DETAILED DESCRIPTION OF THE INVENTION

[0065] Various embodiments of the present invention will be described with reference to the accompanying drawings. It is to be noted that the same or similar reference numerals are applied to the same or similar parts and elements throughout the drawings, and the description of the same or similar parts and elements will be omitted or simplified.

[0066] (Clustering Distribution not Detected by Feature Amount)

[0067] There are various causes of LSI failures, since there are several hundreds of manufacturing devices to manufacture the LSIs. Correspondingly, there are various failure clustering distribution patterns. In order to examine the cause of failure, an algorithm to detect wafers with a specific failure clustering distribution pattern is necessary. Usage of feature amounts obtained by changing the failure distribution pattern to corresponding numerical values allows detection of wafers with a specific failure clustering distribution pattern. The “feature amounts” is defined as the numerical values corresponding to the feature distribution pattern. The use of the term “feature” refers to features of the failure distribution pattern. Examples will be discussed below. However, usage of the feature amounts representing that specific failure distribution pattern does not allow detection of failure distribution patterns, other than the failure distribution pattern of interest. In addition, even when a calculation algorithm that finds various feature amounts and then categorizes the features is prepared in advance for various failure clustering distribution patterns, an unexpected failure distribution pattern cannot be detected. Accordingly, it is desirable to generate feature amounts capable of detecting even an unexpected failure distribution pattern.

[0068] (Importance of Identifying Failure Development Patterns in Lot Units)

[0069] In an LSI manufacturing line of a plant for an LSI manufacturing process, that process is performed for each lot of wafers, where a single lot includes 25 wafers. When the manufacturing device malfunctions, failures that match a specific failure clustering distribution may be developed in each wafer lot. The following feature may be found in the intra-lot failure distribution patterns. When one of two chambers in a manufacturing device malfunctions, failures that match a failure clustering distribution may be developed, for example, on only wafers with even wafer numbers attached to the wafers in the intra-lot processing order. As described above, intra-lot failure development patterns have a close relationship with malfunctions of a manufacturing device. Therefore, in order to identify the cause of a detected failure with failure clustering distribution, not only detection of wafers with a similar failure development pattern, but also detection of lots with a similar failure development pattern is critical. More specifically, failures with a failure clustering distribution are quantified as feature amounts for each wafer, and those feature amounts of failures are also reflected in each lot. This may allow identification of a manufacturing process that is the cause of failures.

[0070] (Wafer Failure Analyzer)

[0071] As shown in FIG. 1, a wafer failure analyzer 2 of an embodiment of the present invention has a bus 3, a generalized feature amount calculation unit 4, a clustering failure identification unit 9, an alignment unit 10, an autocorrelation coefficient calculation unit 11, an inter-wafer correlation coefficient calculation unit 12, a similar wafer detection unit 13, a similar wafer group generation unit 14, a similarity calculation unit 15, a sorting unit 16, a failure category generation unit 17, a failure category determination unit 18, a representative wafer determination unit 19, a representative wafer group determination unit 20, a common device identification unit 21, and an input/output unit 22. The generalized feature amount calculation unit 4 has a partitioning unit 5, an extension unit 6, a failure rate calculation unit 7, and a feature amount Sx/Sy calculation unit 8.

[0072] The wafer failure analyzer 2 may further have a failure information storage unit 35, a process history storage unit 36, an order library storage unit 37, and a report storage unit 38. Alternatively, as shown in FIG. 1, the failure information storage unit 35, the process history storage unit 36, the order library storage unit 37, and the report storage unit 38 may be provided outside the wafer failure analyzer 2, and may be connected thereto.

[0073] The wafer failure analyzer 2 may be a computer. The wafer failure analyzer 2 may be implemented by instructing the computer to execute a sequence described by a program.

[0074] (Lot Failure Analyzer)

[0075] As shown in FIG. 1, a lot failure analyzer 1 of the embodiment of the present invention has the wafer failure analyzer 2. In addition to the wafer failure analyzer 2, the lot failure analyzer 1 further has a representative lot value calculation unit 23, a target wafer detection unit 24, a feature amount per failure categorized wafer calculation unit 25, an inter-lot correlation coefficient calculation unit 26, a similar lot detection unit 27, a similar lot set generation unit 28, a similarity calculation unit 29, a sorting unit 30, a failure category generation unit 31, a characteristic representative lot value determination unit 32, a common device detection unit 33, and an input/output unit 34.

[0076] The lot failure analyzer 1 may be a computer. The lot failure analyzer 1 may be implemented by instructing the computer to execute a sequence described by a program.

[0077] (Wafer Failure Analysis Method)

[0078] As shown in FIG. 2, with a wafer failure analysis method of the embodiment of the present invention, to begin with, the input/output unit 22 receives a target wafer ID in step S1.

[0079] In step S2, the failure information storage unit 35 receives failure information of multiple tests. More specifically, the positions of failures in multiple wafers are input.

[0080] In step 3, the generalized feature amounts g are calculated by the generalized feature amount g calculation unit 4. In other words, the respective wafers are partitioned into multiple sections, and a failure rate is calculated for each section of the wafers. More specifically, a rate of failures existing within each section f is calculated for each of the multiple tests. This failure existing rate f is calculated as a ratio of the number of failure bits developed within each section to the number of all bits existing within the tested section. The first autocorrelation function in terms of the feature amounts Sx and Sy that represent exposure-caused failures is calculated for each section using an exposure cycle period as a lag. Each section is then extended. The number of failures developed within each extended section is used to calculate again the failure existing rate f. In the extended section, the area overlapped with the adjacent section is 60% or less than the area of the original section.

[0081] In step S4, the clustering failure identification unit 9 determines whether there is a failure with a failure clustering distribution. In other words, frequency distribution of the failure existing rate f is generated for each wafer, and a logarithm of this frequency distribution is approximated with a quadratic function. The second-order and the first-order coefficient of this quadratic function are then calculated, and whether there is a clustering failure is determined based on the second-order and the first-order coefficients.

