Bridging Antarctic observations and models with deep learning
Abstract
Deep learning is a promising method to accelerate scientific discoveries in cryospheric sciences. In this talk I will discuss a few recent progress we made using supervised and unsupervised, classical deep learning methods and physics-informed deep learning methods to help understand processes that govern ice-shelf dynamics. Ice shelves play a crucial role in slowing glacier flow into the ocean which impacts the global sea-level rise. The flow of glaciers is governed by the ice rheology, but direct measurement of the rheology of glacial ice is challenging. Physics-informed neural networks (PINNs) have recently emerged as a new class of numerical solver for partial differential equations, leveraging deep neural networks constrained by equations. I'll start with a simple idealized example with synthetic data to demonstrate how PINN can be trained to predict ice hardness, and show the importance of weighing the physics constraints in the presence of data noise. Then I will discuss how PINNs trained with real world data from Antarctica and the 2-dimensional shallow-shelf equations can help discover ice rheology that govern ice-shelf dynamics. We used PINN to solve the governing equations for ice shelves and invert for its viscosity. Our calculation yields flow laws of ice shelves that are different from commonly assumed Glen's flow law, and suggests the need for reassessing the impact of our finding on the future projection of sea-level rise.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.C56A..03L