In situ endpoint control saves chemicals in wet processing
01/01/2001
David Taraci, Jerry Weston, SEZ America, Phoenix, Arizona
Fred Wertsching, Luxtron Corporation, Phoenix, Arizona
overview
Investigation of endpoint detection, using cost of ownership analysis, on a pre-lithography backside film-removal step helped to determine how the capital investment of an endpoint detection system affects the cost/good wafer equivalent. Results showed a 12.5% increase in throughput and 14% savings in process chemicals and consumables. These factors led to an overall reduction in the cost/good wafer equivalent of 10.4% — from $3.46/wafer to $3.10/wafer. The value of adding endpoint control is realized further when increased utilization requires additional tools.
 |
Hard-mask removal on SEZ's Spin-Processor using the Luxtron end point detection system.
In situ process control, such as endpoint detection, is a key to success in maximizing yields on 180nm technology and beyond. However, this increase in control typically requires an increase in capital investment, thus affecting per-wafer-manufacturing costs. For this reason, implementation of in situ process monitoring for many process steps that do not require such strict control is often overlooked. It is possible to benefit from in situ process control while still reducing the overall cost/wafer. In particular, we have used cost of ownership (COO) analysis to look at the benefits of endpoint detection on a pre-photolithography backside film removal step and its effect — measured as the cost/good wafer equivalent (GWE) — on capital investment, chemical consumption, throughput, and other factors.
Doing the tests
In our tests, we gathered process data from a SEZ Spin-Processor 203 for film removal equipped with a Luxtron 1015 [1] for in situ endpoint detection. We made 49-point measurements on 50 monitor wafers with 5000Å of LPCVD Si3N4, using a Rudolph FEIII ellipsometer, to determine wafer-to-wafer (WTW) film thickness variability within a lot.
We used 5000Å LPCVD nitride monitors to gather our experimental data while minimizing handling times, thus producing more accurate etch rate data for COO modeling input. Later, we used a more typical thinner film thickness for actual COO analysis.
 |
We processed the first lot of nitride monitors using 49% HF at 55°C with active endpoint detection to minimize overetch. The endpoint detection wavelength showed reproducible endpoint capability with a characteristic diminishing oscillation that stopped upon breakthrough to bare silicon (Fig. 1a). Post-etch measurements were made on the ellipsometer to ensure complete film removal. We used removal times for the endpoint lot to calculate a mean etch rate. Our data showed that the etch rate diminished linearly over time as the etchant was consumed.
We sent a second lot of monitors through the same process as the first lot without using endpoint detection control, although the endpoint detection system did collect data passively (Fig. 1b.) The amount of overetch this lot received was based on total thickness range across both lots to ensure complete film removal from every wafer; this is typical for critical pre-photolithography and pre-rapid thermal processing steps.
Unlike the first, we processed the second lot using a time correction feature of the spin-processing tool; this allows the diminishing etch rate of the chemistry to be corrected automatically by adding predetermined amounts of time. The time correction table for the second lot was derived from the linear trend line determined from the data of the first.
Figure 1 (p. 96) shows an example of scans obtained from each of the processes described above. It is clear that the overetched lot (using passive data collection) received a notably longer process time, well beyond endpoint.
Because one of our metrics for this process evaluation was to be chemical consumption, we monitored the volume of the etchant in the tank of the spin processor during the etch process. We were now equipped with run time data from the tool's log files, etch rate data, and process chemistry consumption.
Cost analysis
We loaded our collected data into Wright, Williams, and Kelly's Two.Cool 2.3 Cost of Ownership modeling software. Developing a workable COO model requires the application of pre-knowledge regarding typical process flow. Thus, we used a spin process that removes 2000Å of LPCVD Si3N4 from the backside of 200mm wafers. We also used SEZ standard reliability values (i.e., MTBF, MTTR, MTBA, etc.) as published in the spin processor tool specification.
The endpoint system reduces excessive overetch in a tool on a per-wafer basis, thus increasing the number of wafers that can be processed through a given tool/week. Using COO modeling — which gives wafer volumes in wafer starts/week — with a fixed utilization for both tools, wafer starts/week increased 12.5% from 2514 without endpoint to 2829 with endpoint. In our analysis, this increase in volume affected all of the cost drivers/GWE in the COO (see table).
 Figure 2. Cost/GWE at various wafer starts/week with an in situ endpoint detector and a standard overetch process. |
Clearly, a significant benefit from this COO analysis is the reduction of materials and consumables, in this case, process chemistry. In addition, a process that overetches for extended lengths of time unnecessarily adds exposure time of the chemistry to the exhaust, where a good portion is lost. With endpoint detection, there is a 28% increase in the number of wafers/liter and a saving of 14% in the amount of chemical used per day. With the standard overetch, 42 liters/day are being consumed at a rate of 9.37wafers/liter, compared to the use of the endpoint detection system, where 36 liters/day are being consumed at a rate of 12.63 wafers/liter.
