Abstract

As Santa’s elves wrapped the world’s Christmas presents in the final days of 2023, a group of scientists at Google DeepMind kept busy wrapping their own gift to the numerical weather prediction community. GenCast, a diffusion model trained to produce probabilistic global medium-range weather forecasts at 12-hourly, 1-degree resolution, was published to arXiv on 25 December 2023 and represents the latest step forward in a field moving at hurricane pace. Traditionally relying upon physics-based models which solve systems of differential equations describing known atmospheric behaviours and tendencies in discrete cells, weather prediction has seen an influx of attention from the machine learning community in the past two years, from Google DeepMind’s GraphCast to NVIDIA ’s FourCastNet to Huawei’s PanguWeather to Microsoft’s ClimaX. At its core, weather prediction is one of image-to-image translation, or perhaps video frame prediction—where red, green, and blue channels of an input and target image are swapped with physical variables such as temperature, pressure, and wind speed in an observed state and a target state some hours ahead—and otherwise arbitrary pixel space represents a fixed discretization of the globe in latitude and longitude. Thus, it seems natural for the resounding success of diffusion models in conditional image generation to translate to this task. In contrast to previous approaches using ML for weather, GenCast leverages the stochastic nature of diffusion models to produce ensemble forecasts, each generated through conditional denoising of independent noise samples, thereby maintaining physical consistency and avoiding the unrealistic spatial smoothing that comes with the use of MSE as a training strategy for deterministic forecasting models. In this talk, we will present a refresher on numerical weather prediction and diffusion models, dive into the clever details of GenCast and the engineering secrets to its success, compare GenCast with other contemporary work in the ML for weather space, and conclude by highlighting the many opportunities that remain for ML across the geosciences.

Suggested Reading: GenCast: Diffusion-based ensemble forecasting for medium-range weather (https://arxiv.org/abs/2312.15796)