GPU-resident Atmospheric Large Eddy Simulations: Predicting the Weather on Small Scales

Current weather and climate models do not have sufficient resolution to explicitly resolve low cloud evolution, and therefore use statistical models for this purpose. While these models are reasonably successful in forecasting statistical estimates of relevant weather properties like precipitation and cloud amount, any local quantitative prediction (where exactly is a cloud developing) is outside their reach. For this reason, many detailed numerical cloud process studies make use of Large-Eddy Simulations (LES), high-resolution simulations of idealized cases, but performed on a limited domain. Computational limits force LES simulations of appreciable size to typically be performed on supercomputers.

However, we have recently ported our LES code to the GPU, resulting in an entirely GPU-resident Atmospheric LES (called GALES; Schalkwijk et al, BAMS, 2012). The use of the GPU provides a huge computational boost to the possibilities of LES: simulations formerly performed on 30-60 computer cores can now be performed on a single GPU. This is a potential leap in the possibilities of LES application.

At present, this speedup allows GALES to run on a day by day basis in a set-up in which the LES represents a single grid box (about 25km x 25km) of a large-scale weather model (LSM). This set-up this allows GALES to run in a 'predictive' mode, with the LSM providing the current large-scale forcings, and is a prudent test into the possibilities of LES weather prediction. Moreover, the set-up provides an unprecedented three-dimensional high-resolution dataset of daily cloud conditions, structure and evolution. By running in hindcast as well, we have already simulated a full year of real-weather cloud evolution.

A recent grant on computer time on the CURIE-GPU cluster in France allows us to go even further, however. By up-scaling the above set-up, i.e. by simulating with multiple GPUs a matrix of weather model grid boxes, it becomes possible to fully encompass a small country as the Netherlands or an appreciable part of a US state (approx. 400km x 400km with ~256 GPUs), providing rain and cloud field data with high temporal and spatial resolution. This allows us to bridge a part of the gap between large scale models and LES, such that we can study, for example, mesoscale dynamics interacting with the turbulent boundary layer.