Optimizing cluster tool throughput
07/01/1997
Optimizing cluster tool throughput
R.A. Hendrickson, Brooks Automation Inc., Chelmsford, Massachusetts
Case studies have analyzed factors such as wafer-handling options, cluster tool configuration, front-end tool performance, process times, and other activities and evaluated their effects on throughput. These studies have motivated changes in original tool configuration designs, resulting in better throughput performance at less cost than originally anticipated.
Factors that affect throughput
Factors affecting throughput fall into three primary areas:
factory interface (FI) cycle,
process cycle, and
pipeline crossover cycle.
Analyzing the interplay among these factors can help identify bottlenecks and design tools for optimal throughput.
The FI cycle is the time required to present a new cassette of wafers to the process cycle. It includes venting the load lock to atmosphere, exchanging the cassette, and pumping the load lock back to vacuum. In some instances it may include a purge, heating, or cooling cycle within the load lock.
In cluster tools with one load lock, or dual load locks that feed different process paths (dual-independent load locks), the FI cycle time adds directly to the time required to process a cassette of wafers. In tools with dual alternating load locks or dedicated input and output load locks, some or all of the FI cycle occurs in the background.
The length of the FI cycle itself depends on pump-down time, vent time, in-lock process time (purge, heat, cool), exchange time (either cassette exchange or wafer-by-wafer exchange), and load lock batch size (# of slots in cassette, typically 13, 25, or 26).
The process cycle is the time to complete the processing of one cassette load, or load lock "batch." In steady-state operation, cassettes are provided continuously as needed. The process cycle clock starts when the last wafer of cassette n is returned to the load lock and stops when the last wafer of cassette n+1 is returned to the load lock.
Process cycle times may be limited by either robot operations or process operations. When the robot must sit inactive and wait for the process to finish, the tool is process limited. Process-limited tools can improve cycle time by:
shortening the effective process time by adding another process module of the same type,
reducing the actual process time in the particular module with the longest process time, and
decreasing the time or frequency of process module clean cycles.
If a completed process is waiting for the robot to become available, then the tool is robot limited. Robot-limited tools can improve cycle time by:
increasing robot speed,
using a different robot geometry, and
reducing the number of robot operations.
The optimal operation point is near the crossover between robot-limited and process-limited activities. At this point, both the robot and process modules are as close to 100% utilization as possible. When the process cycle is operating near the crossover point, attacking any of the above factors can significantly improve throughput.
The pipeline crossover cycle is a key factor in tools with dual alternating load locks. It is defined by the number of wafers that a cluster tool can hold. The terms "pipeline fill" and "pipeline drain" are often used for situations where wafers are moving into the tool but no completed wafers are yet leaving the tool; or when wafers are still exiting the tool, but no new wafers are available to enter the tool. The pipeline crossover cycle occurs when wafers from one cassette are still draining out of the tool, and wafers from the subsequent cassette are feeding in. Both load locks must be open to the process cycle-one feeding wafers in, the other receiving finished wafers. During this time, neither load lock is available for a factory interface cycle. Factors affecting the pipeline crossover cycle include the number of process modules being used, as well as "batch" sizes for multiwafer process modules.
Secondary factors that may affect throughput include the efficiency of the scheduling algorithms, software communication overhead, and other mechanism actuation times (e.g., slot valve actuation and index times).
If the FI cycle does not limit the tool, then throughput will be proportional to the process cycle. The time available for the FI cycle, on the other hand, will depend on the potential load lock off-line times between pipeline cycles. For example, in a cluster tool with nine process modules, 25-wafer cassettes, and a basic process cycle that can provide one wafer every 60 sec, the pipeline crossover will take 9 min. Thus, each load lock will be offline and available for FI for 16 min. If venting the load lock, exchanging the cassette, and pumping the load lock back to vacuum require more than 16 min, the FI cycle will become the throughput limiting factor.
 |
Figure 1. Effects of robot design on throughput performance of an aluminum reflow cluster tool; a) cluster tool configuration and process flow, b) throughput predictor results. This case study demonstrates the effects of robot design selection. The dual same-side pan robot design provided the highest throughput rates over a wide range of aluminum deposition operating points. As aluminum deposition times increase, throughput is limited by process time rather than robot design.
Measuring effects on throughput
Common methods for calculating cluster tool throughput have included complex spreadsheets with parallel columns. They track activities such as processes, robot moves, valve moves, load lock pumping and venting, etc. This method is time-consuming and becomes increasingly complex as the number of process modules and types increases, and wafer flow paths become more complicated.
An alternate method for the rapid analysis of cluster tool throughput bottleneck factors is now available. A newly developed Throughput Predictor Model simulates cluster tool operations and analyzes a wide range of variables affecting cluster tool throughput. With this new software model, throughput analysis that might take days or weeks to perform with spreadsheet methods can be done within hours.
Designing the Throughput Predictor Model
The Throughput Predictor Model simulates cluster tool configurations ranging from a one-process module tool with a single-wafer input load lock, to a dual load lock 8-sided platform with serial/parallel flow, such as an aluminum reflow tool. ProModel factory simulation software was the basis for a family of animated cluster tools representing 4-, 5-, 6-, 7-, and 8-sided configurations with one or two vacuum load locks and a central wafer transport robot. The model includes a large number of variables such as robot arm motions, process module types, load lock strategies, batch sizes and wafer locations, indexing, and other activities (see table).
 |
The predictor software can simulate both single-wafer and batch-process module types. Multiwafer batch process modules incorporate additional input variables such as batch size, number of wafer locations, and index times. Single-wafer modules that require cleaning also consider frequency of clean, length of clean, cover wafer source (if required), and pre-align source and time (if required).
