Machine learning weather prediction at the operational threshold: architectures, data, evaluation, and the observation-interface frontier

Machine learning weather prediction (MLWP) has moved from research demonstration into operational practice in roughly five years. Global models trained on reanalysis now produce ten- to fifteen-day forecasts at 0.25 degree resolution in seconds, and several match or exceed established numerical weather prediction (NWP) systems on standard verification metrics. This review examines what has been built, what it has been built on, and how it is being evaluated. Six architectural families are distinguished: spherical graph networks, three-dimensional transformers, cascaded multi-horizon systems, diffusion and flow-based generative ensembles, geometric mesh-based ensemble systems, and hybrid models that couple a differentiable dynamical core to learned physics. The data substrate receives detailed treatment, because ERA5 is a model-dependent product rather than a neutral training set, and every model trained on it inherits the analysis increments, parameterised tendencies, and changing observational coverage of that reanalysis. The verification stack is examined with equal care. Root-mean-square error and anomaly correlation are dominated by planetary-scale, slowly evolving flow, which makes them least sensitive to the failures that matter most: spectral deficit, over-smoothing of extremes, displacement error, and regime-dependent breakdown. A four-tier experimental protocol is proposed, combining reforecast benchmarking, out-of-distribution and regime-conditional testing, spectral and conservation diagnostics, and block bootstrap inference. The review then documents the frontier that emerged in 2025 and 2026, in which end-to-end systems ingest observations directly and bypass conventional data assimilation, and foundation models share one backbone across atmosphere, ocean, cryosphere, and atmospheric composition. The argument throughout is that the field's bottleneck has shifted from architecture to the observation interface and to evaluation practice, and that the credibility of MLWP over the next five years will rest on whether its claims survive out-of-distribution, conservation-aware, and statistically defensible scrutiny.