Contributed Session: Experiment Design Applications

Session Chair: Jim Filliben, NIST

Case Studies in Design & Analysis of Computer Experiments

Andrew J. Booker
Research & Technology Dept., Boeing Information & Support Services

Computer simulations are used in the analysis of many products and processes at Boeing. This provides a large and varied arena for the application of design and analysis of computer experiments (DACE) as described in Sacks, Welch, Mitchell and Wynn (Statist. Sci. 4 (1989):409--435). This talk will describe the approach to DACE used at Boeing and, as time permits, present case studies in aerodynamic design, process optimization and electrical power system design. The talk will conclude with suggestions of areas of research which would benefit practitioners.

[Andrew J. Booker, Boeing Information & Support Services, P.O. Box 3707, Seattle, WA 98124-2207 USA; booker@espresso.rt.cs.boeing.com ]

Applying "Incomplete" Selective t-Covering Experimental Designs to Software Testing

I. S. Dunietz
W. K. Ehrlich
B. D. Szablak
Operations Technology Center, AT&T

C. L. Mallows
Statistics Research Dept., AT&T Bell Laboratories

A. Iannino
Pipeline Associates

In performing functional testing on a Software Unit under Test (SUT), we define I relevant factors or parameters, where each factor i takes on q_i settings. The product, Q, of these factors' settings defines a partition of the SUT's input domain. Resource constraints often preclude running one or more test cases constructed from each of the strata in the partition. We analyze (for a specific SUT) the code coverage achieved by test designs that guarantee testing all combinations of test factor settings taken t at a time (i.e., t-covering designs) where tSUT, our results indicate that the code coverage obtained by lower-order t-covering designs (i.e., t<=2) is comparable with that achieved by complete designs, and suggest the possible applicability of lower-order covering designs to software testing.

[W. K. Ehrlich, AT&T, 480 Red Hill Rd., Middletown, NJ 07748, USA; W.K.Ehrlich@att.com ]

Yield Optimization Using Response Surfaces

Timothy Tan
Foo Say Wei
Dept. of Electrical Engineering, National Univ. of Singapore

Yaacob Ibrahim
Dept. of Industrial & Systems Engineering, National Univ. of Singapore

Loh Han Tong
Dept. of Mechanical & Production Engineering, National Univ. of Singapore

Tolerances in circuit components such as resistors and capacitors will invariably affect the circuit design. Variabilities in the responses can lead to some non-conforming designs thereby affecting product quality and the overall manufacturing yield. Tolerances can be treated in the design stage by optimizing the yield. A strategy which combines a simulation-based yield optimization method with response surface is proposed. Using Monte Carlo simulation, the Centre-of-Gravity (CoG) method evaluates the yield and search along the direction of increasing yield. The search however does not converge but tends to oscillate about the optimal solution. A response surface for the yield is then fitted over the oscillating region using points generated from the CoG method. The response surface is then used to optimise the yield. A dramatic yield improvement has been observed on several examples. Several successive fits of the response surface also lead to significant yield improvement.

[Dr Yaacob Ibrahim, Dept. of Industrial & Systems Engineering, National Univ. of Singapore, 10 Kent Ridge Crescent, SINGAPORE 119260 iseyacob@leonis.nus.sg ]

Reliability Improvement of Disk Drives Through Multiresponse Experiments

Vijendra P. Singh
Seagate Technology

Manesh Shah
International Pressure Vessel Inc.

Jagdish S. Rustagi
Dept. of Statistics, Ohio State Univ.

In computer disk manufacturing the following chemical engineering problem arises. The process involves preheating followed by deposition of various layers of metals and overcoats in a series of chambers, through which each disk travels sequentially. This continuous process is controlled by several parameters such as voltage and current, gas flow and current pulse duration in each chamber. The reliability is measured through magnetic parameters on different tracks in sampled disks. The challenge for reliability improvement is to first characterize this process by determining the most critical control variables and their interactions. These variables result in various responses such as the magnetism and disk layer uniformity. The challenge is not only in multiresponse characterization but also in combining historic experimental data with minimum number of additional experiments to obtain the process relationship for optimal reliability determination of the disk. This problem has been solved using the theory of design of sequential experiments. Several other engineering applications in the disk drive industry will be discussed.