Relating West Antarctic surface meteorology to the large-scale atmospheric circulation

A deeper understanding of Antarctic meteorology is one key to improved interpretation of the ever-growing body of ice-core-based paleoclimate records from this region. Automatic weather stations (AWS) currently provide the only year-round, direct measurements of surface weather on the ice sheet. As the spatial coverage of the network has expanded year to year, so has our meteorological database. Unfortunately, many of the records are relatively short (less than 10 years) and/or incomplete (to varying degrees) due to the vagaries of the harsh environment. Recent developments in climate downscaling in temperate latitudes suggest it may be possible to use GCM-scale meteorological data sets (e.g., ECMWF reanalysis products) to both fill gaps in the AWS records and extend them back in time to create a uniform and complete database of West Antarctic surface meteorology (at AWS sites). Research with northern hemisphere mid-latitude data sets has shown that artificial neural network (ANN) techniques support nonlinear downscaling from regional to local scales. For the West Antarctic domain, ANNs are used to predict known AWS surface data (e.g., temperature, pressure) using large-scale features of the atmosphere (e.g., 500 mb pressure height). Once trained, the ANN can predict from the GCM-scale data for periods lacking AWS data (e.g., prior to installation). Thus, surface records can be reliably extended into both pre- and post-AWS periods given appropriate GCM-scale data.

Creating the ideal ANN remains a highly subjective task due to the large number of parameters available and the difficulties associated with optimization on the complex error surface that results. Correlation coefficients and other simple statistical tools, such as RMS error, have allowed us to evaluate the performance of various ANN configurations. For example, 850 mb heights alone predicted pressure better than the combination of 700 and 500 mb heights. Confidence in the optimization of a particular ANN configuration has been increased by choosing the best result from an ensemble of multiple runs of a fixed configuration. This approach has greatly reduced the likelihood of picking an ANN trained with an unrepresentative subset of the input domain.

Results from filling the gaps in an existing, long record and the extension of a short record into the pre-AWS period will be discussed.