| Summary: | Large swaths of the Earth's high-northern latitude land surface consist of permafrost landscapes. Permafrost is a key factor for hydrology, ecology, biogeochemistry, and for human infrastructure. It contains one of the largest soil carbon reservoirs on Earth, which will partially be mobilised upon thaw and will accelerate global climate warming. A quantitative understanding of permafrost thaw is therefore essential. Surface manifestations of permafrost thaw include thermokarst lake growth and drainage, gully formation, and retrogressive thaw slumping. These processes exert important influences on carbon, energy, water, and sediment balances, with implications also on habitats and surrounding ecosystems. The polar regions, where most of the permafrost landscapes are found, have been experiencing enhanced warming due to Arctic amplification. The warming leads to changes in the balance of and feedback between surface processes associated with permafrost thaw. Being able to predict the future of these processes on a pan-Arctic scale under various climate scenarios is therefore important. One widely used tool for modelling energy balance and the freeze-thaw process of permafrost is CryoGrid. We aim to develop a physics-informed machine-learning framework to upscale and accelerate CryoGrid to predict the future evolution of surface processes arising from permafrost thaw. We first calibrate the ability of CryoGrid to reproduce observed surface manifestations of permafrost thaw. Next, we move towards utilising physics-constrained deep-learning techniques to enable the usage of CryoGrid on a pan-Arctic scale to efficiently make various predictions. The resulting predicted observables can be used to constrain or indicate specific climate variables.
|
|---|