Master Thesis Defense by Josephine Gondán Kande

Title: Predicting Greenland Ice Sheet Mass Balance - Exploring Training Strategies for Machine Learning

Abstract:
Climate change, primarily driven by global warming, is impacting various components of our climate system (Arias et al., 2021). A key consequence is the rising sea level, which is largely attributed to the melting of land-based ice masses such as the Greenland Ice Sheet (GrIS). Accurate models of GrIS mass loss are thus needed for predicting sea level rise, however, such models are both complicated and computational expensive to run.

This study investigates the potential of machine learning (ML) models to predict ice sheet mass change, with a focus on XGBoost as the primary algorithm. Three experiments are conducted, training and predicting on different combinations of observational temperature data from Ilulissat and mass loss data from the Grace project. Additionally, the experiments use modeled mass loss from the PISM ice flow model together with the NGRIP ice core temperature record.

The experiments reveal the limited amount of observed data available can give rise to complications and overfitting of the models. Cross Validations is used to prevent that and better train on sparse data. The project also investigate for using Machine Learning on modeled historic and future data to test the model over longer timescales. This is however problematic, as even though the ML model successfully captures the general mass loss trajectory, it struggles to handle variability, particularly when the training data differs significantly from the data used for prediction.

These findings suggest that while ML can provide valuable insights into ice sheet dynamics, its ability to predict under unobserved conditions remains limited, and future work should focus on incorporating more physical processes, higher temporal resolution, and a broader range of training data to improve model robustness and accuracy.

Supervisor: Bo Møllesøe Vinther