IEEE TRANSACnONS ON SEMICONDUmOR MANUFACTURING, VOL. 7, NO. 2. MAY 1994
`
`A Planner and Scheduler for
`Semiconductor Manufacturing
`Hugh E. Fargher, Michael A. Kilgore, Paul J. Kline, and Richard A. Smith
`
`~
`
`117
`
`Abstract- The Microelectronics Manufacturing Science &
`Technology (MMST) project includes two closely related CIM
`subsystems for planning and scheduling wafer production. The
`MMST Planner plans all work release into a factory so as to meet
`stated goals, and predicts work completion dates. The MMST
`Scheduler operates in real-time to determine the sequence of
`lot movements and machine loadings that will be performed on
`the fab floor. Both the Planner and the Scheduler continually
`maintain plans which are up to date with the factory status
`by incrementally replanning for unexpected events. The MMST
`Planner can be used as a decision support tool to rapidly analyze
`the consequences of various manufacturing decisions. Planning is
`performed using a modified beam search algorithm, and is based
`on a time-phased capacity model of the factory. Fuzzy arithmetic
`is used to model the uncertainty inherent in cycle time data.
`The MMST Planner is fully distributed, allowing simultaneous
`users in different parts of the factory. The MMST Scheduler
`uses a heuristic method called Score Tables to develop schedules
`of future events. The Scheduler evaluates event prerequisites to
`determine when to initiate lot transfers and machine loadings,
`and responds to any failures of execution.
`
`I. INTRODUCTION
`HE Computer Integrated Manufacturing (CIM) compo-
`
`T nent of the Microelectronics Manufacturing Science &
`
`Technology (MMST) project formed an important part of
`the final project loo0 (1-K) wafer demonstration. This paper
`describes two closely related CIM subsystems used to manage
`production control, namely the MMST Planner and Scheduler.
`The MMST Planner was used to maintain a wafer release
`plan into the MMST factory, and to predict wafer processing
`completion dates. However, once released, control of wafer
`movement passed to the MMST Scheduler. The MMST Sched-
`uler operated in real-time to determine the sequence of lot
`movements and machine loadings performed on the fab floor.
`This paper also describes some of the results obtained through
`using the MMST Planner and Scheduler subsystems during the
`1 -K wafer demonstration, together with possible future work.
`
`11. MMST PLANNER
`The problem tackled by the MMST Planner is that of
`determining when to release work into a factory so as to
`best satisfy customer requests, given the current manufac-
`turing constraints. With this in mind, the MMST Planner
`Manuscript received August 1, 1993; revised December 22, 1993. This work
`was supported in part by the Air Force Wright Laboratory and the DARF'A
`Microelectronics Technology Office under Contract F33615-88-C-5448.
`The authors are with Texas Instruments, Inc., P.O. Box 655012, Dallas, TX
`75265 USA.
`IEEE Log Number 9400853.
`
`has been designed as a make to order or make to forecast
`planning system, for use in complex job-shop manufacturing
`environments. The system maintains a work release plan,
`which determines when work should be released into the
`factory and, to a given confidence level, when that work will
`complete processing. The work release plan is generated so
`as to avoid starving or overloading bottleneck machines at
`any projected time in the plan. This in turn helps reduce
`work-in-process (WIP) and production cycle-times. Plans are
`incrementally updated in line with a user defined strategy
`which, for example, could be used to give preference to plans
`that meet customer requested due dates over those that simply
`maximize machine utilization. The MMST Planner functions
`as a decision support tool, continually maintaining an up-
`to-date plan and providing rapid analysis of user requests.
