SBIR Phase II: Physics-Informed Machine Learning Emulators to Model Physical Spatio-Temporal Processes for Climate and Weather Risk Forecasting

The broader impact of this Small Business Innovation Research (SBIR) Phase II project is to provide commercially-deployable technology for highly-scalable, spatially-granular, and cost-effective risk predictions of climate-driven events, such as wildfire spread, from real-time to yearly time scales. As insured losses due to wildfires have increased over fivefold in the last decade, the associated risk makes it critical to improve the ability to predict physical and financial impacts at scale. Current predictive technologies used in major industries, like energy and insurance, are based on complex, hand-engineered, and computationally-intensive numerical physics models of climate and weather. In contrast, the proposed technology develops special AI emulator systems that learn the relevant physics and key drivers, including wind and surface hydrology, in wildfires. The proposed system can perform predictions much more efficiently due to a far simpler computational workflow and native AI hardware acceleration. In addition, AI emulators automate the assimilation of vastly higher amounts of remote-sensing and other observational data (e.g., radar measurements from weather satellites or land cover and vegetation data) over numerical models, allowing for increased accuracy, continuous improvement, and dynamic predictions reflecting changing on-the-ground conditions. This Small Business Innovation Research (SBIR) Phase II project addresses the pressing need in the energy and insurance industries to accurately and consistently assess wildfire risk over large geographical regions and at a localized level, on time scales ranging from daily to yearly. The proposed R&D will focus on developing and validating an AI emulator of wildfire spread. This entails 1) developing AI architectures for assimilating observational (remote-sensing) and numerical simulation data on drivers of wildfire at different temporal and spatial scales, including vegetation, soil hydrology, and atmospheric winds; 2) integrating data on historical wildfires and their spread to drive the learning process; 3) conducting extensive verification and validation studies; and 4) developing and deploying APIs and graphical interfaces for accessing AI emulator output on the cloud. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.