Improving Data Assimilation in Numerical Weather Prediction of Winter Cyclones using Artificial Intelligence and Machine Learning

Operational numerical weather prediction (NWP) models frequently struggle with track errors for wintertime cyclones which can lead to life threatening impacts and economic disruptions. Data assimilation is an advanced statistical method for preparing relatively low-resolution, high-fidelity sensor information for these NWP models. This method produces a first-guess field that serves as an estimation of the current state of the atmosphere in operational NWP models, which ultimately produce weather forecasts that are essential tools for forecasters. Despite many advances in the field, shortcomings are still very apparent and NWP models still have prediction issues with common weather occurrences, including those that produce extreme weather. The very large data volumes and computational expense of imperfect data assimilation requires unique solutions to improve NWP predictions. With the advance of machine learning (ML) and artificial intelligence (AI), along with other techniques for processing large volumes of data, the possibilities for improving data assimilation are numerous. This proposed effort will take advantage of recent advances in AI/ML in order to generate better initial conditions for NWP models, which will contribute to improved forecasts of extreme events. This effort will begin by gathering large amounts of data prior to and during major winter weather events to create training datasets for proper orthogonal decomposition (POD) reconstruction which will provide an AI/ML approximation of the natural state based on the data collected. We plan to run NWP models using both traditional data assimilation and with POD dynamic modeling for winter cyclones. The new AI/ML technique should present new approaches for future NASA model development and NASA computational groups.