Modeling snow on sea ice with physics-guided ML
1min
Paper: Environmental Data Science
Snow on sea ice matters for sea-ice thermodynamics, satellite retrievals and downstream integration in climate models but running snow models at scale is expensive.
In this work we develop and test ML emulators of SnowModel, including a physics-guided LSTM where snow-compaction physics is added as a constraint in the loss function that improves accuracy and generalizability across five Arctic regions. The emulator runs faster than SnowModel enabling much cheaper large-area experiments. I had the pleasure of presenting this work at Climate Informatics 2024 in London.