Appendix F - Application of Machine Learning to High-Resolution Earth System Model Data

The NASA Global Modeling and Assimilation Office (GMAO) uses Goddard Earth Observing System (GEOS), a Global Climate Model (GCM), to produce high resolution atmospheric data, such as weather forecast, subseasonal to seasonal forecasts, and climate reanalysis. High quality GEOS-based precipitation estimates are fundamental to improving land surface estimates and for forecasting initial conditions from the GMAO models. In this project, we will adapt a novel deep learning architecture: super resolution deep residual network (SRDRN) to be used by the GMAO for downscaling and bias correcting GEOS-based precipitation estimates. The SRDRN algorithm was constructed by the Co-I Di Tian’s team, which was inspired by a novel super scaling deep learning approach in computer vision field. The SRDRN algorithm deeply exploited full spatio-temporal dependencies of large- and local-scale climate data and therefore better captured local-, small-scale features such as extreme precipitation events compared to the classic downscaling methods. Through transfer learning the trained SRDRN algorithm in one region could also be applied to a different region without additional training. The algorithm was validated through synthetic downscaling experiments and is ready to be adapted for both downscaling and bias correcting real world earth system data. We will downscale hourly MERRA-2 precipitation data using the SRDRN algorithm with the stage IV radar 4-km hourly precipitation observations. The algorithm will be trained and evaluated over different periods and regions in the contiguous United States (CONUS). Given the results found in our synthetic downscaling experiments, we expect the SRDRN algorithm will show outstanding performance on downscaling MERRA-2 precipitation at hourly, daily, and monthly timescale. The trained SRDRN algorithm can be directly used to downscale any GEOS-based precipitation estimates in real time during historical or future periods, and can be used to produce high-resolution, radar observation-corrected GEOS-based precipitation products over any region of the globe without radar observations. Building our existing efforts in deep learning applications for high resolution earth system data, the proposed activities are expected to significantly improve the MERRA-2 precipitation as well as any GEOS-based precipitation and other land surface or near surface estimates in terms of data accuracy, resolution, and latency.

This is a mature technology (TRL 7+) — the realistic path in is usually NASA's Technology Transfer Program: licensing an existing NASA patent, or a Space Act Agreement to use NASA facilities/expertise directly. NASA also runs a startup licensing program with no upfront fee for companies formed to commercialize a specific NASA technology.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.