Improving the Performance of Ice Sheet Modeling Through Embedded Simulation

• Main Author: Dufour, Christopher G
• Format: Text
• Language: unknown
• Published: DigitalCommons@UMaine 2016
• Subjects: ice sheet model; high performance computing; embedded simulation; interpolation; Computer Sciences; Geographic Information Sciences; Numerical Analysis and Scientific Computing; Systems Architecture; Greenland; Antarc*; Antarctica; Ice Sheet
• Online Access: https://digitalcommons.library.umaine.edu/etd/2473; https://digitalcommons.library.umaine.edu/context/etd/article/3542/viewcontent/Dufour_Thesis.pdf

Understanding the impact of global climate change is a critical concern for society at large. One important piece of the climate puzzle is how large-scale ice sheets, such as those covering Greenland and Antarctica, respond to a warming climate. Given such ice sheets are under constant change, developing models that can accurately capture their dynamics represents a significant challenge to researchers. The problem, however, is properly capturing the dynamics of an ice sheet model requires a high model resolution and simulating these models is intractable even for state-of-the-art supercomputers. This thesis presents a revolutionary approach to accurately capture ice sheet dynamics using embedded modeling at a high resolution. Such an approach embeds a high-resolution ice sheet model of a region evolving rapidly within a low-resolution ice sheet model of areas evolving slowly. The embedded model approach was implemented within the Parallel Ice Sheet Model (PISM), a widely used model for the study of large scale ice sheets limited to simulating models in isolation. PISM is limited to simulating ice sheet models in isolation and thus implementing an embedded model requires new synchronization and communication schemes. In this work we analyze the accuracy of our prototype embedded model with respect to directly observed ice velocities. We have shown a stronger correlation to directly observed values, yielding a T-test value of 0.64, compared to a non-embedded model T-test of 0.02.