All Climate Events

Evaluating Last Interglacial Constraints on Future Antarctic Ice-Sheet Mass Losses with a Supervised Machine Learning Technique

Download as iCal file

Wednesday, 25 September 2019, 12:00

Wednesday, September 25, 2019. 12:00PM. Evaluating Last Interglacial Constraints on Future Antarctic Ice-Sheet Mass Losses with a Supervised Machine Learning Technique. Daniel Gilford, Rutgers University. Sponsored by NOAA GFDL. More information here

 

Antarctic ice loss has accelerated in recent decades, and its contributions to global sea level rise could be as high as ~1 meter by the end of the century. But deep uncertainties in ice-sheet physics make it difficult to produce robust projections of sea level rise. The lack of agreement among experts and poor constraints on ice-sheet model parameters exemplify this ambiguity. Past warm periods provide useful analog climates where sea levels were much higher than present, but these paleoclimate estimates are themselves deeply uncertain. It is critical to assess the efficacy of paleoclimate estimates constraining future Antarctic ice-sheet mass losses to (1) improve future projections of sea level rise and (2) inform paleoclimate research and observational goals. Here we use Gaussian Process regression to develop a statistical “emulator” designed to mimic the behavior of an ice-sheet model. Gaussian process modeling is a non-parametric supervised machine learning technique which maps inputs (e.g. ice-sheet model parameters) to target outputs (e.g. sea-level contributions from Antarctica) with explicit quantification of model uncertainties. Emulation is applied over the last interglacial (LIG; a warm period ~125,000 years ago) and a future high greenhouse gas emissions scenario, and is trained on ice-sheet model ensembles constructed by varying the parameters of maximum rate of ice-cliff loss and the coefficient of ice-shelf hydrofracturing. Emulated projections are calibrated with various LIG constraints on Antarctic sea level contributions to find the likely configurations of model parameterizations. Using a Bayesian approach, we quantify uncertainties in future projections from the ice-sheet model if the LIG estimated contributions were well-known. Results also show how specific assumptions about the LIG period (such as those used in past studies) can influence future projections, and highlight the importance of continued observational efforts.

Location  NOAA GFDL, Smagorinsky Seminar Room, Princeton University, Princeton, NJ.