiment of this invention. It is a flowchart of the data mining method using the time group classification | category method of embodiment of this invention. It is data which shows the processing time and yield in each process which applied the data mining method using the time group classification method of Example 1 of this invention. It is explanatory drawing of the change point of a sorting result and time group division | segmentation. It is a graph which shows the result of data mining. It is explanatory drawing of the mining result of the etching process of a polycrystalline silicon layer. It is explanatory drawing of the mining result of the formation process of the 1st layer via | veer. It is explanatory drawing of the mining result of the etching process of the film-forming process of 2nd layer Al wiring. It is explanatory drawing, such as OMP about 4 processes, a unit degree, and a yield difference. It is explanatory drawing of the report of the cause of the yield fall based on FIG. It is explanatory drawing of the regression tree of a data mining result. It is a notional block diagram of a manufacturing process. It is explanatory drawing of the analysis method of the process which becomes a problem. It is a conceptual explanatory drawing of the average yield for every using apparatus. 16 is a regression tree created by performing data mining on the process shown in FIG. 15. It is a box-and-whisker diagram of the yield of four combinations shown in FIG.

Claims (4)

A data mining method characterized by mechanically classifying a processing time including a waiting time in each process using a time classification method, and performing data mining using the classified time group as an explanatory variable. In the above time group classification method, the processing time of each processing unit in each process is automatically arranged in order of magnitude by mechanical processing, and a predetermined number of points where the difference before and after the processing time arranged in the order is large The data mining method according to claim 1, wherein the time group is bundled by taking out and dividing the time group. 3. The method according to claim 1, wherein a median processing time is obtained at the time of classification by the time group classification method, and a processing time larger than a predetermined value with respect to the median is deleted as noise. Data mining method. 4. The process management method according to claim 1, wherein each manufacturing process is managed by performing data mining using a manufacturing yield as an objective variable. 5.

Images