Terrafuse is developing a physics-enabled AI platform for rapidly building AI emulator (surrogate) models of complex physical systems such as climate and weather. These AI emulators learn to behave like numerical physics models (in the sense that they replicate their spatial and temporal statistics), but are orders of magnitude more computationally-efficient to run and scale, and more cost-effective and faster to build. The technology efficiently assimilates simulation data, observational (remote/ground sensing) data, and domain knowledge (conservation laws, constraints), allowing for increased accuracy and geographical scalability. This unified computational workflow allows to quickly and consistently develop models for multiple applications - climate modeling, hydrology, weather prediction etc. - that would otherwise require significant hand-tuning and customization. The emulators are built in a modular fashion from standard operators implementing key functionality of of climate prediction tasks, including spatial and temporal downscaling of gridded data, spatio-temporal modeling of fluid dynamics processes, and modality translation (e.g., numerical simulation to remote-sensing observations). The proposed project develops a proof of concept system built on top of the terrafuse AI emulation platform for use cases of potential interest to Air Force stakeholders, including hyperlocal weather prediction, ultra-fast modeling of turbulence, and natural hazard risk modeling.
