Diagnosing cloud occurrence biases in the AM3 using atmospheric classification

Abstract:

Parameterizations in models attempt to statistically relate large-scale atmospheric variables such as temperature, humidity, and winds with variables that change on scales too small to resolve, such as cloud properties. We use atmospheric classification to establish the relationships between large- and small-scale variables in observations and compare these relationships to those that exist in GFDL’s AM3 model.

Atmospheric states for a region surrounding the ARM Southern Great Plains (SGP) site are created with a clustering technique. Atmospheric state in this context can be thought of as a frequently occurring regional weather pattern. States are defined by ERA-Interim reanalysis data and validated with observations from the SGP cloud radar. We classify the state of the atmosphere every 6 hours, creating a time series of atmospheric state. This time series is used to composite ISCCP data to create joint cloud top pressure – optical depth histograms of cloud occurrence for each state.

Snapshots from the AM3 model are sorted into the observed atmospheric states, and cloud occurrence data from the ISCCP simulator are composited to produce modeled joint histograms of cloud occurrence. Comparison of the observed and modeled cloud occurrence for a single state provides a test of the model parameterization under a particular set of physical conditions. In contrast, comparing the observed frequency of occurrence of the atmospheric states with their occurrence within AM3 tests how well the model reproduces the conditions of the region. Doing so allows us to parse the model’s total bias of cloud occurrence into contributions from the parameterization, and contributions from the distribution of states.

We find that the model lacks high thin cloud under all conditions in the model, but that biases in deep thick cloud are state-dependent. We show that frontal conditions in the model do not produce enough deep thick cloud, while weather patterns associated with isolated convection produce too much. We find that increasing the horizontal resolution of the model improves both of these biases, but also changes the distribution of states, ultimately increasing the total cloud occurrence bias. This creates an interesting question of whether or not the high resolution run performs better than the low resolution.