Development of an Integrated Approach to Augment an NWP model with a Machine Learning Component
Abstract
Approved for Public ReleaseThe proposing research team and their former collaborators have developed a hybrid modeling approach to a,ugment a physics-based numerical model of a complex system (e.g., the atmosphere) with a machine learning (ML) component. This appro,ach, which does not require making changes to the existing numerical model, is called Combined Hybrid-Parallel Prediction (CHyPP). I,n CHyPP, the ML component makes frequent, periodic, interactive changes to the full evolving model solution after learning about the, flow- and location-dependent errors introduced by the numerical model by training on a time series of past (observation-based) anal,yses. In forecast mode, little or no correction is made to a numerical forecast at locations where the ML component learns that the,numerical forecast is likely to be highly accurate. However, where the ML component learns that the numerical forecast is likely to,have a large error, it can make a large correction, or completely replace the numerical forecast with an ML forecast before moving o,n with the forecast for the next few hours. The goal of the project is to prove the applicability of the ChyPP concept to a state-of,-the-art NWP model and to further develop the methodology. The specific model used in the project will be NAVGEM, the current global, NWP model of the U.S. Navy. The goal of the further development of the methodology is to produce a prototype forecast system, in wh,ich the training of the ML component of the hybrid model is integrated with data assimilation. In the envisioned integrated approach,, the ML component of the hybrid model will be updated sequentially at data assimilation times, using real-time observations for its, online training, while the online-trained hybrid model will provide the background (forecasts) for the data assimilation system. Th,e integrated hybrid modeling approach is expected to be particularly beneficial for the analysis and prediction of processes in part,s of the atmospherewhere the model errors are large and for which high-quality (reanalysis-type) training data sets are not readily,available. One such part of the atmosphere, which is also of special interest for the U.S. Navy, is the thermosphere/ionosphere. The, integrated methodology also has the potential to improve the performance of a forecast system for the troposphere and lower stratos,phere, for which high-quality reanalysis data sets are available.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 01, 2022
- Source ID
- N000142212319
Entities
People
- István Szunyogh
Organizations
- Office of Naval Research
- Texas A&M University
- United States Navy