Flexible Graph Construction for Neural Weather Prediction

Neural weather prediction models like neural-lam rely on graph neural networks, but the current graph construction pipeline is locked to rectangular grids — blocking support for irregular data sources like weather stations, satellite swaths, and icosahedral model grids (ICON, MPAS). This project replaces the hardcoded pipeline with a flexible, topology-agnostic architecture spanning two repositories (weather-model-graphs and neural-lam). The core work delivers three things: (1) a two-step mesh architecture that decouples node placement from connectivity, enabling any layout (triangular, prebuilt, density-adaptive) to combine with any connectivity mode; (2) a bridge layer that eliminates 600+ lines of duplicated code by making neural-lam call weather-model-graphs directly, validated through a shared schema contract; and (3) migration to PyTorch Geometric's HeteroData for type-safe, self-documenting graph representation. As stretch goals, the project introduces a graph quality metrics framework for comparing mesh topologies without training, density-adaptive meshing driven by data or prediction error, and spherical coordinate support to fix systematic distortion at high latitudes. I have already contributed the foundational PRs (#81, #91, #92, #123, #258) and authored the two strategic issues (#384, #385) that define this project's architecture.