300mm fab automation can meet advanced defect detection needs
10/01/2001
Yield/productivity SPECIAL REPORT
COVER article
 |
Ralph Spicer, Dadi Gudmundsson, Raman Nurani KLA-Tencor Corp., San Jose, California
overview
Built-in particle inspection is not likely to provide benefit over a comprehensive standalone inspection approach that is facilitated by 300mm fab automation and robotics. Among other justifications, significant time-to-detect benefits may be achievable through optimized automation and robotics support and fab layout without the loss of flexibility. This avoids the added capital cost associated with built-in detection, which may offer little incremental benefit once automation and fab layout are optimized.
Whether or not to integrate defect inspection into process tools is an important decision facing 300mm fab planners. It affects capital procurement strategies, fab automation and robotics, floor planning, and data-systems integration. Built-in inspection is an expensive decision that will be almost impossible to change once selected. The alternative is a comprehensive defect inspection strategy using cell integration: cell process control implemented with standalone inspection tools serving multiple process tools, all linked by automation, robotics, and yield control information systems (Fig. 1).
 |
Capital-cost considerations are substantial. In the built-in case, each inspector is tied to a specific process tool (1:1). By comparison, standalone defect-inspection systems typically serve 4-10 process tools. For this reason, proposed built-in inspectors have been limited to lower-cost systems designed to detect primarily large particles and scratches, as opposed to a more comprehensive range of yield-limiting defects with standalone inspectors.
 |
With these two alternatives in mind, we set out to understand relevant issues surrounding automation of defect inspection for fully automated 300mm fabs. We wanted to see whether there is a justification beyond "gut feel" arguments that favor built-in defect inspection, or does the application of defect inspection to 300mm wafer processing justify the automated alternative?
 |
Process control considerations
There are three very different classes of built-in technology that are being pursued today: integrated metrology, integrated particle inspection, and integrated defect inspection. The differences among them are significant (Table 1), and it is vital to analyze each separately. In particular, it is important not to confuse applications of built-in metrology with those of defect and particle inspection.
 |
Built-in metrology is applicable to advanced process control concepts, such as feedforward (e.g., modifying a recipe based on film thickness measurements) and feedback (e.g., modifying deposition). Defect inspection, on the other hand, does not lend itself to such applications: What would one adjust on an etcher where an "excursion" (i.e., a process out-of-control event on an SPC chart) is detected by the defect inspector?
 Defect types (top to bottom): missing patterns; barrier depressions; bridging; pin hole; and unprinted line segment. |
Equipment productivity arguments for built-in integration are real — fewer wafer moves and faster time-to-inspection for increased process tool uptime through faster fixes. These are important, but it is also important to consider that there are many variables driving yield, and it is the productivity of yielding die, not all die, that drives fab profitability. Two primary yield variables are:
- Defect types — does an inspection strategy have a high probability of detection for defect types that cause yield loss? Will it keep up with as-yet-unknown defect types that will occur as design rules shrink and new processes are introduced? Will it be able to adapt as fab yield learning proceeds?
- Time to detection — does the inspection approach minimize the time to detect a defect "excursion"? Can much of the benefit of integration be achieved through well-planned product flow, automation, and robotics instead?
The 300mm-inspection task
Any robust defect control strategy for 300mm-wafer fabs must be capable of finding both process-induced and tool-induced defect types with a high probability of detection. At advanced design rules, defect densities must decrease to achieve viable production yields [1]. As defect densities fall, the criteria for "excursions" becomes tighter, meaning that an inspection system must be able to detect smaller defects without an undue increase in false alarms. In addition, appropriate inspection equipment typically includes a requirement for multiple-design-rule reuse.
While it could be argued that particle inspection technology is likely to improve in the future, such technology will remain behind needs as advanced design rules and new process materials are introduced. As design rules shrink, fewer "excursions" are caused by people and process tool contamination, making particles less important than process-induced defects; environmental and equipment-induced particle sources are steadily being reduced and do not scale with design rules, as process-induced defects do.
The choice of an inspection approach can have a major impact on a 300mm fab's ability to ramp quickly. This is because problems being solved during ramp are of a very different nature than those solved during high-volume production. Unpredictable, new and unexpected scenarios dominate these problems. An inspection strategy that is effective during ramp must be able to capture defect types that cannot be predicted a priori.
In contrast, once a 300mm fab approaches the theoretical yield limit of the process, it is less likely that new defect types will occur, making defect inspection much more predictable. This might lead one to ask whether the fab should invest in sensitive inspectors for ramp and development and then switch to particle inspection equipment during production, even though non-particle defects would still be present to some degree.
Two trends indicate that more sensitive tools are needed throughout a fab's life. First, the yield "hurdle" that a process must clear prior to ramping to production is rising with a faster ramp, so there is less time to qualify less-capable inspection equipment to check that it finds defect types identified during development. Second, the time between process changes is decreasing, so a typical fab is always in a development-ramp state. So, fabs require ever-better inspection to detect ever-changing defect types.
