#29 from R&D Innovator Volume 2, Number 3          March 1993

Murphy’s Wall:  Dealing with Complexity
by James M. Minor, Ph.D.

Dr. Minor is senior staff researcher at Syntex Research, Palo Alto, California.  He and his team received the 1992 Award for Statistics in Chemistry from the American Statistical Association.

In the "good old days," an expert researcher could adjust a few critical variables, visualize patterns, infer useful concepts, and make discoveries that led to new products.  Today the extreme complexity of research reduces the possibility that this process will succeed.  Too often, dynamic behavior and patterns of noise or chaos are the rule rather than the exception. 

Frequently, scientists unwittingly make research even more complex.  They tend to over-complicate explanations for their research and either oversimplify or ignore the impact of inherent variation and new or unknown effects.

In today's research, "Murphy's law" creates such a fundamental barrier to our study of nature that I call it "Murphy's wall" instead.  The task of researchers is to conquer "Murphy's wall" and reduce their ignorance about a research problem so that it can become a manageable engineering project.  To put it another way, our experiments must eliminate vast amounts of ignorance within reasonable limits of time and resources.

Statistics to the Rescue

In two decades of collaborating with scientists, I've developed statistical methods to surmount these barriers to discovery and at the same time enhance our intuitive and cognitive abilities.  About 10 years ago I developed a strategy that cut research effort by as much as 50 percent. 

The Strategy of Research (SR) creates optimal experimental schedules to reduce research complexities.  The basic procedure is to apply the power of multivariate statistical analysis to scientific research.  SR is an efficient way to advance research from “the bench” to the “bottom‑line.”  I recently reviewed this strategy.¹

Here is some evidence of SR's effectiveness.  The technique shortened a high-performance liquid chromatography of complex samples from many weeks to overnight.  It dramatically improved the sensitivity and selectivity for automatic clinical blood analysis.  And it rescued a color copying business by solving, in six months, critical problems which had plagued their technology for years.

If you test one variable at a time, you will need many experiments before you get significant feedback.  The beauty of SR is that it converts experiments with large, batch‑oriented schedules to an interactive sequence of mini‑schedules (small chunks of experiments). 

While some aspects of research can be made routine—such as collecting, storing and looking at data—research by definition cannot follow a simple formula.  The mini-schedules of SR provide feedback and stimulate an interactive thought process that helps plan for future experiments.  Exploration within these mini-schedules can be merged with a global view of the entire project, while filtering out unanticipated effects or noise.  If necessary, investigators can go back and forth between the essential experiments, all the while increasing their knowledge of the problem.

SR consists of six basic types of testing:

1.  Screening experimental procedures and/or instrumentation for operational problems.

2.  Simple replication to test and establish inherent variability.

3.  Simple ranging to explore and validate limits of operation.

4.  Screening groups of synergistic effects.

5.  Resolution of synergistic effects from important groups.

6.  Diagnosis for missing effects that affect predictions.

But How Does it Work?

To make the process more comprehensible, I'd like to explain how SR helped develop a process for manufacturing a three-layer laser disc.  The disc had a polymethylmethacrylate (PMMA) layer for rigidity.  Bonded to this was a collapsible microporous (cmp) layer.  At the bottom was a thin layer of reflective metal.

To store information on the disc, a laser beam heats the metal, thus collapsing the adjacent cmp structures.  To read the disc, we detect these distortions by reflecting a low-power laser beam from the metal layer through the PMMA and cmp layers.

The disc is constructed by spin-coating a solution of PMMA onto the rigid PMMA substrate.  When it coagulates, the PMMA forms the cmp structure. When the disc is dry, the metal layer is deposited onto it.

The problem was twofold:  How to control the process to get a product—and then how to optimize the process?  If the disc was to satisfy all criteria, 11 physical controlling factors (e.g., cmp polymer type, solvent type, etc.) and several measured properties (e.g., surface evenness, reflectivity) must be considered.  In a sense, the task corresponds to finding a needle in a multi-dimensional haystack.

Without SR, a minimum of 150 experiments would be needed just for the basic data for manufacturing.  Using SR, fewer than 100 experiments identified and validated a synthesis for the high-performance disc.  SR allowed such efficiency because so many possible combinations were actually insignificant, so there was no need to waste experimental time on trivial phenomena.  By testing bundles of interactions, the dominant interactions became obvious, and that guided the selection of future experiments.

By learning about the relationship between performance and the 11 physical process variables, we could determine which process would give best results.  The most promising predictions were converted into experiments, and the results were fed back into the analysis and optimization routines to check for missing effects, update predictions, and home in on even better results.

This project was complicated by large experimental noise, truncated or rounded measurements, malfunctioning equipment, and human error.  As usual, the values of some factors were interdependent—for example,  the maximum permissible spin depended on the viscosity of the melted plastic.

At the end, an outside testing firm verified that the discs surpassed all expectations.  These results supported a successful process patent.  The SR approach produced a wide enough knowledge base to scope out the entire region of the patent application. The patent had broad claims which will be difficult for a competitor to skirt.

Unlike the Berlin wall, Murphy's Wall will not be overturned—if anything, it's likely to get higher.  Even so, multivariate statistical methods help produce better solutions to complex projects—within the time and cost constraints of industry.

¹ChemTech 22, 751, 1992.