Expanding the palette of potential explanations for change and variability in paleoclimatic reconstructions by applying data-science and machine-learning methods to diagnose climate change and climate variability in transient and long-run snapshot simulations

Abstract:

Despite the availability of transient and long-run (century or more) paleoclimatic simulations, and the high temporal resolution of the resulting model output (routinely monthly, but increasingly daily), the use of those data for explaining the climate changes and variability recorded by the ever-increasing number of paleoclimatic indicators (so-called “proxies”) tends to focus only on long-term means and simple measures of variability. This focus makes sense in some ways, because, for example, it is possible to show that large-scale vegetation distributions are explained by the long-term means of a small number of bioclimatic variables. However, observations of present-day vegetation changes clearly reveal the role of interannual climate variations and the ecological disturbances they produce (e.g. flash droughts, wildfires) in effecting those changes. Short-term climatic variations are also important in generating the “signals” recorded by hydrological indicators.

The use of only long-term means or simple measures of variability may in part be technologically motivated, because any simulation has the potential to produce as much surface-climate data as has ever been observed, plus many unobservable variables. Recent developments accompanying the data-science and machine-learning democratization of data analysis have the potential of greatly diversifying the characterization of paleoclimatic simulations by allowing the standard tools of synoptic and dynamic climatology (e.g. PCA/eigenanalysis, composite anomalies and correlation fields) to be applied to long paleoclimatic simulations. Such applications require reframing our view of the simulations as having two components: climate change, as represented by long-term mean differences, and (temporally) local anomalies, or differences from the varying long-term means.

An example implemented in R to perform large-scale eigenanalysis and “rolling” applications of composite anomaly analyses illustrates, for example, that the relative importance of large-scale controls of interannual variations in drought in the midcontinent of North America vary over the Holocene, with evaporative demand decreasing, large-scale subsidence increasing, and moisture flux into the interior showing little variation.