Daniel E. Rivera

Control Systems Engineering LaboratoryDepartment of Chemical and Materials Engineering

Ira A. Fulton School of EngineeringArizona State University

daniel.rivera@asu.edu

Inventory Management in Semiconductor Manufacturing Supply Chains (and Beyond):

Insights Gained from a Process Control Perspective

About the Presenter

• Born and raised in San Juan, Puerto Rico

• Education– B.S. ChE degree from the

University of Rochester (1982)– M.S. ChE degree from the

University of Wisconsin (1984)– Ph.D. from Caltech (1987)

• Positions:– Associate Research Engineer,

Shell Development Company, Houston, TX (1987-1990)

– Associate Professor, Arizona State University, (1990 - present)

Control Systems Engineering Laboratory Projects

• Chemical Process Control.• American Chemical Society-Petroleum Research Fund: “Constrained

Multisine Inputs for Plant-Friendly Identification of Chemical Processes”

• Honeywell Intl. Foundation: “Control Systems Engineering Laboratory”

• Supply Chain Management.• National Science Foundation: “GOALI: Process Control Approaches to

Supply Chain Management in Semiconductor Manufacturing”

• Intel Research Council:“Supply Chain Management Research Using Process Control Approaches”“Improving Short-term Demand Forecasting in Supply Chain Management”

• Behavioral Health.• NIH-NIDA (subcon via Penn State): “Control Engineering Approaches

to Adaptive, Time-Varying Interventions in Drug Abuse Prevention”

http://www.fulton.asu.edu/~csel

Presentation Outline

• Control engineering basics review

• Supply Chain Management (SCM) as a process control problem

• Application to SCM in semiconductor manufacturing

• Adaptive interventions in drug abuse prevention

• Summary and conclusions

What to take with you from this talk

• The transfer of variance from a valuable system resource to a less expensive one is an important outcome of well-designed control systems, in any application setting.

• Both feedback and feedforward strategies are needed in the design of effective control systems for delayed, nonlinear, stochastic plants.

• Process control ideas have significant application in diverse problem settings, for example:

– supply chain management for semiconductor manufacturing, and

– adaptive interventions in behavioral health

• Prepare yourself for life-long learning, since you may very well work on problems you have never imagined (in a not-too-distant future).

Control Engineering

• Control engineering is a broadly-applicable field that spans all areas of engineering:

– Chemical– Electrical– Mechanical and Aerospace– Civil / Construction– Industrial– Biomedical– Computer Science and Engineering

• Control engineering principles play a role in everyday life activities.

Control Engineering (Continued)

Considers how to manipulate or adjust system variables so that its behavior over time is transformed from undesirable to desirable,

• Open-loop: refers to system behavior without a controller or decision rules (i.e., MANUAL operation).

• Closed-loop: refers to system behavior once a controller or decision rule is implemented (i.e., AUTOmatic operation).

Open-Loop (Manual) vs. Closed-Loop (Automatic) Control

Open-Loop “Manual” Closed-Loop “Automatic”

An Improved Closed-Loop System(Dual Climate Control)

An Industrial Process Control Problem

QuickTime™ and aBMP decompressor

are needed to see this picture.

Objective: Use fuel gas flow to keep outlet temperature under control, in spite of occasional yet significant changes in the feed flowrate.

The “Shower” Control Problem

Hot Cold

Controlled: Temperature, Total Water Flow

Manipulated: Hot and ColdWater Valve Positions

Disturbances:Inlet Water Flows,Temperatures

The presence of delay or “transportation lag”

makes this a difficult controlproblem

Feedback and Feedforward Control Strategies

• In feedback control strategies, a controlled variable (y) is examined and compared to a reference value or setpoint (r). The controller issues actions (decisions on the values of a manipulated variable (u)) on the basis of the discrepancy between y and r (e = r - y, the control error).

• In feedforward control, changes in a disturbance variable (d) are monitored and the manipulated variable (u) is chosen to counteract anticipated changes in y as a result of d.

