Flexible Response Surface Modeling
Monday 2 June, 10:20--12:10

Organizer: Doug NYCHKA
North Carolina State University
Moderator: Dennis COX
Rice University

Process Modeling with Nonparametric Response Surface Methods

Sarah W. HARDY, Douglas NYCHKA
North Carolina State University, Raleigh, NC, USA
Perry HAALAND
Becton Dickinson Research Center, Research Triangle Park, NC, USA
A fundamental problem in industrial process development is the deliberate exploration of the different factors that affect product quality. Timely identification of key process variables at the design stage can often have a large impact on product performance and development time. Furthermore, understanding the functional relationship among input process variables enables researchers to optimize the process for specific characteristics. Most statistical methods for estimating response surfaces posit a {\it parametric} polynomial model. A significant drawback of these traditional polynomial models occurs when the true response surface is highly nonlinear with a complex structure. These complex surfaces, however, are frequently encountered in the development of innovative processes based on cutting-edge technology. New methodologies, termed nonparametric response surface modeling, offer techniques for fitting response surfaces to the data without the restriction that the relationship be characterized by a polynomial function. One promising family of estimators are thin plate splines because they are the sum of a polynomial function and a smooth function. This work discusses the decomposition of the thin plate model into its polynomial and smoothing components. The decomposition enables the researcher to test if the smoothing component is significantly improving the predictions from the model. It also allows the researcher to assess the effect of individual variables in the polynomial model component and to fit a thin plate spline with a reduced form of the polynomial component. Another issue addressed is the individual effect of variables on the response surface. This leads to the use of cross-validation for selecting an efficient set of variables to describe the response. It has been the experience at Becton Dickinson that thin plate splines facilitate the optimization of very complex physical, biological and chemical processes and some examples will be given of their use.

[Sarah Hardy, Department of Statistics, North Carolina State University, Raleigh,NC 276905-8203, swhardy@stat.ncsu.edu]


Spatially Correlated Nonlinear Random Coefficients Model in A Semiconductor Manufacturing Application

A.H. HARTFORD, J.C. LU, W.C. HOLTON, S. WILLIAMS, K. W. KIM
North Carolina State University, Raleigh, NC, USA
It is difficult to investigate the semiconductor device design with physical experiments, especially in the case of building the future generation devices. Technology computer aided design (TCAD) simulation tools are commonly used to compare different device design structures. These tools are constructed from physical laws and are deterministic in nature. In our research, the orthogonal array latin hypercube sampling method is used to direct the simulations for 0.18 micron device design studies. The device reliability measurement, transconductance, is a nonlinear function involving derivatives of other measurable variables from device simulations. A random coefficient model is proposed to characterize the nonlinear function, and a stochastic model with a spatial correlation structure is proposed to model the random coefficients. For predicting the device performance at future manufacturing conditions, a measurement error structure is added to incorporate the process variation into the device design study. The Two-stage least squares estimation method and the approximation inference are developed for this new model. Examples from real-life experiments show the potential of our approach in finding a device design which is least sensitive to manufacturing variations.

[J.C. Lu, Department of Statistics, North Carolina State University, Raleigh NC 27695-8203, jclu@stat.ncsu.edu]


Back to JRC program
Back to JRC Home Page