Nof S.Y. Springer Handbook of Automation

Подождите немного. Документ загружается.

Semiconductor Manufacturing Automation 52.3 Equipment Integration Architecture and Control 915

Single-slot

chamber

Multi-slot

chamber

Va cuuming and

pumping buffer

Loadlock

a) b) c)

Fig. 52.3a–c Tool architectures: (a) single-slot cluster tool, (b) multi-slot cluster tool, (c) tool with intermediate buffers

velop the circuit patterns on the wafers that are formed

by exposures to circuit pattern picture images at the

steppers. Process modules for coating and develop-

ing, and accompanying baking and cooling modules

are combined into a track tool with several robots, as

illustrated in Fig. 52.4. Each process step has ﬁve to

ten parallel modules [52.5, 6]. An automated wet sta-

tion also has a series of chemical and rinsing baths for

cleaning wafer surfaces, which are combined by several

robots moving on a rail [52.7]. Recently, EDS processes

for testing devices on wafers are automated to form

a kind of track system. A number of testing tools for

wafer burn-in (WBI) test, hot pretest, cold pretest, laser

repair, and posttest are conﬁgured in series–parallel by

several robots moving on a rail. EDS systems and wet

stations can process several different wafers concur-

rently while most cluster tools or track tools for coating

and developing repeatedly process identical wafers.

Wafers mostly go through a sequence of process

steps in series. For some processes, wafers visit some

process steps again; for instance, unlike conventional

chemical vapor deposition, atomic-layer deposition pro-

cess controls the deposition thickness by repeating

extremely thin deposition multiple times. Therefore,

a wafer reenters the chambers many times. In track sys-

tems, wafer reentrance can be achieved, depending on

the chamber conﬁguration and process recipe. In some

processes, a chamber should be cleaned after a speciﬁed

number of wafers have been processed or when sensors

within the chamber detect signiﬁcant contamination. If

a wafer remains in a chamber after processing, this can

lead to quality problems. This idle time, called wafer

delay, must be bounded, reduced or regulated. Process

times or tasks times are rather constant, but can be sub-

ject to random variation, mostly within a few percent.

There can be exceptional delay, even if only rare, due to

10 coaters

10 cool plates

Loadlocks 10

developers

Interface to

steppers

Robot

Optical

edge bead

removers

5 bakers

Robot

5 hot plates

Module

electronics

Buffer

5 bakers

Robot

Buffer CPx

Buffer

Module

electronics

Fig. 52.4 A track system

abnormal process conditions. A wafer alignment task,

which correctly locates a wafer unloaded from a load-

lock onto a robot arm by using a laser pointing system,

sometimes fails and needs to be retried. Integrated tools

mostly limit intermediate buffers. Therefore, blocking

and waiting are common and even deadlocks can occur.

Reentrance, wafer delays, cleaning cycles, and uncer-

tainty all increase scheduling complexity signiﬁcantly.

Tool productivity by intelligent scheduling and control

is critical for maximizing fab productivity and even af-

fects wafer quality signiﬁcantly.

52.3.2 Tool Science: Scheduling and Control

Scheduling Strategies

There can be alternative scheduling strategies for clus-

ter tools. First, a dispatching rule determines the next

robot task depending on the tool state. It can be con-

sidered dynamic and real time. However, it is hard to

optimize the rule. We are only able to compare per-

Part F 52.3

916 Part F Industrial Automation

formances of heuristically designed dispatching rules

by computer simulation. Second, a schedule can be

determined in advance. This method can optimize per-

formance if a proper scheduling model can be deﬁned.

When there is a signiﬁcant change in the tool situ-

ation, rescheduling is done. Cyclic scheduling makes

each robot and each processing chamber repeat identi-

cal work cycles [52.7,8]. Once the robot task sequence

is determined, all work cycles are determined. Most

academic works on cluster tool scheduling consider

cyclic scheduling. Cyclic scheduling has merits such

as reduced scheduling complexity, predictable behav-

ior, improved throughput, steady or periodical timing

patterns, and regulated or bounded task delays or

wafer delays and work in progress [52.7–10]. In cyclic

scheduling, the timings of taskscan be controlledin real

time whilethe sequenceor workcycle is predetermined.

