Intelligent Preventive Maintenance Scheduling in Semiconductor Manufacturing Fabs SRC/ISMT FORCe:Factory Operations Research Center Task NJ-877 Michael Fu, Director Emmanuel Fernandez Steven I. Marcus Crystal City, VA, December 13-14, 2001 1 SRC/ISMT Factory Operations Research Center CONTENTS •Project Summary - Michael Fu •Optimal Preventive Maintenance Scheduling Model - Emmanuel Fernandez •Generic MIP Model Implementation - Jason Crabtree •Optimal Preventive Maintenance Policy for Unreliable Queueing systems with Applications to Semiconductor Manufacturing Fabs - Xiaodong Yao •“Best Practices” PM Survey -Emmanuel Fernandez •SMITLab and Project Web Page - Jose A. Ramirez •NJ 877 FORCe Kick-off Meeting Presentation 04-26-01 •Papers: •Optimization of Preventive Maintenance Scheduling for Semiconductor Manufacturing Systems: Models and Implementation •A Markov Decision Process Model for Capacity Expansion and Allocation 2 SRC/ISMT Factory Operations Research Center Overview Michael Fu, Ph.D. Professor, School of Business., Univ. Maryland Project Summary 3 SRC/ISMT Factory Operations Research Center Research Plan (1) Develop, test, and transfer software tools for optimal PM planning and scheduling; (2) Research and validate the models, methods and algorithms for software development in (1); (3) Facilitate the transfer of models, algorithms and tools to 3rd party commercial software vendors. 4 SRC/ISMT Factory Operations Research Center Executive Summary • • • • 5 Successful Systems Integration of Software Implementation of PM Scheduling Algorithm at Member Company in 2001 PM Practices Survey Instrument Developed, Distributed, and Preliminary Data Analysis Underway Project Website Up and Running In Progress Research – Generic Implementation of PM Scheduling Algorithm – PM Planning Models (Analytical and Simulation-Based) SRC/ISMT Factory Operations Research Center Project Overview •Review of Project Administration - Industrial Liaisons, Research Personnel - Project Management - Research Progress • Anticipated Results, Deliverables • Task Description, Past Year, Next Year • Technology Transfer • Review/Summary of Research Approach • Research Details • “Best Practices” PM Survey 6 SRC/ISMT Factory Operations Research Center Industrial Liaisons • • • • • • • • • • • 7 Gurshaman S. Baweja, Texas Instruments Incorporated Ben-Rachel Igal, Intel Corporation Mani Janakiram, Intel Corporation Ying Tat Leung, IBM Corporation Marcellus Rainey, Texas Instruments Incorporated Madhav Rangaswami, Intel Corporation Ramesh Rao, National Semiconductor Corporation Man-Yi Tseng, Advanced Micro Devices, Inc. Jan Verhagen, Philips Corporation Sidal Bilgin, LSI Motorola still being decided SRC/ISMT Factory Operations Research Center Research Personnel Faculty: – Michael Fu, Maryland – Steve Marcus, Maryland – Emmanuel Fernandez, Cincinnati Students: – Xiaodong Yao, Maryland (advanced PhD) – Ying He, Maryland (advanced PhD) – Jiaqiao Hu, Maryland (beginning PhD) – Jason Crabtree, Cincinnati (advanced MS) – Jose Ramirez, Cincinnati (beginning PhD) 8 SRC/ISMT Factory Operations Research Center Project Management • Weekly Site Meetings at Maryland & Cincinnati • Weekly Teleconferences UMD & UC • Monthly Teleconferences With Liaisons and PI’s • Project Website – http://www.smitlab.uc.edu/PMResults.htm 9 SRC/ISMT Factory Operations Research Center Anticipated Primary Results (1) Models, algorithms and a suite of software tools will be developed to automate and/or guide in optimally planning, scheduling and coordinating PM tasks for bottleneck tools in a semiconductor fab. (2) Software will be designed for seamless integration with existing commercial discrete-event simulation software packages such as AutoSched AP and WorkStream. 10 SRC/ISMT Factory Operations Research Center Task Description Year 1 - Implementing the PM scheduling algorithm; developing, distributing, and analyzing PM practice survey to drive PM planning models and algorithms; literature review of research on analytical and simulation-based models for PM planning with production considerations. Year 2 - Developing generic implementation platform for PM scheduling algorithm to facilitate possible transfer to 3rd party software provider; developing, testing, and validating PM planning models and algorithms. Year 3 – Implementing PM planning models and algorithms, validating and testing; training workshop to facilitate transfer to 3rd party software vendor. 