Description:

Interest in artificial intelligence or data-driven weather forecasting models has seen impressive growth recently in light of results showing forecasting skill on par with traditional prediction models, but with much shorter runtimes. This development has opened prospects for much larger forecast ensembles, more rapidly updating forecast cycles, and higher resolution forecasts over larger forecast domains than currently possible with traditional NWP models on many current HPC systems with much fewer resources. 

However, with these prospects come the potential HPC challenges of shifting more compute resources to accelerator-based designs as well as managing, post-processing, and disseminating the much greater levels of input and output data expected in much shorter periods of time, potentially straining HPC storage and data transport network systems. Therefore, this session seeks to explore efforts to implement AI NWP models in current HPC environments. Submissions are encouraged to give insight into the model selection process, the development of workflows built around these methods that handle the data management process, and/or the challenges that were encountered while developing and implementing the end-to-end forecast process.