A cluster tool that repeats identical work cycles can

be formally modeled and analyzed by a timed event

graph (TEG), a class of Petri nets [52.12]. Transitions,

places, arcs, and tokens usually represent activities or

events, conditions or activities, precedence relations

between transitions and places, and entities or condi-

tions, respectively. They are represented graphically by

rectangles, circles, arrows, and dots, respectively. Fig-

ure 52.5 is an example of a TEG model for cluster tools.

Once a TEG model is made, the tool cycle time, the

optimal robot task sequence, the wafer delays, and the

optimal timing schedules can be systematically identi-

Unloading complete

robot move

Unloading

from LL

Unloading

Chamber 1 available

Fig. 52.5 A timed event graph model for a dual-armed cluster tool [52.11]

ﬁed [52.7,9,10]; for instance, the tool cycle time is the

maximum of the circuit ratios in the TEG model, where

the circuit ratio of a circuit is the ratio of the sum of the

total times in the circuit to the number of the tokens in

the circuit. For instance, the cycle time of a dual-armed

cluster tool can be derived from the ratio as

max



max

i=1,...,n

+2u +2l+3v

(n +1)(u +l +2v)



where p

, m

, u, l, v,andn are the process time of pro-

cess step i, the number of parallel chambers for process

step i, the unloading time, the loading time, the move

time between the chambers, and the number of process

steps, respectively [52.13].

Schedule Quality

For a cluster tool with a given cyclic sequence, there

can be different classes of schedules, each of which

corresponds to a ﬁring schedule of the TEG model.

A periodic schedule repeats an identical timing pattern

for each d work cycles. When d =1, the schedule is

called steady. In a steady schedule, task delays such

as wafer delays are all constant. In a d-periodic sched-

ule, the wafer delays have d different values, while

the average is the same as that of a steady sched-

ule. The period d is determined from the TEG model.

A schedule that starts each task as soon as the pre-

Part F 52.3

Semiconductor Manufacturing Automation 52.3 Equipment Integration Architecture and Control 917

Robot

PM1

PM2

PM3

PM4

PM5

u vsvsvr u v

svsvr u v

svsvr u vs

Waver delay

Periodicity = 1

rr + 1 r + 2 r + 3

Robot

PM1

PM2

PM3

PM4

PM5

svsvru vsvs

vru vsv

svrvr

svsvru vsvsv

Periodicity = 3

rr + 1 r + 2 r + 3

Robot

PM1

PM2

PM3

PM4

PM5

u v

svsvru vsv

svru vvs

svru v

svsv urv s

rr + 1 r + 2 r + 3

Fig. 52.6a–c Examples of schedules. (a) Steady schedule: a SESS, (b) 3-periodic schedule, (c) irregular schedule

ceding ones complete is called earliest. An earliest

schedule can be generated by the earliest ﬁring rule of

the TEG model that ﬁres each transition as soon as it

is enabled. In other words, an earliest starting sched-

ule need not be generated and stored in advance. The

TEG model with the earliest ﬁring rule can be used as

a real-time scheduler or controller for the tool. There-

fore, an earliest schedule can be implemented by an

event-based control, which initiates a task when an ap-

propriate event, for instance, a task completion, occurs.

Therefore, an earliest starting schedule based on such

event-based control has merits. First, potential logical

errors due to message sequence changes can be pre-

vented. When a tool is controlled by a predetermined

timing schedule, communication or computing delays

may cause a change in a message sequence and a criti-

cal logical error; for instance, a robot may try to unload

a wafer at a chamber before processing at the cham-

ber has been completed and hence when the wafer slot

is still closed. Second, the earliest schedule minimizes

the average tool cycle time, which is the same as the

maximum circuit ratio of the TEG model. Therefore,

the most desirable schedule is a steady and earliest

starting schedule (SESS). For a cluster tool with cyclic

operation, there always exists a SESS. Figure 52.6ais

an example of SESS for the TEG model. A SESS can

be computed in advance using the max-plus algebra

or a kind of longest-path algorithm [52.9] and imple-

mented by an event-based controller based on the TEG

model [52.10,13].

Controlling Wafer Delays

When a tool has a strict constraint on the maximum

wafer delay, as in low-pressure chemical vapor deposi-

tion, coating processes or chemical cleaning processes,

it is important to know whether there exists a feasi-

ble schedule that satisﬁes the constraint. There have

been works on the schedulability of a cluster tool,

that is, the existence of a feasible SESS [52.11, 14].

