Community Talk: Vassili Kitsios (CSIRO)
A machine learning approach to rapidly project climate responses using CMIP data
Abstract
The world economy is currently grappling with how to transition to net-zero emissions later this century. There is hence a need to assess the physical climate risks for a broad range of future economic, and hence emissions, scenarios. A small selection of the possible scenarios is simulated by numerous climate models as part of the international Coupled Model Inter-comparison Projects (CMIP). These simulations provide estimates of the future temperature, rainfall and other properties required to assess physical climate risk. Here, we employ model reduction and machine learning techniques to generate climate simulation output at a fraction of the computational time and cost. The presentation will outline the methodology and validate the method via its ability to reconstruct CMIP data excluded from the learning phase. One can now rapidly estimate the climate response for many more emissions pathways that would ordinarily be possible using CMIP models alone. We refer to this approach as QuickClim. We use QuickClim to reconstruct plausible climates for a multitude of concentration pathways. As expected, higher mean temperatures coincide with higher end-of-century carbon concentrations. However, the uncertainty in the climate variability saturates earlier, in the mid-century, during the transition between current and future climates. For pathways converging to the same end-of-century concentration, the climate is found to be sensitive to the choice of trajectory. In net-zero emission type pathways, this sensitivity is of comparable magnitude to the projected changes over the century.
Please use this thread for further discussion on this talk.