11 SRC/ISMT Factory Operations Research Center Deliverables to Industry 1. Survey of current PM practices in industry (Report) (P:15-DEC-2001) 2. Models and algorithms to cover bottleneck tool sets in a fab (Report) (P:31-MAR-2002) 3. Simulation engine implemented in commercially available software, with case studies and benchmark data (Report) (P:30-SEP-2002) 4. PM planning/scheduling software tools, with accompanying simulation engine (Software, Report) (P:30-JUN-2003) 5. Installation and evaluation, workshop and consultation (Report) (P:31-DEC-2003) 12 SRC/ISMT Factory Operations Research Center Past Year Progress and Accomplishments • Two student internships at member company: - successful implementation of PM scheduling algorithm - tested and validated with ASAP simulation and real data - integrated with MES and PM monitoring databases • PM Practices Survey Instrument Developed and Distributed - thus far, 12 responses from 8 companies and 5 countries - very diverse answers (more details later) • Investigated analytical (MDP models and queueing models) and simulation-based approaches to PM planning problem • Developing generic implementation of PM Scheduling Algorithm and IT implementation 13 SRC/ISMT Factory Operations Research Center Next Year Plans • Complete implementation of generic PM scheduling algorithm - further student internships at member companies to implement, test, and evaluation models and algorithms - begin facilitating possible transfer to software vendors • Complete PM Practices Survey Data Analysis and Report, - make available to member companies - utilize in research on models and algorithms (see next bullet) • Develop MDP and queueing models, in conjunction with simulation-based approaches, for PM planning problem, focusing on bottleneck tool sets in the fab 14 SRC/ISMT Factory Operations Research Center Technology Transfer • Software Developed – MIP model for PM scheduling algorithm in OSL – ASAP cluster tool model to validate PM scheduling algorithm – Subroutines to integrate various databases • Conference Presentations – INFORMS International Conference, 6/01 – Conference on Control Applications, 9/01 – SRC/ISMT FORCe Annual Review Meeting, 12/01 • Publications – X. Yao, M.C. Fu, S.I. Marcus, and E. Fernandez, “Optimization of Preventive Maintenance Scheduling for Semiconductor Manufacturing Systems: Models and Implementation,'‘ Proceedings of the 2001 IEEE Conference on Control Applications, 407-411. 15 SRC/ISMT Factory Operations Research Center Model Overview Emmanuel Fernandez, Ph.D. Associate Prof., ECECS Dept., Univ. Cincinnati Optimal Preventive Maintenance Scheduling Model 16 SRC/ISMT Factory Operations Research Center Project Background Motivations: 1. Scheduling Preventive Maintenance(PM) tasks for cluster tools is a very complicated and challenging job. 2. Many factors, like manpower constraints, projected upstream and downstream WIP, chamber inter-relationships, etc., just to name a few, come into the decision process. 3. A wealth of information is ready for use from the tool monitoring system (TMS), the wafer dispatching system (WDS), etc. Objective: • To provide a computer-aided decision making support tool for cluster tools’ PM scheduling. • Summer Internships: Xiaodong Yao and Jason Crabtree at AMD from June to August, 2001. Faculty visits to industry. 17 SRC/ISMT Factory Operations Research Center Scheduling Algorithm Summary: 1. A mixed integer program (MIP) has been formulated to address the optimization problem for cluster tools’ Preventive Maintenance (PM) scheduling. 2. Four main decision factors in current PM scheduling practice are identified: chambers status vs. tools throughput, PM windows, manpower constraints, and projected WIPs. 3. Optimization’s objective is to maximize profits via tools’ availability, and meanwhile factor into other costs such as from WIP. 4. The algorithm has been designed to work tightly with other information systems like TMS and WDS. 18 SRC/ISMT Factory Operations Research Center Algorithm kernel • Mathematically, the kernel of our algorithm is a mixed integer program, which has been formulated to address the optimal scheduling problem, and can be solved by using a standard optimization package, e.g. IBM OSL Optimization Subroutine Library. • Obtained as the linear and non-stochastic version of MDP model formulation. • The algorithm will be searching for any feasible consolidation of PM tasks on each tool, choosing the “best” schedule in terms of maximizing tools availabilities (or tools throughputs), and meanwhile satisfying manpower constraints, not exceeding maximal WIP limit and trying to reduce overall WIP level. 19 SRC/ISMT Factory Operations Research Center Research Approach The generic form of the problem of interest: min EC , , Where: • μ is a PM policy; • π is a production policy; • E[C] represents the expected total costs. 20 SRC/ISMT Factory Operations Research Center Proposed Framework 21 SRC/ISMT Factory Operations Research Center Model Statement Objective: i I max bi Vi (t ) Ci I i (t ) Cil ail (t ) a (t ) t 1 i 1 l 1 N M uli ail (t ) 1, for those PM tasks that have to be finished within window wli , uli . Subject to: (1) t wli (2) Vi (t ) f i (a i (t ), i (t )) , for i 1 M ,t 1 N. (3) I i (t 1) I i (t ) K i Vi (t ) di (t ) , for i 1 M , t 1 N 1. (4) I i (t ) Li , for i 1 M , t 1 N . (5) r(a(t ), (t )) R(t ) , for t 1 22 N. SRC/ISMT Factory Operations Research Center Model Nomenclature Let N denote the planning horizon, M the number of tools under consideration. The numbering of PM tasks for each tool is from 1 to i , where i represents the total number of PM tasks for tool i. The variables and parameters of our model are as follows. l ai (t ) : Binary