Yield Management Relies More on Integrated Data Handling

		At a Glance
	Fabs must collect and analyze data in a very efficient and integrated manner to not only keep processes and yields under control, but also avoid being smothered by the mountains of data. The data analysis requirements for yield management is increasing exponentially. Products are available to help, but many believe that much more needs to be done.

The complexity of the yield management task will continue to grow as the industry moves on to larger wafers and smaller designs. Design rule scale-down decreases the minimum size of a potential killer defect, increasing the number of potential killers as well as the effort required to find them. With more inspection points, each generating more data, the volume of data to be examined stands to increase exponentially, even with optimized inspection and analysis plans in place.

Overall factory effectiveness is expected to continue increasing as well, calling for faster reactions to out-of-control conditions. The growing amounts of inspection data and faster analysis requirements place a larger demand on the processing power, in terms of algorithms and pure computing speed.

Data handling needs for yield management have grown from simply putting some data in a spreadsheet to running a fully automated, up-to-the-minute data collection and processing system (lead photo). Computer technology and network implementations have matured to the point that fabwide data collection and processing can be done in real time.

Gary Green of Defect & Yield Management Inc. (Bedford, Mass.) said, "In early yield analysis software implementations, most of the emphasis was in connection to the data sources and in allowing users to interrogate datasets using graphical user interfaces. This was basically in keeping with the development concept of 'learning to walk before learning to run.' That is, emphasis was placed on functionality over automation, while still yielding large time savings for engineers in their daily yield analysis routines. The products were more engineering tools than 'production systems,' so that if the software failed, production could still proceed using classic techniques."

An increasingly important part of yield management is data collection and analysis. (Source: Promis Systems)

"However, as the viability of the new data management technology proved itself in daily operation, customers pushed for more real-time functionality, such as SPC with pager and e-mail alarms, automatic uploads of defect data summaries to WIP systems, and yield prediction capabilities. The software was starting to evolve into a more useful production system."

Current usage

Right now, finding the source of a process problem requires the use of a dedicated system connected to, but not necessarily part of, the fab's manufacturing execution system (MES). Nick Ward of Consilium (Mountain View, Calif.) said, "I would say without exception that MES packages, ours included, make no attempt to provide the tools to enable you to do that kind of detailed analysis, perhaps correlation type work. That is not a function of a manufacturing execution system. In fact, we deliberately attempt to offload that to external systems through either reporting or extracts. If you're trying to do some sort of post-engineering work, some kind of commonality study which could involve a very, very great number of views onto the data correlation exercises, that kind of thing can kill a database very, very easily."

The amount of data to be analyzed for yield management purposes is large enough to require a separate system. Ward added, "Detailed analysis needs to be clearly separated from the MES, so that an engineer can release an unbounded commonality search without impacting manufacturing. To that end, you need to be able to have some kind of sandbox area, if you like, that an engineer can do that kind of unbounded analysis work."

Many companies do not wait for a yield problem to occur to perform yield analysis calculations. They have the calculations performed constantly so that they can quickly examine the results. Ray Scobie of Mitel (Kanata, Ontario, Canada) commented on the creation of the company's yield analysis database: "We weren't cheap with the horsepower, and we weren't cheap with memory. If we needed to add 10 tables, for example, in order to have all the standard deviations precalculated at a part level or process level, we would do that so that the user wouldn't have to wait for the computer to calculate it at withdraw time. A lot of the calculations are already resident in the database. That was one of the things that we held as a very high priority."

One consistently expressed need is for a central repository for the data, or at least a unified interface. According to Fourmun Lee of Motorola (Chandler, Ariz.), "Where I see the systems failing right now is that they don't do a good enough job of tying together all the different islands of data. There has been some very good progress in that area, but it's a moving target."

Lee also identified some key considerations for a yield analysis system. "One consideration is being able to access all the different kinds of data from one interface. That's very key. Another item that's very key is being able to understand the relationships between the different kinds of data. The unit probe data tells you one thing, your parametric data might tell you something else and the in-line defect data or the in-line process monitoring data may contain some other information. We need a system to effectively correlate, or relate, the different information that's available in the different types of data." Lee recently presented these ideas in a paper on integrated yield analysis.¹

Available aids

A number of packages have been introduced or updated to tie data sources together, perform analysis and present the results in a useful manner (Fig. 1). Some have added data warehousing functionality to their offerings to help handle the data analysis load. Many of them have licensed a spatial signature analysis (SSA) program that was developed for SEMATECH by Oak Ridge National Laboratories (ORNL, Oak Ridge, Tenn.).