`Processing capacity of the factory is represented using a high
`level capacity model. Plans are generated using an artificial
`intelligence heuristic search technique which ultimately de-
`termines the recommended work release plan. This contrasts
`with the MMST Simulator [l] which may use a given work
`release plan as input to determine measures such as resulting
`production cycle-times. The MMST Planner is typically used
`to provide a rapid analysis of the consequences of operational
`decisions (such as when to release a particular order), in
`terms of how they would affect the factory in its current
`configuration. Again this contrasts with simulation, which
`is typically used to determine the consequences of more
`strategic decisions (such as work release policies or addition
`of machines) by running suites of simulations on differing
`factory configurations.
`
`A . MMST Planner Goals
`The overall goal of the MMST Planner is to provide decision
`support for production planning in complex manufacturing
`environments. Such a tool could enable improved customer
`satisfaction, while making better use of the production re-
`sources available.
`There is considerable evidence to show the importance of
`work release in achieving the typical goals of semiconductor
`manufacturing [2]. Consequently, an important goal of the
`MMST Planner is to continually maintain an up-to-date plan,
`which determines when work should be released into the
`factory, and to predict when that work will be completed. The
`plan does not determine precisely when wafers are processed at
`particular machines, instead, it simply determines that machine
`processing capacity will be available during the time that work
`
`0894-6507/94$04.00 0 1994 IEEE
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 1
`
`

`

`118
`
`IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 7, NO. 2, MAY 1994
`
`is planned to be in the factory. All times are determined to
`within some granularity, defined by a time interval duration
`which is typically a day or shift. A plan horizon is maintained
`to equal a multiple of the chosen time interval duration.
`Although the MMST Planner determines work that may be
`released during each time interval, the MMST Scheduler
`determines when that work is released into the factory on a
`minute-by-minute basis. In this way, the MMST Planner can
`be thought of as shuffling a queue of work outside of the
`factory entrance. Higher priority work is inserted nearer the
`front of the queue, and is therefore planned for release at an
`earlier date.
`Maintaining an up-to-date plan requires a plan representa-
`tion which remains consistent with the current factory status
`and clock-time. To achieve this, the MMST Planner compares
`the planned and actual work progress in terms of the percent
`processing complete. Replanning is then performed on work
`that is ahead or behind plan. In addition, replanning may be
`performed when machines unexpectedly go down or become
`available for processing. Furthermore, the plan representation
`always covers the current clock-time up to the plan horizon.
`The MMST Planner must also model uncertainty when
`planning work release and completion dates. This applies
`particularly to predicting work cycle-times once released into
`the factory. Very often, the best available data is in the form of
`previously observed cycle-time distributions and as a result the
`MMST Planner associates a confidence level with all planned
`work completion dates that it calculates.
`The MMST Planner requires a strategy, which is used
`to guide plan generation. A strategy is composed of an
`ordered list of goals, which are used to define heuristics and
`constraints. Goals are used to sequence planning decisions so
`as to attempt to meet given plan measures, such as meeting all
`due dates or balancing machine utilization. For example, a goal
`to meet all requested due dates may have a heuristic which
`sequences work based on slack to due date. Constraints are
`used to limit plan measures such as planned WIP. A constraint
`to limit planned WIP would prevent generation of a plan that
`exceeds the stated WIP at any point in time. At any point in
`the planning process, decisions are sequenced using the first
`goal on the strategy’s ordered list of goals that is relevant to
`the decision being made. The user may also define multiple
`strategies, but only the active strategy is used at any one time
`to sequence altematives.
`Part of the overall goal of the MMST Planner is to allow
`multiple, simultaneous users of the system, in a distributed
`environment. This is achieved by providing access to the
`MMST Planner from any connected workstation in the factory,
`and by managing the concurrency issues that arise when more
`than one person is using the system (such as when two users
`attempt to update the plan at the same time).
`Finally, for decision support, the MMST Planner must
`provide a way of rapidly exploring the consequences of various
`manufacturing decisions. It achieves this by allowing rapid
`incremental updates to the existing plan to be explored, without
`having to necessarily commit to such updates. Consequently,
`the MMST Planner concentrates on the impact of decisions on
`the factory as it currently exists.
