High-Resolution Environmental Forecasting Through Advanced ML Downscaling Techniques

The US Navy operates in diverse and challenging environments where accurate, high-resolution, timely environmental forecasts are critical for mission success. To support the Navy's ability to anticipate and respond to localized environmental phenomena (e.g., tropical cyclone winds and surge effects, coastal winds, aviation visibility, ocean eddies), there is a pressing need for a capability that can generate skillful, near real-time environmental forecasts with horizontal spatial resolution of less than 1 km and vertical resolution on the order of 10 meters, particularly in the atmospheric boundary layer and upper ocean. However, most current-day Numerical Weather Prediction models are computationally too burdensome to achieve such high resolution in the short processing times that the near real-time paradigm necessitates. Furthermore, there is a scarcity of observational data at these fine scales, complicating model validation and hindering the development of reliable high-resolution forecasting tools.ÝATAC and subcontractor NCAR propose this Phase I SBIR work to address these challenges by leveraging several NCAR-developed and non-NCAR cutting-edge machine learning weather prediction (MLWP) techniques that use downscaling, nesting, and other related methods for generating high-resolution, short-to-medium range MLWP forecasts. These techniques include (1) Enhanced Super-Resolution Generative Adversarial Network-based downscaling, (2) Diffusion-based downscaling, (3) Analog Ensemble-based downscaling, and (4) National Renewable Energy Laboratory (NREL) Super-Resolution for Renewable Energy Resource Data with Climate Change Impacts (Sup3rCC) model. Moreover, we apply existing high-resolution Large Eddy Simulation sets as training datasets for the MLWP models and leverage ATACís big data processing architectures and deep bench of data scientists/analysts to perform expert comparison of the performance of these models and provide summary of strengths/weaknesses of the methods, their operational viability, and relevant metrics. Our Phase I SBIR work rigorously evaluates the above MLWP downscaling techniques and performs a literature survey of other relevant MLWP techniques, identifying their strengths, weaknesses, and best-use scenarios. This comparative analysis combines quantitative metrics (e.g., standard accuracy metrics, spatial coherence metrics, temporal consistency metrics), model explainability analysis (Explainable AI techniques such as Shapely feature importance), robustness testing (using variety of test datasets covering coastal regions, open-ocean environments, etc.), and computational efficiency metrics. This multi-dimensional comparative analysis enables us to determine the most effective approach for near-real time, high-resolution weather forecasting in landfalling tropical cyclone environments, directly addressing the critical need identified in the SBIR problem statement.