GMDD Geoscientific Model Development Discussions GMDD Geosci. Model Dev. Discuss. 1991-962X Göttingen, Germany 10.5194/gmd-2017-276 The Climate Generator: Stochastic climate representation for glacial cycle integration Arif Mohammad Hizbul Bahar 1 Tarasov Lev 2 Hauser Tristan 3 Faculty of Engineering and Applied Science, Memorial University, Canada Department of Physics and Physical Oceanography, Memorial University, Canada Department of Environmental and Geographical Science, University of Cape Town, South Africa 02 01 2018 2018 1 27 Copyright: © 2018 Mohammad Hizbul Bahar Arif et al. 2018 This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/ This article is available from https://gmd.copernicus.org/preprints/gmd-2017-276/ The full text article is available as a PDF file from https://gmd.copernicus.org/preprints/gmd-2017-276/gmd-2017-276.pdf

This paper presents a computationally efficient stochastic approach to simulate atmospheric fields (specifically monthly mean temperature and precipitation) on large spatial-temporal scales. In analogy with Weather Generators (WG), the modelling approach can be considered a <q>Climate Generator</q> (CG). The CG can also be understood as a field-specific General Circulation climate Model (GCM) emulator. It invokes aspects of spatio-temporal downscaling, in this case mapping the output of an Energy Balance climate Model (EBM) to that of a higher resolution GCM. The CG produces a synthetic climatology conditioned on various inputs. These inputs include sea level temperature from a fast low-resolution EBM, surface elevation, ice mask, atmospheric concentrations of carbon dioxide, methane, orbital forcing, latitude and longitude. Bayesian Artificial Neural Networks (BANN) are used for nonlinear regression against GCM output over North America, Antarctica and Eurasia. <br><br> Herein we detail and validate the methodology. To impose natural variability in the CG (to make the CG indistinguishable from a GCM) stochastic noise is added to each prediction. This noise is generated from a normal distribution with standard deviation computed from the 10&thinsp;% and 90&thinsp;% quantiles of the predictive distribution values from the BANNs for each time step. This derives from a key working assumption/approximation that the self-inferred predictive uncertainty of the BANNs is in good part due to the internal variability of the GCM climate. Our CG is trained against GCM (FAMOUS and CCSM) output for the last deglacial interval (22&thinsp;ka to present year). For predictive testing, we compare the CG predictions against GCM (FAMOUS) output for the disjoint remainder of the last glacial interval (120&thinsp;ka to 22.05&thinsp;ka). The CG passes a <q>climate Turing test</q>, an indistinguishability test in analogy with the original Turing test for artificial intelligence. This initial validation of the Climate Generator approach justifies further development and testing for long time integration contexts such as coupled ice-sheet climate modelling over glacial cycle time-scales.