Lee and Park [52.14] propose a necessary and sufﬁ-

cient condition for schedulability, that is, the existence

of a feasible SESS, based on circuits in an extended

Part F 52.3

918 Part F Industrial Automation

version of TEG called negative event graph, which

models the time-window constraints on wafer delays

by negative places and tokens. In fact, schedulability

can also be veriﬁed by the existence of a feasible so-

lution in an associated linear program. However, the

necessary sufﬁcient condition identiﬁes why the time

constraints are violated, and often gives a closed-form

schedulability condition based on the scheduling pa-

rameters such as the process times, the robot task times,

and the number of parallel chambers for each process

step.

Most schedulability analyses assume deterministic

process and task times. When a cluster tool is oper-

ated by a SESS, the wafer delays are kept constant.

However, in reality, there can be sporadic random dis-

ruptions such as wafer alignment failures and retrials

or exceptional process times. In this case, the sched-

ule is disturbed to a non-SESS, in which the wafer

delays ﬂuctuate and may exceed the speciﬁed limits.

However, there are regulating methods that quickly re-

store a disrupted schedule. Kim and Lee [52.15]propose

a schedule stability condition for which a disrupted ear-

liest ﬁring schedule of a TEG or a cluster tool converges

to the original SESS regardless of the disruptions, and

a simple way of enforcing such stability by adding an

appropriate delay to some selected tasks. Therefore, we

can regulate wafer delays to be constant. Such a sta-

bility control method has been proven to be effective

even when there are persistent time variations of a few

percent [52.15]. Even when the process times or the

robot task times vary signiﬁcantly, but only if they are

within a bounded range, schedulability against wafer

delay constraints can be veriﬁed by an efﬁcient algo-

rithm on an associated graph [52.15]. When the initial

timings are not appropriately controlled or a SESS is

disrupted, the earliest schedule converges to a periodic

schedule whose period is determined from the TEG.

Therefore, the wafer delays can be much larger than the

constant value for a SESS. For a given wafer delay con-

straint, even if the schedulability condition is satisﬁed,

that is, a feasible SESS exists, a periodic schedule may

have wafer delays that exceed the limit. Therefore, we

are concerned with whether such a periodic schedule

with ﬂuctuating wafer delays can satisfy the wafer de-

lay constraint. Lee et al. [52.10] proposed a systematic

method for identifying exact values of task delays of

a TEG or wafer delays of a cluster tool for each type

of schedule: steady or periodic, earliest or not. From

the method, the schedulability of periodic schedules,

which occurs when timings are not well controlled, can

be veriﬁed.

Workload Balancing for Tools

In a traditional ﬂow line or shop, the workload of a pro-

cess step is the sum of the process times of all jobs

for the step. The bottleneck is the process step with the

maximum workload. Imbalance in the workloads of the

process steps causes waiting of jobs or work in progress

before the bottleneck. However, in automated manufac-

turing systems such as cluster tools, the workload is

not easy to deﬁne because the material-handling system

interferes with the job processing cycle. To generalize

the workload deﬁnition, we can deﬁne the generalized

workload for a resource as the circuit ratio for the cir-

cuit in the TEG that corresponds to the work cycle of

the resource [52.10,16]; for instance, the workload for

a chamber at process step i with m

parallel chambers in

a single-armed tool is (p

+2l+2u +3v)/m

, because

each work cycle of a chamber requires a wafer process-

ing (p

), two loading tasks (2l), two unloading tasks

(2u), and three robot moves (3v). A robot has work-

load (n+1)(u +l +2v), the sum of all robot task times.

Therefore, the overall tool cycle time is determined by

the bottleneck resource as

max



max

k=1,2,...,n

+2l+2u +3v

(n +1)(u +l +2v)



Imbalance between the workloads or circuit ratios

causes task delays such as wafer delays. In a single-

armed tool, the workload imbalance between process

step i’s cycle and the whole tool cycle is

max



max

k=1,2,...,n

+2l+2u +3v

(n +1)(u +l +2v)



−

+2l+2u +3v)

Notice that each chamber at process step i has cycle

time (p

+2l+2u +3v), while the overall cycle time

at the process step is (p

+2l+2u +3v)/m

. Therefore,

the delay in each cycle of a chamber at process step i

is m

times as long as the workload imbalance at the

process step. Consequently, the average wafer delay at

a chamber at process step i is [52.10]

max



max

k=1,2,...,n

+2l+2u +3v

(n +1)(u +l +2v)



−(p

+2l+2u +3v) .

