TWI588673B - Method for analyzing semiconductor fabrication fault and computer program - Google Patents

Description

Semiconductor process error analysis method and computer program product

The invention relates to a method for error analysis, in particular to a method for analyzing semiconductor process errors.

Process yield is an important indicator in semiconductor manufacturing. The process yield not only represents the level of semiconductor process technology, but also the process yield reflects the cost of the semiconductor process. In short, when the process yield is high, the product per unit area can produce more products, which lowers the cost per unit of product. Therefore, the process yield is attributable to the interest rate of the semiconductor manufacturing plant.

In a semiconductor process, the workpiece is subjected to a series of manufacturing steps using various types of semiconductor machines, such as exposure, chemical mechanical polishing, or chemical vapor deposition. For example, a semiconductor product needs to undergo 500 to 600 processing steps. Process parameters for a single processing step can be volatile or unstable, which can cause defects in the semiconductor product and result in reduced yield. However, it is difficult to find the process parameters that cause defects in such a large number of process parameters.

In view of this, how to select the process parameters that may cause defects as a reference for the operator from a large number of process parameters is the goal that is currently in great need.

The present invention provides a semiconductor process error analysis method and a computer program product, which divides a process record of a workpiece to be analyzed into a good group, a mixed good product, and a defective product group, and calculates a feature value in each group, and then A feature weight is determined according to the characteristic values of the two to select a process parameter that may cause defects.

The semiconductor process error analysis method according to an embodiment of the present invention includes: obtaining a process record corresponding to a plurality of workpieces from a database, wherein the plurality of workpieces comprise a plurality of good products and a plurality of defective products, and the process record includes a plurality of process parameters Calculating one of the feature parameters of each process parameter; and sorting the plurality of feature weights to determine one of the possible process parameters of the defective product, wherein the step of calculating the feature weight of each process parameter comprises: dividing the process record a first group and a second group, wherein the first group is composed of a plurality of good products, and the second group is composed of a plurality of good products and a plurality of defective products; for one of a plurality of process parameters, Filtering a plurality of corresponding products and a plurality of process materials of the plurality of defective products; calculating a first feature value of one of the first groups; calculating a second feature value of one of the second groups; and calculating the first feature value according to the first feature value and The second feature value determines the feature weight.

A computer program product according to another embodiment of the present invention includes a code for performing an above-described semiconductor process error analysis method after loading an electronic device.

The purpose, technical contents, features, and effects achieved by the present invention will become more apparent from the detailed description of the appended claims.

The embodiments of the present invention will be described in detail below with reference to the drawings. In addition to the detailed description, the present invention may be widely practiced in other embodiments, and any alternatives, modifications, and equivalent variations of the described embodiments are included in the scope of the present invention. quasi. In the description of the specification, numerous specific details are set forth in the description of the invention. In addition, well-known steps or elements are not described in detail to avoid unnecessarily limiting the invention. The same or similar elements in the drawings will be denoted by the same or similar symbols. It is to be noted that the drawings are for illustrative purposes only and do not represent the actual dimensions or quantities of the components. Some of the details may not be fully drawn in order to facilitate the simplicity of the drawings.

Please refer to FIG. 1 to illustrate a semiconductor process error analysis method according to an embodiment of the present invention. First, a process record corresponding to a plurality of workpieces is obtained from a database (S10). For example, the processed workpiece can be a wafer, a solar substrate, a liquid crystal panel, or the like. In one embodiment, the process record includes at least a plurality of process parameters known as good and defective products, multiple process parameters of each good or defective product, and multiple process data of each process parameter of each good or defective product. . The present invention analyzes the above process records to screen process parameters that may lead to defective products. For example, the process parameters can be a machine name, temperature, pressure, flow rate, time and other parameters. As mentioned above, a semiconductor product needs to undergo 500 to 600 processing steps. Therefore, the process record contains a large amount of data, and most of the digital electronic data is stored in the electronic device that executes the database software. In other words, the semiconductor process error analysis method of the present invention is executed by an electronic device having an arithmetic function to access data in the database and perform subsequent analysis. For example, the electronic device having the computing function can be a computer.

