We introduce a two-stage pipeline for probabilistic precipitation nowcasting. 1) We develop a latent diffusion model PreDiff. (2) We incorporate an auxiliary knowledge alignment mechanism, allowing the integration of domain expertise.

We propose Earthformer with novel generic building blocks “Cuboid Attention” to explore the design of space-time attention for Earth system forecasting problems, and achieve SOTA performance on two synthetic datasets and two real-world benchmarks.