We note from the well-known queueing formula, Lit-

tle’s law, that the average delay is proportional to the

Part F 52.3

Semiconductor Manufacturing Automation 52.3 Equipment Integration Architecture and Control 919

average work in progress. In a cluster tool, wafer de-

lays are more important than the number of waiting

wafers because of extreme limitation on the wafer wait-

ing space. Wafer delays can be reduced or eliminated

by balancing the circuit ratios. Such generalized work-

load balancing can be doneby addingparallel chambers

to a bottleneck process step, accommodating the pro-

cess times within technologically feasible ranges or

intentionally delaying some robot tasks [52.10,16]. Lee

et al. [52.10,16] proposed a linear programming model

that optimizes such workload balancing decisions un-

der given restrictions. Workload balancing is essential

for cluster tool engineering.

Additional Works

Cluster toolswith cleaningcycles, multi-slots, and reen-

trance present more challenging scheduling problems.

There are some works on using cyclic scheduling for

these problems [52.4, 17,18]. For a tool controlled by

a dispatching rule, we cannot optimize the rule and

identify or control wafer delays. Wafer delays are un-

expected and can be excessively long. Nonetheless,

dispatching rules are inevitable when the scheduling

problem is too complex or involves uncontrollable sig-

niﬁcant uncertainty. Reentrance, cleaning cycles, and

multi-slots contribute signiﬁcantly to scheduling com-

plexity. In general, process times and robot task times

in cluster tools and track equipment are relatively

well regulated and have variations within a few per-

cent, because most processes are designed to terminate

within a speciﬁed time. However, modern adaptive

process control that adapts process control parame-

ters based on real-time sensor information may cause

signiﬁcant time variation. Cleaning based on cham-

ber conditions may occur randomly and hence increase

uncertainty signiﬁcantly. There are some works on

dispatching rules for cluster tools with cleaning and

multi-slots [52.19].

52.3.3 Control Software Architecture,

Design, and Development

In a cluster tool, each processing module or chamber

is controlled by a process module controller (PMC).

The robot, loadlocks, and slot valves at the chamber

are controlled by a transport module controller (TMC).

A module controller receives data from the sensors in

a chamber, and issues control commands to the ac-

tuators such as gas valves, pumps, and heaters. The

module controllers use bus-type control networks called

ﬁeldbuses such as process ﬁeld bus–decentralized pe-

ripherals (PROFIBUS-DP) and control area networks

(CANs) for communication and control with sensors

and actuators. The module controllers are also coor-

dinated by a system controller, called the cluster tool

controller (CTC). A CTC has a module manager and

a real-time scheduler. A module manager receives es-

sential event messages from the PMCs, manages the

states of the process modules, and sends the PMCs

detailed control commands to perform a scheduling

command from the scheduler. Communication between

the PMCs, TMC,andtheCTC usually uses transmis-

sion control protocol/Internet protocol (TCP/IP) based

on Ethernet because they are well-known and accepted

universal standards.

A real-time scheduler monitors the key events from

each PMC and the TMC through the module man-

ager. The events include starts and completions of

wafer processing or robot tasks, which are essential for

scheduling. Then, the scheduler determines the states of

the modules and scheduling decisions as speciﬁed by

the scheduling logic or rules, and issues the schedul-

ing commands to the module manager. Since the wafer

ﬂow pattern can change, the scheduling logic should be

easily changed without much programming work. The

modules are often conﬁgured by a tool vendor to fulﬁll

a speciﬁc cluster tool order. For large liquid-crystal dis-

play (LCD) fabrication, the modulesare often integrated

at a fab to assemble alarge-scale cluster tool. Therefore,

the scheduler should implement the scheduling logic

in a modular way for ﬂexibility when changing logic.

To do this, the scheduling logic can be implemented

by an extended ﬁnite state machine (EFSM) [52.13].

An EFSM models state change of each module and

embeds a short programming code for the scheduling

logic or procedure. The scheduling logic also includes

procedures for handling exceptions such as wafer align-

ment failures, processing chamber failures, robot arm

failures, etc. Figure 52.7 illustrates a typical architec-

ture for communication and control in a cluster tool.

