R I Lithography E
T
I
C
L
E
N
S
P
E
C
T
I
O
N
Progressiveness Kills Quality Arresting Reticle Quality Degradation in Wafer Fabs Carmen Jaehnert, Doris Uhlig, Infineon Technologies Kaustuve Bhattacharyya, Kong Son, Ben Eynon, Dadi Gudmundsson, KLA-Tencor Corp.
DUV lithography has introduced a progressive mask defect growth problem widely known as crystal growth or haze. Even if the incoming mask quality is good, there is no guarantee that the mask will remain clean during its production usage in the wafer fab. These progressive defects must be caught in advance during production in the fabs. The ideal reticle quality control goal should be to detect any nascent progressive defects before they become yield limiting. Therefore, a high-resolution mask inspection is absolutely needed; but, then the big question is: “How often do fabs need to re-inspect their masks?� This article builds on previous work1 to present a realistic mask re-qualification frequency model that has been developed based on the data from an advanced fab environment that is using low k1 lithography. Statistical methods are used to analyze mask inspection and product data, which are combined in a stochastic model.
Background
It was traditionally thought that if a clean mask is delivered from the mask house to the fab, very little can go wrong with the mask quality during its lifetime in fab production, unless there is mishandling, etc., during production usage. In fabs today, however, this concept is no longer valid. Industry data2 shows that mask returns from fabs due to contamination defects are significant. Even if masks arrive from the mask house perfectly clean, over the course of production usage in the fab some of these masks show catastrophic defect growth— commonly known as crystal growth, haze, fungus or precipitate3. This is a progressive defect growth on reticles causing reticlequality degradation over time, which subsequently impacts device yields. This scenario dominates in fabs using DUV lithography4. Traditional ESD defects and migrating defects (from non-critical to critical locations on the mask) also fall under this category, and should also be monitored. As a result, routine mask re-qualification in fabs has 38
Winter 2005
Yield Management Solutions
become a necessity. Yet, developing a suitable re-qualification frequency for all the layers of all mask sets that are in production usage, is not a simple task. In a previous study on finding a cost-effective mask requalification frequency, Vince Samek et al1, developed a statistically-based methodology to plan and/or optimize the use of reticle inspection capacity in the fab. Statistical methods were used to analyze reticle and product data, which were combined, in a stochastic model with financial parameters. The model used the combined information to calculate the cost of different reticle inspection strategies, allowing both capacity planning and allocation optimization of given inspection capacity. At the time of this study, however, the progressive defect growth problem was not as prominent as it is today. Data collection to completely characterize the failure rate, etc. of the problem masks was performed manually and based on a small number of masks (around 40). It became important to extend this work by collecting a much larger volume of data from a few low-k1 fabs that are currently experiencing a progressive defect growth problem.
R
E
T
I
C
L
I
E
N
S
P
E
C
T
I
O
N
Results from masks running the 365 nm lithographic process
Incoming mask inspection (STARlight) defect map
Re-qualification inspection (STARlight) defect map
Figure 1. Progressive mask defect growth in fabs.
In Figure 1, a fab’s incoming inspection record indicates that the mask arrived clean but, after only a month of production usage, the mask showed catastrophic defect growth. If there were a re-qualification frequency established in the fab, it would be helpful to see the numbers of mask inspections within this one month. However, as defect growth is not constant between mask to mask, and lithographic process to process, it is not an easy task to set up the re-inspection frequency for this time period. Thorough data analysis and statistical modeling is required to establish a suitable mask re-qualification frequency. Data collection procedure
All the inspection results from KLA-Tencor’s STARlight™ mask inspection tools are stored on connected data servers. These servers sometimes carry more than one year of inspection data. KLA-Tencor has developed a methodology to extract the stored inspection results from the server to plot the history of inspections in a meaningful format. A set of data extraction, manipulation, and analysis codes was developed to use in conjunction with the data from the server. This enables flexible extraction of only defect data that is important and relevant to the user. This is necessary due to the extremely large volume of data. The extracted data can then be used in a statistical model to develop a suitable sampling plan or to determine mask re-qualification frequency.
The first set of data from one of the fabs shows that there is not much mask Huge Defect Growth - ~ 1 micron contaminants defect growth problem in the 365 nm spreading all over the mask lithographic process. This result was expected, since we know that defect growth increases at lower exposure wavelengths. The data in Figure 2 shows the defect count history for multiple plates used in the 365 nm exposure wavelength over time.
Mask Defect vs. Time (365nm Lithographic Process)
Defect Count, #
After 1 month of production use
0
30
60
90
120
150
180
Days Figure 2. Mask defect versus time at the 365 nm exposure wavelength.
Results from masks running the 248 nm lithographic process Masks exposed at the 248 nm wavelength showed more activity than the previous 365 nm process. As the wavelength decreases, this behavior is expected.