The model also uses robot scheduling algorithms based on the transport module control software used at Brooks. Case studies commonly use a range of robot performances as a baseline for comparing changes in throughput performance due to robot speed with changes due to whichever variable is being analyzed.
Simulating different case study scenarios and altering the variables provided by the model software allows analysis of the various factors contributing to cluster tool throughput. The results can then be used during design of a cluster tool to determine the architecture that will provide maximum throughput and minimize cost.
 |
Figure 2. Comparison of single pan vs. dual pan robot arms on two-module and three-module etch tools; a) etch cluster tool configuration and process flo,. b) throughput predictor results. At an operating point of 40 sec etch process time, large throughput gains can be achieved by using a dual pan vs. single pan robot arm in a two-module configuration. Adding a third process module at this point would not increase throughput enough to justify its cost.
 |
Figure 3. Effect of adding a preheat module to a metal etch and asher cluster tool; a) cluster tool configuration and process flow. b) Without a preheat module and at 180 sec metal etch process time, throughput was limited to 55 wph. c) Adding a preheat module reduced the operating point to 150 sec metal etch process time, and increased throughput from 55 to 62 wph.
Throughput comparison studies
While the most obvious factors affecting cluster tool throughput are process time and robot type, a faster robot will not always increase throughput, nor will adding an additional process module double the throughput. Predictive simulation analysis gives a better understanding of the relationships between robot design, process times, and tool activity variables. These parameters affect not only throughput, but also cost/performance ratios. On more than one occasion, intuitively
observed "bottlenecks" have been corrected in surprising counter-intuitive ways.
 |
Figure 4. Effects of clean times on throughput of a CVD cluster tool; a) tool configuration and process flow. b) Adjusting cool times had little effect on throughput. Significant gains can be made by adjusting clean times.
Robot arm geometries vs. cluster tool architectures. The most cost-effective and efficient robot arm design depends on the particular cluster tool architecture being used. Figure 1 compares throughput for three different robot arm geometries on an aluminum reflow cluster tool. In this type of multistep, serial process module configuration, the performance difference between a single pan and dual opposing pan robot designs was minimal. The dual same-side pan robot achieved a 25% throughput gain compared to the dual opposing pan robot-making the dual same-side pan robot the design of choice for this application.
In Fig. 2, a multiple, parallel etch module configuration, the most significant throughput gain (31%) was realized when changing from a single pan to dual opposing pan robot design. The dual same-side pan delivered an 8% throughput improvement compared to the dual opposing pan design. The dual opposing pan robot provided the most beneficial cost:performance ratio; the modest performance enhancement provided by the dual same-side pan robot may not justify its higher associated costs.
Single vs. dual pan robot arm for a three-process module etch tool. In this case, the user hoped to lower tool costs by choosing a single pan robot arm instead of a dual pan arm for a predetermined three-process module etch cluster tool configuration. The predictive analysis software evaluated not only the differences in throughput between single pan and dual pan robots, but also the amount of throughput gain achieved by having three process modules instead of two.
In fact, using a dual opposing pan robot arm on a two-process module configuration achieved a more significant gain in throughput at a lesser cost than adding a third process module to a single pan robot configuration (Fig. 2). At a process time of 40 sec, adding a third process module in addition to a dual opposing pan robot had no effect whatsoever on throughput. As a result, the user chose a dual opposing pan robot with two process modules. This configuration was significantly less expensive than the original three-module design being considered.
Robot speed vs. process time. Another case study analysis aimed to increase throughput on a metal etch tool that required long process times. The cycle time of a tool can be estimated by adding the process time and the time needed to exchange the processed wafer for a new one. Thus, intuition suggested that increasing robot speed would help to improve overall tool throughput. In this type of situation, however, the robot is often sitting idle while waiting for the process to finish. When the wafer swap time is small in comparison to the process time, minimal throughput enhancement is gained by cutting the transfer time in half.
As an alternative, designers suggested that reducing the process time would have a potentially greater impact. We evaluated the addition of a preheat module to shorten the wafer`s time within the etch module. The preheat module improved throughput significantly, whereas increasing robot arm speed only had a minimal effect (Fig. 3).
Clean cycle vs. cool time for a CVD tool. A typical CVD tool that required a periodic clean cycle rarely operated in a steady state because the sequential shutdown of process modules for cleaning brought the entire tool`s production to a halt. As the clean cycles ended, the first few wafers began piling up for the cool cycle. This behavior suggested that the cooling cycle was the bottleneck and cool times needed to be reduced.
However, simulation analysis demonstrated that changing the cool times would actually have minimal effect. Another study evaluated an alternative approach-altering the clean cycle frequency or length. Concentrating engineering time on the clean cycle rather than the cool cycle results in significant throughput gain (Fig. 4).
Conclusion
Predictive throughput analysis is a cost-effective tool for experimenting with new cluster tool designs. It can be used in the ongoing study of the effects of factors such as process time, batch size, cleaning frequency, robot type, load lock strategies, and more. This critical information can guide investments in new tool design and focus engineering efforts to optimize throughput gains-a large contributing factor in the tool`s overall productivity index.
RUTH ANN HENDRICKSON received her BA degree in physics from Wellesley College and pursued post-graduate work in electrical engineering at Northeastern University. She joined Brooks Automation Inc. in 1989 and has guided its throughput analysis project for the past five years. Hendrickson is currently product manager for etch handling systems applications. Brooks Automation Inc., 15 Elizabeth Drive, Chelmsford, MA 01824; ph 508/262-2400, fax 508/262-2500.