Simulation Software Finds Hidden Capacity

N/Ast fabs lose 5-10% of their potential throughput to scheduling delays, bottlenecks, and other workflow problems, claims Michael Thompson, president of <%=company%> (Bountiful, UT). As the cost of new construction climbs, more and more companies are trying to exploit this hidden capacity.

Improved scheduling, the obvious answer, is easier said than done. With hundreds of manufacturing steps and re-entrant process flows, semiconductor manufacturing defies attempts to find a single optimal schedule. Additional complications from unexpected downtime and changes in product priorities create chaotic behavior that is inherently unpredictable [1]. Most scheduling methods, whether automated or manual, have abandoned the search for an optimal scheduling algorithm in favor of empirical rules based on the Theory of Constraints. This approach, first proposed by Eli Goldratt in 1984 [2], concentrates on minimizing unproductive equipment time, within the constraints of overall manufacturing goals. For example, an environment requiring minimum cycle time, such as an ASIC fab, might schedule lengthy furnace anneals differently from one requiring minimum cost of ownership, such as a memory fab. N/Ast IC companies have developed empirical rules of thumb for things like WIP inventory at a bottleneck machine. These rules of thumb are often specific to a particular work cell, however, and may not serve the needs of the overall fab. AutoSimulations' software couples tools for simulation and analysis of proposed scheduling rules with software to implement these rules consistently throughout the fab.

For example, Line 5 at Samsung's Kiheung (South Korea) facility is a mature memory fab. Interbay automation transports wafers from one work cell to the next, but humans actually load the cassettes into the process tools. The fab has been using Promis Systems' (Toronto, ON) manufacturing execution system (MES) for about five years. AutoSimulations software, installed in early 1997, collects transaction records from the MES system, time stamps them, and places them in a proprietary repository. This repository, a highly compressed chronological record of fab events, can be analyzed without degrading MES performance.

Next, software tools analyze the performance of the entire fab, or of individual work cells, to determine when production did or didn't meet the desired constraints, and why. Simulation tools use the repository's historical data for "what if" analysis: would a proposed rule have prevented or caused a given problem in a given situation? Simulations can also carry existing rules forward through time to anticipate upcoming problems: will the bottleneck at Tool A go down quickly enough, or should an additional tool be devoted to that process?

Finally, real time dispatching (RTD) software enforces the desired scheduling policies. Rod Kirby, AutoSimulations' project manager for the Samsung installation, explained that the decision engine considers all waiting lots for a given task and determines which one should have the highest priority. The RTD software's global perspective allows it to balance the productivity of a particular tool against the overall fab's requirements. For instance, a machine might need to sit idle in order to be ready for an upstream high priority lot.

The dispatching software can implement any rule a human can articulate-it has no preconceived ideas about "good" or "bad" rules. While the software can transfer rules from one fab to another, Thompson said that different fabs usually have different scheduling requirements. The rules generally must evolve around a particular fab's constraints.

In April, 1998, Samsung also installed AutoSimulations software at Line 6 in Kiheung. In addition to the interbay automation used in Line 5, Line 6 uses intrabay automation to transport cassettes from stockers to the process tools.

MES software cannot distinguish between wafers located near the tool or away in a stocker, Thompson explained. In order to consider wafer location in dispatching decisions-for instance to always run the most convenient lot-the software also needed to intercept location data from the material control system. While most AutoSimulations installations only make "what next" decisions for equipment and process lots, the software at Line 6 also makes "where next" decisions for the automation systems.

As heavier 300 mm wafers enter manufacturing, this integration of scheduling and automation will become much more important. Thompson expects RTD software to be involved in almost every material move decision in 300 mm fabs.

By Katherine Derbyshire

For more information: AutoSimulations, Inc., 655 Medical Drive, Bountiful, UT 84010, Tel: 801-298-1398; Fax: 801-298-8186.

References
[1] K. Kempf, "The Concepts of Chaos Applied to Manufacturing Production Scheduling," Proc. AAAI/SIGMAN National Conference, Detroit, MI, 1989.

[2] E. M. Goldratt and J. Cox, The Goal: A Process of Ongoing Improvement, 2nd Rev edition, North River Press, 1994.