A tracksystem has a similar communication and control

architecture.

A SEMI standard, cluster tool module commu-

nication (CTMC), speciﬁes a model of distributed

application objects for module controllers and a CTC,

and a messaging standard between the objects [52.20].

Lee et al. [52.21] also propose an object-oriented ap-

plication integration framework based on a high-level

ﬁeldbus communication protocol and service standard,

PROFIBUS-ﬁeld message speciﬁcation (FMS), which

deﬁnes a messaging standard between manufacturing

equipment based on their object models. They sug-

Part F 52.3

920 Part F Industrial Automation

Job edit

Receipe edit

Alarm manipulation

IO value setting

Complete-

wafer

Complete-

wafer

Ready

Disabled

Clean

Process

Prepare

Prepare-

wafer

Prepare-

chamber

Prepare-

chamber

Prepare-

wafer

Robot FSM

Process module

PM2 FSMPM1 FSM

User interface

Job manipulation

Task command scheduling

Exceptional handling

Event reporting

Exceptional handling

Scheduler

PMC

TMC

Mechanical

Events/

messages

Slot valveLoadlock Robot armProcess

Lot manipulation

Lot controller

MC level task decomposition

MC coordination

Module manager

CTC

Unload

Ready

Prepare

Ready PlacePick

Load

ClosedOpen

Ready

Disabled

Pump-

down

Ready

Disabled

Loadlock

Robot arm

Slot valve

Chuck

Fig. 52.7 A cluster tool controller architecture [52.22]

gest that some object models in CTMC, which were

deﬁned based on a traditional object model for material-

handling systems, need to be modiﬁed to handle the

robot tasks in a cluster tool.

Each time a new cluster tool is developed, the

scheduling logic and a CTC application should be in-

tegrated and extensively tested. However, tool testing

and veriﬁcation involve difﬁculties. First, a real tool is

expensiveand hence cannotbe tied upfor extensive test-

ing. Second, testing with a real tool can be hazardous

due to mechanical or space restrictions. Third, since the

dynamics of a real cluster tool is slow, it takes signiﬁ-

cant time to test the system. Finally, it is often difﬁcult

to recognize subtle logical errors by observing opera-

tional behavior of a real tool. Therefore, the CTC and

scheduler need to be tested in a virtual environment

such as a virtual cluster tool (VCT), in which the pro-

cess modules and the transport modules are replaced by

their emulators [52.22]. The emulators receive control

commands from the scheduler through the module man-

ager and/or the module controllers, and create messages

for events such as process completions or robot task

completions at appropriate times. The process times can

be accelerated for initial rough-cut testing. Tool engi-

neers examine the sequence of the events generated at

the CTC or module controllers, and detect an anomaly.

Such veriﬁcation takes several days or weeks and is te-

dious. Some errors are hard to recognize and are often

missed. Joo and Lee [52.22] propose the use of event

sequence ﬁnite statemachines for automaticerror detec-

tion, which is basically identicalto a ﬁnite state machine

except that, when an event other than allowed ones at

a state occurs, an error is assumed. They detected sev-

eral unexpected logical errors, including logical errors

caused by message sequence changes due to communi-

cation delay. Most tool simulators, such as ToolSim by

Brooks Automation, focus on performance evaluation

of a conﬁgured tool rather than high-ﬁdelity model-

ing and veriﬁcation of tool operation and messaging

between a CTC and module controllers.

Part F 52.3

Semiconductor Manufacturing Automation 52.4 Fab Integration Architectures and Operation 921

52.4 Fab Integration Architectures and Operation

52.4.1 Fab Architecture and Automated

Material-Handling Systems

In modern 300mm fabs, wafer cassette-handling tasks

for interbay moves as well as intrabay ones are au-

tomated. In order to save the footprint and secure

human operator access for equipment maintenance or

exception handling, overhead transport (OHT) systems

are mostly used. Traditional automated guided vehicle

(AGV) or rail-guided vehicle (RGV) systems have been

replaced by OHTs. In order to reduce particle contam-

ination risk, tasks of loading and unloading wafer at

tools are automated by using a new wafer carrier, the

front open uniﬁed pod (FOUP), and a standard me-

chanical interface (SMIF). Processing tools are often

enclosed in a minienvironment with extreme cleanness.

