Screening for Fuel Economy: A Case Study of Supersaturated Designs in Practice

ABSTRACT

A successful use of supersaturated design and analysis is demonstrated through a case study completed at the Lubrizol Corporation. In the study, a 28-run supersaturated design is used to screen the effects of more than 70 possible model terms (linear effects, quadratic effects, interactions, and measured covariates) on engine motor oil coefficient of friction (COF). Of the over 70 model terms of interest, 50 are two-way linear interactions. A Lubrizol-developed model-averaging technique known as Bayesian variable assessment (BVA) is used to identify the important high-level factors and model terms from the experiment.

This study is unique in the literature due to complications in multiple factor levels, physical correlations and constraints on the factors, curvature, and the desire to screen for a large amount of interactions. The test results are subject to common cause variation and unknown special causes such as operator error and test instrument error.

Due to time and cost constraints, supersaturated designs are necessary to screen for phenomena such as gasoline-powered engine fuel economy. Based on the results from a 10-run follow-up experiment, the use of the supersaturated design analyzed using BVA is concluded to be a success in this case study.

KEYWORDS:

• Bayesian variable assessment
• design and analysis of supersaturated design
• Markov chain Monte Carlo

ACKNOWLEDGMENT

We are grateful to the reviewers for their thorough and constructive comments, which have led to a significant improvement in the final version.

Notes

Note that Main Effects Include Non-Linear Main Effects and Measured Covariates.

^a Sometimes reported in the literature as smallest effect identified.