Simulation of Arctic sea ice within the DeepMIP Eocene ensemble: Thresholds, seasonality and factors controlling sea ice development

The early Eocene greenhouse climate maintained by high atmospheric CO₂ concentrations serves as a testbed for future climate changes dominated by increasing CO₂ forcing. In particular, the early Eocene Arctic region is important in the context of future CO₂ driven climate warming in the northern polar region and associated shrinking Arctic sea ice. Here, we present early Eocene Arctic sea ice simulations carried out by six coupled climate models within the framework of the Deep-Time Model Intercomparison Project (DeepMIP). We find differences in sea ice responses to CO₂ changes across the ensemble and compare the results with available proxy-based sea ice reconstructions from the Arctic Ocean. Most of the models simulate seasonal sea ice presence at high CO₂ levels (≥ 840 ppmv = 3× pre-industrial (PI) level of 280 ppmv). However, the threshold when sea ice permanently disappears from the ocean varies considerably between the models (from <840 ppmv to >1680 ppmv). Based on a one-dimensional energy balance model analysis we find that the greenhouse effect likely caused by increased atmospheric water vapor concentration plays an important role in the inter-model spread in Arctic winter surface temperature changes in response to a CO₂ rise from 1× to 3× the PI level. Furthermore, differences in simulated surface salinity in the Arctic Ocean play an important role in the control of local sea ice formation. These differences result from different implementations of river run-off between the models, but also from differences in the exchange of waters between a brackish Arctic and a more saline North Atlantic Ocean that are controlled by the width of the gateway between both basins. As there is no geological evidence for Arctic sea ice in the early Eocene, its presence in most of the simulations with 3× PI CO₂ level indicates either a higher CO₂ level and/or an overly weak polar sensitivity in these models.