Professional Documents
Culture Documents
Caterina Genua
Silvestro Giuffrida
Salvatore Rinaudo
ST Microelectronics
Stradale Primosole, 50
95121 Catania (Italy)
Abstract: In the highly competitive semiconductor industry, shop floor controls have
recently become fundamental for production controllers and industrial engineers. Despite
the emphasis put on SFS techniques, the impact on the actual operations of the fab may
not reflect controllers’ expectations. A framework has been developed which provides
them with means of analyzing the production flow, through statistical analysis of
equipment, operators and flow linearity parameters, and through data collection for
historical comparisons between simulated and real activity of the fab, which can be
helpful for documenting the improvements on manufacturing performance offered by the
use of shop floor controls. Copyright © 2006 IFAC
1. INTRODUCTION the real activity of the fab does not adhere to the
dispatching decisions in a systematic way. Even though
In the past few decades shop-floor scheduling has various knowledge-based and genetic algorithms do exist,
been recognized as an important tool in improving the inadequate performance of shop floor control tends to
manufacturing performance. Simulating the activity represent a problem in most industrial enterprises (Sloan,
of the fab through a mathematical model allows 2003; Scherer, 1998).
industrial engineers and production controllers to
perform capacity, what-if analysis and planning and Therefore, it is important to create solutions that help to
to offer dispatching indications to people working on bridge the gap between theory and practise. An analysis
the shop floor. framework has been developed offering industrial
engineers and controllers the possibility to measure the
The semiconductor manufacturing cycle is one of the impact of the shop floor scheduling policies on the real
most difficult processes to model. In a wafer activity of the fab, through evaluation and monitoring of
fabrication environment, people have to cope with many parameters indicating fab performance.
complex reality, with randomness in yields, reworks
and resources unavailability, rapidly changing
products and technologies, reentrant flows, time 2. BACKGROUND
constrained operations, batching… Moreover,
traditional research does not take into account many
2.1 Literature review
potentials present in real world production, e.g.
operators’ experience, motivations and Many examples of shop floor control study can be found
qualifications. In other words, while researchers are in different manufacturing environments. The approaches
still studying the effects of different control policies can be differentiated according to the performance
and industrial engineers and production controllers measures to optimize (globally or locally), the control
are trying to adopt sophisticated shop floor tools to policies (scheduling and dispatching techniques) included
improve production performance, the impact of these in the model, the detail of the model used to reproduce the
studies and control measures is not as big as fab activity in a simulator.
expected, and it is not uncommon to discover that
equipments, products, operations…), retrieved from
As for the semiconductor world, some authors claim Workstream through Cobol extractors or APF reporter,
that a reduction in mean cycle time is needed to processes all these information and creates a set of reports
optimize performance, others target reduction in regarding lots and equipments status changes during the
cycle time variance in order to minimize lateness, simulation horizon. The outputs provided by the simulator
thus respecting on time delivery (Lu, et al., 1994). are reprocessed through a set of procedures (called STAP
Critical Ratio Criterion also targets maximized on toolbox) to build some reports for analysis or dispatching
time delivery by prioritizing lots based on their purposes.
current position, due date and remaining cycle time
(Lee and Chen, 1997; Kim, et al., 1998). The fabs whose results will be presented in this paper are
a pilot line for technology development and a production
Another important criterion is line balancing, i.e. line, which operate 24 hours per day, with work teams
targeting the linearity of product flow, in order to that change every 8 hours. An 8-hours period of work is
avoid bubbles of WIP (Work in Process, total called “shift”. The only exception is represented by
number of wafers present in the manufacturing Sunday night, when there is no regular personnel
processing line) to form and to equilibrate workload operating. Three simulations per day are run, in order to
in the different tools (Shen and Leachman, 2003). produce a set of reports about fab activity in the following
Dispatching and scheduling schemas involving a 8 hours shift. Everyday these data are collected to build an
weighted average of various performance measures historical repository for reporting and performance
have been developed (Dabbas and Fowler, 2003; analysis purpose.
Kim, et al., 2003; Lee and Chen, 1997).
Some examples will also be given of a new performance
As far as the model detail is concerned, if interested measure, called Dynamic XFactor, that has been
in an extensive what-if analysis, the researcher can introduced recently and that has begun to be monitored
choose to include all the necessary details, possibly for another ST Microelectronics production fab.
integrating uncertainty analysis (Shen and
Leachman, 2003), while for Simulation Based Real
Time Scheduling (Kim, et al., 2003) and fast- 3. GANTT CHARTS FOR PRODUCTION FLOW
rescheduling flexible solutions (Liao, et al., 1996) it
is important to reduce the complexity of the model to In the following, some examples will be given of the
minimize response times. engineering data analysis that can be performed using our
framework, focusing on some of the performance metrics
As for other manufacturing environments, shop floor and related aspects treated extensively by Genua, et al.
control and throughput analysis case studies can be (2005).
found in Gupta and Sarawgi (2002), Norton (1996),
Li (2004).
