The use of emulators for uncertainty characterization and integrated modeling
Abstract: In this talk, I will present some recent work that my colleagues and I, at the Joint Global Change Research Institute/Pacific Northwest National Laboratory, have focused on, in the general area of climate model emulation. The aim of my talk is not to go into minute details of our work, but rather to give a flavor to the TIES audience of a range of questions, related to uncertainty characterizations, that provide motivation for the application of data-driven statistical and machine learning methods.
In the first example, an emulator of climate model output is developed to serve as the Earth system component in an integrated modeling framework of impacts and mitigation scenarios. (I will also briefly describe such framework that attempts to “close the loop” between Earth and human system modeling.) This emulator exploits the rich archives of Earth system model output that the Coupled Model Inter-comparison Project provides at regular intervals of several years (the most recent being CMIP6). The CMIP effort collects the state-of-the-art simulations of future scenarios by multiple Earth system (climate) models. The CMIP catalogue, however, is limited to a handful of scenarios since climate model simulations are computationally expensive and only a few (well-spaced) can be sensibly prescribed to modeling centers without stretching their resources. Emulators like the one I’m going to present, STITCHES, are developed to fill-in the gaps in between the available scenarios. This part will take up most of the talk but if time is left, I will quickly describe two more applications of the emulator concept. One uses machine learning tools to simulate daily temperature and precipitation fields mimicking a climate model that the ML tools have been trained on. This work is in the name of enlarging the sample size of daily climate model output to better characterize extremes’ behavior. Lastly, I may be able to talk briefly about the use of statistical surrogate models to estimate the relationship between model parameters and model output (when the model is a complex Earth system model that can be run only a few times by varying its parameters because of the same computational constraints alluded to earlier) in order to identify the best parameter settings according to some validatory metric. These best parameter settings often do not coincide with, but lie between those settings that were actually run by the climate model, making the role of the emulator key to identify them.