Next, one of the feature parameters of each process parameter is calculated (S20). It should be noted that each process parameter described herein refers to a process parameter that the operator suspects or is interested in, and is not necessarily all process parameters of a semiconductor product. It can be understood that the feature weights can also be calculated for all process parameters if necessary.

Please refer to FIG. 2 to illustrate the steps of calculating the feature weights of each process parameter. First, the process record is divided into a first group and a second group, wherein the first group is composed of a plurality of good products, and the second group is composed of a plurality of good products and a plurality of defective products (S21). Then, the corresponding good products and the plurality of process materials of the defective products are screened out (S22). Referring to FIG. 3, process data of a process parameter of a first workpiece and a second workpiece is displayed, which includes at least a time stamp and a parameter value at that time. It should be noted that in the subsequent analysis step, the first workpiece may be a good or a defective product, and the second workpiece may be a good or a defective product. In order to easily obtain the corresponding process data between the first workpiece and the second workpiece, the plurality of process materials corresponding to the process parameters may be first translated to the same reference. For example, the time of the first workpiece and the second workpiece to perform the process parameter can be shifted to 0 seconds, and the other process data is also relatively translated, as shown in the column of zero return time in FIG. Then, the corresponding plurality of process materials with the closest time between the first workpiece and the second workpiece are taken out. For example, referring to FIG. 3, the process data of the first workpiece at 0 seconds corresponds to the process data of the second workpiece at 0 seconds; the process data of the first workpiece at 1 second corresponds to the second workpiece at 1 second. Process data at the time. Due to the lack of the process data of the second workpiece at 2 seconds, the process data of the first workpiece at 2 seconds corresponds to the process data of the second workpiece with the closest time, that is, the process data of the second workpiece at 2.5 seconds. Because of the lack of the process data of the second workpiece at 3 seconds, the process data of the first workpiece at 3 seconds also corresponds to the process data with the closest workpiece time, that is, the process of the second workpiece at 2.5 seconds. data. The process data of the first workpiece at 4 seconds corresponds to the process data of the second workpiece at 4 seconds; the process data of the first workpiece at 6 seconds also corresponds to the process data of the second workpiece with the closest workpiece time, that is, the second workpiece Process data at 7 seconds.

It should be noted that when the distance between the first workpiece and the second workpiece is calculated subsequently, the number of process materials of the two should be consistent. In an embodiment, when the difference between the number of process data between different workpieces exceeds a predetermined value, the corresponding plurality of process materials of the workpiece are excluded, and an error record is recorded.

Referring to FIG. 2, next, one of the first feature values of the first group is calculated (S23). In one embodiment, calculating a plurality of first distances between each good product and other good products, and calculating a first coefficient of variation (ie, a ratio of the variance to the average value) of the plurality of first distances, wherein the first group is The first eigenvalue of the group. For example, there are 100 pieces of all workpieces, of which 5 are defective and the remaining 95 are good. In step S23, a first distance between 95 pieces of good products and a first coefficient of variation of the first distances are calculated. The calculation of the distance is as follows: First, the square of the difference between the corresponding plurality of process data is calculated to obtain a plurality of intermediate values. Taking the embodiment shown in FIG. 3 as an example, as described above, the process data of the first workpiece at 0, 1, 2, 3, 4, and 6 seconds respectively correspond to the second workpiece at 0, 1, 2.5, 2.5, 4, 7 seconds of process data. The squares of the difference between the plurality of pieces of process data corresponding to the first workpiece and the second workpiece are 1, 1, 1, 9, 1, and 1, respectively. Then, the square root of the sum of the plurality of intermediate values is calculated to obtain the first distance. For example, the sum of the square roots of 1, 1, 1, 9, 1 and 1 is 3.741, which is the distance between the first workpiece and the second workpiece. In step S23, a plurality of first distances are obtained between the plurality of good products, so that the first coefficient of variation of the first distance can be calculated.