Design and operation of the architecture and AMHSs

of such fully automated fabs should be optimized to

maximize throughput and reduce the cycle time while

minimizing capital investment. An AMHS itself can

be a bottleneck due to a limited number of vehicles

and congestion on transport rails. Transport routes are

not so ﬂexible and should be considered as a limited

resource. Therefore, in some 300mm fabs, even crit-

ical metrology steps are skipped in order to reduce

excessive vehicle trafﬁc and the cycle time. Schedul-

ing and dispatching systems are not yet well designed

to handle such fully automated fabs. Control software

such as manufacturing execution systems (MESs), ma-

terial control systems (MCSs), equipment controllers,

and schedulers as well as the AMHS architecture are

Stocker

Empty

OHT

Loaded

OHT

Equipment

Interbay

loop

Intrabay

loop

Fig. 52.8 An overhead transport system

not yet as intelligent and ﬂexible as human operators,

who make adaptive and intelligent decisions depending

on the situation. There still remain many challenges to

smart and efﬁcient fully automated fabs. Figure 52.8 il-

lustrates a typical OHT system layout, which consists

of intrabay and interbay loops. There are works on opti-

mal design of OHT networks, optimal number of OHTs,

and performance analysis [52.23,24].

Automated material-handling systems mostly have

limited handling capacity and ﬂexibility due to re-

stricted paths and limited number of vehicles. There-

fore, stockers or waiting places have been mandatory

solutions for such problems. Stocking wafer cassettes

at a bay involves signiﬁcant delay due to prior wait-

ing cassettes and handling operations. Therefore, in

some 300mm fabs, a desire to minimize the deliv-

ery cycle time led to attempts to combine several bays

into a larger cell by eliminating bay-stockers in or-

der to enforce direct delivery. However, this may cause

signiﬁcant OHT congestion and blocking, and hence

throughput degradation. Nonetheless, direct delivery

is one of the key technological challenges for next-

generation 450mm fabs [52.25]. To achieve the goal of

direct delivery, we need quite different architectures of

fabs and material-transfer systems. A solution might be

to mimic a transfer line or a conveyor system, where

wafer cassettes go through a signiﬁcant number of pro-

cess tools without intermediate stocking. Such a system

is called aninline system. Oneof themost seriousdisad-

vantages of inline systems is lack of ﬂexibility. In future

fabs, lot sizes will continue to shrink. Therefore, con-

Part F 52.4

922 Part F Industrial Automation

ﬂicting goals of ﬂexibility and direct delivery should

be resolved. LCD fabs, for which material transfer has

been fullyautomated from the early stage due to manual

handling difﬁculties, tend to introduce inline systems

for more process steps as the panel size increases con-

tinually. A future 450mm fab may also resemble an

LCD line [52.25]. Stocker racks may be extensively lo-

cated in parallel to the inline system [52.25]. Several

alternatives for future fab and material-handling system

architectures are now being discussed [52.26].

Traditionally, AMHSs have been scheduled and

controlled separately from job scheduling. That is,

wafer processing jobs are scheduled disregarding the

limited capacity of the AMHS, and then material-

transfer tasks, which are requested from the job

schedule executor such as a real-time dispatcher, are

separately planned and controlled by a material control

system (MCS), that is, the AMHS controller. However,

such decoupling is not so effective for modern inte-

grated systems where job scheduling is signiﬁcantly

restricted by the AMHS, and vice versa. Interaction be-

tween job schedulesand material transfercontrol should

be considered, or they should be simultaneously sched-

uled as in cluster tools. MCSs have been engineered

by AMHS vendors and are managed by automation

engineers in fabs. However, job scheduling has been

done by production management or control staffs. In

the future, the two staff groups should better collabo-

rate to tightly couple job scheduling and AMHS control.

As fab technologies evolve, material-handling require-

ments become more challenging. SEMI has updated

a roadmap for AMHSs for future fabs [52.27].

ERP

MES

Cell/system controller

Equipment controllerEquipment controller

MMI (man–machine

interface)

Control

logic

drivers

GEM

SECS-II

HSMS

(TCP/IP)

MMI (man–machine

interface)

Control

logic

drivers

GEM

SECS-II

SECS-I

(RS232)

UI (user interface)

Cell control logic

VFEI (virtual

factory equipment

interfaces)

MES

lib-

rary

SECS-II

HSMS

(TCP/IP)

SECS-I

(RS232)

Open object interface framework

(CORBA/DCOM/OPC-based application object interface)

Fig. 52.9 Communication architecture for fab automation

52.4.2 Communication Architecture

and Networking

SEMI communication standards have been widely used

in fabs to reduce system integration efforts [52.28].