Hot Cold

Controlled: Temperature, Total Water Flow

Manipulated:Hot and ColdWater ValvePositions

Disturbances:Inlet Water Flows,Temperatures

Controller

F TTemp. setpoint

Actuators

Flow setpoint

Sensors

Shower Problem: Automatic Feedback Control

Closed-Loop Feedback Control “Block Diagram”

C = ControllerP = Plant Model/“Transfer Function”Pd = Disturbance Model/“Transfer Function”

Controlled:Temperature,

Total Water Flow

Manipulated: Hot and ColdWater Valve

Positions

Disturbances:Inlet Water Flows,

TemperaturesReference:

Desired Temperature, Total Water Flow

C

+

r ec = r - ym u

d

n

y

P-++

Pd

ym

sensor noise

-20

-10

0

10

20

0 500 1000 1500 2000 2500 3000 3500 4000

Measured Output

Time[Min]

-10

0

10

0 500 1000 1500 2000 2500 3000 3500 4000

Input

Time[Min]

Open-Loop(Before Control)

Closed-LoopControl

Temperature Deviation(Measured Controlled Variable)

Hot Water Valve

Adjustment (Manipulated

Variable)

From Open-Loop Operation to Closed-Loop Control

The transfer of variance from an expensive resource to a cheaper one is one of the major benefits of engineering process

control

Supply Chain Management

• A supply chains consist of interconnected entities (e.g., factories, warehouses, and retailers) which transform ideas and raw materials into delivered products and services

F

Factory

W

Warehouse

R

Retailer

Motivation

• In the modern economy, products do not simply compete against other products; supply chains compete against other supply chains.

• Billions of dollars in potential savings by eliminating supply chain inefficiencies (PriceWaterHouseCoopers, 2000; Kempf, 2004)

• An effective SCM system can– Improve an enterprise’s agility to respond to market upturns

(and downturns) – Increase revenue while reducing manufacturing and transportation

costs.– Eliminate excess inventories and reduce safety stocks– Lower lead times and improve customer satisfaction

The Business Literature Can Inspire a Control Engineering Approach

• The “bullwhip” effect (Lee et al., "Information Distortion in a Supply Chain: The Bullwhip Effect", Management Science 43(4) 546, 1997); demand distortion caused by variance amplification of orders upstream in the supply chain

• This and similar terminology highlight issues relating to stability and performance of a dynamical system, which merit a control-oriented approach.

• Not strictly an engineering/scientific problem: financial, organizational, and social issues come into play in this problem.

“Bullwhip” Effect

Supply Chain Inventory Management as a “Level” Control Problem

LT

ORDER DECISIONS/STARTS

Demand

Meet demand (with forecast possibly given f days beforehand) for a node with day production (or order fulfillment) time and d delivery time.

CTLproduction time; also known as throughput time)

d delivery time)(Disturbance)

Starts (Manipulated)

Net Stock (Controlled)

Feedback-Only Inventory Control Problem

LT

CTL

Demand

In the feedback-only control problem, ordering decisions are calculated based only on perceived changes to “level” (e.g., net stock or equivalent variable).

(Disturbance)

Starts (Manipulated)

Net Stock (Controlled)

production time)

d delivery time)

Single Node Inventory Problem Combined Feedback/Feedforward Control

LT

LIC

Demand

Demand Forecast(known f days beforehand)

production time)

d delivery time)

(Disturbance)

Starts (Manipulated)

Net Stock (Controlled)

In the combined feedback/feedforward problem, a demand forecast is used for feedforward compensation.