`
`B. MMST Planner Approach
`The planning algorithm used within the MMST Planner is
`described in detail elsewhere [3], [4]. This section gives only
`a brief overview of the algorithm.
`The plan representation used within the MMST Planner
`has been devised so as to model the projected work load
`within the factory, while allowing incremental updates at any
`time to account for changing circumstances, such as new
`orders or machine failures. Allowing incremental updates for
`new orders enables rapid feedback to customers concerning
`feasible ship dates, without having to wait for a daily or
`weekly periodic plan update. Allowing incremental updates for
`machine failures provides an early warning for work that may
`be late. The plan representation is based on the processing
`capacity of machine groups within the factory, divided into
`contiguous time intervals of arbitrary duration. The use of
`contiguous time intervals is referred to as a time-phased
`representation, and allows a plan to be represented up to some
`predefined horizon to any level of detail. Consequently, the
`planning algorithm is used to determine work committed to
`each machine group during each time interval. In practice
`plan generation is influenced by the bottleneck machines, a
`feature shared by other planning approaches [5]. By using a
`time-phased capacity model, bottlenecks for a particular time
`interval can always be identified, since bottlenecks may vary
`over time.
`The planning algorithm is divided into two parts, that of
`determining the sequence of work to be planned (given its
`requested due-date, customer priority, etc.), and incorporating
`the required work into the plan (given the current machine
`group commitments, type of planning request, and any con-
`straints on which time interval the work may be planned
`for). Ultimately, any update to the existing plan (including
`replanning due to machine failure) can be tackled in this way.
`The MMST Planner strategy determines the sequence of work
`to be planned. To incorporate the required work, the MMST
`Planner algorithm uses a beam search similar to that used
`within other artificial intelligence planning systems such as
`ISIS [6]. However, unlike ISIS the beam width grows with
`search depth and uses a simple backtracking scheme to search
`within the beam. The search algorithm is ultimately searching
`for available processing capacity, over existing time intervals,
`to incrementally update the plan to accommodate any change
`in work commitments.
`The plan representation must also model the uncertainty
`inherent in the entire production process cycle-time, since such
`cycle-times often form the best available data for planning. Un-
`certainty is modeled by reinterpreting the plan representation
`in terms of fuzzy sets, an approach which has been previously
`used within FSS [7]. However, whereas FSS generates an
`overall cycle-time distribution from all contributing process
`cycle-time distributions, the MMST Planner performs the
`reverse operation. Comparisons with simulation have shown
`the accuracy of this reverse operation [3]. The result can
`be used to determine the degree of membership of work
`commitment for each machine group within each time interval,
`which reflects the expected accuracy of the projected work
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 2
`
`

`

`FARGHER et al.: A PLANNER AND SCHEDULER FOR SEMICONDUCTOR MANUFACTURING
`
`I I9
`
`load. Typically, projections become less certain the further
`they are made into the future. The final result is that planned
`work completion dates are computed to some confidence
`level. The overall cycle-time distribution can also be used to
`determine an earliest and latest expected completion date.
`
`C . MMST Planner Architecture
`One of the requirements of the MMST Planner architecture
`is to operate in a distributed environment. To satisfy this
`requirement the MMST Planner can be accessed from any
`connected workstation in the factory, and used to view the
`current plan, plan new work or perform what-if analysis.
`Clearly, determining the user’s authority is important. Few
`users would have the authority to accept updates to an existing
`production plan, while many may have the authority to view
`the current plan. Experience has shown that while the main
`users of the MMST Planner work from within their office, it
`is important that they be able to use the same features on a
`workstation on the factory floor.
