Spaceborne lidar observations from the Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite are used to evaluate cloud amount and cloud phase in the Community Atmosphere Model version 5 (CAM5), the atmospheric component of a widely used state‐of‐the‐art global coupled climate model (Community Earth System Model). By embedding a lidar simulator within CAM5, the idiosyncrasies of spaceborne lidar cloud detection and phase assignment are replicated. As a result, this study makes scale‐aware and definition‐aware comparisons between model‐simulated and observed cloud amount and cloud phase. In the global mean, CAM5 has insufficient liquid cloud and excessive ice cloud when compared to CALIPSO observations. Over the ice‐covered Arctic Ocean, CAM5 has insufficient liquid cloud in all seasons. Having important implications for projections of future sea level rise, a liquid cloud deficit contributes to a cold bias of 2–3°C for summer daily maximum near‐surface air temperatures at Summit, Greenland. Over the midlatitude storm tracks, CAM5 has excessive ice cloud and insufficient liquid cloud. Storm track cloud phase biases in CAM5 maximize over the Southern Ocean, which also has larger‐than‐observed seasonal variations in cloud phase. Physical parameter modifications reduce the Southern Ocean cloud phase and shortwave radiation biases in CAM5 and illustrate the power of the CALIPSO observations as an observational constraint. The results also highlight the importance of using a regime‐based, as opposed to a geographic‐based, model evaluation approach. More generally, the results demonstrate the importance and value of simulator‐enabled comparisons of cloud phase in models used for future climate projection.

While the radiative influence of clouds on Arctic sea ice is known, the influence of sea ice cover on Arctic clouds is challenging to detect, separate from atmospheric circulation, and attribute to human activities. Providing observational constraints on the two‐way relationship between sea ice cover and Arctic clouds is important for predicting the rate of future sea ice loss. Here we use 8 years of CALIPSO (Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations) spaceborne lidar observations from 2008 to 2015 to analyze Arctic cloud profiles over sea ice and over open water. Using a novel surface mask to restrict our analysis to where sea ice concentration varies, we isolate the influence of sea ice cover on Arctic Ocean clouds. The study focuses on clouds containing liquid water because liquid‐containing clouds are the most important cloud type for radiative fluxes and therefore for sea ice melt and growth. Summer is the only season with no observed cloud response to sea ice cover variability: liquid cloud profiles are nearly identical over sea ice and over open water. These results suggest that shortwave summer cloud feedbacks do not slow long‐term summer sea ice loss. In contrast, more liquid clouds are observed over open water than over sea ice in the winter, spring, and fall in the 8 year mean and in each individual year. Observed fall sea ice loss cannot be explained by natural variability alone, which suggests that observed increases in fall Arctic cloud cover over newly open water are linked to human activities.

While the representation of clouds in climate models has become more sophisticated over the last 30+ years, the vertical and seasonal fingerprints of Arctic greenhouse warming have not changed. Are the models right? Observations in recent decades show the same fingerprints: surface amplified warming especially in late fall and winter. Recent observations show no summer cloud response to Arctic sea ice loss but increased cloud cover and a deepening atmospheric boundary layer in fall. Taken together, clouds appear to not affect the fingerprints of Arctic warming. Yet, the magnitude of warming depends strongly on the representation of clouds. Can we check the models? Having observations alone does not enable robust model evaluation and model improvement. Comparing models and observations is hard enough, but to improve models, one must both understand why models and observations differ and fix the parameterizations. It is all a tall order, but recent progress is summarized here.