Machine Learning Downscaling Capability for Environmental Forecasts

Rapid progress has been made recently in the development and application of artificial intelligence/machine learning alternatives for weather analysis and prediction (MLWP). For example, the European Centre for Medium-Range Weather Forecasts recently launched the Artificial Intelligence Forecasting System as an operational service and plans to make ensemble forecasts available along with data-driven sub-seasonal (extended-range) forecasting. Similar efforts are on the way in the U.S. and other major weather forecast centers.ÝÝThe Navy seeks to leverage these advances in Machine Learning (ML) methods, environmental prediction, and downscaling/super-resolution techniques to develop new capabilities for targeted local-scale and tactical-scale short-range to medium-range forecasts. Many uses of environmental predictions like tropical cyclone wind and surge effects, electro-optical or acoustic propagation in the boundary layer, coastal winds, aviation visibility, ice-edge circulations, ocean eddies, navigation through sea ice, etc., require knowledge of small-scale features to achieve the most useful forecasts, reduce operational risks, and deliver tactical advantages to navy end-users.ÝThe solution we propose is to hybridize sub-seasonal/seasonal global data-driven MLWP and physics-based forecasts (NWP) by developing and testing ML models, using unique neural transforms, generative diffusion, super-resolution, closure schemes, and physics-informed network architectures to deliver local area downscaled ocean, ice, and atmosphere fields ensemble predictions at scales below 1km. For this purpose, in Phase I we will test and augment solutions we developed (e.g. during the Reduced Order Modeling STTR - N6833520C0567) to deliver high-resolution (less than 1-km horizontal grid spacing and high-fidelity in the vertical) short-range to medium-range downscaled forecasts from a lower resolution set of historical data, observations, and forecast fields. These methods include deep neural operators for high-resolution MLWP constrained with NWP-based Principal Component transforms, interpretable and generalizable neural closure modeling for complex dynamical systems, and compressed sensing solutions. During Phase 1, we will also compare existing neural operators, vision transformers, generative diffusion models, graph neural networks, and physics-informed architectures for downscaling. In Phase 2, we will implement into prototypes the new downscaling ML models and neural architectures that follow the rules of physics, employ information theory to select the most useful data to train these models, develop novel neural data assimilation schemes in latent and dynamically reduced spaces, derive and accelerate generative diffusion models for super-resolution and data assimilation, and implement differentiable software for high-resolution multi-dynamics downscaling optimized for GPU training and integration.Ý