Atmospheric classification at Darwin

Poster PDF

Description

Unlike numerical weather prediction models, general circulation models do not make predictions of what the weather will be at a particular point in time, thus making direct moment-to-moment comparison to observations impossible. Instead, observations and model output must be temporally averaged over long periods to create statistical distributions of both observations and model output, which can then be compared to each other. While comparing these averages is often effective in determining the presence of errors in the model, it generally provides little insight on the source of the errors. Not being able to identify when or under what conditions the errors occur makes it difficult to understand which physical processes are not being properly represented in the model and, in turn, what corrective measures should be taken. To avoid this problem, we classify observations into a number of large-scale atmospheric states, as defined by a neural network pattern recognition program. Averaging of both observations and model output is then done within each category. With this method, when errors appear between the observed and modeled averages for a particular atmospheric state, the physical circumstances which produced the error are known, helping to identify the physical processes which are insufficiently represented by the model. Previous work by Marchand and coauthors (2006, J.A.S. and 2009, J. Climate) has shown that the approach works well over the U.S. Southern Great Plains. However, the classifier struggled to find statistically meaningful states during the summer. This raised questions as to the effectiveness of the technique for convective atmospheres. Here, we apply the classification to Darwin, Australia, and explore the sensitivity of the method to a variety of inputs. We use two years of ECMWF reanalysis data for a region surrounding Darwin as our input to the classifier, and then create vertical cloud occurrence profiles for each state using data from the vertically pointing millimeter radar at the Department of Energy ARM Climate Research Facility site at Darwin as a diagnostic tool. In particular, we explore the importance of domain size, horizontal resolution, and input variable selection to the states created. We show that, with only minor variations, this method produces a robust set of states independent of the precise input configuration.