State-of-the-art ice-sheet models are computationally expensive to run. A single simulation may require hours to days of computing time, making it difficult to explore uncertainties in future climate forcing, model parameters, and physical assumptions. This is a major obstacle when estimating the possible contribution of ice sheets and glaciers to future sea-level rise.

The goal of this project is to develop a fast emulator that reproduces the behaviour of a complex ice-sheet model. Once trained on a limited ensemble of model simulations, the emulator can be used to predict ice-sheet evolution under a wide range of future climate scenarios at a fraction of the computational cost.

Unlike purely statistical or ML approaches, the emphasis here is on developing physically motivated emulators that incorporate our understanding of glacier and ice-sheet dynamics. The aim is to build simple yet realistic models that capture the essential relationship between climate forcing and ice loss.

Possible approaches include:

  • Developing low-order models of ice-sheet mass evolution, for example dMdt=f(M,T,P,…) \frac{dM}{dt} = f(M,T,P,\ldots) where ice loss depends on the current ice mass and climate forcing.
  • Using scaling relationships between ice-sheet mass, volume, area, thickness, and length to constrain model behaviour.
  • Separating accumulation (snowfall) and ablation (runoff and melt) processes rather than treating climate forcing as a single variable.
  • Training and validating emulators against large model intercomparison projects such as ISMIP or GlacierMIP.
  • Developing regional emulators for different sectors of Greenland or Antarctica.
  • Building glacier-scale emulators and combining them to estimate large-scale ice loss.

The project may involve a combination of:

  • Physical modelling.
  • Statistical inference and calibration.
  • Machine learning.
  • Time-series analysis.
  • Sea-level projection and uncertainty quantification.

A successful emulator could enable thousands or millions of model evaluations, making it possible to characterize uncertainties that are currently computationally inaccessible with full ice-sheet models.

Background and inspiration

Students

  • Jakob
  • Nikolaj