Machine learning to improve climate models and projections
Applications of machine learning to climate modeling have witnessed a boom over the the last couple of years. Those applications range from improving physical process representation, better using massive datasets such as remote sensing information, as well as postprocessing of large climate mode outputs. In this talk, we will present two major opportunities for machine learning in climate modeling: 1. Using machine learning to better represent subgrid (i.e. smaller than the model grid) physical processes informed by high-resolution simulations or observations and 2. Using machine learning to constrain multi-model climate model projections and to better understand the climate system.
In the first part of the talk, we will show how machine learning can be used to represent several key physical processes of the Earth system such as convection or turbulence. We will highlight physical lessons learnt from those statistically informed parameterizations and limits of standard physically-based parameterization, such as mass flux approaches for convection, which cannot be corrected by parameter tuning alone but rather require a rethinking of the structural form of the equations as they miss some key processes. We will discuss challenges related to extrapolation to unseen regimes and strategies to address those issues by merging machine learning with physical knowledge and physical invariances.
In the second part of the talk, we will highlight how machine learning can be used to constrain projections of future climate with observations. Those techniques perform better than multi-model means or standard emergent constraints. In addition, the use of causal discovery and causal inference can go beyond standard model evaluation and analysis approaches and help better highlighting the physical mechanisms at play.