Although the state of the art in simulation of the Earth system continues to advance, the outputs of global circulation models (GCMs) remain too coarse for direct application to many climate impacts problems. A variety of methods are available to downscale GCM outputs to higher resolution, ranging from simple statistical adjustment to nested dynamical modeling and machine learning approaches. These methods often disagree, and it is challenging to evaluate their relative credibility, especially when comparing different types of method (e.g., statistical vs dynamical).

We define credible results as those where phenomena and processes are coherent; i.e, if it's raining, the conditions that lead to rain (uplift of atmospheric moisture, etc.) need to be present. To evaluate a variety of downscaling methods, we consider precipitation occurrence in the Southern Great Plains, at a point near the DOE ARRM site in Oklahoma during the wettest month, May. In this context, the dominant process leading to precipitation occurrence is the advection of moisture north from the Gulf of Mexico.

Using composite synoptic conditions on wet vs dry days, we can determine whether the results of downscaling GCM precipitation are consistent with its driving processes. This approach provides a strategy for process-based added-value analysis of all types of downscaling methods that goes beyond simple metrics of statistical similarity.

We have downscaled precipitation from 3 different GCMs using 2 regional climate models (RegCM4 & WRF), a machine learning method (U-Net CNN), and 4 statistical methods of varying complexity. Using this approach to compare them with one another and the raw GCM results, we find that when the performance of the GCM is good, all downscaling methods inherit its credibility and can produce reasonable results as well. When the GCM's performance is poor, dynamical methods can mitigate regional circulation errors, where the other methods cannot. We also find that the more sophisticated non-dynamical methods perform worse on bad inputs than simpler methods.