Reproduce “CoDiCast: Conditional Diffusion Model for Global Weather Prediction with Uncertainty Quantification” on OpenResearch

About this paper

Jimeng Shi, Bowen Jin, Jiawei Han, Sundararaman Gopalakrishnan, Giri Narasimhan

CoDiCast introduces a conditional diffusion model for global medium-range weather forecasting, delivering accurate predictions with robust uncertainty quantification at a computationally efficient cost. It addresses the limitations of both traditional numerical weather prediction and existing machine learning-based approaches by providing inherent probabilistic outputs.

Key insights

• Conditional diffusion models naturally capture and quantify forecast uncertainty by generating an ensemble of plausible future scenarios from different initial noise samples during inference.
• Effectively integrating historical weather observations through a pre-trained encoder and cross-attention allows the diffusion model to produce accurate predictions consistent with past states.
• The inherent probabilistic and generative capabilities of DDPMs can bridge the gap between deterministic MLWP and computationally expensive NWP uncertainty quantification.

Summary by alphaXiv.