| Popis: |
Recent years have seen substantial performance-improvements of deep-learning-basedweather prediction models (DLWPs). These models cover a large range of temporal andspatial resolutions—from nowcasting to seasonal forecasting and on scales ranging fromsingle 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 theshort-to-mid-term forecasting ranges, we review several recent, best-performing modelswith respect to critical design choices. We emphasize the importance of self-organizinglatent representations and inductive biases in DLWPs: While NWPs are designed to sim-ulate resolvable physical processes and integrate unresolvable subgrid-scale processes byapproximate parameterizations, DLWPs allow the latent representation of both kinds ofdynamics. The purpose of this review is to facilitate targeted research developments andunderstanding of how design choices influence performance of DLWPs. While there isno single best model, we highlight promising avenues towards accurate spatio-temporalmodeling, probabilistic forecasts and computationally efficient training and infer |
