Application of machine-learning methods to the climate of the past millennium

Reconstructing and analysing the climate of the past millennium has traditionally involved using statistical methods to calibrate annually resolved proxies. It also increasingly requires analysing large data sets from ensembles of long climate simulations and paleoclimate reanalysis. The accurate annual dating of most proxies and the increasingly large data sets make machine-learning methods an attractive tool to re-calibrate proxy records and investigate the causality of past climate variability, e.g. extreme events. The available log climate simulations also offer a pre-training data set for other machine-learning applications in climate research, for which the observational records are usually too short.

 

In this talk, I will present a few examples of the application of machine-learning methods to these goals.  Climate reconstructions based on annually resolved proxies can now be produced with methods (Gaussian Process Regression or Long Short Term Memory Networks) that can better preserve the statistical properties of the target variable, like the past amplitude of variations and serial autocorrelation. Causality analysis of past variability episodes, including extremes, can be investigated in climate simulations with  Random Forest and Layerwise Relevance Propagation in neural networks. Finally, data assimilation methods, which blend proxy and model data into a single reconstruction, can be augmented with methods of the family of K-Nearest-Neighbour, thereby also providing an attribution of past climate episodes to one external forcing.