Next, one of the second feature values of the second group is calculated (S24). For example, the defective product is mixed into the good product to calculate a second distance between each other, that is, each good product and a plurality of second distances of the defective product and the other good products and the defective product are calculated, and one of the plurality of second distances is calculated. The second coefficient of variation, which is the second characteristic value of the second group. Taking the foregoing as an example, the second distance between the 100 pieces of the workpiece including the good product and the defective product, and the second coefficient of variation of the second distances are calculated. The calculation method of the second distance and the second coefficient of variation is as described above, and will not be described herein. Finally, the feature weight is determined according to the first feature value and the second feature value (S25). It will be appreciated that the inherent nature of the process parameters may affect the first eigenvalue as well as the second eigenvalue. For example, a process parameter that is subject to greater variation may result in a relatively large feature value, but does not necessarily represent a process parameter that is subject to large variations that may result in product defects. Conversely, a relatively stable process parameter may result in a relatively small feature value, but it does not necessarily mean that a relatively stable process parameter does not cause product defects. Therefore, feature weights need to take into account the characteristics of the process parameters. In an embodiment, the smaller the first feature value and the second feature value are, the smaller the feature weight is. Conversely, the larger the first feature value and the second feature value, the larger the feature weight.

For example, referring to Table 1, when the first characteristic value of the first group is small, the process parameter is relatively stable. Therefore, the second feature value of the second group mixed into the defective product is determined according to the first characteristic value. It is divided into multiple intervals and gives each interval a corresponding feature weight. For example, suppose there are three process parameters A, B, and C, and the second feature value is 5, wherein the first feature value of the process parameter A is in the range of 0-10, and the first feature value of the process parameter B is In the interval of 11-15, the first characteristic value of the process parameter C is in the interval of 16-20. According to Table 1, although the second characteristic values of the process parameters A, B, and C are all 5, the corresponding feature weights are 10, 5, and 2, respectively. It can be understood that the intervals listed in Table 1 are merely exemplary, and the operator can set an appropriate interval according to the actual process parameter characteristics. Table 1         <TABLE border="1" borderColor="#000000" width="_0002"><TBODY><tr><td> first eigenvalue</td><td> second eigenvalue</td><td> Feature weights</td></tr><tr><td> 0-10 </td><td> 0-2 </td><td> 2 </td></tr><tr>< Td> 3-4 </td><td> 5 </td></tr><tr><td> 5-6 </td><td> 10 </td></tr><tr>< Td> ... </td><td> ... </td></tr><tr><td> 11-15 </td><td> 0-4 </td><td> 2 </td> </tr><tr><td> 5-8 </td><td> 5 </td></tr><tr><td> 9-12 </td><td> 10 </td> </tr><tr><td> ... </td><td> ... </td></tr><tr><td> 16-20 </td><td> 0-10 </td> <td> 2 </td></tr><tr><td> 11-15 </td><td> 5 </td></tr><tr><td> 16-20 </td> <td> 10 </td></tr><tr><td> ... </td><td> ... </td></tr></TBODY></TABLE>

Referring again to FIG. 1, the feature weights of the plurality of process parameters are sorted for the operator to determine one of the defective process parameters (S30). Taking the foregoing embodiment as an example, the process parameters A, B, and C are sorted to 10, 5, and 2. Therefore, for the process parameters A, B, and C, the process parameter A is more likely to be the process parameter of the product.

In the foregoing embodiment, the first distance between all the good products (95 pieces) and the second distance between all the workpieces (100 pieces) including the good product and the defective product are calculated, and thus the calculation amount is large. In an embodiment, when calculating feature weights, a plurality of representative products can be used as a basis for calculating the first distance and the second distance. For example, 10 representative products can be screened out from 95 good products as the basis for the subsequent calculation of the first distance and the second distance. It should be noted that different process parameters may have different representative products.

