A review of deep learning for weather prediction

Recent years have seen substantial performance-improvements of deep-learning-based
weather prediction models (DLWPs). These models cover a large range of temporal and
spatial resolutions—from nowcasting to seasonal forecasting and on scales ranging from
single to hundreds of kilometers. DLWPs also exhibit a wide variety of neural architec-
tures and training schemes, with no clear consensus on best practices. Focusing on the
short-to-mid-term forecasting ranges, we review several recent, best-performing models
with respect to critical design choices. We emphasize the importance of self-organizing
latent representations and inductive biases in DLWPs: While NWPs are designed to sim-
ulate resolvable physical processes and integrate unresolvable subgrid-scale processes by
approximate parameterizations, DLWPs allow the latent representation of both kinds of
dynamics. The purpose of this review is to facilitate targeted research developments and
understanding of how design choices influence performance of DLWPs. While there is
no single best model, we highlight promising avenues towards accurate spatio-temporal
modeling, probabilistic forecasts and computationally efficient training and infer