Machine Learning Weather Prediction: Models, Verification and Limits

Between 2022 and 2025 a series of neural networks trained on four decades of reanalysis data began to match, and on many headline scores to beat, the best physics-based global weather forecasts, at a small fraction of the running cost. This review describes that change for an information technology readership and reports no new experiment. We begin with the numerical baseline and the ERA5 reanalysis on which nearly all of the models are trained. We then follow the deterministic models built on convolutions, neural operators, transformers and graph networks, the probabilistic and hybrid systems that came after them, and precipitation nowcasting from radar and satellite as a neighbouring problem. Benchmarks and published scores are then set beside the critical studies, which found forecasts that blur as lead time grows, weak physical balance at small scales, missing error growth from tiny perturbations, underestimated wind peaks in individual storms and poorer skill on record-breaking extremes. The same body of work also found realistic dynamical behaviour, and we report both sides. Compute cost, explainability and the dependence on numerical analyses for training and initial states close the review.