Kernel-based Bayesian Inference for Paleo-Climate Reconstruction
Please login to view abstract download link
Paleo-climate reconstruction aims at estimating how Earth’s climate, in particular surface temperature, has changed over timescales that extend far beyond the period covered by instrumental observations. Our region of study is the East Antarctic Plateau, where the temperature evolution of even the last two centuries is still largely unknown. We have obtained temperature profiles from ice boreholes in the region of interest. These are "proxies" for the past temperature evolution, since they are the result of a heat transfer from the surface into the ice layers. To recover past temperatures, we need to solve an inverse problem, in which the forward model is a heat transfer equation, where the top boundary condition over time corresponds to the to-be-recovered surface temperature. We have developed a kernel-based Bayesian inference framework for the temperature reconstruction task. It models surface temperature using a kernel-based ansatz. We obtain the posterior distribution using a Markov chain Monte Carlo ensemble sampler. In my talk, I will introduce the application, have a detailed look at the modeling of the inverse problem and will report about the latest results of our kernel-based inference approach.
