AI Merged With Human Knowledge Produces the Best Possible Weather Forecasts

Abstract

A new approach called cross-attractor transforms (Agarwal et al., 2025, https://doi.org/10.1029/2024gl110472) aims to improve weather forecasts by using neural networks to learn optimal maps between nature and imperfect numerical weather prediction (NWP) models. Unlike the latest generation of machine learning weather prediction (MLWP) models, this approach leverages prior knowledge via the known governing equations and learns only what is needed to map between that imperfect numerical model and the target system being forecasted (e.g., real-world weather). This approach draws from the same underlying principles of dynamical systems theory and chaos theory that have been the foundation of operational NWP for the last half century, and extend upon machine learning based post-processing efforts. The results show that an imperfect numerical model enhanced by the cross-attractor transforms have the potential to outperform both MLWP models and NWP models post-processed with ML, highlighting the value in merging prior knowledge with data-driven ML methods.