Design of Experiments

Design of Experiments


Gosset

Neil Sloane of the Math Center has developed Gosset, which is billed as the world's most powerful program for constructing experimental designs. The variables may be discrete or continuous, discrete variables may be quantitative or qualitative, continuous variables may range over a cube or a ball (or both). The variables may be required to satisfy linear equalities or inequalities, so mixtures and constraints present no difficulty. The model to be fitted may be any low degree polynomial. The user specifies the number of observations, and the design may be required to include a specified set of points. The following types of designs can be requested: A-, D-, E-, or I-optimal designs, block designs, designs with protection against loss of one trial, and designs with correlated errors. Applications so far include optimizing the production of wafers for integrated circuits; placing laser beams for treatment of tumors in the brain; growing protein crystals; designing a cellular substrate used in catalytic converters; and designing coated photographic paper.

IC Fabrication

Anne Freeny and others have recently concluded a study of an etching process of SiO2 as part of the qualification of an integrated vapor phase clean module. The statistical aspects of the study included experimental design and data analysis to identify levels of the process parameters that minimize wafer-to-wafer variability. This work used two important tools developed in the Math center, the Gosset program for optimal experimental design and the S language for data analysis and statistical modeling.

Fiber Manufacturing

The first step in the modified chemical vapor deposition process for producing optical fiber is to deposit layers of glass on the interior walls of a tube. Changes in the speed of a moving torch affect uniformity in the thickness of the layers. Scott Vander Wiel , Daryl Pregibon and others have developed models for predicting how the glass deposition will vary along the length of the tube. They use reciprocal torch speed as the input to a transfer function modeled as the exponential of a cubic spline. To improve deposition uniformity, they have used the fitted model to design a new torch speed profile that has nearly on-target predictions. The model is necessarily a simplification of reality and initially it did not extrapolate well. Iterating through experimentation and refitting, however, produced a torch speed trajectory that produces substantially more uniform glass deposition. The methodology and software will enable engineers to design torch speed profiles for new fiber products and to update them for existing products.

Robust Design Experiments with Dynamic Characteristics

Vijay Nair and others have developed a structured data analysis strategy for robust design experiments with dynamic characteristics. The difference between conventional static robust design experiments and dynamic ones is that the former attempts to optimize a process to attain a single target value whereas the latter has a varying target that depends on a signal or input to the system. Such experiments are being increasingly used to design systems that are robust to uncontrollable variation over a range of signal factors. The cornerstone of the data analysis strategy is a pair of newly developed graphical methods, sensitivity-standard deviation plots and gamma plots. These are used to identify appropriate measures of dispersion that define a performance measure suitable for optimization in the robust design paradigm.