Fig. 1. Many commercially available packages provide integrated data analysis and visualization. (Source: Knights Technology)

The SSA program separates defect clusters from random defects and compares the cluster patterns to signatures stored in a library. For example, once a spin-on glass (SOG) streak pattern has been identified and put into the library, the program can alert process engineers to the possibility of a streak problem when that pattern is found in the in-line optical inspection data. Corrective action can be initiated minutes after the streaking problem surfaces. The main usefulness of the program is in isolating, classifying and sourcing problems that cause clusters and systematic distributions of defects on the wafer. Random defects still account for the majority of yield difficulties.² Programs that use SSA provide it as an attractive addition to their capabilities.

YieldManager from Knights Technology (Sunnyvale, Calif.), an Electroglas company, is designed to collect data from a wide variety of sources, including defect review, binsort, lot history, bitmaps, parametrics, metrology and spatial signatures. It also provides connectivity to Windows-based programs and VARS. Knights recently added its Q-Yield database analysis product, originally developed by Quadrillion Corp. (Dunrobin, Ontario, Canada), to its yield management products.

ADE Yield Management Solutions (formerly LPA Software, South Burlington, Vt.) offers DefectAnalyzer as a part of its yield management system. It performs spatial analysis, two- and three-dimensional mapping and generates wafer map galleries. It also does overlay analysis of lost, common and adder defects.

The Quest defect data management system from KLA-Tencor (San Jose, Calif.) is designed to work with a wide variety of defect detection systems, review stations and SEMs. It performs spatial analysis and overlays of defect data and bitmap failures; identifies repeating defects; provides statistical process control (SPC) with alarms; correlates defects and binsort yield; and tracks defects by layer, class and size.

Defect & Yield Management has developed a data analysis system to connect to Quest and other data sources. Called the NeuralNet Engineering Data Analysis, or NEDA, system, it is designed to connect work-in-progress (WIP), bin, bit, metrology, parametric and defect data sources in an interactive manner.

SAS Institute (Cary, N.C.) provides data warehousing and decison support tools, as well as database and statistical processing capabilities for a number of fabs. SAS also provides a toolkit to help integrate databases kept at different locations.

The DMS-II system from Inspex (Billerica, Mass.) collects data in real time from inspection stations and integrates them with data from review stations, probe-and-test systems and programs to predict yields. It uses a client-server relational database to correlate defects to yields in real time.

Parallel Visual Explorer (PVE) from IBM Semiconductor Manufacturing Solutions (Middleburg, Va.) is a data mining tool that manipulates quality and yield data to enable visual detection and correction of process problems. Developed at IBM's Yorktown, N.Y., research center, it is used by the company's Microelectronics Division and other manufacturers to correct process deviations not identified by classic statistical process control (SPC) techniques.

The HyperView system from Heuristic Physics Laboratories (HPL, Milpitas, Calif.) is an integrated yield analysis and prediction system designed to access, compress, visualize and analyze data from in-line monitors and test equipment. It is a distributed system that performs analysis in the background and generates charts, reports and alarms for engineering use.

Increasing complexity

According to a projection in the 1997 National Technology Roadmap for Semiconductors,³ the number of transistors per microprocessor is expected to double every two years. Isolating the source of a process problem is a task thats complexity is also expected to double every two years (Table 1). The amount of data that fabs must deal with to keep processes under control is already very large, and it is expected to increase exponentially. On this problem, Mitel's Scobie commented, "I think that that plight can be a little bit unbelievable."

Table 1. Device and Process Complexity³

	1997 250 nm	1999 180 nm	2001 150 nm	2003 130 nm	2006 100 nm	2009 70 nm	2012 50 nm
Transistors per microprocessor (x 106)	11	21	40	76	200	520	1400
Process steps	350	380	420	450	500	550	600
Fault isolation complexity factor (x 109)	3.8	8	17	34	100	290	640
Defect sourcing complexity trend	13	2.13	4.33	8.93	263	743	1703

Shrinking the design rules will itself increase the number of potential killer defects, because particles that used to be too small to destroy a device will no longer be harmless. Also, the process steps become more sensitive to a larger set of parameters, making not only processing itself more difficult, but also tracking a problem to its source.⁴ By the year 2012, it is estimated that problems found in 50 nm processes will be 170 times more complex to track to their source than those in 0.25 µm (250 nm) processes.

