The goal of this research is to explore how deep learning can be combined with in situ data analysis and visualization.  In situ feature tracking can mitigate the risk of lost insight by identifying features of interest to save, which could greatly reduce the amount of data that needs to be saved.  Deep learning has the potential to accurately and efficiently detect and track features online.  We plan to investigate the following aspects:

• How to use in situ workflow tools to stitch together scientific simulations and state-of-the-art deep learning frameworks
• How to adapt and tune deep learning algorithms for feature tracking in scientific simulations
• Are deep learning algorithms scalable and efficient enough to keep up with the data producing rate of the simulation

We will develop and test prototypes for various applications including plasma sciences, superconductivity, earth systems, and fluid dynamics.  The target platforms include both NVidia GPU and Intel KNL based systems.  We are also expecting two to three paper publications in collaboration with current and past summer students.