October 9, 2013

Quantifying Uncertainty at Scale

Tiffany Trader

University of Texas at Austin researcher George Biros has received a $2.85 million grant from the Department of Energy to develop methods for estimating uncertainty in large-scale computer simulations. The project has three main thrusts of particular interest to the DOE: the melting of continental ice sheets in Antarctica; complex fluid flows (such as what is observed in potential algae biofuels); and complex multiscale models.
Prototype simulation of dynamics of Antarctic ice sheet

Prototype simulation of dynamics
of Antarctic ice sheet

The modeling and simulation of complex natural or engineered systems can require billions of parameters, each of which involves a degree of uncertainty, notes Biros, a mechanical engineering professor at the Institute for Computational Engineering and Sciences (ICES). “Estimating the overall uncertainty of the outcome can be quite challenging,” he added.

Biros explains that the mathematical structure that underlies simulations of physical systems is well-understood; it’s the input values that are the source of uncertainty. The larger the scale of the simulation, the more uncertainty is introduced. Even small unknowns can have a big impact on accuracy.

One way the effects of uncertainty can appear is the cone-like shape of predicted hurricane paths.

“You typically see the cone opening as you look into the future,” Biros said, “which means whatever small perturbation that you have gets amplified so you’re more and more uncertain about the future.”

So far uncertainty research has mainly been relegated to small-scale systems with not very many parameters using software that can be run on laptops. Biros and his team are studying more sophisticated models, which require the advanced processing power of supercomputers.

While this work has direct implications for energy applications, it can also serve as a model for other complex systems. If this project is successful, Biros believes that with modifications, it should be transferable to other systems.

Well-known in HPC circles, George Biros is a two-time recipient of the Gordon Bell Prize. Awarded by the Association for Computing Machinery, the prize has been referred to as the Nobel Prize for supercomputing.