weatherbench

Probabilistic Downscaling of Climate Variables Using Denoising Diffusion Probabilistic Models

Simplified, reproducible reimplementation of DeepMind's GraphCast weather model using graph neural networks (JAX · Haiku · Jraph). Runs on a single GPU, with honest negative-results reporting and detailed failure-mode analysis of a heavily scaled-down forecasting model.

Joint-embedding atmospheric world-model research for transferable, uncertainty-aware weather representation learning.