`The MMST Planner is able to run in a distributed envi-
`ronment using two different processes: the server and the
`User Interface (UI) process. At any point in time there is
`one server process running and zero or more U1 processes,
`all running on one or more workstations. The server process
`runs continuously, while U1 processes are started and stopped
`depending on when users need access to the planner. The
`server process is responsible for maintaining the continuous
`presence of the planning system, and automatically replan-
`ning whenever the manufacturing environment significantly
`changes from the production plan assumptions. Examples
`of such changes include work release, work completion, or
`unexpected machine failure. While automated interactions go
`directly to the server, human interaction with the planner is
`performed using the U1 process. Examples of such interaction
`include order entry, browsing the current plan, planning new
`orders, and performing what-if analysis. Since multiple pro-
`cesses can run simultaneously, concurrency issues are handled
`by preventing users from committing plan changes that violate
`other committed changes, and informing users of any plan
`changes that have recently occurred.
`The server process also monitors the comparison between
`planned and actual work progress in the factory, therefore
`providing some feedback between the MMST Scheduler and
`Planner. If actual work progress deviates more than some
`user defined tolerance from planned progress, the work is
`automatically replanned. This provides a warning to system
`users that particular work may be deviating from its original
`planned completion date.
`D. MMST Planner User Interaction
`The MMST Planner enables the system user to explore
`the consequences of various manufacturing decisions, without
`having to necessarily commit to them. Decision support is
`divided into two types: implicit what-if and explicit what-
`i f analysis. The two types of decision support reflect two
`different uses of the system. Implicit what-if is limited to ana-
`lyzing the consequences of planning new work, or replanning
`current work, in the factory. Once performed, the updated plan
`
`can be either accepted by the system user (in which case it
`becomes the current factory production plan) or rejected (in
`which case all updates are discarded). For this reason, implicit
`what-if planning requests are limited to those for which the
`system user has the authority to execute. Explicit what-if
`analysis is used to analyze the consequences of a variety of
`operational decisions, such as when to bring a machine down
`for maintenance, or whether to continue processing particular
`work over a weekend. However, plans generated using explicit
`what-if analysis may not be accepted through the MMST
`Planner user interfaces, since they typically imply production
`decisions which are outside the authority of the planner system
`user.
`Implicit what-if analysis can be used to plan new work with
`a variety of commands. The commands either plan new work
`non-disruptively (which guarantees that no existing planned
`release or completion dates will be affected) or disruptively
`(which makes no such guarantee). Disruptive planning requests
`are those which effectively reshuffle the queue of orders
`outside the factory door, while non-disruptive requests simply
`slip a new order into that queue. Furthermore, plan requests
`either attempt to plan work release so as to complete on a
`particular day (which invokes a backward planning algorithm)
`or plan release regardless of completion dates (which invokes a
`forward planning algorithm). Fig. 1 shows the MMST Planner
`screen that results from an implicit what-if request. In this case,
`plans are displayed in tabular form (as opposed to a Gantt
`chart), showing the current plan (lower), and the modified
`plan (upper) on the same screen. Notice that the system user
`can then compare the two plans before deciding whether
`to accept or reject the modifications. The four smaller lists
`in between the plan tables show other information such as
`successfully planned work, work which could not be planned
`due to capacity constraints, any work dislodged during the
`planning process (and not planned back in) and all remaining
`unplanned work.
`Explicit what-if analysis is divided further into two
`and extensive analysis. As the name suggests,
`types-brief
`brief analysis is performed faster, but provides less information
`to the system user. For example, brief analysis may be used
`to determine which planned order items would be affected
`if a particular machine was to go down for 36 hours, but
`would not determine by how much their planned release or
`completion dates are pushed back. Extensive analysis, which
`employs full-scale incremental replanning, would determine
`precisely by how much the planned dates are pushed back.
`The advantage in having both brief and extensive analyses
`available is that system users can quickly browse the overall
`effects of a wide variety of decisions using brief analysis,
`before evaluating particular decisions in more detail using
`extensive analysis.