Mask Defect vs. Time (248nm Lithographic Process)
The following data is extracted from the last six months of production run results stored on the server. Except at the 365 nm wavelength, most of the data was comparable between the two fabs. Therefore, the majority of the time, only the data from one fab is shown. If there is a significant difference, the other fab’s data is also shown. Y-axes for all the charts below are on the same scale.
Defect Count, #
Evaluation results
0
30
60
90
120
150
180
Days Figure 3. Mask defect versus time at the 248 nm exposure wavelength.
Winter 2005
www.kla-tencor.com/magazine
39
R
E
T
I
C
L
E
I
N
S
P
E
C
T
I
O
N
Darkfield Masks Mask Defect vs. Time
Mask Defect vs. Time
Defect Count, #
Defect Count, #
(193 nm Lithographic Process)
0
30
60
90
120
150
180
0
Days
30
60
90
120
150
180
Days
Figure 4. Mask defect versus time in 193 nm wavelength of exposure
Figure 6. Darkfield masks remain stable in defect count over time.
light.
Results from masks running the 193 nm lithographic process In Figure 4, it is clearly visible that masks running at the 193 nm exposure wavelength become defective at a much higher rate than masks running at lower wavelengths.
The above charts are plotted against time. However, a wafer counter was also used for many masks of interest, and that data will be used in the statistical model.
Mask Behavior After Cleaning
Results between clearfield and darkfield masks Figures 5 and 6 illustrate that clearfield masks (such as gate/poly, metal level, etc.) show faster degradation with an increase in defect count, while defect counts for the darkfield masks (such as contact, implant level, etc.) remain stable over time. Defect Count, #
Gate Level 1 Gate Level 2 Mask sent back for cleaning
Clearfield Masks Mask Defect vs. Time
0
30
60
90
120
150
180
210
Days Defect Count, #
Figure 7. Defect growth repeats after cleaning.
Stochastic reticle inspection cost model 0
30
60
90
120
150
180
Days Figure 5. A large number of clearfield masks showing fast defect growth.
Results from masks that are cleaned after defect detection After a mask reaches a critical defect growth state, the mask is generally sent back for cleaning. We observed that when the mask returns to the fab after cleaning, it becomes more susceptible to defect growth. Figure 7 illustrates a typical example. 40
Winter 2005
Yield Management Solutions
A stochastic reticle inspection cost model has been developed in conjunction with UC Berkeley that allows the total reticle inspection cost to be estimated1. The total reticle inspection cost is defined as the sum of the cost of reticle inspection and the cost of yield loss due to the printing of un-detected reticle defects. Given a reticle failure rate and inspection frequency (in addition to other inputs), the reticle inspection cost model provides an estimate of the cost due to yield loss which is then added to the reticle inspection tool cost of ownership (CoO) to obtain the total reticle inspection cost for a layer.
R
E
T
I
C
L
E
I
N
S
P
E
C
T
I
O
N
• Reticle failure probability is a function of the number of wafers exposed and the probability of reticle failure after “n” exposures have a mixed geometric distribution.
$
$ Frequency
Material at risk cost
$ Frequency
+
Inspection cost
Frequency
=
• Other products/layers can use the stepper while the reticle in question is being inspected. This means that reticle inspection has a negligible effect on stepper throughput.
Total cost
Figure 8. The combination of material at risk cost and inspection cost reveals how an optimum reticle inspection plan and capacity can be achieved.
Model features and assumptions
The main features of the reticle cost model: • Accounts for reticle failure detection ability by reticle inspection tools. • Accounts for reticle failure detection ability through tool monitoring of stepper (includes test-wafer transit time and queuing delay in front of wafer inspection tool). • Accounts for reticle failure detection at end of line, i.e. at probe. • Accounts for cost of tool monitoring, i.e. test wafer cost, test wafer cleaning cost, and hours of wafer inspection tool use. • Recovery of yield through re-work of un-etched wafers can be included. • Calculates hours of inspection tool use per week for both reticle and wafer inspection tools. Reticle inspection cost model assumptions: • Reticle failure is defined as the time point where a defect on the reticle starts to be printed on wafers, and the defect that appears on the wafer is a killer. • Once a reticle starts printing a defect, it continues printing until the defect (on the reticle or on the test wafer) is detected. • When a reticle fails, only one die in the field will be killed. If there is only one die in the field, then the yield hit is 100 percent; if there are two die in the field, the yield hit is 50 percent, and so on.
