Data-driven emulation of melt ponds on Arctic sea ice

Abstract. The genesis, development and disappearance of melt ponds on sea ice are complex, insufficiently understood, and driven by metre-scale mechanisms unseen by numerical models. Melt pond formation is thus parametrised with substantial uncertainty. Despite melt ponds playing a major role in sea ice thermodynamics, melt pond parametrisations have traditionally not been included into GCMs, and for instance do not play a role in the IPCC assessments of when sea ice will disappear from the Arctic in summer.

Previous research showed it was possible to learn a state-of-the-art physical parametrisation of melt ponds (from model data) using neural networks. The emulator was suitable for use in a thermodynamic sea ice model (the Icepack model) and ran stably for about ten years. In this study, we follow on from that work and develop a novel data-driven parametrisation of melt pond fraction. Using pan-Arctic satellite observations and reanalysis data we show that it is possible to learn and predict to a sufficient degree the melt pond fraction as determined from Medium Resolution Imaging Spectrometer (MERIS) and Ocean and Land Colour Instrument (OLCI) data, the target in our supervised learning setup, from well chosen observationally based inputs. Our deep learning emulator has been intentionally designed to be pointwise with the perspective of being suitable for incorporation within physical models of sea ice such as Icepack.

Our results prove the concept that it is possible to learn parametrisations directly from data for sea ice thermodynamical processes. In doing so our work provides a viable emulator of melt ponds for use in GCMs and demonstrates a route for developing and further refining observationally based data-driven emulators of melt ponds that are ready for implementation in GCMs. We also briefly discuss these future avenues for advancing on this work and further developing data-driven emulators of melt ponds.

Furthermore, the results show the difference in modelling melt ponds over different Arctic regions. In particular, it highlights the critical importance and need for better observations and understanding in the marginal ice zone which will be crucial for future impacts.