Speaker: Maria Molina (Univ Maryland)
Topic: Machine Learning-based Predictability Assessment and Bias Correction of Subseasonal Precipitation

For the field of subseasonal-to-seasonal (S2S) prediction (timescales of two weeks to two months), skillful prediction of precipitation remains very difficult. Predictability stemming from atmospheric initial conditions is substantially reduced beyond approximately two weeks and the ocean generally does not offer added predictability until a trajectory reaches the seasonal timescale. Imperfect initial conditions and model systematic errors also contribute to the difficulty of deterministic initialized forecasts. Ensemble forecasting has helped assess forecast spread in relation to initial condition errors, but the high cost of running global initialized forecasts precludes the creation of many ensemble members. These challenges motivate the use of machine learning methods for S2S prediction. Two approaches to S2S prediction research using the Community Earth System Model version 2 (CESM2) will be highlighted: (1) a predictability study and (2) a bias correction approach. The first study focuses on assessing the predictability of North American weather regimes, which are persistent large-scale atmospheric patterns that can imprint on surface anomalous precipitation. Various Earth system components, such as the atmosphere and land, will be used to assess contributions to predictability. The second study focuses on the use of machine learning models for offline bias correction of S2S forecasts of global precipitation. The machine learning model architectures include image-to-image approaches (e.g., U-Net), which enables learning of spatial patterns and displacement errors in CESM precipitation fields using various convolutional and pooling layers. Additionally, recent work highlighting the importance of culturally-relevant science communication and trends in the representation of Hispanic and Latinx people in STEM will be shared.