Thursday, January 23
Mesa Lab Chapman Room
University of Chicago
This talk considers two different statistical approaches to producing possible future climates. The first is climate model emulation, which seeks to develop a simple statistical model that can reproduce some of the outputs of a GCM under conditions (forcing scenarios, parameterizations) for which the GCM has not been run. Here, we focus on emulating regional temperature and precipitation at the annual time scale as a function of CO2 trajectory. Both the input (CO2 trajectory) and the outputs (annual regional-scale temperature and precipitation) are time series. It is important to use an emulator that respects this time series structure and I will describe a fairly simple regression model that does so and is able to reproduce model output reasonably well with very few (sometimes one is enough) runs of the GCM as a training set.
Emulating GCMs is one thing, but producing accurate simulations of future climates is another and is of greater interest for at least some purposes (such as impact assessments). Directly using the output of GCM for this purpose is unwise and it is common to combine observational data and GCM output to produce future climate simulations. One can either use the observational record to modify the GCM output or can use GCM output to modify the observational record. We advocate the latter approach. I will describe a spectral method for modifying observed daily temperatures that allows one to account for possibly different changes in variability at different time scales in a natural and coherent way.
This talk represents joint work with Elisabeth Moyer, Stefano Castruccio, David McInerney, Robert Jacobs, Feifei Liu and William Leeds.