Please refer to FIG. 4 to illustrate the steps of screening representative products. First, a plurality of good products are selected as a representative good group (S41). For example, two of the 95 good products are selected as representative good groups. Next, a first average value of the distances between all the good products in the representative good product group is calculated (S42). The distance between good products is calculated as described above and will not be described here. There are only 2 good products in the representative good group, so the first average is the distance between 2 good products. Next, one of the representative good product groups is added to the representative good product group (S43), and a second average value of the distances between all the good products in the representative good product group is calculated (S44). For example, select one good product from the remaining 93 good products to join the representative good product group, and a second average of the distance between the three good products. The second average is compared to the first average, and if the second average is less than the first average, the newly added good is retained in the representative good group. If not, discard the farthest product in the representative good group (S45). Assuming that the second average is currently smaller than the first average, the newly added product remains in the representative good group. At this time, the representative good group has 3 good products. Repeat steps S42-S45, and all the good products go through the above screening steps. Finally, the representative good group only retains a relatively concentrated and small number of representative products, which can greatly reduce the calculated system load and time.

A computer program product according to an embodiment of the invention includes a code. The code can be loaded by an electronic device (such as a computer) to perform the semiconductor process error analysis method shown in FIG. 1, FIG. 2 and FIG. The detailed description of the semiconductor process error analysis method of the present invention is as described above, and will not be described herein.

In summary, the semiconductor process error analysis method and the computer program product of the present invention divide a process record of a workpiece to be analyzed into a good group, a mixed good product, and a defective product group, and calculate each element in each group. The distance and the coefficient of variation of the distance are used as feature values, and then a feature weight is determined according to the feature values of the two groups to filter out process parameters that may cause defects. Preferably, the analytical method of the present invention can reduce the calculated system load and time by screening representative products.

The embodiments described above are only intended to illustrate the technical idea and the features of the present invention, and the purpose of the present invention is to enable those skilled in the art to understand the contents of the present invention and to implement the present invention. That is, the equivalent variations or modifications made by the spirit of the present invention should still be included in the scope of the present invention.

S10-S30‧‧‧Steps
S21-S25‧‧‧Steps
S41-S45‧‧‧Steps

1 is a flow chart showing a semiconductor process error analysis method according to an embodiment of the present invention. 2 is a flow chart showing the steps of calculating feature weights in accordance with an embodiment of the present invention. Figure 3 is a schematic diagram showing the corresponding process data of one of the two workpiece process parameters. 4 is a flow chart showing the steps of screening a representative good group according to an embodiment of the present invention.

S10-S30‧‧‧Steps

Claims (20)

