Shayn McGregor: Tropical variability modelling:
Thanks Ben, totally agree. At least in the land surface context, empirical / machine learning approaches have been super useful for defining how well a model is utilising the information itâs been given in its parameters and inputs⌠and I think thatâs a really defensible basis on which to define whether a model is doing well or not. Itâs obviously only one way that one might decide on performance expectations a priori, but I really like it because itâs not an application-specific definition of what makes a model âgood enoughâ.
The question of coupler vs soft coupling via ins/outs/restarts has been briefly discussed. In principal the coupler approach would be the way to go. However this is unlikely to be the method chosen initially because
- weâd have two sets of technical developments ongoing at the same time
- coupling between the land (CABLE/JULES) and other components of ACCESS is currently either within the same executable (UM including UKCA) or via the âcouplingâ section of the UM and the UMâs state vector array (in the case of communication with MOM) implying weâd have to manipulate those parts of the system of model systems as well.
The ambition (at the moment) is to use the soft coupling approach initially, clearly identify what information needs to be passed between components and get something working, and then transition to a coupler methodology. But that is on a longer-than-CMIP7 timeline being realistic.