Assessing Climate Modeling Uncertainties in the Siberian Frozen Soil Regions by Contrasting CMIP6 and LS3MIP

Abstract. Climate models and their land components still show pervasive discrepancies in frozen soil simulations. Contrasting the historical runs of seven land-only models of the Land Surface, Snow, and Soil Moisture Model Intercomparison Project (LS3MIP) with their Coupled Model Intercomparison Project Phase 6 (CMIP6) counterparts allows quantifying the contributions of the land surface parameterization scheme and the atmospheric forcing to the discrepancies. The simulation capabilities are assessed using observational data from 152 sites in Siberia and reanalysis data. On average, the 0.2-m soil temperatures in the CMIP6 simulations are 5.4 °C colder than the observations if the simulated soil temperature drops below -5 °C. The LS3MIP simulations are even colder with a bias of -6.7 °C. In the winter months (December, January, and February), the LS3MIP ensemble diversity in 2-m temperature is less than the CMIP6 diversity (0.8 °C vs 3.2 °C). In contrast, the diversity of winter 0.2-m soil temperatures is larger in the LS3MIP ensemble (5.7 °C) than in the CMIP6 ensemble (3.6 °C). For permafrost sites, the spatial correlation of the simulations of winter soil temperature against observation is not better than 0.7, and spring/autumn spatial correlations of snow depth are less than 0.75 for the CMIP6 models. The biases of 2-m temperature have a different sign and are amplified in magnitude compared to the biases of the soil temperatures, especially below 0 °C. Four of the climate models and their land components underestimate the snow insulation effect. We conclude that land surface models struggle to well simulate soil temperatures and snow depth under low-temperature conditions. The CMIP6 models tend to compensate for errors in their land component by errors through errors in the atmospheric model component. In shallow snow depth (0 to 0.2 °C) cases, all models show between 1 to 8 °C less air-soil temperature difference than in situ data. Therefore, a better representation of surface-soil insulation is essential for improvements in frozen soil land modeling.