Probabilistic Inference of Ice Sheet Basal Temperature with Thermal Modelling and Radar Attenuation

Accurate prediction of ice sheet mass balance requires robust understanding of basal conditions, particularly the ice-bed interface temperature. However, thermal modeling predictions of basal temperature are limited by uncertainties in boundary conditions and sparse in-situ validation data.

Ice-penetrating radar wave attenuation has emerged as a promising large-scale proxy for depth-averaged ice temperature. We present three complementary methods to integrate observed attenuation rates with thermomechanical modeling for improved basal temperature estimation: (1) gradient-assisted MCMC coupled with a fast 1.5D enthalpy model for exact Bayesian inference, (2) Gaussian Process Regression combined with 3D enthalpy model ensembles for exact Bayesian inference, and (3) generative AI integrated with 3D enthalpy model ensembles for approximate Bayesian inference. This multi-method approach offers flexibility in balancing computational demands, inference accuracy, and output continuity.

Application to radar attenuation data from the Amundsen Sea Embayment, West Antarctica, reveals widespread thawed conditions near Pine Island Glacier contrasting with heterogeneous basal conditions upstream of Thwaites Glacier. A pronounced basal temperature gradient between these glaciers suggests a significant flow boundary. This radar-and-model-informed basal temperature field represents a crucial step toward assimilating novel observational constraints and improving sliding mechanics in ice sheet models.