A Parsimonious Deep Learning Earth System Model for Sub-Seasonal and Seasonal Forecasting
Abstract
Approved for Public Release. We propose to further develop our coupled atmosphere-ocean deep learning model into a full deep learning earth-system model (DLESM) by adding land and cryosphere components. Our model is very parsimonious currently using just 9 forecast variables at each horizontal grid location. We will expand the number of prognostic fields, but strive to keep the set of forecast variables relatively small, including additional fields as necessary to accurately represent the full earth system. A key focusof the proposed research is to develop novel analysis tools to explore tipping points and sensitivities in the coupled earth-system, such as conditions that trigger blocking events, El Nino, or changes in other patterns of low-frequency variability.A unique aspect of machine learning is that some end-user variables that are tightly coupled with the weather can potentially be incorporated intoa DLESM in a way that both improves sub-seasonal and seasonal weather forecasts, but also predicts quantities of immediate interestto end users. As an example, including normalized differential vegetation index as a prognostic variable in the DLESM might not only provide important information about latent and sensible heat fluxes between the atmosphere and the vegetated surface, but also yield direct forecasts of plant vitality of interest to farmers and agriculture. The traditional approach for improving global earth-system models is to continually increase numerical resolution and incorporate more detailed representations of all the relevant physical processes. If it proves possible, as preliminary results suggest, to simulate the evolution of large-scale perturbations in the earth s atmosphere, oceans, or vegetation with a much more parsimonious approach than that in current numerical models, the impact on geoscience would be enormous.A key application of our improved DLESM will be the improvement of sub-seasonal and seasonal forecasts of the atmospheric and oceanic states. Such information can ensure that supplies appropriate for periods of unusual heat, cold, flood or drought are available. Improved forecasts can provide advanced information about the likelihood of stormy, calm, cloudyor clear weather. Such intelligence can help the Navy tune its all-weather capabilities to the conditions most likely to occur during multi-week campaigns.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 08, 2024
- Source ID
- N000142412528
Entities
People
- Dale R. Durran
Organizations
- Office of Naval Research
- United States Navy
- University of Washington