[0082] In step S5, the order in which the alignment unit 10 aligns the failure existing rate f and the lag width are stored in an order library. In this order, the rates of failures existing within the respective sections f that are relevant to one another are aligned sequentially. The generalized feature amounts g are aligned based on the order library. The failure existing rates f are aligned for each wafer based on the orders stored in the order library, so as to form a waveform.

[0083] In step S6, the autocorrelation coefficient calculation unit 11 calculates the autocorrelation coefficients for the formed waveform based on the designated lag width stored in the order library.

[0084] In step S7, the inter-wafer correlation coefficient calculation unit 12 represents the degree of similarity between wafers in terms of the failure existing rates f as a numerical value. This numerical value may be correlation coefficients Y for all target wafers in terms of the failure existing rates f. Alternatively, this numerical value may be calculated by Fouriertransforming waveforms of the failure existing rates f and then comparing the first spectra of waveforms with each other. In this case, the numerical value is an agreement level between the spectra. Alternatively, the numerical value may be calculated using the maximum entropy method.

[0085] In step S8, the similar wafer detection unit 13 detects wafers to-be-paired with the correlation coefficient Y which greater than a certain threshold.

[0086] In step S9, the similar wafer group generation unit 14 generates similar wafer groups S each including wafers with similar failure distributions by detecting for each wafer a corresponding wafer which has a larger correlation coefficient Y than a predetermined threshold.

[0087] In step S10, the similarity calculation unit 15 calculates a similarity R among the similar wafer groups S, the ratio of the number of wafers belonging to both a similar wafer group and another similar wafer group to number of wafers belonging to either a similar wafer group or another similar wafer group. This similarity R among the similar wafer groups S allows extraction of other similar wafer groups which exceed a predetermined threshold.

[0088] In step S11, the sorting unit 16 sorts the similar wafer groups S by size.

[0089] In step S12, the failure category generation unit 17 groups the similar wafer groups S in the order of size, and generates a failure categories C. More specifically, the first failure category is configured by the similar wafer group and the other similar wafer groups that are extracted for that former similar wafer group, which are sorted in a decreasing order of the number of wafers belonging to each similar wafer group.

[0090] In step S13, the failure category determination unit 18 determines the failure category C to which each wafer belongs.

[0091] In step S14, the representative wafer determination unit 19 determines a representative wafer for each failure category C.

[0092] In step S15, the representative wafer group determination unit 20 determines the representative wafer group of the failure categories C.

[0093] In step S16, the common device identification unit 21 identifies, from process history manufacturing steps and manufacturing devices, the devices that are commonly used to manufacture the wafers belonging to the same failure category C.

[0094] In step S17, the input/output unit 22 outputs a failure analysis report.

[0095] The wafer failure analysis method can be represented by a wafer failure analysis program sequence, which can be executed by a computer. The wafer failure analysis method can be implemented by having the computer execute the wafer failure analysis program sequence.

[0096] (Lot Failure Analysis Method)

[0097] As shown in FIG. 3, with a lot failure analysis method of the embodiment of the present invention, to begin with, a target lot ID is input via the input/output unit 34 in step S21.

[0098] In step S22, the target wafer detection unit 24 finds a target wafer ID from a process history.

[0099] In step S23, the wafer failure analyzer 2 implements the wafer failure analysis method.

[0100] In step S24, the feature amount per failure categorized wafer calculation unit 25 calculates the feature amounts Ci (w) in terms of wafer w in failure category i (i corresponds to C in Example 1). More specifically, the ratio of the number of wafers belonging to the failure category i with the numerical value representing the degree of similarity among wafers that is greater than the predetermined threshold to the total number of wafers belonging to the failure category i is calculated as a feature amount Ci (w) for each wafer. In addition, the failure category I, to which no wafers belong, is set to zero.

[0101] In step S25, the representative lot value L calculation unit 23 calculates the representative lot value L for a wafer lot. At least one of average lot value, a wafer failure rate per lot, the intra-lot maximum value, degree of even/odd-caused inhomogeneous distribution, degree of first/latter half-caused inhomogeneous distribution, degree of wafer number-caused inhomogeneous distribution, or a periodic regularity is used as a feature amount Ci (W) for a wafer w for calculation of each representative lot value L.

[0102] In step S26, the representative lot value L calculation unit 23 aligns the representative lot values L so as to form a waveform for each lot. More specifically, the representative lot values L for a lot are sorted in decreasing order, so as to form a reference waveform.

[0103] In step S27, the representative lot value L calculation unit 23 calculates an error sum of squares from the waveform sorted in the corresponding order to the reference waveform formed for each lot and that reference waveform.

[0104] In step S28, the inter-lot correlation coefficient calculation unit 26 calculates a coefficient which represents the degree of similarity in representative lot value L between lots as a numerical value. This numerical value may be a correlation coefficient r in a representative lot value L among all of the lots. Alternatively, the numerical value may be calculated by Fouriertransforming the waveform of the representative lot values L and then comparing the second spectra with each other. Alternatively, this numerical value may be found using the maximum entropy method.

[0105] In step S29, the similar lot detection unit 27 detects lots to-be-paired with a correlation coefficient rij equal to or greater than the threshold of the correlation coefficient rij.

[0106] In step S30, the similar lot set generation unit 28 detects another lot with a correlation coefficient rij, which represents the degree of similarity between lots, equal to or greater than a predetermined threshold for each lot. The single lot and the detected other lots form a lot set S with a similar failure development pattern.

[0107] In step S31, the similarity calculation unit 29 calculates as the similarity R among the similar lot sets S, the ratio of the number of lots commonly belonging to those two similar lot sets S to the number of lots belonging to either of those sets S. Other lot sets with a similarity R from among the similar lot sets S equal to or greater than a predetermined threshold are detected.

