Comprehensive climate models typically include atmosphere, ocean, sea ice, and land components and the coupling between them. These models can also incorporate land ice, atmospheric chemistry and terrestrial and marine biogeochemistry, enabling carbon cycle simulations. Early climate models were developed in the 1970s and have increased in complexity over the years, with more process interactions, more sophisticated parameterizations of subgridscale processes, and higher spatial resolution being incorporated over time. These models have been skillful at predicting anthropogenic climate change, and even early models accurately simulated aspects of the spatial pattern of warming. However, there is still considerable uncertainty associated with model structure, and climate models which incorporate different parameterizations can differ greatly in many characteristics of future projected change. Because of this, it is imperative that there are continued developments and improvements of these modeling systems. Bringing new approaches, such as Machine Learning, to this challenge has the potential to rapidly accelerate progress.
Studies across M²LInES are using scientific and interpretable Machine Learning to gain new insight on parameterization development across the atmosphere, ocean, and sea ice systems. Development and testing of these parameterizations is underway in component model configurations and work is planned to incorporate these into a number of climate models. This includes, among others, improved parameterizations of:
- The simulated conductive heat fluxes through sea ice (Zampieri et al, 2024 ),
- Ocean mixing processes,
- Moist convection in the atmosphere
