ML-Based Downscaling for High-Resolution Atmospheric Forecasting

The Navy requires relocatable forecasting tools capable of producing high-resolution environmental intelligence at tactically relevant scales, particularly within the atmospheric boundary layer. Current numerical weather prediction (NWP) models, while valuable at mesoscale resolutions, lack the spatial fidelity and vertical resolution needed for forecasting in complex, coastal, or contested environments where phenomena such as sea breezes, low-level jets, and terrain-induced flows can significantly impact mission operations. This proposal aims to demonstrate the technical feasibility of using generative machine learning (ML) to downscale coarse-resolution NWP forecasts (such as HRRR) into high-resolution fine-scale atmospheric fields at 10-100?m horizontal and approximately 10?m vertical resolution. The solution integrates the open-source Sup3r framework, developed by the National Renewable Energy Laboratory (NREL), with high-fidelity training data generated from Aeris’ in-house GPU-accelerated large-eddy simulation (LES) model, JOULES. Paired JOULES-HRRR datasets enable supervised training of ML models to recover sub-kilometer structure from operational NWP inputs, grounded in physically realistic simulations. Sup3r supports a modular architecture for generative modeling, enabling flexible integration of diverse ML architectures. During Phase I, we will configure and evaluate several candidate models, including convolutional neural networks (CNNs), generative adversarial networks (GANs), conditional diffusion models, latent diffusion models (LDMs), and flow-based models. These will be assessed for spatial accuracy, boundary-layer fidelity, uncertainty quantification, and inference performance. Sup3r’s compatibility with PyTorch allows rapid iteration across model types and supports multi-variable outputs such as wind, temperature, and humidity. The Phase I effort includes the creation of a representative JOULES-HRRR training dataset for a Navy-relevant littoral domain, adaptation of Sup3r for this domain, and prototyping of a reduced-scale downscaling pipeline focused on a single architecture. A structured comparative analysis will inform model selection for Phase II. All activities are aligned to a Navy use case involving a carrier strike group operating near an island chain, an environment where small-scale environmental effects directly impact forecasting accuracy and mission planning. By enabling fast, relocatable, high-resolution downscaling from operational forecast inputs, this approach directly addresses the Navy’s need for ML-enabled forecasting under the Machine Learning Weather Prediction (MLWP) initiative and broader METOC modernization priorities.