24AI Downscaling Model Guidance to Support the Gridded Forecast Process in Areas of Complex Terrain in OCONUS

National Weather Service (NWS) meteorologists at weather offices around the US have been producing gridded official forecasts for the National Digital Forecast Database (NDFD) for roughly 20 years. More recently, the National Blend of Models (NBM), which uses statistical post processing to produce deterministic and probabilistic guidance fields for NDFD forecast elements, has been used by meteorologists as a starting point for their gridded forecasts. This allows forecasters to focus more on the meteorology of the forecast message, and on providing decision support services, rather than on manual grid editing. The NBM is still maturing, and there are significant issues in the NBM in the Alaska and Pacific Regions owing to the complexity of the terrain, inability of model guidance and analyses to model or capture significant weather features, lack of in-situ data and radar, limited GEO satellite information at higher latitudes, and representative issues with ice / atmosphere coupling at higher latitudes. This requires meteorologists in these regional offices to spend significantly more time manually editing grids than at most CONUS offices. Given that improvements to model resolution, new models, and an increase in observational data are not likely to occur in the next several years, the goal of this study is to investigate new techniques to downscale NBM model guidance to produce more realistic results in complex terrain. Improved guidance for these forecasts will significantly reduce the amount of time meteorologists spend on manual correction and allow them more time to work on other NWS goals. Downscaling (inferring high-resolution information from low-resolution information) problems are common in meteorology, and various statistical and dynamical techniques can be used to downscale fields from weather models around complex terrain and low-observation areas. Artificial Intelligence (AI) deep-learning methods such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) have shown great potential in meteorological downscaling problems when enough training data exists. Using archived NDFD data from a 4 year timespan and available wind observations as the ground truth, we demonstrate a case study in training an AI model to produce initial guidance for meteorologists creating forecasts in the Alaska region. A data-processing flow is detailed and established, and preliminary results from several deep-learning architectures are given, highlighting the improvements over the NBM guidance and comparing the model results against the ground truth using various quantitative verification scores.