CAREER: Dimensionality reduction of glacier and ice-sheet processes using deep learning: Development, and evaluation

About This Grant

The current generation of ice-sheet models underpredicts observed rates of mass loss of the Greenland Ice Sheet, in part because these models are inherently incomplete representations of complex systems. This work aims to reduce this bias by increasing the complexity of glacier processes that are represented in the next generation of ice-sheet models. Addition and refinement of specific glacier processes to ice-sheet models will improve the accuracy of future sea-level projections, helping world communities plan for future infrastructure and societal consequences of sea-level change. However, adding specific glacier processes to ice-sheet models raises the computational demand of the already severely resource-constrained operation of those models. This project addresses this obstacle by using artificial intelligence (AI) techniques to produce high-efficiency representations of complex glacier processes that slot into ice-sheet models without substantially increasing their computational costs. The project condenses three specific glacier processes into efficient modules, produces datasets through AI-based analyses of remote-sensing products to fine-tune and evaluate those modules, and evaluates their efficacy in improving ice-sheet model predictions. The modules can be readily applied within any of the 15 plus active ice-sheet models. These outcomes are expected to improve the accuracy of forecasts of future sea-level rise. The project facilitates critical quantitative thinking about future worlds by the next generation of students to prepare them to enter the workforce. It accomplishes this through extensive contact with maturing thinkers in high schools and universities through research–education integration across Earth sciences and AI, emphasizing crucial transferrable skills across science, technology, engineering, and math (STEM). Although the glaciology community has a robust understanding of sub-grid-scale glacier processes, such as crevassing and the development of glacier hydrologic systems, there are computational barriers for their inclusion into ice-sheet models. Notably, ice-sheet models typically use a resolution on the scale of kilometers, which is considerably coarser than required to resolve, for example, the formation and evolution of crevasse fields that collect and drain surface meltwater. This work uses parameterization, deep learning, and AI to represent key glacier processes at a level for incorporation into the next generation of ice-sheet models. This work connects three specific processes – crevassing, structural glaciology, and englacial and subglacial hydrology – across the separate realms of process-scale development and large-scale ice-sheet modeling. The connection of these glacier processes into ice-sheet models ensures that resource investments into ice-sheet modeling will deliver on the promise to improve sea-level forecasts. This research capitalizes on existing process-scale models, remote sensing products, and deep-learning techniques to synthesize specific glacier processes into the ice-sheet models that make sea-level projections. This project connects deep learning and AI in Earth science to high schools and develops the STEM workforce by training university students. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.