Ramin Bostanabad
University of California, Irvine
NASA’s physics-based models that simulate large-scale hydro-mechanical processes have dramatically increased our capabilities in understanding and predicting socioeconomically important phenomena such as droughts, floods, and hurricanes. However, these complex models rely on effortful assimilation processes to produce reliable predictions, are computationally very expensive, and embody a wide range of uncertainty sources. To address these pressing challenges, we leverage the recent technological advancements in machine learning to develop novel methods that harness the full potentials of NASA’s multiscale physics models. Our innovations take a fundamentally different approach for combining observational datasets with model predictions that reduces computational costs, eliminates manual tuning, and improves accuracy. These contributions uniquely complement the cyberinfrastructure foundations and massive observational datasets that NASA has heavily invested in over the past few decades.