Computational challenges of studying uncertainty in numerical simulations

Alejandro Ribes,
Seminar

CT

Argonne National Laboratory
Building 240/Room 4301
Host Tom Peterka
Event Sponsor Mathematics and Computer Science Division

Multiple simulation runs (sometimes several thousand) are required to compute sound statistics for global sensitivity analysis. Current practice consists of running all the necessary instances with different set of input parameters, store the results to disk, often called ensemble data, to later reading them back from disk to compute the required statistics. The amount of storage needed may quickly become overwhelming, with the associated long read time that makes statistical computing time consuming. To avoid this pitfall, scientists reduce their study size by running low resolution simulations or down-sampling output data in space and time. Today petascale and tomorrow exascale machines offer compute capabilities that would enable large scale sensitivity studies. But they are unfortunately not feasible due to this storage issue. In this talk we will explore this problem and discuss novel approaches that may be used in the future.

BIO

Alejandro Ribes graduated in computer science (bachelor’s and master’s) from the Universitat Jaume I, Spain. He also holds a master’s degree in image processing and computer vision from Université de Nice Sophia-Antipolis, France; and received a Ph.D. in multispectral imaging applied to fine art paintings, from the Ecole Nationale Supérieure des Télécommunications, Paris, France.
Alejandro Ribes was also a postdoctoral fellow at the French Atomic Energy Commission, Orsay, France, working on parallel MRI reconstruction. During this postdoc he was appointed as a lecturer at the Computer Science Department of Ecole Polytechnique, Palaiseau, France, where he taught for two years. Alejandro also worked in MRI technology as a visiting scholar at the National Yang-Ming University, Taipei, Taiwan.

From 2009, Alejandro Ribes holds a permanent position at the Research & Development Department of EDF, a major European electric power company. He is responsible for an activity involving the visualization of complex and large industrial data, normally issued from numerical simulations of physical processes. He also collaborates with Université Pierre et Marie-Curie (Paris) by lecturing on opto-electronics.