[0108] In step S32, the sorting unit 30 sorts the similar lot sets S by size.

[0109] In step S33, the failure category generation unit 31 groups the similar lot sets S in decreasing order of size. A failure category G is configured with the similar lot sets S and other similar lot sets S detected with respect thereto in a decreasing order of the number of lots belonging to the similar lot sets S.

[0110] In step S34, the characteristic representative lot value determination unit 32 calculates the first total sum of the numerical values, each representing the degree of similarity in representative lot value L among all of the lots belonging to the failure category G, and the second total sum of numerical values, each representing the degree of similarity among lots when one component is excluded from the representative lot values L. One of the representative lot values L that provides the largest difference between the first and the second total sum is calculated. The calculated representative lot value L is determined as the representative lot value Lm that is the most characteristic in the failure category G.

[0111] In step S35, the common device identification unit 33 identifies, from the process history, a manufacturing device by which the lots with a large representative lot value Lm, which is the most characteristic in the failure category G, are unevenly processed. A manufacturing step and a manufacturing device commonly used to manufacture the lots belonging to the failure category G are identified from the process history.

[0112] In step S36, the input/output unit 34 outputs a report to the report storage unit 38.

[0113] The lot failure analysis method can be represented by a lot failure analysis program sequence, which can be executed by a computer. The lot failure analysis method can be implemented by making the computer execute this program.

[0114] As mentioned above, according to the failure analysis method of the embodiment, failure distribution developed in each wafer can be automatically classified in units of wafer and in units of lot.

[0115] According to the failure analyzer of the embodiment, failure distribution developed in each wafer can be automatically classified in units of wafer and in units of lot.

[0116] According to the failure analysis program of the embodiment, failure distribution developed in each wafer can be automatically classified in units of wafer and in units of lot.

EXAMPLE 1

[0117] In Example 1, the wafer failure analyzer 2 in FIG. 1 and the wafer failure analysis method in FIG. 2 are described. In Example 1, wafer failure distribution is represented by a lot of numerical value groups, and wafers with similar failure distributions are automatically identified.

[0118] In step S1, a target wafer ID is input to the input/output unit 22.

[0119] In step S2, failure information as shown in FIGS. 4 through 7 is input from the failure information storage unit 35. The failure information includes wafer failure position data, which is identified within a wafer indication region 41 and a failure indication region 42. This failure information is formed based on the FBM of the failure bits in certain semiconductor memory products, which have been tested after completion of a wafer manufacturing process, and stored in the failure information storage unit 35. The wafer indication region 41 shows an existing wafer region. The failure indication region 42 shows failure bit positions. It is found that the failure information includes failure bits with random distribution and failure bits with clustering distribution.

[0120] Moreover, it is determined that the wafers in FIGS. 4 and 7 include failure bits with clustering distribution at the outer regions of the wafers. Furthermore, it is determined that the wafer in FIG. 5 includes failure bits with clustering distribution at the center of the wafer. In FIG. 6, it is determined there are failure bits with clustering distribution at the lower right of the wafer. In addition, the wafer in FIG. 6 has failure bits that are distributed at regular intervals, and those intervals agree with an exposure cycle period in an exposure process. Accordingly, the failure bits in FIG. 6 are exposure-caused failures. In this way, it is determined that there are various failure distribution patterns for clustering failures. From the view of the cause of failures developed with clustering distribution, it is expected that the failures in FIGS. 4 and 7 may be caused by a manufacturing step or a manufacturing device in the same manufacturing process. On the other hand, the failures in FIGS. 4 through 6 may be caused by a different manufacturing step or manufacturing device, respectively. In order to identify the cause of failures developed with clustering distribution, wafers with similar failure clustering distribution should be identified, and a manufacturing step or a manufacturing device commonly used for the identified wafers should be identified based on the process history of the wafer manufacturing processes stored in the process history storage unit 36.

[0121] In step 3, the generalized feature amounts g are calculated by the generalized feature amount g calculation unit 4. In order to quantitatively represent the similarity, that is, whether or not the failure clustering distributions have similarity, the failure clustering distributions are represented by the generalized feature amounts g, which are a lot comprised of numerical value groups.

[0122] To begin with, a wafer is partitioned into multiple sections. As shown in FIG. 8, a boundary line 47 is provided (½)r that is spaced from the center of the wafer along the radius; a boundary line 48 is provided (¾)r that is spaced from the center of the wafer along the radius; and a boundary line 49 is provided, so as to partition chips, which are deployed in the outermost region of the wafer so as to contact the wafer edge, and other chips located interiorly of the wafer edge; where r denotes the radius of the wafer. Those three boundary lines 47 through 49 partition the wafer indication region 41 into ring-shaped regions 43 through 46.

[0123] Next, eight boundary lines 61 through 68 are provided so that the wafer is divided into 45 degree angles. Those eight boundary lines 61 through 68 partition the wafer indication region 41 into eight fan-shaped regions 51 through 58.

[0124] As shown in FIG. 10, sections in FIGS. 8 and 9 are integrated into a total of 104 defined sections. For example, section A is defined as the logical product of section 44, which is located between (½)r and (¾)r spaced from the center along the radius, and section 58, which is located between an angle of 315 degrees and an angle of 360 degrees. Section B is defined as the logical product of a wafer edge section 46, which extends along the radius, and sections 51 through 54, which extend between an angle of 0 degrees and an angle of 180-degrees. Similarly, the other sections can be defined as the logical product of sections partitioned along the radius and sections partitioned with a fixed angle.