While old tools are connected only by RS-232 ports,

modern tools have Ethernet connections. The semicon-

ductor equipment communication standard I (SECS-I)

and high-speed message standard (HSMS) deﬁne data

standards on RS-232-based serial communication and

TCP/IP communication over Ethernet connection, re-

spectively. SECS-II deﬁnes messaging standards. The

generic equipment model (GEM) and virtual factory

equipment interfaces (VFEI) are object-based appli-

cation interface standards for equipments and factory

control applications, respectively. The overall com-

munication architecture is summarized in Fig.52.9.

AMHSs use ﬁeldbus or control networks, either open

or proprietary.

As advanced process control (APC) technology for

real-time process sensing and real-time adaptive con-

trol becomes widespread, there is increasing demand

on high-speed real-time communication technology, be-

yond the current communication architecture, in order

to process massive process sensing data in real time.

52.4.3 Fab Control Application Integration

The most critical application for factory integration

is a manufacturing execution system (MES). Its ba-

sic functions are to monitor equipment, send recipes,

and keep track of wafers or other auxiliary materials

such as photomasks. Quality monitoring and scheduling

functions tend to be performed by separate appli-

cations from specialized vendors. MES applications

should be easily and reliably integrated with equip-

ment control applications. Traditionally, MESsused

middleware based on message queueing to reliably

process massive event messages from many equip-

ments. No messages should be lost and the response

time should be controlled. Therefore, such messages

from many different tools are queued and the mes-

sage queues are served by reasonable queueing or

service policies for load balancing and response time

control. Such message-based communication and inte-

gration require signiﬁcant application work to integrate

MES applications with equipment control applica-

tions. An application designer should understand all

low-level messages and their required sequence for

logical interaction between the MES and equipment

controllers. Debugging, veriﬁcation, and modiﬁcation

Part F 52.4

Semiconductor Manufacturing Automation 52.4 Fab Integration Architectures and Operation 923

are not easy. An alternative approach is object-

based application integration. Each equipment and an

MES application have a model of constituent ob-

jects, which specify the functions and informational

states. Then, interactions between an MES and equip-

ment are implemented by method calls or service

requests between their corresponding objects. The com-

mon object request broker architecture (CORBA)is

a middleware solution for facilitating application inte-

gration and interaction between such distributed objects

and managing objects and services. MES application

designers can conveniently make use of the high-

level services of the objects in equipment control

applications as well as common MES application ob-

jects. Detailed messaging sequences are handled by

the methods of the objects that provide the relevant

services. SEMI proposed an object-based MES appli-

cation design standard, called the computer-integrated

manufacturing (CIM) framework. SEMI also devel-

oped a standard object model for control applications

of process equipment, called the object-based equip-

ment model (OBEM). There have been concerns about

whether CORBA can work reliably and fast enough

for modern fab environments that generate massive

amounts of real-time data. However, MES vendors

have successfully implemented CORBA-based MES

solutions, for example, IBM’s SiView and AIM Sys-

Application components/objects

CORBA

OBEMOBEM

MES

Basic components/

objects

...

OBEM

(Object

based

equipment

model)

Machine

module

Machine

port

Machine

Equipment

control

applications

Equipment

resource

Equipment

resource

Part

location

Material

location

Process

capability

OBEM/CORBA

Intermediate

components

Physical equipment

User interface

Equipment

resource

Clock Carrier part

Fig. 52.10 Object-based interaction for MES and equipment control applications

tem’s NanoMES. Figure 52.10 illustrates object-based

interaction.