`Fig. 2 shows the MMST Planner screen used to perform
`brief analysis. Notice that the particular analysis just performed
`is that of determining which order items are affected by taking
`the AVP526ChamberSpec machine down for 12 hours, the
`result being displayed in the scrollable list (“Order Items
`Affected”). Other brief analysis query types are listed on the
`screen in Fig. 2.
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 3
`
`

`

`I
`
`120
`
`IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. I, NO. 2, MAY 1994
`
`................................................................................................................................................................................................................................................................(cid:16)(cid:1)
`PlanWORKS: Plan Requestor Results: Planned order Items
`...............
`
`ModHied Pian
`
`(Planned - When Posslble) (created 6/6/93; 10:16 am)
`
`I
`
`Current P ~ M
`
`Fig. 1.
`
`Implicit what-if results screen.
`
`Performing a single extensive analysis may require more
`than one screen, since each such analysis may consist of a
`whole group of planning requests. For example, the analysis
`may determine the effects of taking one of the Coater and
`Implanter machines down for maintenance over the next 2
`days, while opening the factory to work on all critical order
`items over the following weekend in an attempt to make up the
`lost time. Separate MMST Planner screens are used to define
`the time period to take the machines down, as well as the time
`period to open the factory. Once all planning requests have
`been defined, the analysis is actually performed by selecting
`‘analyze’ on the main planner extensive analysis screen. The
`result of the analysis is a screen similar to Fig. 1, used to
`compare plans before and after a request has modified the
`production plan. However, the difference is that the modified
`plan may not be accepted in this case. It can only be browsed,
`with a view to warning other personnel of the effects that
`would result if they decided to place the Coater and Implanter
`down, together with working on critical order items over the
`following weekend. If the Coater and Implanter were then
`actually placed down for maintenance, the MMST Planner
`would be informed of the change in machine status and would
`perform an automatic replan so as to maintain consistency
`with the current factory status.
`
`E. MMST Planner Results
`During the MMST 1-K wafer demonstration the MMST
`Planner was used to perform continuous and off-line planning
`operations. Continuous planning operations included all day-
`to-day tasks, such as order entry, planning and work release, as
`well as planner server execution. Off-line planning operations
`were performed by the system developers, as and when
`required, to assist in any decision making.
`The day-to-day tasks were typically performed by the fac-
`tory plan manager. Order creation and planning were generally
`performed one day prior to the day corresponding to the
`required release date for the order. One or more orders
`would be created, planned and prepared for release at a time.
`Preparing the orders ahead of time ensured that new work
`was ready to start when the MMST Scheduler was prepared to
`begin processing the work. In addition to the tasks performed
`by the users of the MMST Planner, the Planner server was
`available 24 hours a day. The server monitored progress of
`work in the factory and supported the work release operation.
`As part of the day-to-day tasks, daily reports were produced
`detailing lot progress. The reports compared planned wafer
`moves against observed wafer moves, for various subsets of
`work. Both tabular and graphical versions of the report were
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 4
`
`

`

`1
`
`FARGHER er al.: A PLANNER AND SCHEDULER FOR SEMICONDUCTOR MANUFACTURING
`
`121
`
`Analysis Selectlons
`
`Wpmal: 1 AVP526ChamberSpec
`
`I
`
`s-Date:
`
`1 P 6 / 0 7 / 9 3 ~ m n H l a r y Y
`
`--: pq" om
`
`-44Idyais Results
`onler lhmJ Affected
`MMST 11 3
`MMST122
`MMST.14 1
`
`mwn-
`
`-:
`
`pm
`
`+minutes
`4$ hours
`
`Fig. 2. Brief analysis screen.
`
`generated and were used by managers to determine whether
`work was progressing adequately.
