Exploit high performance computing for data analysis

CISL is exploring the use of HPC systems for statistical analysis in the climate and atmospheric sciences for a variety of topics. The goal is to enable types of analyses that were previously computationally not feasible or would have taken too much time to be of practical use. One example is a Bayesian model for climate change detection and attribution that we developed recently in collaboration with academic researchers. Detection and attribution is a commonly used multivariate regression methodology to identify if changes in observational trends can be detected beyond natural variability, and if so, if they can be attributed to human actions. One of the issues in current detection and attribution methods is that certain quantities are used in an ad hoc way, e.g., the choice of truncating the empirical orthogonal functions (EOFs), and the uncertainty associated with their estimation is not incorporated properly. The Bayesian framework addresses these issues.

In the spirit of developing statistical methods that rely on the effective use of modern computational infrastructure, we have developed a computationally efficient, parallelizable Bayesian detection and attribution model and created a corresponding software implementation. The newly developed method does not require a specific EOF truncation, but rather evaluates all truncation options and then performs Bayesian model averaging to combine them. The key developer at NCAR is now also the chair of the Climate Change Detection and Attribution Working Group as part of the Mathematical and Statistical Methods for Climate and Earth Systems Year at the Statistical Applied Mathematical Sciences Institute.

The research has been published as Katzfuss, M., D. Hammerling, and R.L. Smith, 2017: A Bayesian hierarchical model for climate change detection and attribution. Geophys. Res. Lett., 44, 5720–5728, doi:10.1002/2017GL073688.

This research is made possible through NSF Core funding.