3DoF Internal Model Control Results(random unforecasted demand at t = 90)

Feedback-only Combined FB/FF

f = 20, d = 2, f = 1, r = 1, d = 1, nr=1, nd=3, nff=2

The ASU-Intel SCM Project TeamInvolves multiple faculty and graduate students from various departments

in Engineering and Mathematics

• Dept. of Mathematics, CLAS:– Professors Dieter Armbruster, Matthias Kawski, Christian Ringhofer and Hans

Mittelmann; Eric Gehrig (Ph.D. student), Dominique Perdaen, Ton Geubbels (Visiting Researchers from TU-Eindhoven, The Netherlands).

• Chemical Engineering, Fulton School:– Prof. Daniel E. Rivera; Wenlin Wang and Jay D. Schwartz (Ph.D. students),

Michael D. Pew (UG student), and Asun Zafra Cabeza (Visiting Researcher from the University of Seville, Spain)

• Computer Science and Engineering, Fulton School– Prof. Hessam Sarjoughian; Donna Huang and Weilong.Hu (Ph.D. students)

• Intel collaborators:– Karl G. Kempf, Kirk D. Smith, Gary Godding, John Bean, Mike O’Brien

Proposed Architecture

strategicplanning

inventoryplanning

tacticalexecution

simulation

The Outer LoopProblem

The Inner Loop Problem

Validation

Prediction

goals

goals

limits

Semiconductor Manufacturing Process

Fluid Analogy for Single Fab/Test1, Assembly/Test2 and Finish Nodes

Modeling Issues and Challenges

• The manufacturing process displays long throughput times (TPT) which are stochastic and nonlinearly dependent on load

• Yields are also stochastic

• There is an error between the forecasted and actual demand, which is also stochastic

• Additional problem features include package dynamics, stochastic splits in die properties, and multi-factory issues involving cross-shipments, shared capacity, and correlated demands.

Fab/Test Manufacturing Node Dynamics

Load

Ou

tsS

tart

sL

oad

Time

Th

rou

gh

pu

t T

i me

(Inventory Levels,WIP)

(Actual Demand)

(Future Starts)

(Forecasted Demand)

(Previous Starts)

Model Predictive Control

Model Predictive Control Advantages

• Ability to handle large multivariable systems

• Ability to enforce constraints on manipulated and controlled variables

• Effective integration of feedback, feedforward controller modes; ability to incorporate anticipation

• Novel formulations (such as hybrid MPC) enable the application to systems involving both discrete-event and continuous variables.

Case Study: Assembly/Test2 Stochastic Split Problem

The outcome of the Assembly/Test2 process is stochastic in terms of the number of fast and slow devices that result.

Fast devices can be used to make high speed products (C37). Slow devices can be used to make low speed products (C39).

C35

C36

I10

C37

I20

C39

I21

E1E2E3D1D2D3

C38

M10

M20

M30M30 M30

C41

I31

M40C40

I30

M40

C90

Slow devices

Fast devices

X

Num

ber

of D

ie

Speed

Fab/Test1

Assmbly/Test2

Fin/Pack

Case 2: No Move SuppressionA/T2 Load

Finishing Load

F/T1 Starts

Reconfiguration Starts

CW (Fast)

CW (Slow)

Case 2: With Move Suppression [10 10 10 0 10]

A/T2 Load

Finishing Load

F/T1 Starts CW (Fast)