[0125] Next, as shown in FIG. 11, the failure indication region 42 in FIG. 4 and positions of failures in the sections in FIG. 10 are compared to find a failure bit existing rate f_i for each of 104 sections i. The failure bit existing rate f_i is calculated using the following Expression (1), where nr_i denotes the number of all bits belonging to a section i, and nf_i denotes the total number of failure bits developed in a section i.

f_i=nf_i /nr_i (1)

[0126] Where i of the section i denotes a number assigned to each section.

[0127] For example, failure bit 71 is located in section 1, which is the logical product of sections 44 and 54. When the number of all bits belonging to section 1 is ten, the failure bit existing rate f₁ is 1/10or 0.1. Failure bit 72 is located in section 2, which is the logical product of sections 43 and 51. When the number of all bits belonging to section 2 is ten, the failure bit existing rate f₂ is 1/10 or 0.1. Failure bit 73 is located in section 3, which is the logical product of sections 43 and 56. When the number of all bits belonging to section 3 is ten, the failure bit existing rate f₃ is 1/10 or 0.1. Failure bits 74 through 78 are located in section 4, which is the logical product of sections 46 and 58. When the number of all bits belonging to section 4 is ten, the failure bit existing rate f₄ is 5/10 or 0.5. Failure bits 79 and 80 are located in section 5, which is the logical product of sections 46 and 57. When the number of all bits belonging to section 5 is ten, the failure bit existing rate f₅ is 2/10 or 0.2. Failure bits 74 through 80 are located in section 6, which is the logical product of section 46 and sections 57 and 58. When the number of all bits belonging to section 6 is twenty, the failure bit existing rate f₆ is 7/20 or 0.35. Failure bits 74 through 78 are located in section 7, which is the logical product of section 46 and sections 51 and 58. When the number of all bits belonging to section 7 is twenty, the failure bit existing rate f₇ is 5/20 or 0.25. The generalized feature amounts g are configured with the failure bit existing rates f_i.

[0128] Wafer tests for searching failure bits are performed for various kinds of electrical characteristics. The wafer test includes, for example, a function test that determines whether or not each bit functions as memory, or a margin test that determines whether or not an operating time or an electric current value satisfies the product standard even when each bit functions as memory. The failure existing rate f in Expression (1) is calculated for each wafer test and each wafer. As shown in FIGS. 12A through 12D, feature amount values or the values of the generalized feature amounts g are aligned in a row in a fixed order among wafers, and each waveform shown by a polygonal line is formed for each wafer. Those waveforms may be regarded as the waveforms of the generalized feature amounts g. Therefore, quantification of the similarity between wafers can be thought of as quantification of the similarity of the waveforms of the generalized feature amounts g. It can be seen from FIGS. 12A and 12D that the waveforms of the generalized feature amounts g for wafers in FIGS. 4 and 7 are similar to one another.

[0129] Note that the waveforms of the generalized feature amounts g change depending on how to align components gi of the generalized feature amounts g. In Example 1, selecting each section in turn from the adjacent sections, which is the logical product of sections 43 and 51, in a counterclockwise and spirally proceeding order, each corresponding component gi of the generalized feature amounts g is aligned. Regarding alignment for different wafer tests, to begin with, the components gi of the generalized feature amounts g for the same wafer test are aligned such that they are positioned next to each other. As to aligning differing wafer tests, similar wafer tests are arranged next to each other. This allows reduction in high-frequency components of the waveform.

[0130] Next, as shown in FIG. 13, the section that is the logical product of sections 44 and 58, for example, is partitioned into minute segments 121, each having a size matching the exposure cycle period, in order to detect features of the exposure-caused failures. The directions of the exposure cycle periods are set to x and y axes, respectively.

[0131] In addition, as shown in FIG. 14, the section in the wafer in FIG. 6, which is the logical product of sections 43 through 45 and sections 57 and 58 is divided into minute segments 121, each having a size matching the exposure cycle period. The directions of the exposure cycle periods are set to x and y axes, respectively. More specifically, as shown in FIG. 14(b), a projection profile q_j of the number of failure bits on the y axis is calculated. In addition, as shown in FIG. 14(c), a projection profile p_j of the number of failure bits on the x axis is calculated; where j denotes a segment number.

[0132] Next, as shown in FIGS. 15A and 15B, the autocorrelation functions Racx and Racy for the projection profiles p_j and q_j are calculated, respectively. The autocorrelation function Racx is calculated using Expressions (2) through (4) by shifting the projection profile p_j by a lag kx, which equals the exposure cycle period, and then finding correlation of the projection profile p_j with the shifted projection profile; in the same manner, the autocorrelation function Racy is calculated using the same expressions by shifting the projection profile q_j by a lag ky, which equals the exposure cycle period, and then finding correlation of the projection profile q_j with that shifted projection profile. 1

[0133] Calculation for the y axis is performed in the same manner using Expressions (2) through (4).

[0134] If there are exposure-caused failures, the projection profiles p_j and q_j have a periodicity that equals the exposure cycle period. The autocorrelation function Racx becomes a local maximum value when the projection profile p_j is shifted by the lag corresponding to the exposure cycle period; in the same manner, the autocorrelation function Racy becomes a local maximum value when the projection profile q_j is shifted by the lag corresponding to the exposure cycle period. As shown in FIGS. 15A and 15B, sx and sy denote the exposure cycle periods converted by a corresponding number of segments, respectively, and Racx(sx) and Racy(sy) denote the feature amount components Sxi and Syi that represent exposure-caused failures in each section, respectively; where i denotes a section number. Note that this calculation is also performed for each wafer test and for each wafer. The feature amount components Sxi and Syi are then added to the waveforms in FIGS. 12A through 12D for each wafer, configuring the generalized feature amounts g. In other words, a set of feature amount components Sxi and Syi for each wafer, which represent a failure bit existing rate fi and an exposure-caused failure for each section obtained by each wafer test, is called the generalized feature amount g. Hereafter, gi denotes the elements of the generalized feature amounts g, and are called generalized feature amounts. The suffix i of an element gi in the generalized feature amounts g can take number up to Ng, which is shared by a failure bit existing rate fi and feature amount components Sxi and Syi that represent exposure-caused failures, and is capable of being counted up or down.

