Hindcasting to measure ice sheet model sensitivity to initial states
Validation is a critical component of model development, yet notoriously challenging in ice sheet modeling. Here we evaluate how an ice sheet system model responds to a given forcing. We show that hindcasting, i.e. forcing a model with known or closely estimated inputs for past events to see how well the output matches observations, is a viable method of assessing model performance. By simulating the recent past of Greenland, and comparing to observations of ice thickness, ice discharge, surface speeds, mass loss and surface elevation changes for validation, we find that the short term model response is strongly influenced by the initial state. We show that the thermal and dynamical states (i.e. the distribution of internal energy and momentum) can be misrepresented despite a good agreement with some observations, stressing the importance of using multiple observations. In particular we identify rates of change of spatially dense observations as preferred validation metrics. Hindcasting enables a qualitative assessment of model performance relative to observed rates of change. It thereby reduces the number of admissible initial states more rigorously than validation efforts that do not take advantage of observed rates of change.