Proceedings of the 2004 Winter Simulation Conference R .G. Ingalls, M. D. Rossetti, J. S. Smith, and B. A. Peters, eds. ABSTRACT In this research, we investigate how well Weibull, Gamma, and special bimodal distribution are suited as an alternative to the exponential distribution approach in the stochastic modeling of machine downtimes and times between failures. We also discuss the question whether sampling shop-floor data should not only include first order statistics, but also measures that allow to monitor and model the variability of the equipment and processes and even the correct distribu- tion of the data. 1 INTRODUCTION A typical semiconductor manufacturing facility contains up to 1000 various machines and tools. Besides the complexity of handling this vast amount of equipment, there are several other factors that make production planning and control in this environment particularly difficult. (Cf. (Schömig and Fowler 2000) and (Uzsoy et al. 1992) for a thorough summary of these factors and shop- floor control problems in semiconductor manufacturing.) Unpredictable machine downtimes are believed to be the main source of uncertainty in the semiconductor manufacturing process. Obviously, downtimes are a severe problem, since production capacity is lost and the flow of material is disrupted. The reliability of semiconductor manufacturing equipment is unusual from a number of standpoints. Despite of all efforts to tune and calibrate machines to an optimum performance, they are still subject to random failures. The failure of equipment or processes is often not a hard failure in the sense that something obviously breaks or goes wrong; but rather, a soft failure in which the equipment begins to produce out of the tolerance region. For this reason, the equipment usually completes a lot or batch prior to being taken out of service for repair which often involves more tuning, calibration, and test rather than component replacement. Since some wafer fabrication tools, such as ion implanters, may be down 30-40% of the time, the impact of periods of unavailability on production control as well as overall productivity is tremendous. Hence, appropriate modeling of equipment and process failures is a must to derive meaningful output performance measures. The SEMI E10 and E58 standards provide a framework for sampling machine-level data in the semiconductor industry. Downtime is a period of time during which the equip- ment is not in a condition to perform its intended function. This period does not include any portion of time, where the equipment or the entire facility are not scheduled to per- form fabrication. Generally, it is distinguished between scheduled and unscheduled downtimes. A scheduled downtime occurs, when the equipment is not available to perform its intended function due to planned events such as preventive maintenance, production test, change of consumables, and machine setup for run- ning a different process. All of these procedures are clearly separable and planned in their respective process. Also in- cluded are test run times for the required subsequent re- qualification and re-approval. Waiting times resulting from delays in the process are also included. Unscheduled downtimes are periods of time during which the equipment is not in a condition to perform its in- tended function due to an unplanned event. Examples are: technical failures, unplanned measures to secure operation, unplanned shut down of supply infrastructure. These events interrupt equipment operation. In resolving these interruptions (Interrupts) they are distinguished as follows based on timing and personnel requirements: An assist is an unplanned interruption that occurs dur- ing an equipment cycle if all three of the following condi- tions apply: (1) The interrupted equipment cycle is re- sumed through external intervention (e.g. by an operator), (2) there is no replacement of a part, other than specified consumables, (3) there is no further variation from specifi- cations of equipment operation. An assist usually lasts not longer than 6 minutes. A failure, however, is any un- planned interruption or variance from the specifications of equipment operation other than assist. 2 MODELING EQUIPMENT DOWNTIMES Obtaining the averages of uptimes and downtimes are suffi- cient when these time periods are assumed to be exponen- tially distributed. This is the prevalent assumption in reli- MODELING TOOL FAILURES IN SEMICONDUCTOR FAB SIMULATION Oliver Rose Institute of Computer Science University of Würzburg Würzburg, 97074, GERMANY.