|
QUEST team working to put
uncertainty theory to practice
Posted December 12, 2011
Habib Najm of the Department of Energy’s Sandia National Laboratories discusses uncertainty quantification. Najm leads a collaboration called QUEST, for Quantification of Uncertainty in Extreme Scale Computations.
Part of a series.
Each element of any computer model comes with uncertainty – imprecision springing from indefinite data, poorly understood physical properties and inexactness in the model’s recreation of reality. The big question: How can scientists calculate the degree of possible error in their models? Exploring that question lies at the heart of uncertainty quantification, known to the experts simply as UQ.
Because UQ requires large computing capabilities, calculating error in many scientific computations today remains more theory than practice, says Habib Najm, a distinguished member of the technical staff at Sandia National Laboratories’ Livermore, Calif., site.
The advent of exaflops-capable computers – roughly a thousand times faster than today’s supercomputers – later this decade may provide the extra power needed for UQ on today’s simulations. UQ also will be important in understanding how much confidence to place in the highly detailed models exascale computing will enable.
For that reason, UQ is a major part of the Department of Energy Advanced Scientific Computing Research office’s exascale-computing portfolio, which includes a project Najm leads called QUEST, for Quantification of Uncertainty in Extreme Scale Computations. Sandia’s QUEST collaborators are Los Alamos National Laboratory, Johns Hopkins University, Massachusetts Institute of Technology and the universities of Southern California and Texas at Austin.
Defining uncertainty
Uncertainty often is described as a distribution of values within which a particular observed value could fall. To quantify it in predictions from a simulation, scientists start by identifying sources of uncertainty in parameters that make up the model – the variables that affect the output. Next, they typically perform sensitivity analysis, identifying the parameters whose uncertainty affects predictions the most. This uncertainty then can be propagated through a model to reveal how it changes a simulation’s output.
