Fully nonlinear data-assimilation in climate models

Abstract:

While climate modelling requires huge computational resources, initialising climate models using data assimilation is even more demanding. As the climate system is highly nonlinear both through nonlinear dynamics and strong feed backs the data-assimilation methodology has to be nonlinear too. Furthermore, it has been realised that one best forecast is not that useful and proper uncertainty quantification is essential for advances in the field. Both requirements point to fully nonlinear ensemble techniques, such as particle filters. Recently particle filters have been generated that allow for efficient ensemble members for climate model initialisation, by generating proper samples from the posterior probability density function in huge dimensional spaces.

Another issue is to connect the complex climate model to the data-assimilation code. We have developed a very efficient framework to do this using only MPI communication between separate model and data-assimilation executables, called EMPIRE. This framework allows for very fast connection of any complex model to state-of-the-art ensemble data-assimilation methods.

We will show an example of use of this new methodology to the HadCM3 climate model, which has over 2 million model variables, using the EMPIRE framework. We will discuss timings and efficiency, as well as some of the physical results. Finally, we will discuss the coupling with the very high resolution UM vn8.2 with close to 300 million model variables.