Recently, the service-oriented architecture (SOA)

has been increasingly popular for business and enter-

prise applications [52.29]. Business processes tend to

change frequently to cope with business requirement

changes, and to be distributed over the Internet. There-

fore, more ﬂexibly composable services are deﬁned and

called as needed to form a new business process. Ob-

jects are considered to have too small granularity to

be used for business processes [52.29]. Further, dis-

tributed objects technology such as CORBA and the

distributed component object model (DCOM) are not

easy standards to work with, because it is difﬁcult to

integrate object applications that were developed by

different people at different places on different plat-

forms at different times. Furthermore, CORBA and

DCOM are not widely understood by software en-

gineers and control and automation engineers. Web

services have been open standards for easily integrating

applications distributed on the Internetby using extensi-

ble markuplanguage (XML)-based open standardssuch

as simple object access protocol (SOAP), web services

description languages (WSDLs), and universal descrip-

tion, discovery, and integration (UDDI), and standard

web protocols such as XML, hypertext transfer proto-

col (HTTP), and transmission control protocol/Internet

Part F 52.4

924 Part F Industrial Automation

protocol (TCP/IP). Therefore, SOA based on web ser-

vices can provide open standards for easily integrating

distributed factory applications with proper granular-

ity. Therefore, some fabs or vendors for MESsor

fab management applications are also now considering

SOA-based design. However, it should be studied more

whether SOA really makes sense for factory applica-

tions in terms of reliability and real-time performance.

52.4.4 Fab Control and Management

Fab operation is highly complicated due to the com-

plex process steps and the massive number of lots in

progress. One of the most crucial fab control applica-

tions is a real-time dispatcher that keeps track of the

lots and equipment states, and determines which lots

will be processed at which tools. It uses dispatching

or scheduling rules that are proven to be effective for

fab operation. The rules may be developed and tested

for each fab through extensive simulation in advance.

The essential function of a dispatcher is to process mas-

sive amounts of job and equipment data reliably and

quickly, and compute a dispatch list quickly. A dis-

patcher sends a scheduling command to the MCS and

the process equipment directly in an automated fab,

whereas in a manual fab human operators perform the

job of loading tasks as speciﬁed in the dispatch list.

ERP

Demand planning

Production planning

Scheduler/dispatcher

Equipment management

system

(EEES, e-Diagnostics, ...)

Quality/yield

management system

(SPC, APC, ...)

MES

(Manufacturing execution system)

– WIP tracking, equipment monitoring, command/control

Middleware

– Transaction processing, event handling, recovery

Order promising/

order management

Master planning

Integrated

equipment

controller

Equipment

controller

AMHS

controller

Communi-

cation

network

APS system

(Advanced planning & scheduling)

Factory

operating

system

Fig. 52.11 A fab control system architecture (EEES – engineering

equity extension service, SPC – statistical process control)

An alternative scheduling approach to the dispatch-

ing rules is to have a separate scheduler that determines

an appropriate work-in-progress level for each process

step by using a dynamic lot ﬂow model and then de-

termines an optimal schedule for each process step

separately under the restriction of the ready times and

the due dates that are imposed by the schedule of other

process steps. Frequent rescheduling is needed to cope

with changes in fabs. Even in this case, the dispatcher

retains the basic functions except for the scheduling

function, and may change the schedule from the sched-

uler by local rules depending on the fab state. This

approach has potential for further improving fab per-

formance. However, there should be more experimental

studies on whichapproach ismore effective fordifferent

fab management environments.

A production planning system or supply-chain plan-

ning system determines daily production requirements

for key process stages to meet order due dates or de-

mand forecasts while minimizing inventory level. The

system also considers binning due to random yields

and capacity constraints. Other important fab control

applications include yield management systems and ad-

vanced planning and scheduling (APS) systems. An

overall fab control application architecture is summa-

rized in Fig.52.11.

In spite of extensive literature on fab scheduling,

control, and management, there still remain many is-

sues, including how the dispatching and scheduling

systems, and scheduling rules should be developed

to fulﬁll complex scheduling requirements for fully

automated 300mm fabs or future 450mm fabs, in

which AMHSs will be more strongly coupled with job

scheduling for direct delivery, and lot deﬁnition and job

ﬂows will change signiﬁcantly.

52.4.5 Other Fab Automation Technologies

Fab automation aims at an autonomous factory that re-

liably and intelligently produces high-quality wafers.

As quality requirements have become stricter and

the cost of attaining this quality has increased, fabs

have developed quality-sensitive automation technolo-

gies. Advanced process control (APC) technology

includes fault detection and classiﬁcation (FDC) and

run-to-run (R2R) control [52.30]. FDC makes use

of statistical methods such as multivariate analysis,

or intelligent computing or data-mining technologies

such as neural networks or rules, in order to de-

tect early any anomaly in process control that will

cause signiﬁcant quality problems, classify the prob-

Part F 52.4