Maryland's Factory Sciences Research Project
Informal Update - July 1, 1998
Contact: Michael Fu (mfu@rhsmith.umd.edu)

This report covers the period from the contract start (October 1) through June 30, 1998.
 

Contents:

1. Virtual Center Interactions

2. Current SRC Mentor List

3. Activities

4. Research Tasks Progress Report
 

1. Virtual Center Interactions

Collaboration within the Factory Sciences Virtual Center included the following:

2. SRC Mentors:

Our set of mentors for the period was the following:
 
Edward Rietman  Lucent 
Lance Solomon  Intel 
Lori Jones  TI 
Vic Palmeri  AMD 
Gerald Feigin  IBM – left IBM in May
Laurie Goldstein  Motorola – assigned in June

3. Activities

Project-related activities of team members included the following:  

4. Research Tasks Progress Report

Task Title: Markov Decision Processes for Integrating Life Cycle Dynamics into
Fab-Level Decision Making (task leaders, Drs. Steven Marcus and Michael Fu)

Summary:

In the planning and scheduling of a semiconductor manufacturing fab, there has been relatively little research that takes into account fab life cycle dynamics.  We propose a finite-horizon transient Markov decision process (MDP) model that integrates product life cycle dynamics and that will provide decision support on such critical operational issues as when to add additional capacity and when to convert from one type of production to another.  One specific case of dynamics to be treated in this research is that of technology shrink: a technology generation typically involves several evolutionary versions, starting with a more relaxed one to bring up manufacturing yield and proceeding through incremental --- though demanding --- enhancements  (“shrinks”) that increase product performance and market value.  This business requirement, driven by time-dependent price changes for each shrink and the criticality of time-to-market with the next version, drives crucial operational challenges: how to devote --- in an environment of substantial risk --- valuable resources (tool and process capacity) to bring-up of the next shrink while maintaining the existing product's yield and profitability.

Status:

One postdoctoral fellow, Dr. Shalabh Bhatnagar, and one graduate student, Ying He, have been working with the task leaders on this project.  A preliminary draft of the working MDP model formulation has been circulated to interested parties at IBM, Motorola, and Texas Instruments, with feedback received from IBM thus far. (We would be happy to distribute a copy of this draft, about 10 pages in length, to any interested parties in SRC.)  In parallel, we are investigating computationally efficient methodologies for numerically solving large-scale MDPs, including recent literature in neuro-dynamic programming.  Some of the resulting work will be presented at the INFORMS Seattle meeting this coming October.
 

Title: Response Surface Models for Incorporating Unit Processes into Fab-Level Operational Decision Making (task leaders, Drs. Jeffrey Herrmann and Gary Rubloff)

 

Summary:

The research is aimed at bridging the gap that currently exists between modeling at the process level and operations at the fab level.  It will demonstrate through a concrete example  (a subfactory for tungsten plug technology) an approach that integrates operational level models and process level models for the purpose of qualitatively and quantitatively assessing how process level improvements and changes benefit fab-level production objectives.  The approach incorporates process response surface models into the operational modeling framework, providing substantially more insight and capability than current practice, which uses only fixed process parameters that are set based on optimization at the process level, performed in isolation from operational impact.   Sensitivity analysis methods that integrate operational models and unit process response surface models and evaluate how process parameters (not just process performance) affect fab operations will be developed.  These methods will help identify where improvements in processes would have the most impact on fab operational efficiency, and also where more data and better models are needed.  The proposed research will complement previous results in planning and scheduling.   By linking the process level to the operations level, existing optimization schemes can be enhanced by incorporating the adjustment of certain process parameters.
 

Status:

Several graduate students, Naranjan Chandrasekaran, Brian Conaghan, and Zhiping Shi, have been working with the task leaders on this project.  Starting this summer, one undergraduate student, Quan Nguyen, will be joining the team.  Currently, the main focus is the tungsten plug subfactory, concentrating on the processes in the tungsten plug step: contact clean, liner deposition, and tungsten (W) plug deposition, all performed in cluster tools.  We have integrated a response surface (RSM) model for the W CVD process into a deterministic simulation model for a Novellus multistep process module.  With this integrated model, we can determine how changes to the process parameters (e.g., temperature, pressure) affect the module performance (e.g., throughput).  In addition, we have added this to a queueing model that calculates the expected waiting times for given failure and repair rates.   We will be presenting a paper at this year's AVS conference to describe the current results.

We are now building response surface models (RSMs) for the other processes (e.g., TiN PVD and CVD) and constructing simulation models for different cluster tool architectures    (e.g., Novellus Concept II and Applied Centura) to evaluate the consequences of other process-operations interactions.  We will build RSMs from empirical data, dynamic simulations, and analytical models.   We will use Lee Schruben's cluster tool simulator to    model the tools and extend it by integrating process models.  For each cluster tool, we will describe how throughput and cycle time change as the process parameters and equipment    design parameters vary. This will measure each tool's operational sensitivity and predict the impact of process changes.  In addition, we will find regions of acceptable performance and compare tool architectures.