Data-driven Medium-range Weather Prediction Achieves Comparable Skill to Dynamical Models. But What Does It Mean?

Abstract:

In this talk I will present our latest work on predicting synoptic weather several days ahead with a deep neural network. Our models achieve state-of-the-art results on the WeatherBench challenge in predicting geopotential, temperature and precipitation. Compared to dynamical model baselines, our data-driven models, trained on 5.625 degree resolution data, are comparable in skill to the IFS model run at T63 (approx. 1.9 degree) resolution. To achieve this skill pretraining on climate model simulations was essential. Further we show that the skill of the model scales with resolution. While it is hard to directly compare the skill of the neural networks and the dynamical models, we claim that, given enough data, there is no fundamental reason that data-driven methods cannot be as good as any physics-based model. However, the required amount of data to compete with operational NWP models is likely very large and beyond what is available currently.