Ingests raw weather station observations from CSV and API sources (SynopticPy, WeatherReal-Benchmark) and converts them into ML-ready NetCDF structures. It solves a recurring data problem—normalizing heterogeneous point observations into a uniform format—and is extracted from the forecasting pipeline with minimal internal coupling.

Builds sparse bipartite connectivity graphs from latitude/longitude coordinates using H3 hexagonal grid hierarchies, with caching for repeated use. This is a reusable spatial-data pattern for any GeoAI or mesh-based GNN project requiring efficient spherical graph construction from raw coordinates.

Generates connectivity between regular latitude/longitude grids and icosahedral meshes, producing edge indices and local coordinate rotation matrices. The spherical-geometry logic is self-contained and applicable to other earth-system ML models that need irregular mesh-to-grid transfer.

Provides isotropic Gaussian noise generation for diffusion model training, including noise-level conditioning and denoising parametrization. This is a clean, reusable component for probabilistic generative modeling on gridded or meshed atmospheric data.

Loads ERA5 reanalysis from WeatherBench2 cloud storage, applies normalization and target noise augmentation for diffusion training. It demonstrates a robust pattern for streaming large geophysical datasets with reproducible transforms, though the storage path is somewhat domain-specific.