Transcript Slide 1
GOALI: Process Control Approaches to Supply Chain Management
in Semiconductor Manufacturing
NSF Grant Number: DMI- 0432439
D. E. Rivera
H.D. Mittelmann
K.G. Kempf
Department of Mathematics and Statistics
Decision Technologies
College of Liberal Arts and Science
Intel Corporation, Chandler, AZ
H.S. Sarjoughian
Control Systems Engineering Laboratory
Arizona Center for Integrative M&S
Department of Chemical and Materials Engineering
Department of Computer Science and Engineering
Ira A. Fulton School of Engineering
Arizona State University, Tempe, AZ
Arizona State University, Tempe, AZ
Investigate the use of approaches based
on process control for mid-level planning in
semiconductor manufacturing supply chain
networks, under simulated settings.
Approach:
•Develop strategies using Model Predictive
Control (MPC) contextualized for the
requirements of semiconductor
manufacturing supply chains.
•Develop optimization and process
modeling techniques that support
combined tactical and strategic decisionmaking for SCM.
•Develop a prototype
simulation/optimization environment
supporting discrete process flow and
optimization decision models.
decision –
configure,
pack, ship,
where &
when
Fabrication Starts (a control point)
Fab/Test1 (a manufacturing system)
Test1 Outs and Transport
C1
M10
M20
test1
test2
pack
I20
11
2.1
vendor1
1
1.1 si 2
3
Finish Starts (a control point)
Finish (a manufacturing system)
Finish Outs and Transport
M30
finish
assembly
Components Warehouse (an inventory storage)
I30
- variable tpt
faster
good slower
- die
not good
- variable tpt
faster
good slower
- product
not good
- variable
demand
(volume and
time)
Fab
2
Demand (over time)
D3
•Originates from the chemical process industry
•Optimization-based receding horizon algorithm
•Readily incorporates hard and soft constraints
•Robustness to model mismatch and uncertainty
achieved through proper choice of tuning
D2
D1
t
(Forecasted Demand)
T1-2
14
2.3
Fab
3
9
T1-3
10
P2
6.2
41
3.6
Asm
2
7.
3
7.4
pp
42
46
vend
p 8
45
7.5
p
38
7.6
29
T2-2
4.2
3.5
3.7
22
26
23
27
44
5.2
36
Fin2
Box
2
F
16
vendo
r5
3.8 pp
17
Blue = Intel
Red = Material
Subcontractor
Green = Capacity
Subcontractor
vendor
6
3.9 ram
3.10 3.11
3.12 Asm
3
31
32
D
T2-3
4.3
6.3
Decision Algorithm
Validation
inventory
planning
simulation
limits
Decision
Model
Decision
Execution
Engine
Projection
Model
Projection
Execution
Engine
Prediction
tactical
execution
Fab/Test1 Node Response
(Previous Starts)
The Inner Loop Problem
(Future Starts)
40
35
3.5
6
vendor2
6.1
7.2
43
30
strategic
planning
(Inventory Levels,
WIP)
(Actual Demand)
37
C
vendor4
25
19
P1,P
2
15
P1,P
2
5
13
Box
1
Design and Implementation
Approach
The Outer Loop Problem
goals
Modeling Challenges:
8
3.4 pp
vendor3
21
12
P2
M40
Model Predictive Control:
P1
Fin1
39
24
3.3 pp
T1-1
A
4
1.2 si
7
P1
Shipment (a control point)
C4
goals
•Long lead times with nonlinear
dependence on load
•Stochastic throughput time and yield
•Stochastic and potentially erroneous
customer demand
Fab
1
2.2
fabrication
5.1
34
3.2
20
Semi-Finished Inv (an inventory storage)
C3
33
4.1
B
Assembly Starts (a control point)
Assm/Test2 (a manufacturing system)
Test2 Outs and Transport
C2
T2-1
28
7.1
18
Die/Package Inventory (an inventory storage)
I10
Asm
1
3.1
vend
7
Research Objectives:
decision – how
many wafers to
start into which
factory when
decision – how
many of which die
to put into which
packages in which
factory when
A Representative
Semiconductor
Mfg Network
System
Interoperability
To efficiently manage a large-scale
supply chain in semiconductor
manufacturing using hierarchical decision
and simulation techniques.
The Semiconductor Mfg Process
A Fluid Approximation of A Simple
Semiconductor Mfg Network
Model
Composability
Problem:
Projection Algorithm
• Model composability refers to composition of models – e.g.,
Linear Programming Decision Model and Discrete Event
System Simulation Projection Model
• System interoperability refers to the interoperation of execution
engines – e.g., ILOG Solver and DEVSJAVA Simulator
Semiconductor Supply-Chain Network System
MPC calculation steps:
Step 1: Prediction of future inventories using a state estimator based
on a nominal model of the supply chain, and relying on current and
previous information of the system variables (inventories, customer
demand, forecasts and starts). An adjustable filter gain (fa) is selected to
tune for system stochasticity and uncertainty.
Customer Demand
Step 2: Optimization of current and future starts according to the
following objective function:
Monthly Projection
(weekly buckets)
Strategic
(Linear Programming)
Decision
Tactical
(Model Predictive Control)
Process Flow
Weekly Projection
(daily buckets)
Experimental Configuration
• Performance consideration
computing node
– data size (bytes)
– complexity of computation (possibly with large iteration)
– network communications (bits per second)
MPC
LP
DEVSJAVA
Data
Storage
Network
Physical Process Flow
(daily)
• large data sets
• medium/large data sets
• large data sets
• complex computation
• light computation
• light computation
• large data sets
• complex computation
Current and Future Efforts:
Case Study
Sinusoidal demand and stationary forecast: fa=0.01
• Examined a three-node semiconductor
manufacturing network composed of Fab/Sort,
Assembly/Test and Finish/Pack nodes.
•Develop a software design and implementation
enabling distributed simulation and optimization
model executions
• Evaluated the effect of demand anticipation and filter
gain selection on problems involving erroneous
forecasts and periodic demand.
•Explore computational efficiency and tuning of the
MPC decision algorithms using:
•Model Predictive Control offers a flexible framework for
achieving operational goals under conditions of
nonlinearity, stochasticity, forecast error, and
uncertainty.
Stationary demand and forecast: fa=0
•Examine the interaction and optimal configuration
between the inner and outer loops that define the
strategic and tactical plans.
•Simultaneous Perturbation Stochastic
Approximation (SPSA) search methods and
•Bilevel Nonlinear Programming
Erroneous demand and forecast: fa=0.04
Sinusoidal demand and forecast: fa=0.01
•Develop multi-modeling scheme and distributed simulation for
integrating DEVSJAVA, Matlab, and OPL Studio environments
•Enable simulation of complex process flows, decision making,
and their interactions for a representative Intel semiconductor
supply-chain network system
Significant Results:
•A novel prototype MPC-based algorithm for semiconductor
mfg supply chains has been demonstrated on benchmark
problems.
•An approach for synthesizing discrete-event simulation
models and MPC models has been prototyped.
Broader Impact:
The MPC-based control and simulation framework will offer
new ways towards hierarchical decision-making in large-scale
enterprise systems for discrete-parts manufacturing.
•Establish the conceptual basis for the efficient,
scaleable solution of the representative
semiconductor mfg. network presented above.
For more information regarding this project, contact
D.E. Rivera at [email protected] or visit our website at
http://www.fulton.asu.edu/~csel/Publications.htm