CSSI Elements: Community software for data-driven inference of ice-sheet physics

Earth’s ice sheets are melting at an accelerating rate, contributing to sea-level rise that threatens coastal communities worldwide. Predicting how much and how quickly sea levels will rise requires understanding the physical properties that control ice flow, such as the slipperiness of the ground beneath glaciers and the stiffness of the ice itself. These properties cannot be measured directly and remain among the most significant uncertainties in ice-sheet science. This project develops openly available software that uses artificial intelligence coupled with high-fidelity physics models to infer these hidden properties from satellite and airborne measurements. The software is designed to lower barriers to participation in ice-sheet science for researchers. The project also invests in education through annual summer schools and conference workshops, training a new generation of scientists at the intersection of geoscience and artificial intelligence. This project transforms a specialized, high-barrier modeling technique into a flexible tool that allows a broader community of students and researchers to test new scientific ideas and incorporate new measurements as they become available, helping to uncover new insights into ice-sheet dynamics. This project develops version 2.0 of the DIFFerentiable neural-network solver for data assimilation of ICE shelves using the JAX Python library, (DIFFICE-jax 2.0). DIFFICE-jax 2.0 is a community cyberinfrastructure for scalable, physics-informed machine-learning inversion across the Antarctic ice sheet. Building on prior software for regional ice-shelf inversion, the system extends these capabilities to the full Antarctic ice sheet through two main advances. First, a multistage neural network training strategy addresses the limitations of standard neural networks in representing multiscale features, enabling large-scale inversion. Second, the integration of full Stokes physics, basal friction inference, and transient modeling makes a broad spectrum of inversion tasks accessible to the community. The software allows users to incorporate new physics and observational data through flexible formulations, reducing the technical barrier of inversion and accelerating scientific exploration. The resulting open-source Python package includes comprehensive documentation, reproducible example notebooks, and a peer-reviewed software publication. Community engagement is supported through annual tutorials, workshops, and a dedicated online discussion forum for troubleshooting, user support, and ongoing contributions beyond the lifetime of the project This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Research, Innovation, Synergies and Education section in the Directorate of Geosciences. 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. NSF Award ID: 2608400 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Ching-Yao Lai | Institution: Stanford University, STANFORD, CA | Award Amount: $600,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2608400 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2608400.html