Editing IPCC:AR6/WGI/Chapter-Atlas (section)

==== Atlas.11.2.3 Assessment of Model Performance ====

<div id="h3-64-siblings" class="h3-siblings"></div>

Evaluating simulated temperature and precipitation is problematic in the Arctic due to sparse weather station observations. The lack of reliable observed precipitation datasets for the Arctic thus makes it ''very unlikely'' to be able to evaluate objectively the skill of models to reproduce precipitation patterns ( [[#Takhsha--2018|Takhsha et al., 2018]] ).

The CMIP5 models reproduce the observed Arctic warming over the past century ( ''medium confidence'' ) ( [[#Chylek--2016|Chylek et al., 2016]] ; [[#Hao--2018|Hao et al., 2018]] ; [[#Huang--2019|Huang et al., 2019]] ). The simulated mean Arctic warming for 1900–2014 averaged over 40 CMIP5 models is 2.7°C compared to the observed values of 2.2°C (NASA GISS data smoothed using a 1200-km radius) or 1.7°C (using a 250-km smoothing radius) ( [[#Chylek--2016|Chylek et al., 2016]] ). However, there are large inter-model differences in the simulated warming which ranges from 1.2°C to 5.0°C. Although the CMIP5 models reproduce the spatially averaged observed warming over the past 50 to 100 years, the pattern is different from that of observations and reanalysis ( [[#Xie--2016|Xie et al., 2016]] ; [[#Franzke--2017|Franzke et al., 2017]] ; [[#Hao--2018|Hao et al., 2018]] ). Zonal mean temperature trends in the CMIP5 models overestimate the warming in the cold season over high latitudes in the Northern Hemisphere ( [[#Xie--2016|Xie et al., 2016]] ). Overall, the amplified Arctic warming in recent decades is overestimated by CMIP5 models ( [[#Huang--2019|Huang et al., 2019]] ). Possible reasons are modelled sea surface temperature biases and an overestimated temperature response to the Arctic sea ice decline. Furthermore, some models, which have a warm or weak bias in their Arctic temperature simulations, closely relate the Arctic warming to changes in the large-scale atmospheric circulation. In other models, which show large cold biases, the albedo feedback effect plays a more important role for the temperature trend magnitude. This implies that the dominant simulated Arctic warming mechanism and trend may be dependent on the bias of the model mean state ( [[#Franzke--2017|Franzke et al., 2017]] ). Compared to CMIP5 models, [[#Davy--2020|Davy and Outten (2020)]] found lower biases in CMIP6 models’ representation of sea ice extent and volume with improved extents linked to a better seasonal cycle in the Barents Sea.

Rapid temperature changes, such as the pronounced increase of 2°C yr <sup>–1</sup> during 2003–2012 over the Kara and Barents seas in March is well captured in Arctic CORDEX simulations ( [[#Kohnemann--2017|Kohnemann et al., 2017]] ). The models show adequate skill in capturing the general temperature patterns ( [[#Koenigk--2015|Koenigk et al., 2015]] ; [[#Matthes--2015|Matthes et al., 2015]] ; [[#Hamman--2016|Hamman et al., 2016]] ; [[#Cassano--2017|Cassano et al., 2017]] ; [[#Brunke--2018|Brunke et al., 2018]] ; [[#Diaconescu--2018|Diaconescu et al., 2018]] ; [[#Takhsha--2018|Takhsha et al., 2018]] ), but tend to show a cold temperature bias which is largest in winter and depends on the reference dataset. [[#Cassano--2017|Cassano et al. (2017)]] showed a large sensitivity of the simulated surface climate to changes in atmospheric model physics. In particular, large changes in radiative flux biases, driven by changes in simulated clouds, lead to large differences in temperature and precipitation biases.

The CMIP5 models perform well in simulating 20th-century snowfall for the Northern Hemisphere, although there is a positive bias in the multi-model ensemble relative to the observed data in many regions ( [[#Krasting--2013|Krasting et al., 2013]] ). Lack of sufficient spatial resolution in the model topography has a serious impact on the simulation of snowfall. The patterns of relative maxima and minima of snowfall, however, are captured reasonably well by the models.