[0135] Next, execution of steps S4 through S6 is omitted. In step S7, the inter-wafer correlation coefficient calculation unit 12 calculates the correlation coefficient Y among all target wafers in terms of the generalized feature amounts g. As shown in FIGS. 16A through 16C, inter-wafer correlation can be found from the scatter diagrams in which the generalized feature amounts g for differing wafers are plotted. In FIG. 16A, there is no correlation between the components gi of the generalized feature amounts g for the wafers in FIGS. 4 and 5. In addition, in FIG. 16B, there is no correlation between the components gi of the generalized feature amounts g for the wafers in FIGS. 4 and 6. On the other hand, in FIG. 16C, there are strong correlations between the components gi of the generalized feature amounts g for the wafers in FIGS. 4 and 7. In order to quantify those correlations, the correlation coefficients rij between wafers i and j in terms of the generalized feature amounts g are found using Expressions (5) through (8). 2

[0136] where, gk and hk denote the components of the generalized feature amounts for the wafers i and j, respectively.

[0137] As a result, the correlation coefficient r12 between wafers w1 and w2 in FIGS. 4 and 5 is 0.02. The correlation coefficient r13 between wafers w1 and w3 in FIGS. 4 and 6 is 0.03. The correlation coefficient r14 between wafers w1 and w4 in FIGS. 4 and 7 is 0.92.

[0138] In step S8, the similar wafer detection unit 13 sets the threshold to 0.8. If the correlation coefficient rij is greater than the threshold of 0.8, it is determined that those wafers i and j have similar failure distributions. As shown in FIG. 17, the correlation coefficients rij that are greater than the threshold of 0.8 are r14 and r41. Other correlation coefficients rij are less than the threshold of 0.8.

[0139] Conventionally, usage of only a specific failure distribution is allowed and a feature amount calculation algorithm dedicated thereto must be individually defined. On the other hand, in Example 1, the similarity of failure distribution among wafers can be quantified using the correlation functions rij in terms of the generalized feature amounts g regardless of the type of failure distribution. In addition, wafers with similar failure distributions can be automatically detected.

[0140] In the latter half of this Example 1, wafers with similar failure distributions are grouped according to similarity in failure distribution among the wafers, which is determined using the generalized feature amounts g, and failure categories are automatically generated.

[0141] Conventionally, failure categories have been generated by a human conducting visual inspection of the FBM and finding failure distributions that frequently occur. Automation of failure category generation allows configuration of a system that automatically outputs a report of problematic failures occurring at a manufacturing plant for semiconductor devices.

[0142] Sets of wafers, each having similar failure distributions, are generated from all wafers manufactured at a plant within a certain fixed period, such as one day or one week, based on the quantified similarity in failure distribution between two wafers.

[0143] To begin with, in step S9, the similar wafer group generation unit 14 generates similar wafer groups S, each having similar failure distributions. The similar wafer groups S are generated by identifying another wafer that is similar to each wafer.

[0144] For example, regarding wafer w1 in FIG. 17, a wafer similar to wafer w1 is the wafer w4. Consequently, a similar wafer group S1 configured with two wafers w1 and w4 is generated as shown in FIG. 18. Regarding wafer w2, there is no wafer that is similar to wafer w2. Consequently, a similar wafer group S2 configured with only a single wafer w2 is generated. Similarly, regarding wafer w3, there is no wafer that is similar to wafer w3. Consequently, a similar wafer group S3 configured with only a single wafer w3 is generated. Regarding wafer w4, the wafer similar to wafer w1 is wafer w4. Consequently, a similar wafer group S4 configured with two wafers w4 and w1 is generated.

[0145] In step S10, the similarity calculation unit 15 calculates the similarity Rij between the similar wafer groups Si and Sj. The similarity Rij is defined by the rate of wafers in the similar wafer groups Si and Sj correlating with each other. In other words, the similarity Rij is defined as the ratio of the number of paired wafers correlating with each other to the total number of paired wafers belonging to the similar wafer groups Si and Sj. The threshold for the similarity Rij is set to 0.5. It is determined that the similar wafer groups Si and Sj, having similarity Rij which is equal to or greater than the threshold of 0.5, are similar to each other. Note that the similarity Rij may be defined as the ratio of the number of wafers belonging to both the similar wafer groups Si and Sj to the number of wafers belonging to at least either the similar wafer group Si or Sj.

[0146] For example, the similarity R14 between the similar wafer groups S1 and S4 is the ratio of paired wafers (w1 and w4) correlated with each other to paired wafers (w1 and w4) belonging to the similar wafer groups S1 and S4, which is 1/1 or 1.0. As shown in FIG. 18, since the similarity R14 is 1.0, which is greater than the threshold of 0.5, it is determined that the similar wafer groups S1 and S4 are similar to each other. Similarly, since the similarity R41 is 1.0, which is greater than the threshold of 0.5, it is determined that the similar wafer groups S4 and S1 are similar to each other.

[0147] The similarity R12 between the similar wafer groups S1 and S2 is 0/1 or zero because there is no paired wafer correlating with each other against the paired wafers (w1 and w4) belonging to the similar wafer groups S1 and S4. Since the similarity R12 is zero, which is less than the threshold of 0.5, it is determined that the similar wafer groups S1 and S2 are not similar to each other. Similarly, since the similarity R21 is zero, which is less than the threshold of 0.5, it is determined that the similar wafer groups S2 and S1 are not similar to each other. Similarly, the similarities R13, R31, R42, R24, R43, R34, R23, and R32 are zero. Since each of the similarities R13, R31, R42, R24, R43, R34, R23, and R32 is less than the threshold of 0.5, it is determined that the paired similar wafer groups (S1 and S3), (S3 and S1), (S4 and S2), (S2 and S4), (S4 and S3), (S3 and S4), (S2 and S3), and (S3 and S2) are not similar.