• Beta risk (the probability of missing a yield-hit signature) is the parameter that describes the ability of reticle inspection tools and wafer inspection tools to catch a reticle failure. Beta risks can be calculated with tools such as Sample Planner5. Analysis objective and reticle failure rate
The most important and most difficult aspect of the analysis is to establish an estimate for reticle failure rate. Reticle failure events are low frequency events that justify a relatively high inspection rate due to the high yield impact they have. Not only is the failure rate often low, but the failure rate is also a function of factors such as pattern density, which affects both defect formation on the reticle and the impact of printed defects on the circuit. This prevents all reticles from being thrown into one category in order to find a mean time to failure. Instead, the reticles need to be categorized based on several characteristics, leaving a decreasing number of reticles in each category from which to obtain a valid estimate. It is also important to identify the analysis objective when attempting to estimate the reticle failure rate input. Using an average failure rate and then optimizing the inspection frequency to obtain the lowest total inspection cost is based on the assumption that some loss of material to a reticle failure event can be tolerated and compensated for with subsequent production. There are cases when the loss of many lots cannot be tolerated, as this puts a promised delivery date at risk, which may subsequently affect future business opportunities with the customer in question. In these cases, a few reticle failure data points (not averaged with non-failing reticles) are enough to establish what can be expected in the worst case. Analysis based on these failure rates will reveal cost-effective inspection frequencies in situations where risk of any material loss needs to be minimized. This method of worst-case analysis is easier than establishing the true failure rate and then establishing an inspection frequency that meets constraints on material put at risk, since that would require considerably more data. Winter 2005
www.kla-tencor.com/magazine
41
R
E
T
I
C
L
I
E
N
S
P
E
Re-qualification frequency
After analyzing the data, we found that masks could be divided into four risk categories: (i) Masks that are cleaned due to a previous defect problem (the most susceptible to mask defect growth) (ii) Clearfield masks running 193 nm lithography (iii) Clearfield masks running 248 nm and 365 nm lithography (iv) Darkfield masks (the least susceptible to mask defect growth) Figure 9 illustrates the cost versus inspection interval model for each of these categories using all the above data. This model tries to predict the statistical best frequency for re-inspection (mask re-qualification frequency) for each mask category. The point where the total cost is the lowest should be the best re-qualification frequency for each category. Conclusion
In advanced fabs, a carefully developed reticle inspection strategy must be implemented with the goal of minimizing the mean time to detect (MTTD) any defect growth resulting from prolonged reticle use. For this study, we attempted to identify an appropriate frequency of periodic inspection of reticles at high resolution. We analyzed thousands of masks over a six-month period to
Total cost $ / wk
DRAM Fab Total cost vs. inspection interval
(i) (ii) (iii) (iv) 0
100
200
300
400
500
600
700
800
900
1000
Inspection interval (# of wafers) Figure 9. Cost versus inspection inter val model for four categories in both LOGIC and DRAM fabs.
42
Winter 2005
Yield Management Solutions
C
T
I
O
N
understand their defect behavior at different process levels and exposure wavelengths. We found that the 193-nm lithographic process causes most of the defect growth versus the 248-nm lithographic process that showed much lower defect growth. We also observed that clearfield masks are more prone to failure than darkfield masks. The data showed that after a mask is identified with defect growth and is cleaned, it is more susceptible to defect growth under subsequent exposures. The statistical model can incorporate a layer-specific defect growth threshold (a worst case scenario is discussed in the previous section) and it can then predict a re-qualification frequency in terms of the number of wafer exposure intervals. Acknowledgements
The authors would like to thank Dr. G. C. Stein for writing some excellent data extraction, manipulation, and analysis codes for this project. The authors would also like to thank the following KLA-Tencor engineers for their help on this project: Steve Mylroie for the quick development of a faster data extraction code for the server, Bar Houston for loading large databases on factory servers, Paul Yu for developing a data visualization macro for this project, and Michael Lang for his help in obtaining production data. This article was previously published in the SPIE Proceedings, vol. 5567. References 1. V. Samek, B. Shiffler, W. Tomlinson, D. Gudmundsson, J. Merritt, R. Nurani, J. G. Shanthikumar, Cost Effective Reticle Quality Management Strategies in Wafer Fabs, The 10th Annual IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop, September 8 - 10, 1999. 2. K. Kimmel, Mask Industry Assessment: 2002, Introductory paper, BACUS Symposium on Photomask Technology, 2002. 3. K. Bhattacharyya, William Volk, Brian Grenon, Darius Brown and Javier Ayala, Investigation of reticle defect formation at DUV lithography, BACUS Symposium on Photomask Technology, 2002. 4. B. J. Grenon, C. R. Peters, and K. Bhattacharyya, Tracking down causes of DUV sub-pellicle defects, Solid State Technology, June 2000. 5. R. R. Williams, D. Gudmundsson, R. Nurani, R. Stoller, M. Chatterjee, A. Seshadri, S. Shanthikumar, “Challenging the Paradigm of Monitor Reduction to Achieve Lower Product Costs�. The 10th Annual IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop, September 8 - 10, 1999.