decision variables for PM task l on tool i at time t, (1: to do PM; 0: not to do PM). a (t ) i : Action vector for all tasks on tool i. a (t ) : Action vector for all tasks on all tools. a i (t ) (t ) . They i : Accompanying status vector for PM task vector are used specifically for those PM tasks whose duration are two (t 1) a i (t ) days. Specifically we have i for all i. (t ) : Accompanying vector for all tools. V (t ) i : Availability of tool i at time t. K i : Wafer throughput coefficient for availability of cluster tool i. 23 SRC/ISMT Factory Operations Research Center Model Nomenclature I i (t ) : Workload level (total in queue and process) for tool i at time t. d i (t ) : Workload from upstream operations for tool i at time t. bi : Weighted profit coefficient for availability of cluster tool i. CiI : Cost coefficient for inventory on tool i. C il : PM cost for performing PM tasks l on i. Li : Workload up limit for tool i at any time. f i (a i (t ), i (t )) : Availability function for tool i under PM status vector a i (t ) and i (t ) . r(a (t ), (t )) : Resource function, computing resource required by tasks status vector a (t ) and (t ) . R (t ) : Total resources available for PM at time t. 24 SRC/ISMT Factory Operations Research Center Finding an optimal schedule • An optimal schedule will be chosen with maximal profits from tools availabilities (or throughputs) among all feasible schedules. A feasible schedule should satisfy: • Any scheduled PM tasks should be in its predefined PM window. • On any day, the total number of resources (e.g. headcount of maintenance technicians) required by that day’s schedule should not exceed the projected available resources. • On any day, the WIP level effected by projected incoming WIP and PM schedule on each machine should not exceed a predefined WIP limit, i.e. it will not schedule a PM task on a certain day when large WIP is expecting to arrive. A schedule with low WIP levels is preferred by the algorithm. 25 SRC/ISMT Factory Operations Research Center Simulation Study 26 • A preliminary simulation study of comparison between a model-based schedule and a reference schedule using AutoSched AP has been conducted. Real fab data was used throughout the study. • The simulation study showed model-based schedule outperforms the reference schedule in general. Some significant improvement on tools’ availability was recorded in the study, for example, 14% improvement for a cluster tool was observed in the simulation. SRC/ISMT Factory Operations Research Center Jason Crabtree M.S. Student, Univ. Cincinnati Generic MIP Model Implementation 27 SRC/ISMT Factory Operations Research Center Generic Implementation Objective: To create a generic model and IT implementation based on experience thus far (models, internships, customizations) 28 SRC/ISMT Factory Operations Research Center Generic Implementation • The MIP model is designed to be very robust, thus it can handle a large variety of tools without changing the actual formulation of the model • The generic model can be separated into three facets: – Inputs – Scheduling algorithm – Outputs • Implementation of the generic model then consists of: – Formulating input data and defining interfaces to collect input data – Formatting input data into proper form for chosen MIP solver – Handling of solution data output from solver 29 SRC/ISMT Factory Operations Research Center Inputs and Outputs A set of tools Initial schedule Planning horizon Projected Incoming WIP Optimization Scheduling Model/algorithm Optimized Schedule Estimated Availability Estimated WIP Input Data: • A Tool family, which includes all relevant tools of interest, for example, a group of cluster tools in thin films module • Tool parameters, like throughput rate, maximal WIP level (limit), inventory cost if applicable 30 SRC/ISMT Factory Operations Research Center Input Data (cont.) • Monitored items, i.e. PM tasks of interest on all chambers • Information about PM’s duration (MTTR), manpower requirement, cost etc. • Chambers scenarios and their effects on whole tool’s throughput (availability) • Planning horizon • PM tasks initial schedule, each associated with a time window (warn-date, due-date and late-date) • Projected incoming WIP in the planning horizon • Projected manpower (maintenance technicians) available in the planning horizon 31 SRC/ISMT Factory Operations Research Center Outputs • Optimal schedule, users can see the optimized schedule and the initial one in TMS for comparison. • An estimated availability of each tool in family will be presented along with an optimal schedule. • An estimated inventory level of each tool in family will be presented along with an optimal schedule. 