Emulating & Analyzing Two Climate Model PPEs using a Simplified Additive Gaussian Process Emulator

Emulating & Analyzing Two Climate Model PPEs using a Simplified Additive Gaussian Process Emulator with Dr. Qingyuan Yang

Abstract: We present a method that emulates and analyzes climate model perturbed parameter ensembles (PPEs). The method is simplified from additive Gaussian Process. It ranks and identifies the importance of individual parameters and parameter groups during the training, which increases the interpretability of the emulator and helps improve our understanding of the climate model. The method is applied to the NASA GISS GCM and CESM CAM PPEs. The performance of the method is comparable to that of the Neural Network. Important parameters and their interactions are successfully isolated and identified with their impacts on the target variables quantified. The relationship between the parameters and target variables in the two PPEs is dominated by the effects of individual parameters and parameter pair interactions. Our analysis suggests that finding the sensitive parameters and using them for training an emulator is a lot more important than determining values of the emulator hyperparameters. We point out the limitations of emulating one variable at a time and emulating all variables all at once separately. Findings from this work might have universal implications for other climate model PPEs.

Qingyuan Yang: Qingyuan Yang (Yang) is a post-doc researcher at LEAP, Columbia University. He works on developing and applying different machine learning algorithms for the emulation and parameter estimation of climate models. Qingyuan Yang worked as a research fellow at Nanyang Technological University. He received his PhD and Master degrees from University at Buffalo (SUNY) and bachelor degree from Nanjing University.