The ocean plays a fundamental role in setting the Earth's temperature and climate, particularly on decadal and longer time scales as it has absorbed 90% of anthropogenic carbon dioxide emissions and the additional radiative heating caused by these emissions. Transient climate change is thus fundamentally linked to how the ocean transports heat, salt, and biogeochemical tracers such as carbon dioxide and nutrients.
Much of this transport is done by mesoscale eddies, the "weather" of the ocean, which transport, mix, and stir tracers in the ocean and account for over half of the kinetic energy in the ocean.
The radius of these mesoscale eddies varies from 200 km near the equator to 10 km near the poles. Climate models are too coarse and cannot represent the effects of any but the largest mesoscale eddies adequately and thus their effects must be parameterized. The parameterization of mesoscale eddies has been the focus of intense research, however a good parameterization remains elusive and it still remains the largest source of uncertainty in ocean models.
We plan to tackle the problem using some new and powerful tools Julia tools. A couple of approaches we had in mind:
- Implement a mesoscale eddy parameterization in an ocean model and estimate the (hyper)parameters as a function of latitude, longitude, forcing, etc. to maximize the accuracy of the parameterization globally. Maybe through Bayesian inference or adjoint sensitivity analysis?
- Use a neural network to parameterize the effects of mesoscale eddies in the partial differential equations that describe the ocean's circulation.
References:
- Griffies et al. (2015). "Impacts on Ocean Heat from Transient Mesoscale Eddies in a Hierarchy of Climate Models", Journal of Climate 28, 952–977. https://doi.org/10.1175/JCLI-D-14-00353.1
- Bates et al. (2014). "Rationalizing the Spatial Distribution of Mesoscale Eddy Diffusivity in Terms of Mixing Length Theory", Journal of Physical Oceanography 44, 1523–1540. https://doi.org/10.1175/JPO-D-13-0130.1
- Reichstein et al. (2019). "Deep Learning and Process Understanding for Data-Driven Earth System Science", Nature 566, no. 7743, 195–204. https://doi.org/10.1038/s41586-019-0912-1
- Innes et al. (2019). "A Differentiable Programming System to Bridge Machine Learning and Scientific Computing", arXiv:1907.07587 [Cs]. http://arxiv.org/abs/1907.07587
- Rackauckas et al. (2019). "DiffEqFlux.jl - A Julia Library for Neural Differential Equations", arXiv:1902.02376 [cs, stat]. http://arxiv.org/abs/1902.02376