This outline of the 300mm-inspection task implies that a robust tool-monitoring defect inspection approach must incorporate learning from higher-sensitivity line-monitor inspection systems. This allows tool monitors to be tuned for new defects introduced by process changes or design rule shrinks. Experience during recent fab startups suggests that without this vital feedback loop, it is impossible to sustain yield learning over time [2]. Thus, a key element of defect control will be eliminated if a 300mm fab sets out to rely on a built-in particle inspection strategy.
Operational aspects of built-in inspection
Built-in defect inspection also has operational issues. Besides the capital cost associated with purchasing a built-in inspector for each process tool, built-in inspection adds to each tool's footprint, reducing a fab's output/unit area.
False defect alarms can be a major consideration. Operators and engineers must respond to each report of an out-of-specification process variation. Sometimes, fundamental technologies used for particle detection are readily confused by process-induced pattern variation on wafers, creating a false alarm. This situation might easily lead to an intolerable distraction on process operations or, worse still, to process operations that ignore alarms or "dumb down" recipes to prevent them, thus completely negating the benefits of built-in inspection.
Reliability is also a concern. With a particle inspector dedicated to a process tool, a failure in the inspector leads to a downed process tool or skipped inspections. Unless MTBFs and MTTRs of built-in particle inspectors can be raised to levels many times better than that which can be achieved today, the increased number of inspection systems in a fab implies that downtime would be a much more common occurrence. A standalone approach, on the other hand, allows lots to be routed around downed inspection systems or process tools.
 Figure 2: Modeling shows financial benefit when using standalone defect inspection with optimized automated material handling compared to built-in particle inspection. |
Studies of defect control in typical 200mm fabs indicate that average time between lot processing and defect inspection can be eight hours or more [3]. Thus, an average defect problem may affect yield across several wafer lots before it is rectified. One of the primary perceived benefits of built-in inspection is that this time is reduced significantly, since wafers can be inspected soon after processing. Built-in inspection is not the only way to reduce time, however. For example, reducing process-to-inspection transport and queue time from eight hours to 30 min via cell integration may gain much of the benefit of built-in inspection while retaining standalone inspection flexibility. Modeling automated 300mm multilevel transport, intrabay robotic shuttles, and optimized tool placement for automated transport access has indicated that this improvement is feasible. For example, PRI Automation and Arizona State University simulated such an automation approach for a proposed 300mm fab and found that average transport time was <4 min [4].
Modeling integrated vs. standalone
We used Sample Planner 3 (SP3) to analyze inspection tool value. SP3 provides a comprehensive model of process control in a fab. It comes from an industry-academia effort that began in 1994. The model is regularly validated and calibrated using data collected from actual fab operations [5, 6].
 Figure 3. An out-of-control event that is not detected by the defect inspector increases yield loss dramatically. |
In our work, we modeled the effects of defect type capture and inspection delay, such as that caused by transit, queuing, and inspection times, on the overall yield loss for various process steps. Figure 2 shows our results for a metal etch where curves show the value of inspection for a 5000 300mm-wafer/week logic fab compared to no inspection at this step. Clearly, value increases as time to detection decreases. However, the increased value is much more pronounced when the processing cell is integrated with comprehensive defect inspection tools, since lower-capability built-in inspectors miss some defects entirely, negating the benefit of decreased inspection delay. In fact, these data show that the cost of missed defects is so substantial that, even with an 8-hr inspection delay, the standalone inspection approach provides a 15% yield benefit over built-in particle detection (i.e., $1.0 million/year vs. $0.85 million/year).
 |
In our analysis, we assumed a 3:1 standalone inspector-to-process tool ratio and a built-in inspector capital cost at 10% that of a comprehensive standalone inspector. We then asked whether it made sense to attempt to build in a comprehensive inspector.
This shifted the top curve downward (thin line in Fig. 2), and so guided our conclusion that reduction of inspection delay via cell integration gave the best economic benefit.
Our analysis presented in Fig. 2 only quantifies the value of prompt detection of "excursions." The benefits of fast ramping due to accelerated yield learning [2] would further tilt results toward comprehensive standalone inspectors.
Financial impact of "excursions"
Consider two production situations: A process where a built-in system finds a defect or process deviation (Fig. 3a) and when it is does not (Fig. 3b). Before the event occurs, production lots are being processed normally. When the event, say an etch chamber problem, occurs, wafers continue through the chamber, creating lots that will have to be partially or completely scrapped. This occurs until the related problem is detected, confirmed, and processing is stopped in the chamber. Then, the tool is serviced, during which time no product is processed. Finally, the chamber is confirmed fixed and processing resumes.
