Climate models are an essential tool for understanding future climate and preparing for the challenges of climate change. While these models are far from perfect, they are constantly being evaluated and updated. To ensure that future simulations are as accurate as possible, the emphasis of this evaluation is on the processes that make up the model and not the model as a whole. This is because compensating errors can result in a model that compares well with observations, but only because errors in different processes are canceling each other out. Climate model evaluation is a continuous effort with errors in different processes continually discovered and then removed.

Clouds have been consistently identified as poorly simulated in climate models and are one of the largest sources of uncertainties in their projections. The poor simu- lation of clouds drives several different errors in climate models such as the long standing excess of solar radiation over the Southern Ocean. This error is known as the Southern Ocean radiative bias has been specifically linked to the representation of cloud phase in the models. This thesis is focused on better understanding the relationship between model cloud representation and the Southern Ocean radiative bias.

The primary approach we use to examine the model cloud representation is a ma- chine learning technique known as self-organizing maps. We use this algorithm to define a set of clusters which represent the different types of clouds that occur within the Moderate Resolution Imaging Spectroradiometer (MODIS) dataset. These clus- ters form the basis for all of the research in this thesis. Strong model biases in the occurrence rates of the different clusters are immediately detected with one cluster changing from a 13% occurrence rate in the observations to 43% in the model. These clusters also allow us to link model biases directly to the representation of particu- lar cloud types. This approach is primarily used to analyze the relationship between clouds and radiative fluxes in the model. From this analysis we are able to identify which cloud types contribute the most to radiative errors over the different regions of the globe.

After initially identifying model biases that were related to the clusters, our focus shifted to better understanding the behaviour of the clusters in both the model and observations. First we introduced additional statistical techniques, that are focused on understanding the variability within the clusters. The majority of the clusters are shown to capture their constituent measurements well, with clusters 5 and 6 standing out as exceptions. This analysis is refined with the usage of sub-clusters which are identified for each of the clusters. A novel statistical measure known as the subsom entropy is introduced to further analyze the sub-clusters.

Additional data is also introduced to explore the relationship between cloud phase and the clusters. This includes reflectivity-altitude histograms from the CloudSat dataset and phase specific cloud fraction values from the MODIS dataset. The CloudSat data shows results that strongly agreed with our earlier interpretation of the clusters, therefore independently validating our interpretation. Major model is- sues simulating the phase specific cloud fraction are also identified. Both the MODIS and CloudSat data show that the ice cloud is restricted to three of the clusters, how- ever the model is unable to replicate this restriction. This shows that the model has fundamental issues replicating the relationship between cloud type and ice phase cloud.

To investigate the relationship between the identified radiative biases and ice phase biases, several different model runs are inter-compared. Each of these model runs has a modified set of ice phase parameterizations and by contrasting them we hope to identify which of the parameterizations best captures reality. This analysis is based around using the clusters to calculate the mean error and estimate the mag- nitude of compensating errors. While only minor differences are identified between the parameterizations, these comparisons allow us to better understand the model biases. For example, the global biases are shown to be comprised of large compen- sating errors while over the Southern Ocean compensating errors are shown to play a less dominant role.