[0148] Next, in step S11, as shown in FIG. 18, the sorting unit 16 sorts the similar wafer groups Si in decreasing order of the number of elements or the number of wafers. The similar wafer groups S1 and S4, each including two wafers, are determined as first and second groups, respectively. The similar wafer groups S2 and S3, each including a single wafer, are determined as third and fourth groups, respectively.

[0149] In step S12, the failure category generation unit 17 groups the similar wafer groups in decreasing order of sorted ranking of each similar wafer group, while referencing the determination of similarity based on similarity Rij. To begin with, the similar wafer group S4 determined to be similar to the similar wafer group S1 with the highest sorted ranking is grouped together therewith. A failure category C1 is assigned as an identifier to these similar wafer groups S1 and S4.

[0150] The highest ranking of the similar wafer group Si excluding the similar wafer group S1 and the similar wafer groups Si grouped in that similar wafer group S1 is the similar wafer group S2. Regarding the similar wafer group S2, there is no similar wafer group Si not assigned with the failure category C1 and determined as being similar to that similar wafer group S2. Therefore, the similar wafer group S2 alone configures a group, and a failure category C2 is assigned as an identifier to that similar wafer group S2. In the similar wafer groups Si not grouped yet, the highest ranked similar wafer group Si is the similar wafer group S3. Regarding the similar wafer group S3, there is no similar wafer group Si not assigned with the failure category C1 or C2 and determined as being similar to that similar wafer group S3. Therefore, the similar wafer group S3 alone configures a group, and a failure category C3 is assigned as an identifier to that similar wafer group S3. This grouping is effective since there are many cases where the upper ranked similar wafer groups Si include almost the same wafers as elements.

[0151] In step S13, the failure category determination unit 18 groups the similar wafer groups Si, and then makes the sets of wafers except for shared wafers among each failure category C assigned group belong to the failure categories C1, C2, C3. The rate of the number of wafers correlating with one another in terms of the generalized feature amounts g to number of wafers belonging to the failure categories C1, C2, or C3, is found for each wafer. If this rate is equal to or greater than a predetermined threshold, each wafer is determined to belong to a corresponding failure category C. As a result, failure categories C to which each wafer belongs are determined. Accordingly, a single wafer may belong to multiple categories C.

[0152] Specifically, the threshold is set to 0.4. Wafer w1 correlates with a single wafer w4 in terms of the failure category C1. Since the failure category C1 includes two wafers w1 and w4, the rate of the number of wafers is 1/2 or 0.5. Since this rate of wafer w1 of 0.5 is larger than the threshold of 0.4, wafer w1 belongs to the failure category C1. On the other hand, wafer w1 does not correlate with other wafers in terms of the failure categories C2 and C3. Therefore, the rate of the number of wafers is zero. Since this zero rate of wafer w1 is smaller than the threshold of 0.4, wafer w1 does not belong to the failure categories C2 and C3.

[0153] Wafer w4 correlates with a single wafer w1 in terms of the failure category C1. Since the failure category C1 includes two wafers w1 and w4, the rate of the number of wafers is 1/2 or 0.5. Since this rate of wafer w4 of 0.5 is larger than the threshold of 0.4, wafer w4 belongs to the failure category C1. On the other hand, wafer w4 does not correlate with other wafers in terms of the failure categories C2 and C3. Therefore, the rate of the number of wafers is zero. Since this zero rate of wafer w4 is smaller than the threshold of 0.4, wafer w4 does not belong to the failure categories C2 and C3.

[0154] In step S14, the representative wafer determination unit 19 determines as a representative wafer, one of the wafers belonging to each failure category C and correlates with the largest number of wafers in terms of the generalized feature amounts g, For example, of the wafers w1 and w4 in the failure category C1, wafers w1 and w4 correlating to each other are determined to be representative wafers.

[0155] In step S15, the representative wafer group determination unit 20 includes in a failure category representative wafer groups a wafer correlating with the representative wafer in the failure categories C in terms of the generalized feature amounts g. For example, wafers w1 and w4 correlating with the representative wafers w1 and w4 in the failure category C1 are determined as a representative wafer group.

[0156] In step S16, the common device identification unit 21 identifies a manufacturing step and a manufacturing device commonly used to manufacture wafers belonging to the same failure category C from a process history stored in the process history storage unit 36. For example, the manufacturing step and the manufacturing device commonly used to manufacture wafers w1 and w4 belonging to the failure category C1 are identified.

[0157] Finally, in step S17, the failure categories C generation result is reported from the input/output unit 22 to the report storage unit 38. As shown in FIG. 19, the reported content includes the number of wafers and failure distribution in the representative wafer or the representative wafer group, which are shown in decreasing order of the number of wafers.

[0158] As mentioned above, in Example 1, the similar failure distribution is highly accurately identified, and the failure categories C are automatically generated from failure information regardless of the failure distribution patterns. These failure categories allow immediate identification of a clustering distribution of failures based on the representative wafer failure distribution. In addition, a manufacturing step and a manufacturing device at a wafer manufacturing plant that cause a clustering distribution of failures to occur can be identified quickly. This allows enhancement of the wafer yield.

[0159] As described above, according to the failure analysis method of Example 1, failure distribution developed in each wafer can be automatically classified for each wafer. According to the failure analyzer of Example 1, failure distribution developed in each wafer can be automatically classified for each wafer.

EXAMPLE 2

[0160] In Example 2, the failure categories C, which are sets of wafers, each having similar failure distributions, are generated from all of the wafers manufactured at a plant within a certain period such as one day or one week. Example 2 also employs the wafer failure analyzer 2 in FIG. 1 and conforms to the flowchart of FIG. 2.

