Exploring the robustness of a surrogate-based ice flow model calibration

Abstract

When simulating ice sheets using numerical models, model parameters have a great impact on the ice flow, velocity, and response to external forces, making them key factors in accurately predicting ice sheet behavior. In order to have confidence in modeled predictions of ice sheets, we must first have confidence in the physical models we use to simulate them. In order to have confidence in these physical models, we must first have confidence in our calibrations of these ice dynamics parameters. Researchers have been training neural networks to emulate expensive ice sheet models. These surrogate ice sheet models are designed to take ice flow parameters as input, and to output ice surface speed fields that closely resemble modeled fields that physically based ice sheet models would calculate. In this way, these surrogate models act as a computationally inexpensive alternative to map model parameter values to the resulting calculated ice speeds. These surrogates have been implemented and leveraged for Bayesian statistically based approaches to ice flow parameter calibrations that would otherwise be intractable using a high fidelity ice sheet model. By examining the methods we use to train these surrogates, and the extent to which the architecture of these surrogates affect their performance, we can have more confidence in our calibrations that make use of them. Here we present an analysis of the methods with which we train surrogate models as a means of calibrating important ice flow parameters. We focus on determining desirable characteristics for data over which the neural networks are trained, as well as the architecture of the surrogates themselves.