A semiconductor process error analysis method, the method comprising: obtaining a process record corresponding to a plurality of workpieces from a database, wherein the plurality of workpieces comprise a plurality of good products and a plurality of defective products, and the process record comprises a plurality of process parameters Calculating one of the feature parameters of each of the process parameters, the step comprising: dividing the process record into a first group and a second group, wherein the first group is composed of the plurality of good products, The second group is composed of the plurality of good products and the plurality of defective products; for one of the plurality of process parameters, the plurality of good products corresponding to the plurality of good products and the plurality of defective products are screened; a first feature value of the first group; calculating a second feature value of the second group; and determining the feature weight according to the first feature value and the second feature value; The feature weights are sorted to determine one of the possible process parameters for the defective product. The semiconductor process error analysis method of claim 1, wherein the smaller the first feature value and the second feature value, the smaller the feature weight. The semiconductor process error analysis method of claim 1, wherein the feature value is larger as the first feature value and the second feature value are larger. The semiconductor process error analysis method of claim 1, wherein the first characteristic value is a first coefficient of variation of one of a plurality of first distances of each of the good products and the other good products in the first group; the second The feature value is a second coefficient of variation for each of the good products in the second group and the second and second plurality of distances of the defective product and the other defective product. The semiconductor process error analysis method of claim 4, wherein the calculating the first distance and the second distance comprises: calculating a square of a difference between the corresponding plurality of process data to obtain a plurality of intermediate values; And calculating a square root of the sum of the plurality of intermediate values to obtain the first distance or the second distance. The semiconductor process error analysis method of claim 1, wherein the step of screening the plurality of process data comprises: translating the plurality of process materials corresponding to the process parameters to the same reference; and according to the good product or the defective product The plurality of process materials are extracted, and the corresponding plurality of process materials having the closest time are taken out. The semiconductor process error analysis method according to claim 1, wherein the difference between the number of the plurality of pieces of process data between the workpieces exceeds a predetermined value, the corresponding plurality of process materials are excluded, and an error is recorded. recording. The semiconductor process error analysis method of claim 1, wherein the step of calculating the feature weight further comprises: screening a plurality of representative products from the plurality of good products, and calculating the feature weights by the representative Good products are the basis of calculation. The semiconductor process error analysis method of claim 8, wherein the step of screening the representative product comprises: selecting a plurality of the good products as a representative good group; calculating all the good products in the representative good product group a first average value of the distance; adding one of the representative good product groups to the representative good product group, and calculating a second average value of the distance between all the good products of the representative good product group; Determining whether the second average value is less than the first average value, and if so, retaining the newly added good product in the representative good product group; if not, discarding the farthest product in the representative good product group; The calculating the first average step, the calculating the second average step, and the determining step until all of the good products have been screened. The semiconductor process error analysis method of claim 1, wherein the workpiece is a wafer, a solar substrate, or a liquid crystal panel. A computer program product comprising a code for performing a semiconductor process error analysis method after loading an electronic device, wherein the step of the semiconductor process error analysis method comprises: obtaining one of a plurality of workpieces from a database a process record, wherein the plurality of workpieces comprise a plurality of good products and a plurality of defective products, and the process record includes a plurality of process parameters; and calculating one of the feature weights of each of the process parameters, the step comprising: dividing the process record into a first group and a second group, wherein the first group is composed of the plurality of good products, the second group is composed of the plurality of good products and the plurality of defective products; One of the parameters, the corresponding plurality of good products and the plurality of process materials of the plurality of defective products are filtered; calculating a first characteristic value of the first group; and calculating a second characteristic of the second group And determining, according to the first feature value and the second feature value, the feature weight; and sorting the plurality of feature weights to determine a possible process for causing the defective product Number. The computer program product of claim 11, wherein the smaller the first feature value and the second feature value, the smaller the feature weight. The computer program product of claim 11, wherein the first feature value and the second feature value are larger, the feature weight is larger. The computer program product of claim 11, wherein the first characteristic value is a first coefficient of variation of one of a plurality of first distances of each of the good products and the other good products in the first group; the second characteristic value And a second coefficient of variation of each of the good products in the second group and the second distance between the defective product and the other good product and the defective product. The computer program product of claim 14, wherein the calculating the first distance and the second distance comprises: calculating a square of a difference between the corresponding plurality of process data to obtain a plurality of intermediate values; and calculating The square root of the sum of the plurality of intermediate values to obtain the first distance or the second distance. The computer program product of claim 11, wherein the step of screening the plurality of process materials comprises: translating the plurality of process materials corresponding to the process parameters to the same reference; and depending on the good product or the defective product The process data of the pen is taken out, and the corresponding process data corresponding to the time is the closest. The computer program product of claim 11, wherein the difference between the number of the plurality of pieces of process data between the workpieces exceeds a predetermined value, the corresponding plurality of process materials are excluded, and an error record is recorded. The computer program product of claim 11, wherein the step of calculating the feature weight further comprises: screening a plurality of representative products from the plurality of good products, and calculating the feature weights by using the representative product The basis of calculation. The computer program product of claim 18, wherein the step of screening the representative product comprises: selecting a plurality of the good products as a representative good product group; calculating a distance between all the good products in the representative good product group a first average value; adding one of the representative good product groups to the representative good product group, and calculating a second average value of the distance between all the good products of the representative good product group; Whether the second average value is less than the first average value, and if so, retaining the newly added good product in the representative good product group; if not, discarding the farthest product in the representative good product group; and repeating the The first average step, the second average step, and the determining step are calculated until all of the good products have been screened. The computer program product of claim 11, wherein the workpiece is a wafer, a solar substrate, or a liquid crystal panel.