32 SRC/ISMT Factory Operations Research Center Algorithm Flow Chart Begin .ini file; .tool file; .item file; System Initialization and selecting a machine Family Specifying a planning horizon TMS database Dispatch report Resource Data File Reading in TMS database, performing data filtering Reading in projected WIP from WDS Reading in projected resource Generating consolidated tasks vector set {v} Computing availability loss and resources requirement for each task vector. Generating MIP model instance in a standard format Chamber configuration SIMPLEX and Branch-and-Bound algorithms are used in the default solver Invoking OSL default solver to solve the MIP model Parsing model solution and interpreting the result to users TMS End 33 SRC/ISMT Factory Operations Research Center System Architecture RPC(Remote Procedures Call) Connection Data Preprocessing Model generating model.mps TMS WDS 34 Solution Parsing Front-end Optimization Solver (OSL) Back-end SRC/ISMT Factory Operations Research Center System Interfaces User Interface: • Tool family is selected by user. • Planning period dates are entered by user. • Manpower schedule is input by user. Interface with TMS: • The TMS PM data file is searched by the software for PM items due within user-defined planning window. • A secondary “optimal” PM data file is updated by the software with the optimal dates from the optimization. Interface with WDS: • A projected WIP data file is extracted from WDS periodically. • The software will use updated WIP data each time when running the optimization algorithm. 35 SRC/ISMT Factory Operations Research Center Using the Software Begin Select Tool Group Confirm PM Items Enter Planning Dates LP Solved On Remote Server Enter Manpower Schedule Run Optimization Confirm Optimal Schedule Update PM Records End 36 SRC/ISMT Factory Operations Research Center Conclusions • The MIP model is very robust and can handle a wide range of tools • Implementation of the generic model consists of formulating and gathering the required input data, passing data to solver, and handling the solution data • The complexity of the final software package is currently being looked into. Keeping the software small promotes speed and allows customization but requires that software be tailored for each implementation. A larger software package enhances ease of integration but may limit robustness of software. 37 SRC/ISMT Factory Operations Research Center Xiaodong Yao Ph.D. Student, Univ. Maryland Optimal Preventive Maintenance Policy for Unreliable Queueing systems with Applications to Semiconductor Manufacturing Fabs 38 SRC/ISMT Factory Operations Research Center Modeling Framework for PM planning and scheduling Two-stage hierarchical framework under development: • • 39 In the 1st stage (higher level), the objective is to investigate the structure of optimal policies for PMs, (for example, PM window with optimal parameters t* and Dt*), in the presence of stochastic dynamics of machine failure and demand processes. In the 2nd stage (lower level), the objective is to determine exact time to perform PM tasks, taking into consideration the inter-dependency of PMs, resource constraints, short-term projected WIP, etc. A mixed integer program (MIP) has been successfully developed, and implemented in a specific environment. SRC/ISMT Factory Operations Research Center Model (Semi-)Markov Decision Processes (MDP) model on 1st stage: • • • • • • • 40 Machines: queueing systems, fab: queueing network Incorporating system “operating” states (e.g. WIP level, queue length) and “technical” states (e.g. deterioration degree of machines) in PM policy Observable vs. unobservable machine states. Unobservable states can be estimated based on information from SPC (statistic process control) or number of wafers produced since last “renewal” state. Stochastic process for machine’s deterioration, e.g., controlled Markov chain. Independent demand process Random times (non-negligible) for PM tasks Cost structure: PM cost, operating cost, and inventory holding cost, etc. SRC/ISMT Factory Operations Research Center Structural Optimal Policy Structure of optimal policy: • • • • • 41 To be investigated with the formulation of MDP models Monotonicity, e.g. control-limit form and/or switching curve, is promising and interesting Effect of system “operating” states on PM policy to be studied. Comparison study between the derived optimal policy and (t*, Dt*) policy Derivation of parameterized policy SRC/ISMT Factory Operations Research Center Simulation-based Optimization Simulation-based optimization for parameterized policy: • • • 42 Monte Carlo simulation is effective for large optimization problems. Efficient gradient estimation will be employed in stochastic approximation algorithm in search for optimal policy parameters, e.g. threshold values in control-limit policies. Unbiased gradient estimators for system performance w.r.t. structural policy parameters, e.g. (t*, Dt*) will be obtained. SRC/ISMT Factory Operations Research Center Emmanuel Fernandez, Ph.D. Associate Prof., ECECS Dept., Univ. Cincinnati “Best Practices” PM Survey •PM Practices Survey Instrument Developed and Distributed •Thus far, 12 responses from 8 companies and 5 countries •Both commonality and divergence in different issues 43 SRC/ISMT Factory Operations Research Center Jose A. Ramirez Ph.D. Student, Univ. Cincinnati SMITLab and Project Web Page http://www.smitlab.uc.edu 44 SRC/ISMT Factory Operations Research Center