CAREER: Multifidelity Modeling and Search Using Adaptive Field Prediction

点击翻译按钮获取中文摘要

A variety of high-fidelity simulations are available to engineers in support the analysis and optimization of engineered systems. However, the computational demands of these simulation codes often mean advanced computational design techniques, such as uncertainty analysis and optimization under uncertainty, are not used to their fullest potential. This Faculty Early Career Development Program (CAREER) project supports fundamental research to advance techniques for simulation-based engineering systems design with a goal of making their application practical on a wider variety of important engineering problems. The project will result in new understanding about how engineers can utilize rich information from simulation results to accelerate computational applications such as uncertainty analysis and optimal design under uncertainty. The techniques target engineering applications that exhibit complex physics, such as aerodynamics, electromagnetics and mechanical structures. New methods pioneered in this project will impact society through more rapid and reliable design of complex engineered systems across domains such as transportation, energy harvesting, weather forecasting, and communication. Educational initiatives of this project focus on instruction and curriculum development for advanced computational design techniques. This includes a new short course on computational design for undergraduate students at Iowa State University, creation of an online hub to make advanced simulation-based design techniques accessible to students and practitioners around the country, and organization of mini-symposia on computational design.This research pioneers a novel class of methods for using the field responses of simulations to construct improved multifidelity models in the context of advanced computational design techniques such as uncertainty analysis and optimization under uncertainty. The process of extracting and adapting physics-based information encoded in the field responses of models of varying degrees of fidelity will be achieved by combining metamodeling techniques and machine learning, as well as the development of novel adaptation techniques. The new methods and algorithms will be derived and rigorously characterized through computational experiments with structural, electronic, and fluid systems case studies. Additionally, the results will provide an understanding of the impact of model correlations and the mechanisms controlling the growth of the computational cost. This will enable the creation of new and unique methods for the automated setup of multifidelity models and allow us to address problems of higher complexity than what is currently possible.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.