[0161] To begin with, in step S1, a target wafer ID to be subjected to failure analysis is input. In Example 2, wafers to be manufactured at a plant within one day are subjected to failure analysis.

[0162] In the following, steps S2 through S9 are the same as those in Example 1.

[0163] In step S10, the similarity Rij between the similar wafer groups Si and Sj is calculated. The threshold for the similarity Rij is set to 0.5. It is determined that the similar wafer groups Si and Sj with a greater similarity Rij than the threshold of 0.5 are similar to each other.

[0164] For example, as shown in FIG. 20, the similarity R106 between the similar wafer groups S10 and S6 is greater than the threshold of 0.5, and thus they are marked with a symbol, such as a circle, and determined to be similar to each other. Since the similarity R102 between the similar wafer groups S10 and S2 is less than the threshold, they are marked with a symbol, such as ‘x’, and determined not to be similar dissimilar to each other. Similarly, since the similarity R108 between the similar wafer groups S10 and S8 is greater than the threshold, they are marked with a symbol of a circle and determined to be similar to each other. Since the similarity R109 between the similar wafer groups S10 and S9 is less than the threshold, they are marked with a symbol, such as ‘x’, and determined not to be similar dissimilar to each other. Since the similarity R103 between the similar wafer groups S10 and S3 is less than the threshold, they are marked with a symbol, such as ‘x’, and determined not to be similar dissimilar to each other. Since the similarity R62 between the similar wafer groups S6 and S2 is less than the threshold, they are marked with a symbol, such as ‘x’, and determined not to be similar dissimilar to each other. Since the similarity R68 between the similar wafer groups S6 and S8 is greater than the threshold, they are marked with a symbol, such as a circle, and determined to be similar to each other. Since the similarity R69 between the similar wafer groups S6 and S9 is less than the threshold, they are marked with a symbol, such as ‘x’ and determined not to be similar dissimilar to each other. Since the similarity R63 between the similar wafer groups S6 and S3 is less than the threshold, they are marked with a symbol, such as ‘x’ and determined not to be similar dissimilar to each other. Since the similarity R28 between the similar wafer groups S2 and S8 is greater than the threshold, they are marked with a symbol, such as a circle, and determined to be similar to each other. Since the similarity R29 between the similar wafer groups S2 and S9 is less than the threshold, they are marked with a symbol, such as ‘x’, and determined not to be similar dissimilar to each other. Since the similarity R23 between the similar wafer groups S2 and S3 is greater than the threshold, they are marked with a symbol of a circle and determined to be similar to each other. Since the similarity R89 between the similar wafer groups S8 and S9 is less than the threshold, they are marked with a symbol of ‘x’ and determined not to be similar to each other. Since the similarity R83 between the similar wafer groups S8 and S3 is greater than the threshold, they are marked with a symbol of a circle and determined to be similar to each other. Since the similarity R93 between the similar wafer groups S9 and S3 is less than the threshold, they are marked with a symbol of ‘x’ and determined not to be similar to each other.

[0165] Next, in step S11, the similar wafer groups Si are sorted in decreasing order of the number of wafers belonging to the groups or in the order of S10, S6, S2, S8, S9, and S3, as shown in FIG. 20.

[0166] In step S12, the similar wafer groups Si that are similar to the similar wafer group S10 are grouped in decreasing order of the sorted ranking from the similar wafer group S10, while referencing the similarity and the dissimilarity indicated by a symbol of a circle or a symbol of ‘x’, respectively. Regarding the similar wafer group S10, the similar wafer groups S6 and S8 those are similar to the similar wafer group S10 are grouped. A failure category C1 is assigned to these similar wafer groups S10, S6, and S8 as an identifier.

[0167] Of the similar wafer groups Si not grouped together with the similar wafer group S10, the similar wafer group S2 includes the greatest number of wafers. The similar wafer group S3, which is similar to the similar wafer group S2 in the similar wafer groups Si not grouped in the similar wafer group S10, is grouped together with the similar wafer group S2. A failure category C2 is assigned to these similar wafer groups S2 and S3 as an identifier. Since the similar wafer group S8 is similar to the similar wafer groups S2 and S10, and the similar wafer group S10 has a greater number of wafers, the similar wafer group S8 is grouped together with the similar wafer group S10. This grouping is effective since there are many cases where almost all of the wafers are shared among the upper ranked similar wafer groups Si.

[0168] Except for the similar wafer groups Si that have already been grouped, the highest ranked similar wafer group Si is the similar wafer group S9. There is no similar wafer group Si similar to the similar wafer group S9 except for the similar wafer groups that have already been grouped. Therefore, the similar wafer group S9 alone configures a group, and a failure category C3 is assigned to the similar wafer group S9 as an identifier.

[0169] In step S13, similar wafers are grouped together, and the sets of wafers resulting from excluding shared wafers among each group are then determined as failure categories C1, C2, C3, respectively. The rate of the number of wafers correlating to each other in terms of the generalized feature amounts g to number of wafers belonging to the respective failure categories C1, C2, or C3, is found for each wafer. If this rate is greater than a predetermined threshold, the respective wafers belong to one of the corresponding failure categories C.

[0170] In step S14, a wafer correlating with the largest number of other wafers in terms of the generalized feature amounts g is selected from the wafers in each of the failure categories C and determined as a representative wafer.

[0171] In step S15, the wafer correlating with the representative wafer in each failure category C in terms of the generalized feature amounts g is determined to form a representative wafer group.

[0172] In step S16, a manufacturing process and a manufacturing device commonly used to manufacture wafers belonging to the same failure category C are identified from a process history stored in the process history storage unit 36.