All together, the cost drivers/GWE lead to an overall reduction in cost/GWE of 10.4% — from $3.46 to $3.10/wafer; 33% of this savings comes from reduced consumable materials (e.g., process chemicals).
Tool utilization
Utilization is a tremendous factor in the overall COO for capital equipment. For the past 18 months, chipmakers delayed new wafer fabs and expansion of existing facilities while they waited for clear signs of stronger growth in IC markets. Now, though, capacity is getting a bit strained, with fab capacity utilization running above 90% for the IC industry [2]. Capital equipment in a fabrication facility that is not being fully utilized is not paying for itself. However, constraints such as scheduled and unscheduled maintenance downtime, engineering time, etc., prevent any piece of equipment from reaching 100% utilization.
 |
When production calls for an increase in wafer starts/month, it is often necessary to add capital equipment, which in turn reduces total utilization of all like tools. For example, if production demand on a single tool already has that tool running at 80% utilization (with a realized maximum of 95% utilization) and an increase in demand dictates a utilization increase that pushes total utilization over 95%, a second tool is required. The result is that the two tools then run at very low capacity, sending cost drivers/GWE to excessive levels.
As a viable alternative, our work has shown that the addition of in situ endpoint detection — for a capital investment much less than adding a new tool — can reduce cycle time by preventing unnecessary overetch and directly producing an increase in wafer throughput. This maximizes the use of a single tool and can lead to a substantial reduction in the overall cost/wafer (Fig. 2).
The sharp increases in cost/wafer in Fig. 2 indicate points at which an additional system must be added to meet required throughput levels. When comparing the overetching process to an endpoint process, the staggering effect becomes apparent. Should required throughput fall into these regimes, the addition of an endpoint system with minimal investment could eliminate the need for another tool, greatly reducing the cost/wafer.
Factoring in reworks
Our COO anlysis did not take into account any reworks for the standard overetch. The endpoint detection system will adjust for any film thickness that may go beyond the standard deviation that was calculated and adjust for drop in chemical effectiveness over time. The standard overetch formula and the time correction table, however, do not always allow for wafers with thicknesses outside the established parameters. This would create the need for some rework. Using an estimated 5% rework rate leads to a drop in GWE/week in the COO analysis (see Fig. 3).
The main reason for using in situ endpoint detection in wafer processing is for process control, which decreases chemical consumption and improves cycle time. The dispense time is shortened and less chemical is consumed/wafer. Cycle time improves because the endpoint detection system stops the process as soon as the film has been cleared off the wafer. This shortened cycle time increases throughput and decreases chemical consumption.
 Figure 3. Cost/GWE at various throughputs/week with an in situ endpoint detector, standard overetch process, and standard overetch process with 5% rework rate. |
This increase in throughput lowers COO on a per-wafer basis. This can be somewhat deceiving when looking at cost drivers on COO management. For example, depreciation of capital equipment is lower per wafer for the endpoint than for the standard overetch, when, in fact, the total cost of the capital equipment for the endpoint detection system is more than the standard overetch. This is explained with the increase in throughput. The endpoint detection system allows for more wafer starts/week, therefore, the depreciation is allocated over more wafers/year than the standard overetch, which lowers the cost/wafer.
Conclusion
Using in situ endpoint detection with spin processing provides tighter process control. In addition, such an equipment setup provides a 12.5% increase in throughput and, significantly, a chemical saving that allows for a 28% increase in the number of wafers/liter and a saving of 14% in the amount of chemical used per day. Such fundamental and incremental improvements, while often difficult, can be extremely beneficial: an increase in production as small as 1% can potentially result in increased sales of $200,000-$300,000/month for a semiconductor manufacturer [3]. Depending on a company's process, the actual numbers will vary, but clearly endpoint detection systems possess benefits other than just producing tighter yields. n
References
- The Model 1015 has been replaced with the Optima 9225, but the two are functionally identical.
- J.R. Lineback, "Chip production gear: After three bad years, how big will the turnaround be?" Semiconductor Business News, Jan. 1, 2000.
David B. Taraci received his BS in physical science from Northern Arizona University. He is field applications group leader for SEZ America, 4824 S. 40th St., Phoenix, AZ 85040; ph 602/437-5050, fax 602/437-4949, [email protected].
Jerry T. Weston received his BS in marketing from Arizona State University. Weston is a member of the inside sales support team at SEZ America and determines cost of ownerships.
Fred Wertsching received his MS in solid state electrical engineering and his BS in biomechanical engineering, both from Arizona State University. He is a senior applications engineer at Luxtron Corp., 3201 N. 16th St., Suite 7, Phoenix, AZ 85016; ph 602/604-0430, fax 602/604-0922, [email protected].