Daniel E. Rivera Control Systems Engineering Laboratory Department of Chemical and Materials Engineering Ira A. Fulton School of Engineering Arizona State University 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 ( ) –Associate Professor, Arizona State University, ( 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”

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 a BMP 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 Controlled: Temperature, Total Water Flow Manipulated: Hot and Cold Water Valve Positions Disturbances: Inlet Water Flows, Temperatures The presence of delay or “transportation lag” makes this a difficult control problem

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.

Ho t Cold Controlled: Temperature, Total Water Flow Manipulated: Hot and Cold Water Valve Positions Disturbances: Inlet Water Flows, Temperatures Controller F T Temp. setpoint Actuators Flow setpoint Sensors Shower Problem: Automatic Feedback Control

Closed-Loop Feedback Control “Block Diagram” C = Controller P = Plant Model/“Transfer Function” Pd = Disturbance Model/“Transfer Function” Controlled: Temperature, Total Water Flow Manipulated: Hot and Cold Water Valve Positions Disturbances: Inlet Water Flows, Temperatures Reference: Desired Temperature, Total Water Flow C + r e c = r - y m u d n y P - ++ PdPd ymym sensor noise

Open-Loop (Before Control) Closed-Loop Control 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. CTL  production 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, n r =1, n d =3, n ff =2

The ASU-Intel SCM Project Team Involves 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 strategic planning inventory planning tactical execution simulation The Outer Loop Problem The Inner Loop Problem Validation Prediction 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 Outs Starts Load Time Throughput Time

(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 E1 E2 E3 D1 D2 D3 C38 M10 M20 M30 C41 I31 M40 C40 I30 M40 C90 Slow devices Fast devices X Number of Die Speed Fab/Test1 Assmbly/Test2 Fin/Pack

Case 2: No Move Suppression A/T2 Load Finishing Load F/T1 Starts Reconfiguration Starts CW (Fast) CW (Slow)

Case 2: With Move Suppression [ ] A/T2 Load Finishing Load F/T1 StartsCW (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 SuppressionWith Move Suppression Fast Device Backlog Slow Device Backlog Fast Device Backlog Unfilled Orders: 0.34% Unfilled Orders: 2.38% Unfilled Orders: 7.41% Unfilled Orders: 4.62%

C36 C40 I10 C45 I20 C43 I21 E1 E2 E3 D1 D2 D3 C44 M10 M20 M30 C50 I31 M40 C49 I30 M40 C90 C37 C41 I11 C46 I23 C48 I22 D1 D2 D3 E1 E2 E3 C47 M11 M21 M31 C52 I31 M41 C51 I33 M41 C91 C38 M50 I51 C35 M51 I50 C39 C42 “Combination” Problem

E Blue = Intel Red = 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 Visits If PF(t) is “High” then Monthly Home Visits If PF(t) is “Acceptable” then No Visits Decision Rules Clinical Judgment Intervention I(t) Process Disturbances + + Tailoring Variable Estimation Reliability/ Measurement Error + + Goal Review Interval 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)

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).