Our open-source JAX package, DIFFICE.jax (DIFFerentiable neural-network solver data assimilation of ICE shelves) is released. We are actively expanding the code's applications to other datasets and welcome collaborations!
diffice-jax.readthedocs.io/en/latest/in...

Special thanks to Bryan Riel for writing a beautiful perspective on our paper, summarizing the key nuances of both our scientific findings and algorithmic advances.
www.science.org/doi/10.1126/...

Result: We find the constitutive law of glacial ice near grounding zones follows power laws but with varying exponents, which can impact grounding line stability. We also develop a method to detect and infer ice's anisotropic viscosity—a long-hypothesized property of glacial ice.

Method: We inject only the physics that we believe are correct, e.g. conservation of momentum, into the ML training, and leaving the uncertain physics, e.g. the constitutive law, to be determined through optimization against observations.
Details documented in our 60-page SI.