Arctic CORDEX RCMs reproduce the dominant features of regional precipitation patterns and extremes (e.g., [[#Glisan--2014|Glisan and Gutowski, 2014]] ; [[#Hamman--2016|Hamman et al., 2016]] ). Due to their higher spatial resolution, RCMs simulates larger amounts of orographic precipitation compared to reanalyses. Overall, the simulated precipitation is within the reanalysis and global model ensemble spread, but the Arctic river basin precipitation is closer to observations ( [[#Brunke--2018|Brunke et al., 2018]] ). However, [[#Takhsha--2018|Takhsha et al. (2018)]] show that the RCMs’ precipitation bias highly depends on the observational reference dataset used.

The annual mean precipitation pattern of ensemble global atmospheric simulations with a high horizontal resolution agrees well with the observations, with precipitation maxima over the Greenland and Norwegian seas ( [[#Kusunoki--2015|Kusunoki et al., 2015]] ). However, the simulated Arctic average annual precipitation shows a positive bias with excessive precipitation over Alaska and the western Arctic ( [[#Kattsov--2017|Kattsov et al., 2017]] ).

Regarding the Greenland Ice Sheet (region GIC), modelled surface mass balance (SMB) has decreased since the end of the 1990s ( [[#Fettweis--2020|Fettweis et al., 2020]] ). A multi-model intercomparison study ( [[#Fettweis--2020|Fettweis et al., 2020]] ) emphasized a simulated positive mean annual SMB of 338 ± 68 Gt yr <sup>–1</sup> between 1980 and 2012, with a decreasing average rate of 7.3 ± 2.0 Gt yr <sup>–2</sup> , mainly driven by an increase in meltwater runoff. [[#Mouginot--2019|Mouginot et al. (2019)]] stated that SMB played a strong role in the ice-sheet mass loss, where SMB dominated in the last two decades. [[#Mottram--2019|Mottram et al. (2019)]] found that SMB processes dominate the ice-sheet mass budget over most of the interior, highlighting that the ice sheet is a contributor to global mean sea level rise between 1991 and 2015. More specifically, SMB models have improved ( [[#Fettweis--2020|Fettweis et al., 2020]] ; [[#Hanna--2021|Hanna et al., 2021]] ) due to increased availability and quality of remotely sensed ( [[#Koenig--2016|Koenig et al., 2016]] ; [[#Overly--2016|Overly et al., 2016]] ) and in situ observations ( [[#Machguth--2016|Machguth et al., 2016]] ; [[#Fausto--2018|Fausto et al., 2018]] ; [[#Vandecrux--2019|Vandecrux et al., 2019]] , 2020). [[#Fettweis--2020|Fettweis et al. (2020)]] showed that the models’ ensemble mean provides the best estimate of the present-day SMB relative to observations. This is the case for the patterns in all seven regions (regional division after [[#Mouginot--2019|Mouginot et al., 2019]] ) apart from the SE accumulation zone where large discrepancies in modelled snowfall accumulation occurred where the spread can reach 2-m water equivalent per year. [[#Montgomery--2020|Montgomery et al. (2020)]] confirmed this, highlighting that RCMs (MAR and RACMO) are underestimating accumulation in south-east Greenland and that models misrepresent spatial heterogeneity due to an orographically forced bias in snowfall near the coast. Further, for north-east Greenland, [[#Karlsson--2020|Karlsson et al. (2020)]] found RCMs underestimate snow accumulation rates by up to 35%. The regional time series show that SMB has been gradually decreasing in all seven regions (1979–2017), although the trend is less strong in central-eastern and south-eastern regions. In the south-west, north-east and north-west, SMB turns negative or close to zero after 2000 and remains above zero in other regions ( ''medium confidence'' ) (Figure Atlas.3 0).

<div id="Atlas.11.2.4" class="h3-container"></div>

<span id="atlas.11.2.4-assessment-and-synthesis-of-projections"></span>