A Multi-element Generalized Polynomial Chaos Method for Modeling Uncertainty in Flow Simulations
With the computational fluid dynamics field reaching now some degree of maturity, we pose the more general
question in this proposal of how to model and propagate uncertainty in flow systems, and how to formulate
efficient and robust algorithms for the corresponding stochastic incompressible and compressible Navier-Stokes equations. To this end, we propose a multi-element generalized Polynomial Chaos (ME-gPC) method that can overcome the
difficulties associated with long-term integration, stochastic bifurcations and strong nonlinearities
typically encountered in unsteady fluid dynamics problems. To evaluate this approach we consider three
representative applications involving biological flows, transitional flows, and shock dynamics that reflect
uncertainty in constitutive laws, sensitivity to the free stream random fluctuations, and random roughness.
Specifically, we will extend the pioneering ideas of Norbert Wiener on polynomial chaos and our previous work on
Wiener-Askey expansions in order to handle stochastic inputs with arbitrary probability distributions,
either continuous or discrete. This fundamental development, in turn, will allow us to introduce a subdomain
decomposition of the random space, thus resolving adaptively large levels of localized random fluctuations and even
stochastic discontinuities. In fact, discretizing the random space with N elements within which we employ
generalized polynomial chaos (gPC) expansion of order p results in a very fast convergence of the form
N^(-2p+2).
This possibility for dual path of convergence, either increasing N or p, is similar to the h-p convergence of
spectral element methods that the PI has been developing in the last twenty years for deterministic problems.
Moreover, the ability to handle arbitrary PDFs will open the possibility of "racking" he PDF of the
solution output - instead of the stochastic input as we currently do - thus updating the gPC trial
basis on-the-°y and correspondingly reducing significantly errors associated with long-term integration
due to strong flow nonlinearities.
The proposed work will have significant and broad impact as it will set the foundations of data assimilation
and rigorous sensitivity analysis in computational fluid dynamics. It will establish, for first-time, a
composite error bar in flow simulations that goes beyond numerical accuracy and includes uncertainties in
operating conditions, the physical parameters, and the domain. The proposed approach will affect fundamentally
the way we design new experiments and the type of questions that we can address, while the interaction between
simulation and experiment will become more meaningful and more dynamic. This, in turn, will find its way in the
design of flow systems equipment and will provide a rigorous reliability framework.
On the education front, the new knowledge will contribute towards understanding nonlinear systems subject to
noise, fundamentals in stochastic dynamics, data assimilation, and design under uncertainty. Currently, there
is very little in our curricula on such subjects where the "deterministic" thinking prevails. We plan to
incorporate these new ideas in engineering and applied mathematics courses we teach at Brown. Outside of the
classroom, undergraduate students, through Brown's UTRA (Undergraduate Teaching and Research Assistantships)
program, will be involved in the research projects, either during the academic year or the summer. For
pre-college students, we will expand the Artemis program at Brown University which is designed to engage
young women from the Providence area in computer and computational sciences. We will use immersive flow
visualizations at Brown's CAVE as an opportunity to educate students about simulation, predictability,
and other issues of computational science and applied mathematics.
|