Master thesis defense by Øjvind Andreas Winton

Title: Scientific machine learning for discovering basal dynamics of Greenland outlet glaciers

Abstract: Basal dynamics of outlet glaciers are essential in predicting the Greenland
ice sheets’ contribution to sea-level rise. Despite this, they are poorly constrained by observations and generalizable models. This study combines physical models and machine learning to discover relations for basal stress
from data through new system identification and parameter estimation techniques under the shallow shelf approximation. The most generalizable model discovered was τ ∝ u −1/2 + 1.5us, which explains 30% of the variance of a spatio-temporal extrapolation test data set. The most generalizable extended power-law formulation was τ ∝ u 0.31s 0.45 explaining 24% variance. 
Fitting parametric models to each glacier individually yielded models that explain 77% of the variance in temporal extrapolation. The models for individual glaciers generally have negative exponents for velocity, contrary to the commonly assumed positive exponent. No relation between basal stress and meltwater was identified. This study lays the foundation for further
attempts to discover ice-flow dynamics from data.

Supervisors: Aslak Grindsted (NBI, University of Copenhagen), Allan P.
Engsig-Karup (DTU Compute) and Sebastian B. Simonsen (DTU Space)