`Historical data provided the opportunity to examine the
`accuracy of plans generated by the MMST Planner. Fig. 3
`shows a graph of planned and actual remaining wafer moves
`for a subset of lots run during the demonstration, where one
`wafer move corresponds to the completion of one processing
`step, for one wafer, on one machine. The subset chosen in this
`case corresponded to larger lot sizes which ran a greater than
`three day cycle time. The labeled dark line (labeled at each
`point) shows the planned remaining wafer moves on a day by
`day basis. The unlabeled grey line shows actual progress of the
`lots through the factory. By comparing agreement between the
`observed and planned lines, a measure of the MMST Planner's
`accuracy can be determined. One measure is to calculate the
`mean percent difference between planned and actual remaining
`wafer moves for a given time period. This example results
`in a measured mean difference of approximately 9%. For
`simplification, Fig. 3 shows a snapshot of the plan generated on
`March 24th and does not illustrate the effects of subsequent
`replanning.
`
`100
`
`0
`24
`March '93
`
`25
`
`26
`
`27
`
`28
`
`29
`
`30
`
`31
`
`Day st.r(lng
`
`1
`Apdl'93
`
`2
`
`Fig. 3. Planned and actual wafer moves.
`
`In addition to using the MMST Planner as the production
`planner during the demonstration, it was also used to perform
`several types of analysis in an off-line planning mode. Off-
`line analysis resulted in the generation and evaluation of
`plans that were not intended to be used as production plans;
`instead, they were generated to answers specific questions
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 5
`
`

`

`122
`
`IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 7, NO. 2, MAY 1994
`
`posed by various managers during the demonstration. The
`reason the analysis was performed off-line was because many
`of the requests required user interfaces that did not exist at
`the time. Consequently, the system developers implemented
`and executed additional code specifically to answer these
`questions. Requests for off-line analysis broadly fell into four
`categories: flow, production, resource usage and wafer move
`analysis. Typical examples included: flow analysis to calculate
`the number of steps per day needed to meet given cycle time
`multipliers, and production analysis to determine the effect of
`various product mixes. Most of the off-line requests can now
`be executed from the MMST Planner UI.
`To find the impact the MMST Planner had on the decision
`making process during the MMST 1-K wafer demonstration,
`users of the system were interviewed. Their feedback indicated
`that the system was used to make both operational and
`strategic decisions. An example operational decision was to
`determine suitable time periods for preventative maintenance
`on particular machines. An example strategic decision was to
`determine the product mix. In addition, daily reports provided
`a speedometer which was used to determine if weekend shifts
`were needed.
`
`F. MMST Planner Future Work
`The MMST 1-K wafer demonstration provided an excellent
`initial test for the MMST Planner. Installation of the MMST
`Planner to both intemal and external beta customers has
`already begun as part of the commercialization effort. Another
`important area of work will concentrate on model validation.
`Customer installations will provide the opportunity to collect
`historical data to compare with results generated by the MMST
`Planner. In addition, simulations will be performed to validate
`plans generated using the MMST Planner model. Other future
`developments for the MMST Planner include an improved
`search algorithm which would dynamically manage planned
`utilization during plan generation. This would enable the
`plan representation to model time intervals of an increasing
`duration. The justification for this is that near term planning
`should be performed with finer granularity (e.g., daily) than
`long term planning (which could use weekly or monthly time
`intervals). A further development includes generalizing the
`MMST Planner to address world-wide planning for multiple
`factories.
`
`111. MMST SCHEDULER
`In most wafer fabs, operators monitor the fab situation
`in order to make scheduling decisions. In the MMST CIM
`system, the Scheduler provides an automated capability for
`making scheduling decisions in real time on the fab floor.
`The Scheduler is responsible for factory scheduling; i.e., the
`movement of lots between machines and the sequence of
`lot processing in the fab. By contrast, the MMST Machine
`Control system [8] is responsible for machine scheduling; i.e.,
`wafer handling and the sequence of wafer processing within
`a machine.
`
`A. MMST Scheduler Goals
`The design of the MMST Scheduler was motivated by the
`following goals of the MMST 1-K wafer demonstration:
`fabricate double-level metal CMOS devices in less than
`three days;
`minimize processing errors;
`eliminate the use of paper in the fab.
