Climate Generator (Stochastic Climate Representation: 120 ka to present year)

Abstract

I present a computationally efficient stochastic climate model for large spatiotemporal scales (example, for the context of glacial cycle modelling). In analogy with a Weather Generator (WG), the model can be thought of as a Climate Generator (CG). The CG produces a synthetic climatology conditioned on various inputs. Inputs for the CG include the monthly mean sea surface temperature field from a simplified Energy Balance Model (EBM), surface elevation, surface ice, carbon dioxide, methane, orbital forcing, latitude and longitude. The CG outputs mean monthly surface temperature and precipitation using Bayesian Artificial Neural Networks (BANN) for non-linear regression. The CG is trained against the results of GCMs (FAMOUS and CCSM) over the last deglacial (22 ka to present). For validation, CG predictions are compared directly against the 120 ka to 22.05 ka interval of FAMOUS results that were not used for CG training. The stochastic noise is added to each prediction by generating the random normal distribution with mean from the ensemble networks for a single guess and Standard deviation computed from 10th and 90th percentile of the BANN predictive distribution for each time step. For the CG trained against FAMOUS, I show the predictive errors (relative to FAMOUS) are comparable to the difference between FAMOUS and the CCSM.