Enabling Design under Uncertainty Through Surrogate Modeling
Speaker: Felipe A. C. Viana, Ph.D. Mechanical Engineer, Probabilistics Laboratory, GE Global Research
Abstract: The use of surrogate modeling in engineering design has progressed remarkably since the seminal paper by Sacks et al. (“Design and Analysis of Computer Experiments,” Statistical Science, Vol. 4, No. 4, 1989). The goal is constructing an approximation of the response of interest (also known as surrogate model, metamodel or response surface approximation) based on a limited number of expensive simulations. Most practitioners in the optimization community are familiar with at least the traditional polynomial response surface methodology. However, more sophisticated methods, such as artificial neural networks, support vector regression, and Gaussian process, are becoming increasingly popular. This seminar will illustrate two ways in which surrogate modeling enables design under uncertainty.
First, a surrogate-driven framework for reliability-based design optimization is presented. In addition to being used for probability of failure estimation, surrogate modeling also helps minimizing the number of high-fidelity simulations. This surrogate-based framework has shown to be (a) capable of handling highly non-linear design spaces, (b) able to scale with parallel computing, and (c) robust to incomplete or failed simulations. Second, a statistical approach for characterizing prediction uncertainty of high-fidelity models is presented. The Bayesian formulation of the Gaussian process is used to fuse information from limited amount of simulations and experimental data. The framework has been successfully used to quantify uncertainty due to (a) model parameters, (b) number of simulations and experiments, and (c) discrepancy between the simulation code and the actual physical system.
Bio: Dr. Felipe A. C. Viana is currently a member of the Probabilistics Laboratory at GE Global Research. His work on design optimization under uncertainty, engineering reliability, Bayesian methods, and fusion of simulations and experiments has been applied to GE’s new designs and fielded systems (especially turbo machinery, electrical power systems, and oil and gas systems). Dr. Viana has earned two doctorate degrees, one in Aerospace Engineering from the University of Florida (2011) and one in Mechanical Engineering from the Universidade Federal de Uberlandia (2008). Throughout his graduate studies, Dr. Viana has developed and applied probabilistic design methods to problems such as optimization of passive vibration control mechanism using piezoelectric resonant shunted circuits, identification of landing gear model parameters (EMBRAER provided experimental data), probabilistic design optimization of a composite laminate, and design of a bendable unmanned air vehicle wing under uncertainty. Dr. Viana has published 17 journal papers and 43 conference papers (which together amount for approximately 600 citations). Dr. Viana has also served as reviewer in top journals and conferences (AIAA Journal, Structural and Multidisciplinary Optimization, Journal of Mechanical Design, Engineering Optimization, among others).