Nonlinear paleoclimatology: Reconstructions in west Antarctica

Proxy histories from ice cores, tree rings, and other indicators are required to extend understanding of climate change to before instrumental records. Available observational and instrumental records are too limited spatially and temporally to fully characterize natural variability, particularly in the polar regions. A suite of highly adaptive, nonlinear tools based on artificial neural networks (ANNs) assembled here provides improved knowledge of West Antarctic climate and paleoclimate, and can be applied to many other regions and proxy types.; Available histories of West Antarctic surface conditions away from coastal stations are primarily from automatic weather stations (AWS), but are short and discontinuous owing to logistical difficulties. Feed-forward ANNs (FF-ANNs) and free-atmosphere reanalysis data from the European Centre for Medium-Range Weather Forecasting (ECMWF ERA-15) have been used to extend and fill the AWS temperature and pressure data for the period 1979–1993, producing a more-consistent and more-accurate climate history than was previously available. Six West Antarctic AWS sites from the Ross Ice Shelf (Lettau, Elaine, Marilyn, Ferrell) and from elevation on the ice sheet (Siple Station, Byrd) are included in this database.; To better understand the free-atmosphere conditions recorded at the surface, ANN-based self-organizing maps (SOMs) were used to objectively extract the generalized patterns or archetypes from the multidimensional, nonlinear, and noisy data sets without supervision. This classification groups years with similar patterns and also identifies transitional states not in the input data. Anomalous years such as the warm conditions of 1980 and 1988 are evident in surface conditions and are linked to large-scale free-atmosphere features including shifts in the position and strength of the Amundsen Sea Low.; The clear linkage between surface conditions and SOM “archetype” states allows reconstructions of local and large-scale conditions from both direct (e.g., AWS) and proxy (e.g., ice-core) data. FF-ANNs were trained to predict SOM states from high-resolution records of accumulation rate and ice-core chemistry. Within the ERA-15 time span, the predictions show skill, and for older times different combinations of proxy indicators produce similar predictions. Greater skill appears likely if additional ice-core data and longer reanalysis records were available.