Radar forward modelling as a precursor for statistical inference

Projections of the future development of the Antarctic Ice Sheet still exhibit a large degree of uncertainty due to difficulties in constraining parameters of ice-flow models such as basal boundary conditions. Deriving better estimates of these parameters from radargrams could greatly improve model accuracy, but integration of inferred radar attributes into ice-flow models is not yet widespread.

Here, we develop a radar forward modeling framework that is geared to train a machine learning workflow (likely simulation-based inference) to extract radar attributes such as the internal stratigraphy and basal boundary conditions (e.g., frozen vs. wet) from radar data. The workflow starts with ice-dynamic forward models predicting physically sound stratigraphies and internal/basal temperatures for synthetic flow settings using shallow ice, shallow shelf and higher order ice-flow models. This is then used as input to the radar simulator (here gprMax), which calculates the radar image produced by such a stratigraphy. To do so, we match the synthetic permittivities of the modeled stratigraphy with statistical properties known from ice-core logging data and prescribe temperature dependent attenuation via an Arrhenius relation. gprMax is optimized for acceleration using GPUs which can be efficiently employed when solving the FDTD discretized Maxwell equations. Currently, 200 m wide and 500 m deep sections can be simulated on a single NVIDIA GeForce RTX 2070 Super graphics card within 390 minutes. The runtime can be substantially improved in a HPC environment. In order to obtain radar simulations comparable with observations, we also add system specific noise and contributions from volume scattering with variable surface roughness. Here, we focus on 50 MHz pulse radar for which we have many observational counterparts. However, the workflow is designed to encompass multiple ice-dynamic settings and different radar frequencies.

The application of physical forward models will result in physically meaningful radargrams which are indistinguishable from observations. This provides a tool to create datasets for training machine learning workflows for inference without the limitations of hand-labeled data.