Speaker
Description
The representation of subgrid processes is a source of uncertainty and a limitation of global models. Cloud Resolving Models (CRM) can model some of these processes more accurately. However, in terms of computational costs, running convection permitting CRMs on a global scale is prohibitively expensive. Here we present some early results of our work where deep neural networks are used to learn the coarse-grained temperature and moisture physics tendencies produced from a series of CRMs nested within the Met Office Unified Model. We use our machine learning (ML) algorithms in as single column models where the outputs from one timestep are the inputs for the next timestep. Running ML models in this way shows both its strengths and limitations. Once trained, ML models are cheaper to run, but then can also suffer from drifts where the ML model diverges from the physics “truth” provided by the coarse-grained CRM data. We discuss how these limitations can be addressed, plus some of the challenges that lie ahead for Machine Learning in the Atmospheric Sciences.
