NextGen Space Weather Modeling Framework Using Physics, Data Assimilation, Uncertainty Quantification and GPUs

Abstract:

We have been recently awarded a major NSF/NASA grant from the SWQU program to develop the NextGen Space Weather Modeling Framework that will employ computational models from the surface of the Sun to the surface of Earth in combination with assimilation of observational data to provide optimal probabilistic space weather forecasting. The model will run efficiently on the next generation of supercomputers to predict space weather about one day or more before the impact occurs. The new project will concentrate on forecasting major space weather events generated by coronal mass ejections (CMEs). Current space weather prediction tools employ first-principles and/or empirical models. While these provide useful information, their accuracy, reliability and forecast window need major improvements. Data assimilation has the potential to significantly improve model performance, as it has been successfully done in terrestrial weather forecast. To allow for the sparsity of satellite observations, however, a different data assimilation method will be employed. The new model will start from the Sun with an ensemble of simulations that span the uncertain observational and model parameters. Using real time and past observations, the model will strategically down-select to a high performing subset. Next, the down-selected ensemble will be extended by varying uncertain parameters and the simulation continued to the next data assimilation point. The final ensemble will provide a probabilistic forecast of the space weather impacts. While the concept is simple, finding the optimal algorithm that produces the best prediction with minimal uncertainty is a complex and very challenging task that requires developing, implementing and perfecting novel data assimilation and uncertainty quantification methods. To make these ensemble simulations run faster than real time, the most expensive parts of the model need to run efficiently on the current and future supercomputers, which employ graphical processing units (GPUs) in addition to the traditional multi-core CPUs. The main product of this project will be the Michigan Sun-To-Earth Model with Quantified Uncertainty and Data Assimilation (MSTEM-QUDA) that will be made available to the space physics community with an open source license. We will describe the main concept of the project and our initial progress.