`The function of the Scheduler within the MMST CIM system
`is to translate general instructions about fab operations into
`more specific instructions that can be executed by other
`CIM subsystems and fab personnel. For example, the MMST
`Specification subsystem might show that “gate pattem ex-
`pose’’ is step 23 in fabricating wafers of a particular device.
`This general instruction is translated by the Scheduler into a
`machine-loading instruction: “perform gate pattem expose for
`Lot 57 on Stepper #2 next.” The where and when that were
`left unspecified by the Specification system have been filled in
`by the Scheduler based on the current fab situation. This more
`detailed instruction is executable by the combined efforts of
`operators and the Machine Control system to accomplish step
`23. The Scheduler notifies the operators through user interfaces
`where lots should be taken for processing, which supports
`paperless manufacturing. The Scheduler also sends the lot id
`and detailed processing specifications to the Machine Control
`software for Stepper #2 to ensure the correct processing is
`performed.
`The goal of minimizing processing errors led us to adopt
`a predictive approach to scheduling and reject a dispatch rule
`approach. Dispatch rules are a common technique for wafer
`fab scheduling (e.g., [9], [lo]). First-In-First-Out is an example
`of a simple dispatch rule. When a machine becomes free,
`dispatch rules decide at that time which lot to load; i.e., there
`is no schedule of future loadings. However, it is common to
`have time constraints between certain processing steps, e.g.,
`cleanups and depositions. An approach to coping with these
`problems is to “look before you leap” by developing a schedule
`for future steps and checking that the constraints are satisfied
`before executing the first step. This is not an option for a pure
`dispatch rule approach.
`Although the MMST Scheduler does not yet cope with time
`constraints, it does develop a schedule of future events for each
`lot and each process resource. In the 1-K wafer demonstration,
`these schedules covered the next two hours (the duration is
`a configurable parameter), and indicated the estimated times
`(hr:min) for the start and completion of events. Schedules were
`developed using a new heuristic scheduling method, called
`Score Tables, to make decisions about which lot to assign next
`to process resources.
`A predictive schedule can become obsolete when lots go on
`hold or machines go down. In such cases, it is necessary to
`reschedule to accommodate the unexpected event. Automatic
`rescheduling is currently an active area of scheduling research
`[ 111. The approach we adopted was simply to flush the sched-
`ule for the affected lots and reschedule. This was satisfactory
`in the 1-K wafer demonstration, where there were typically
`less than ten lots in the fab at any one time. More efficient
`altematives may be required to cope with additional lots.
`
`Authorized licensed use limited to: Don Zhe Nan Wang. Downloaded on July 24,2021 at 02:26:29 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner STMICROELECTRONICS, INC.,
`Ex. 1027, IPR2022-00681, Pg. 6
`
`

`

`FARGHER et al.: A PLANNER AND SCHEDULER FOR SEMICONDUCTOR MANUFACTURING
`
`123
`
`MMST SCHEDULER
`
`GENERATOR
`
`Fig. 4. MMST Scheduler architecture.
`
`TABLE I
`
`ACTIONS hVOLVED IN DISPATCHING A LOT BETWEEN hfACH1NF.S
`SCHEDULER
`MACHINE 1
`Finish lot and
`quest removal
`
`I Move to WIP rack
`
`Ask operator to take
`>I to WIP rack
`operator logout lot
`Request next lot
`
`3 Reserve location for
`lot
`
`7 operator login lot
`
`9 Commit lot to Ma-
`chine 2
`
`12 Move to Mschine:
`
`8 Request next lot
`
`10 Ask operator to
`proccsr lot now
`11 operator ogracs
`
`13 Show lot bound
`for M.chinc 2
`15 Opamor logout lot
`
`14 Ask operator to get
`IotfrOmWIP~Ck
`
`16 Operator login