Diagnosing Cloud Biases in the GFDL AM3 Model With Atmospheric Classification. Issue 23 (4th December 2017)

• Record Type: Journal Article
• Authors: Evans, Stuart; Marchand, Roger; Ackerman, Thomas; Donner, Leo; Golaz, Jean‐Christophe; Seman, Charles
• Is Part Of: Journal of geophysical research. Volume 122:Issue 23(2017)
• Page Start: 12, 827
• Page End: 12, 844
• Publication Date: 2017-12-04
• Subjects: clouds -- model evaluation -- clustering -- classification -- GCMs -- weather patterns; Atmospheric physics -- Periodicals; Geophysics -- Periodicals; 551.5
• DOI: 10.1002/2017JD027163
• Languages: English
• ISSNs: 2169-897X

Abstract

Abstract: We define a set of 21 atmospheric states, or recurring weather patterns, for a region surrounding the Atmospheric Radiation Measurement Program's Southern Great Plains site using an iterative clustering technique. The states are defined using dynamic and thermodynamic variables from reanalysis, tested for statistical significance with cloud radar data from the Southern Great Plains site, and are determined every 6 h for 14 years, creating a time series of atmospheric state. The states represent the various stages of the progression of synoptic systems through the region (e.g., warm fronts, warm sectors, cold fronts, cold northerly advection, and high‐pressure anticyclones) with a subset of states representing summertime conditions with varying degrees of convective activity. We use the states to classify output from the NOAA/Geophysical Fluid Dynamics Laboratory AM3 model to test the model's simulation of the frequency of occurrence of the states and of the cloud occurrence during each state. The model roughly simulates the frequency of occurrence of the states but exhibits systematic cloud occurrence biases. Comparison of observed and model‐simulated International Satellite Cloud Climatology Project histograms of cloud top pressure and optical thickness shows that the model lacks high thin cloud under all conditions, but biases in thick cloud are state‐dependent. Frontal conditions in the model do not produce enough thick cloud, while fair‐weather conditions produce too much. We find that increasing the horizontal resolution of the model improves the representation of thick clouds under all conditions but has little effect on high thin clouds. However, increasing resolution also changes the distribution of states, causing an increase in total cloud occurrence bias. Plain Language Summary: Models generally struggle to simulate clouds. Identifying the processes that cause errors is an important step toward improving them. We define a set of weather patterns for a region in the Great Plains that represent different physical processes and evaluate a climate model's cloud occurrence for each of those patterns. The model underpredicts cirrus clouds for all patterns. For thick clouds, however, the model overpredicts during fair‐weather conditions and underpredicts during stormy conditions. These errors tend to balance each other out. When the model resolution is improved, it does a better job of predicting thick clouds for most weather patterns, but now, the errors no longer balance each other. The result is that the model with better resolution has a worse overall prediction of thick clouds, despite better predictions for most individual patterns. Evaluating models by pattern, rather than just the overall total, helps to identify when there are underlying improvements that might be missed otherwise. Doing so may be valuable for future efforts toward improving the simulation of clouds in models. Key Points: Clustering defines a set of atmospheric states for the great plains that are valuable tools for evaluation of model cloud occurrence The NOAA/GFDL AM3 model does not produce enough cirrus but has counterbalancing errors in thick cloud occurrence Increasing model horizontal resolution improves state cloud properties but removes counterbalancing errors and increases the overall bias