Future analysis needs

In order to maintain expected levels of productivity and yield, process problems must be tracked to their source in a timely manner to recover from an excursion. Trends must also be recognized in a timely manner to maintain or improve baseline yields. Part of the roadmap coverage of yield analysis was to indicate the time frame in which trends and problem sources must be identified, as well as the kinds of technology required to meet those time frames (Table 2).

Table 2. Data Analysis Needs for Sourcing Problems³

	1997 250 nm	1999 180 nm	2001 150 nm	2003 130 nm	2006 100 nm	2009 70 nm	2012 50 nm
Time required to source problems	days	days	days	hours	hours	hours	hours
Time required to recognize trends	weeks	days	days	hours	hours	hours	hours
Information sources for automatic data analysis	spatial analysis	time analysis	time analysis	merge	improve	improve	improve
Standardization of defect data output formats	extend	extend	extend	new	new	new	new
Standard architecture for data transmission/storage	proprietary architecture	develop standards	adopt	apply	apply	apply	apply
Fuse/integrate process and defect data from different tools	single inline tools	in/off-line tools	in/off-line tools	extend	extend	extend	extend
Feedback for automatic process control	manual	open loop	open loop	mixture	closed loop	closed loop	closed loop
Solutions exist Solutions being pursued No known solution

Currently, some process problems can be sourced within minutes, but on the whole, it can take days to trace them to their source. According to the 1997 Roadmap, by the time process technology migrates from 150 nm to 130 nm, problems will have to be sourced within the space of a day. Recognizing trends is an activity that can be accomplished now within the space of a month. That time will have to be reduced to days for 180 nm processes, and again down to hours for 130 nm processes and below.

The information sources for automatic data analysis referred to in Table 2 are the data products that programs have to be able to produce. Ken Tobin of ORNL said, "Right now, we're doing a lot of things with spatial analysis. Let's look at the wafer product, and let's look at the way things are organized on the wafer product and try and tie those things back to manufacturing tools. Well, what about all these in situ sensors? What about even looking at the spatial data over time, and do time analysis? We want to develop that over time. We want to merge that together." Currently, spatial analysis is being automated using ORNL's SSA program. In the next two technology generations, time analysis will have to be automated as well, and merged with existing automatic analyses by 2003.

Since the data used in yield analysis will come from an even wider variety of sources, defect data output format standards will have to be extended.

Currently, many manufacturers of inspection, data storage and analysis equipment use their own proprietary architectures to transmit and store the data. Many of the yield analysis programs advertise that they can handle the various formats. The size of the data handling job is expected to become large enough in the 180 nm technology generation that uniform standards will have to be developed.

Using the analysis information to control processing will also have to become more automated. Right now, once a problem with a tool has been identified, it is still essentially a manual process to notify the MES to remove the tool from the flow, pull the tool off line and perform maintenance. It is possible that real time, in situ monitoring and process adjustment will have to occur within the tool itself by the 100 nm technology generation.

Other considerations

A company that maintains multiple fabs, or builds additional fabs, needs to use knowledge gained in one facility in all of its facilities. Some vendors are working on web implementations and MES level solutions to making information available to all their facilities.

One rift in overall yield analysis that is in the process of closing is that between wafer processing and packaging facilities. In most cases, these two functions are performed at different locations and often in different countries. If a large drop in yield occurs between finished wafers and packaged products, it is difficult to determine if the source of the problem is at the fab or the packaging foundry. Merging or integrating the databases between them is still either an in-house or a custom job. The chip traceability standardization work at SEMI is expected to solve some of these problems.

Conclusion

The job of managing yields in fabs is very involved, to the point where the current level of automation is barely enough to keep yields under control. A number of products are on the market that help tie all the data sources together and perform the necessary analysis. Yield management will become much more complex as process technology progresses. The sophistication and sheer power of yield management systems must increase rapidly to enable an acceptable yield level.

References

1. Fourmun Lee, "Advanced Yield Enhancement: Integrated Yield Analysis," Advanced Semiconductor Manufacturing Conference and Workshop, pp. 67-75.

2. R.K. Nurani, R. Akella, and A.J. Strojwas, "In-Line Defect Sampling Methodology in Yield Management: An Integrated Framework," IEEE Transactions on Semiconductor Manufacturing, vol. 9, no. 4, November 1996, pp. 506-517.

3. 1997 National Technology Roadmap for Semiconductors

4. C.J. McDonald, "Copy EXACTLY! A Paradigm Shift in Technology Transfer Method," Advanced Semiconductor Manufacturing Conference and Workshop, September 1997, pp. 414-417.