Juliacon 2024

We present ODINN.jl, a global glacier evolution model leveraging the Scientific machine learning and foreign language calling capabilities of Julia. We developed a modelling framework capable of handling Universal Differential Equations of glacier ice flow, solving hundreds of thousands of ODEs in parallel. By exploiting Julia’s SciML differentiable programming capabilities, we demonstrate how we can perform functional inversions to learn underlying physical laws of geophysical processes, while assimilating heterogenous, noisy and sparse observations. Moreover, thanks to the use of PyCall.jl, we can do so by re-using data processing and management tools from the Open Global Glacier Model and Xarray in Python. Furthermore, the use of these tools allows us to recycle existing mass balance models (i.e. a source in a differential equation) previously calibrated in Python that are incorporated in the numerical solver. This enables us to build complex hybrid models on top of currently existing infrastructure in Python, leveraging each ecosystem’s strengths.

With ODINN.jl, we show how the future of geoscientific models will gravitate towards flexible modelling frameworks, exploiting mechanistic models based on domain knowledge with machine learning models capable of learning from new massive Earth observations. Moreover, the multi-language capabilities of Julia enable a smooth interoperability of these new tools with already established frameworks, providing a faster track to new methods, while enabling existing communities to benefit from them.