The dynamics of weather systems are among the most complex physical phenomena on Earth, and each day, countless decisions depend on accurate weather forecasts, from deciding whether to wear a jacket or to flee a dangerous storm. Until recently, the dominant approach for weather forecasting was “numerical weather prediction” (NWP), which involves solving the governing equations of weather using supercomputers. In the last year, deep learning methods have surpassed NWPs at deterministic global weather forecasting and are showing promising results in probabilistic forecasting as well.
This talk will focus on the use of Graph Neural Networks(GNNs) for weather forecasting. First, we will give a brief overview on how GNNs are related to Finite Element Methods used in traditional models, leveraging some of their useful inductive biases(Alet et al. ‘19). Then, we will delve into two recent works on GNNs for weather forecasting: 1. GraphCast (Lam et al. ‘23), which used GNNs for state-of-the-art deterministic weather forecasting at 0.25 degrees of resolution, as well as predicting tropical cyclones tracks and 2. GenCast (Price et al. ‘23), which combined ideas from GraphCast and diffusion models for probabilistic weather forecasts at 1 degree better than the best physics model.
Sources:
- Graph Element Networks: adaptive, structured computation and memory (Alet et al., ICML ‘19)
https://arxiv.org/abs/1904.09019
- GraphCast: learning skillful medium-range weather forecasting (Lam et al., Science ‘23)
https://arxiv.org/pdf/2212.12794.pdf
- GenCast: diffusion-based ensemble forecasting for medium-range weather (Price et al., arxiv ‘23)
https://arxiv.org/abs/2312.15796
Link to join virtually: https://cam-ac-uk.zoom.us/j/81322468305
This talk is being recorded.