CW (Slow)Reconfiguration Starts

69.3% variance reduction

51.5% variance reduction

98.9% variance reduction

43.2% variance reduction

4.7% variance reduction

Customer Service Comparison

No Move Suppression With Move Suppression

Fas

t D

evic

e B

ackl

og

Slo

w D

evic

e B

ackl

og

Slo

w D

evic

e B

ackl

og

Fas

t D

evic

e B

ackl

og

Unfilled Orders: 0.34%

Unfilled Orders: 2.38%

Unfilled Orders: 7.41%

Unfilled Orders: 4.62%

C36

C40

I10

C45

I20

C43

I21

E1E2E3D1D2D3

C44

M10

M20

M30M30 M30

C50

I31

M40C49

I30

M40

C90

C37

C41

I11

C46

I23

C48

I22

D1D2D3E1E2E3

C47

M11

M21

M31M31 M31

C52

I31

M41C51

I33

M41

C91

C38M50

I51

C35M51

I50

C39 C42

“Combination” Problem

E

vendor1

vendor2

Fab2

P1,P2

Fab1

P1

Fab3

P2

T1-3

P2

T1-2

P1,P2

T1-1

P1

Asm3

Asm2

Asm1

vendor3 vendor4

vendor5 vendor6

1.1 si

1.2 si

2.1

2.2

2.3

3.1

3.2

3.4 pp3.3 pp

3.7

3.8 pp 3.9 ram

3.10 3.11

3.12

3.5

3.6

3.5

A

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

26

27

24

25

28

29

30

31

32

Box1

7.2

7.1

Box27.6

7.5

F

43

7.4

pp

45

44 vend8

46

7.3

pp

T2-3

T2-1

T2-2 Fin2

Fin1

4.3

5.2

5.133

34

ven

d7

4.1

4.2

6.2

6.1

6.3

C

B

D

35

36

37

38

39

40

41

42

Blue = IntelRed = Mat. Sub.

Green = Cap. Sub.

= Inv Hold

= Prod Mfg= Transport

= Mats Mfg

A “Small” Semiconductor Mfg Problem

Adaptive Interventions

• Adaptive interventions individualize therapy by the use of decision rules for how the therapy level and type should vary according to measures of adherence, treatment burden and response collected during past treatment.

• Adaptive interventions represent an important emerging paradigm for prevention and treatment of chronic, relapsing disorders, such as drug and alcohol abuse, depression, hypertension, obesity, and many other maladies.

• Also known as stepped care models, dynamic treatment regimes, structured treatment interruptions, and treatment algorithms.

• Based on the Fast Track Program (a multi-year intervention designed to prevent conduct disorders in at-risk children).

• Parental function (the tailoring variable) is used to determine the frequency of home visits (intervention dosage) according to the following decision rules:

- If parental function is “low” the intervention dosage should correspond to weekly home visits,

- If parental function is “average” then intervention dosage should correspond to bi-weekly home visits,

- If parental function is “high” then intervention dosage should correspond to monthly home visits.

Home Counseling-Parental Function Intervention

If PF(t) is “Low” then Weekly Home Visits If PF(t) is “Medium” then Bi-Weekly VisitsIf PF(t) is “High” then Monthly Home VisitsIf PF(t) is “Acceptable” then No Visits

Decision Rules

Clinical Judgment

Intervention

I(t)Process

Disturbances

++

Tailoring VariableEstimation

Reliability/MeasurementError

+

+

Goal

ReviewInterval

Estimated Parental Function PF(t)

Outcomes

Parental Function Feedback Loop Block Diagram*(to decide on home visits for families with at risk children)

*Based on material from Collins, Murphy, and Bierman, “A Conceptual Framework for Adaptive Preventive Interventions,” Prevention Science, 2004.

Parental Function Feedback-Only Control Problem

LT

CTL

Depletion

In the feedback-only control problem, intervention dosages are calculated based only on perceived changes to “inventory” (parental function PF(t)).

D(t) (Disturbance)

I(t) (Manipulated)

PF(t) (Controlled)

PF Factor

0102030405060708090

100

1 11 21 31Months

PF LevelPFFactorGoal

Recommended Intervention Dosage

0

1

2

3

1 11 21 31Months

Intervention Dosage

Summary and Conclusions

• The transfer of variance from a valuable system resource to a less expensive one is an important outcome of a well-designed control system, in any application setting.

• Both feedback and feedforward strategies are needed in the design of effective control systems for delayed, nonlinear, stochastic plants.

• Process control ideas have significant application in diverse problem settings, for example:

– supply chain management for semiconductor manufacturing, and

– adaptive interventions in behavioral health

• Prepare yourself for life-long learning, since you may very well work on problems you never imagined (in a not-too-distant future).