Machine Learning for Stochastic Parametrisation in Atmospheric Models

How sensitive is Earth’s climate to increases in CO2 and other greenhouse gases, and how soon will the 2 degree climate target, a key goal of the Paris climate agreement, be breached? These questions remain highly uncertain largely due to small-scale atmospheric processes that are climatically important yet difficult to represent in climate models, such as cloud formation (Schneider et al., 2024). One way of representing model uncertainty is via stochastic parameterizations, which introduce perturbations in the numerical simulations.  Such schemes have been shown to substantially improve the reliability of weather forecasts and are widely used in forecasting centres around the world. They are also being explored by climate modelling centres, but progress is hindered by their poor conservation properties. However, existing stochastic schemes often represent uncertainty in a rather ad hoc and simplistic way. Ideally, perturbations should be state-dependent, e.g. larger in atmospheric conditions which are associated with more uncertainty. In this project, we develop a full ML sub-grid parameterization with built-in stochasticity that is directly learned from data, using a physically informed neural network architecture based on stochastic recurrent neural networks. This project uses the ClimSim dataset designed for developing machine-learnt parameterizations for climate models. The project aims to obtain a new state-of-the-art in both ML parameterisation and stochastic parameterisation, with the potential to significantly reduce and better characterize model uncertainty in both weather and climate models.