In 2023, a Google DeepMind model called GraphCast outperformed the world's best operational forecasting system on 90% of 1,380 verification targets, producing a 10 day global forecast in under a minute. The result, reported by Lam et al., 2023 in Science, confirmed that a model trained purely on historical reanalysis data could beat physics-based numerical weather prediction on its own terms, at a fraction of the cost.
The convergence is not confined to one architecture. Around the same time, Huawei's Pangu-Weather reached a comparable conclusion by a different route. Bi et al., 2023, in Nature, showed that three-dimensional deep networks with Earth-specific positional priors could match or exceed the ECMWF's Integrated Forecasting System across every tested variable, including cyclone tracking. Two independent teams had arrived at the same finding from different designs: deterministic AI forecasting had closed the gap with decades of numerical modelling.
What deterministic models could not offer was uncertainty. GenCast addressed this directly. A conditional diffusion model described by Price et al., 2025, again in Nature, generates 50 member ensembles of 15 day global forecasts in eight minutes, with greater skill than ECMWF's operational ensemble, ENS, on 97.4% of evaluated targets. The gap left to close was no longer accuracy but the probabilistic reasoning that emergency planning and energy markets depend on, one that values the risk of outliers more then the certainty of averages.
Aurora then extended the same logic beyond weather. Bodnar et al., 2025 trained a single foundation model on more than a million hours of heterogeneous geophysical data and fine-tuned it to outperform operational tools on air quality, ocean waves and cyclone tracks, at far lower cost. The frontier had moved from beating one benchmark to generalising across a whole family of Earth system problems.
The one paper here that pushes towards climate rather than weather runs into a different wall. NeuralGCM, a hybrid model combining a differentiable dynamical core with learned physics, tracks climate metrics for multiple decades and reproduces realistic cyclone frequency. Yet Kochkov et al., 2024 are explicit that the model does not extrapolate to substantially different future climates, since its learned components were never trained on the regimes that warming will produce. AI has not solved the climate problem so much as revealed exactly where the weather problem ends and it begins.