Figure 3 details losses that occur: Cost of yield loss (Ly) = the time to detect the defect (td) x the cost of the yield loss (cy)/ hour. Cost of process tool productivity loss (Ltp) = the amount of hours it takes to repair the chamber and bring it back into production (ttp) x the capital cost of the tool/hour. Typical 200mm fabs experience an average td of 8 hrs and ttp of 16 hrs in cases where the defect is caught.
 |
Often, yield loss caused by an "excursion" far outweighs process tool productivity loss. So, while it may appear that an obvious built-in detection strategy would be to set a capital cost target for a process tool's built-in inspection system (e.g., 20% of the process tool's cost), this is not the correct cost-minimizing approach, which should instead take into account the cost of yield loss for the process step. For example, studies indicate that a typical inspection investment is higher for early than for later metal layers, due to tighter design rules and higher defectivity, even though the exact same process tools are being used [5]. One side-effect of building inspection into a process tool is the absence of this kind of investment flexibility.
Great leverage can be obtained by reducing td via reducing the time it takes to make a tool-up-or-down decision. By removing the possibility of incurring queue time at standalone inspection, time to inspection can be reduced dramatically. This is often the touted attractiveness of built-in inspection. Quick time to inspection does not equate to quick td, however, since a missed "excursion" is usually detected days later, when the lot is inspected by a more sensitive line-monitor inspection, by the next process step's inspection, or by back-end final electrical tests.
A central part of our analysis was an estimate of pd for various process steps. We used historical root cause and benchmarking data combined with a survey of process experts to create an exhaustive list of defect types and frequencies typically encountered at various process steps. We then assessed the capability of various inspection approaches against these defect types (Table 2 is a subset of our assessment). We found that a limited-capability built-in inspector missed a significant number of defects at state-of-the-art process rules, reducing pd considerably when compared to today's standalone tool monitor inspectors.
We then loaded the resulting probabilities of detection (pd) and inspector throughputs into the SP3 model; this allowed us to model more accurately interactions between probabilities of detection, excursion frequency, sampling strategy, integration approach, and real-world issues such as transit and queue times.
We modeled a range of inspection times, from the historical average of 8 hrs to 15 min for built-in particle detection.
Conclusion
A built-in particle detection strategy is not likely to provide benefit over a comprehensive standalone inspection approach that is facilitated by 300mm fab automation and robotics. We have reached this conclusion through the following observations:
- The defect detection capability of a robust, comprehensive inspection strategy appears to outweigh the time-to-inspect benefits of built-in inspection. Quick inspection does not lead to quick event detection if the inspector cannot see the defects that cause the event.
- Significant reduction in time-to-inspection may be achieved through optimized automated robotic wafer and lot handling and fab layout without the loss of flexibility and added capital cost associated with built-in inspection. Built-in inspection, per se, has little incremental benefit once automation and fab layout are optimized.
- The choice of inspection strategy must include provisions for future trends, such as new metal and dielectric defect types, and the growing importance of staying within process margins and design and process interactions as a source of defects.
When all of these considerations were taken into account, our analysis showed that building particle inspection into a process tool is not necessarily the most cost-effective strategy for 300mm fabs. These observations suggest that built-in particle inspection may not be the future trend that conventional wisdom might suggest.
Acknowledgments
We thank Wayne McMillan, Anantha Sethuraman, Paul Marella, and Sanjay Tandon of KLA-Tencor for their work in this study.
References
- C.H. Stapper, "Fact and Fiction in Yield Modeling," Microelectronics Journal, Vol. 20, No. 1-2, 1989, pp. 129-151.
- N-S Tsai, "TSMC's Transition to 300mm," an interview in Yield Management Solutions, KLA-Tencor, Spring 2001.
- T. Esposito, et al., "Automatic Defect Classification: A Productivity Improvement Tool," Conference proceedings IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop, Boston, Sept. 1997, pp. 269-276.
- I. Paprotny, R. Gaskins, J. Yngve, "Applying Conservative Distributed Simulation to a Large Scale Automated Material Handling Design," Society for Computer Simulation Intl Western Multiconf., San Diego, Jan. 23-27, 2000.
- R. Williams, et al., "Challenging the Paradigm of Monitor Reduction to Achieve Lower Product Costs," 10th Annual IEEE/Semi Advanced Semiconductor Manufacturing Conference and Workshop, Santa Clara, CA, Sept. 8-10, 1999.
- R. Williams, et al., "Optimized Sample Planning for Wafer Defect Inspection," IEEE Intl Symp. on Semiconductor Mftg, Santa Clara, Oct. 11-13, 1999.
Ralph Spicer received his BS in electrical engineering from MIT and his MBA from Stanford University Graduate School of Business. He is a product marketing director in the Wafer Inspection Group at KLA-TencorCorp., 1 Technology Dr., Milpitas, CA; ph 408/875-5359, fax 408/571-2915, [email protected].
Dadi Gudmundsson received his BS in industrial engineering from University of Arkansas and MS in industrial engineering and operations research from University of California at Berkeley, where he is pursuing his PhD. He is a research scientist for KLA-Tencor.
Raman Nurani received his BS in mechanical engineering from Indian Institute of Technology, his MBA from Loyola University, and his MS and PhD in manufacturing management from Carnegie Mellon. He is manager of the Statistical Methods Group at KLA-Tencor.