[0173] Finally, in step S17, the failure categories C generation result is output as a report. As shown in FIG. 21, the reported content includes the number of wafers and failure distribution in the representative wafer or the representative wafer group in decreasing order of the number of wafers belonging thereto.

[0174] As mentioned above, in example 2, a similar failure distribution is identified very accurately, and the failure categories C are automatically generated from failure information of a lot of wafers tested at a plant within one day regardless of failure distribution patterns. This failure category allows identification of failures with clustering distribution based on the representative wafer failure distribution. In addition, a manufacturing step and a manufacturing device at a wafer manufacturing plant that cause failures with clustering distribution to occur can be immediately identified. This allows enhancement of the wafer yield.

[0175] As described above, according to the failure analysis method of Example 2, failure distribution developed in each wafer can be automatically classified in units of wafer and also in units of lot. According to the failure analyzer of Example 2, failure distribution developed in each wafer can be automatically classified in units of wafer and also in units of lot.

EXAMPLE 3

[0176] In Example 3, the sections for calculating the generalized feature amounts g in step S3 of FIG. 2 are moved, expanded, and/or contracted. This allows improved accuracy in detection of wafers with similar failure distributions.

[0177] As shown in FIGS. 22 and 23, failure distributions in two wafers are extraordinarily similar to each other in that failures tend to gather at the outer regions of the wafers. It is thought that the cause of failures in those wafers in FIGS. 22 and 23 may be the same; therefore, it should be determined that those wafer failure distributions are similar to each other. However, from detailed examination based on step S3 in FIG. 2, it can be seen that the wafer in FIG. 22 has failures with clustering distribution gathering at the outer region of the wafer, in particular, in section C, and part of the failures exist in section D. In the wafer of FIG. 23, clustering failures grouped at the outer region of the wafer exist in sections C and E. The correlation coefficient r with respect to the generalized feature amounts g between those wafers is 0.76. Since the correlation coefficient r is less than the threshold of 0.8, it is determined that the wafer failure distributions in FIGS. 22 and 23 are not similar to each other.

[0178] This is because the sections are defined systematically, and failures crossing the boundary between those sections even slightly may considerably affect the generalized feature amounts g depending on the failure distribution pattern near that boundary line.

[0179] Therefore, all sections are moved, expanded, and/or flexibly contracted. More specifically, section C is moved, expanded and/or contracted within a certain range that allows the ratio of the overlapped area between section C and the adjacent sections to the original area of section C to be less than a fixed value. Regarding section C capable of being moved, expanded, and/or contracted, the area of section C that allows a failure existing rate f to be the greatest when moving, expanding, and/or contracting that section C is determined as a defined area for section C.

[0180] Movement, expansion, and/or contraction of sections may be carried out along the radius and/or with respect to the center angle of the wafer. The ratio that designates the range of movement, expansion, and contraction is set to 60%. As shown in FIG. 24, section C can be widened by 60% along the radius of the wafer. In addition, section C can be widened by 60% or a 27-degree angle with respect to the center angle of the wafer.

[0181] As shown in FIG. 25, section C is expanded to the range including an additional 109-degree angle in the case of the failure distribution of FIG. 22. As shown in FIG. 26, section C is moved to the next section up to the range including an additional 110-degree angle in the case of the failure distribution of FIG. 23. According to the movement and expansion of this section, the correlation coefficient r between the wafers in terms of the generalized feature amounts g is 0.92. Since the correlation coefficient r is larger than the threshold of 0.8, it is determined that the failure distributions in the wafers of FIGS. 22 and 23 are similar to each other.

[0182] Note that when general feature amounts g corresponding to the case where the smaller the value, the more failures exist are used, the sections should be moved, expanded, and/or contracted so that the generalized feature amounts g can be the local minimum values.

[0183] As described above, flexible movement, expansion, and/or contraction of the sections allows reduction in the number of false determinations of similarity between failure clustering distributions.

[0184] As mentioned above, according to the failure analysis method of Example 3, failure distributions in each wafer can be automatically classified in units of wafer and even in units of lot. According to the failure analyzer of Example 3, failure distributions in each wafer can be automatically classified in units of wafer and even in units of lots.

EXAMPLE 4

[0185] With Example 4, in steps S5 and S6 in FIG. 2, wafers with similar failure distributions are identified using the generalized feature amounts g defined in Example 1. For this identification, the alignment of the components in the generalized feature amounts g is taken into account.

[0186] The use of steps S5 and S6 allows omission of steps S7 and S8 of Example 1. In this case, similarity between wafers can be determined whether or not both of the autocorrelation coefficients are large and the difference between the autocorrelation coefficients is equal to or less than a fixed value. In short, if wafer failure distributions are similar to each other, the difference between the autocorrelation coefficients for each wafer is small. On the other hand, if wafer failure distributions are not similar to each other, the difference between the autocorrelation coefficients for each wafer is large.

[0187] Regarding the generalized feature amounts g as a waveform, each waveform differs depending on the alignment order of the components in the generalized feature amounts g. Positioning relevant components next to one another allows formation of a waveform with regularity. Which components are relevant to one another depends on the failure distribution patterns.

[0188] For example, Example 4 is used for wafers with failures expending radially from the wafer center, as shown in FIG. 27.

[0189] In step S5, the alignment unit 10 stores as the order library in the order library storage unit 37 in FIG. 1, a set of an alignment order 112, which indicates the order in which the generalized feature amounts g are spirally aligned counterclockwise as shown in FIG. 28A, and a lag width to be used to calculate an autocorrelation coefficient. The lag width, which is an element of the set with the alignment order 112, is set based on the number of each component relevant to the respective sections 43 through 46. In addition, a set of an alignment order 113 through 115, which indicates the order in which the generalized feature amounts g are radially aligned as shown in FIG. 29A, and a lag width to be used to calculate the autocorrelation coefficients, is stored as an order library in the order library storage unit 37. The lag