3.1 Linearity and Global Process Flow
2.2 Case Study Overview Moves/EOH per shift. Moves represent the number of
transactions towards the next operation (which is a step of
The motivation and data for this paper come from the the process flow) performed in a shift. EOH (End of
work realized for an R&D fab and a production fab Hand) is the number of wafers that at the end of a given
of ST Microelectronics. “Gantt Charts for Production shift are in process on equipment or waiting for the next
Flow” framework, by collecting data related to operation, constituting the work-in-process (WIP).
various metrics, by statistical reprocessing of real
values regarding fab equipment, operators and the Visualizing the global MOVES/EOH flow for a given
whole production flow, and by comparing collected technology/product/process group in a given date range,
data to the results coming from a simulator of the with a differently colored and shaded table allows to
manufacturing process, offers the possibility to detect immediately some operations where the linearity is
evaluate and monitor production performance, as violated.
well as the influence of shop floor decisions on
Non-linearity in the moves distribution could indicate a
practice.
period of unavailability of some resource or the
For data collection, the information system offered movement from batch equipment to a serial one, with the
by MESs (Manufacturing Execution Systems) is formation of queues in front of a station or station family
fundamental to retrieve every kind of information which can also be encountered in the EOH distribution for
which is related to the production process (Qiu and the same time bucket.
Zhou, 2004). In our case, the MES system is
Consilium Workstream, and the data are extracted An increasing number of EOH in certain areas of the
using Brooks APF (Advanced Productivity Family) process flow likely indicates a bottleneck station, meaning
that parts are excessively stocked in front of some
reporting system.
equipment.
The mathematical simulator used in the previously
Figure 1 shows an example of Moves linearity flow for a
mentioned fabs is Brooks Autosched AP. The
particular process group and for a period of two months.
simulator accepts inputs regarding fab status (WIP,
The real cycle time can be retrieved by MES computing
for each lot the difference between its start date and end
date.
Fig. 2. EOH flow example. Reducing the cycle time reduces the work-in-process and
the capital invested on it. This tool can also be used to
detect variance in cycle-times, analyzing cycle time
distribution. Reducing variances in cycle-times, caused by
the randomness of arrivals and processing times, it is
possible to predict safely when a wafer will complete
production and exit the plant. This allows an improved
ability to meet due-dates reliably.
Fig. 3. EOH flow detail. 3.2 Equipment operation distributions and statistics
Theoretical and Real Cycle time. The cycle time, Information related to equipment operation can be
also known as “manufacturing lead time” is the time extracted from the MES system or from automation
taken by wafers in their passage from the entry to the database for long time horizons to compute statistics.
exit of the plant. This quantity can be computed
theoretically according to Little’s Law, which can be Process time. The data related to the time spent by a
applied to every system in equilibrium in which station or a group of stations to perform a process on a
customers arrive, spend time and depart. Assuming given lot, together with the information on the number of
that is the average arrival rate, W the average time a wafers present in that lot, can be collected to form a
customer spends in the system and L the average statistical distribution.
number of customers in the system, L can be
computed using the formula: Arrival rate. The number of lots per hour that arrive in
front of a station is also of interest in the study of
L = λW equipment operation. These values are collected
(1) periodically and statistics are built.
which in terms of Cycle Time, WIP and throughput Down time. Performance metric that takes into account
can be expressed as the state transactions of a station are MTBF (Mean Time
Between Fails, time between two consecutive fails, figure
WIP 7), MTTR (Mean Time To Repair, time needed to repair a
Cycle _ time =
throughput station, figure 8), down percentage
(2) (down%=MTTR/MTBF), availability, percentage of time
a station is available for processing (Av=100-down%).
An example can be found in fig. 9.
maintenance operation, unscheduled failure repairing and
qualification), ENG (process characterization and
engineering, equipment evaluation), NO MAT (material
needed to process wafers is not available), NO OP (no
operator or support functions available, including breaks,
lunches and meetings), NO WAF (no wafers are available
Fig. 7. MTBF to be processed, SETUP (job preparation, e.g. change of
recipe, mask, tooling, test program,… between two
regular production phases, production tests, process
control, engineering runs associated to manufacturing lot).
Based on the quantity of wafers processed in given time
shot by the station, and on the theoretical process time
Fig. 8. MTTR coming from AP simulator events library, the percentage
of time spent by the station in a status which is not
classifiable into the known inefficiencies stated above is
called UNKNOWN TIME. It is an important performance
measure which has to be monitored and minimized.