Innovim meteorologists and computer scientists will be presenting the latest results from a NOAA Office of Oceanic and Atmospheric Research (OAR)-sponsored Small Business Innovative Research (SBIR) grant, “Landfalling Event Atmospheric River Neural Network (LEARN2): an AI Machine Learning Numerical Weather Prediction Post-Processing Tool for Advancing Weather and Climate Prediction” (https://agu.confex.com/agu/fm22/meetingapp.cgi/Paper/1050031, in session SY35A – Community Modeling and Open Innovation to Advance Earth Prediction Systems). The session provides “an opportunity for the broad research community to share information about their latest developments and how these innovations are advancing the capabilities of community modeling systems. Topics in this session will include discussions on the motivation and process by which the community can work together to explore, validate, and integrate all aspects important to advancing weather and climate prediction. Open innovation spans everything from observation impact, model code, and software engineering to experimental design and computing architectures, post-processing, ensemble design, including model uncertainty.” Session conveners include Jose-Henrique Alves, NOAA OAR, Neil Jacobs, University Corporation for Atmospheric Research and Louisa B Nance, NCAR. This research was led by Dr. Philip Ardanuy, INNOVIM’s Chief Scientist, and Dr. Sibren Isaacman, Associate Professor of Computer Science, Loyola University Maryland.
The main phenomena driving rain events in the western United States during the rainy season (October-April) are Atmospheric Rivers. Accurately forecasting rainfall from these events is critical due to the impacts they have on the water supply as well as the flooding and debris-flow risks they pose. LEARN2 is a neural-network-based decision support tool that considers forecast guidance from NOAA/National Centers for Environmental Prediction (NCEP’s) Global Ensemble Forecast System (GEFS) and the European Center for Medium-Range Weather Forecasting Ensembles (ECMWF), remotely sensed fields, and several subseasonal-to-seasonal teleconnection indices, to produce extreme rainfall predictions. In LEARN2, products are run through several neural networks and novel genetic and “voting” algorithms that leverage the power of ensemble forecast systems to predict whether rainfall will exceed thresholds of interest at discrete points in a grid for lookaheads of up to 10 days. Using F1-score as our primary metric, LEARN2 predictions outperform both the GEFS and ECMWF, delivering 1-2 additional days of lead time for skillful extreme rainfall prediction on independent data.