Supercomputing for Multi-Resolution Gaussian Process Modeling

• Activity Number: 87
• Type: Invited
• Date/Time: Sunday, August 3, 2014 : 8:30 PM to 10:30 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Dorit Hammerling*+ and Nathan Lenssen and Douglas Nychka and Stephan R. Sain
• Companies: NCAR and NCAR and NCAR and NCAR
• Keywords: Spatial modeling ; Multi-resolution ; Supercomputing

Abstract

Increasingly large spatial data set with nonstationary covariance have created the need for computationally efficient multi-resolution Gaussian process models. The model discussed here, LatticeKrig, combines the representation of a field using a multi-resolution basis with statistical models for processes on a lattice. The main principle is to expand the two dimensional spatial field in a sequence of basis functions that are organized on regular grids of increasing resolution. The basis coefficients have a stochastic structure, which is modeled using a Markov random field and results in a computationally efficient model applicable to large spatial data sets. In addition to a computationally efficient modeling framework, the availability of supercomputing facilities creates further opportunities to employ parallelization to address questions in the geo- and climate sciences, which involve massive spatial data sets. We show examples how parallelization can be efficiently employed to address such questions using the National Center for Atmospheric Research's supercomputing facilities.