AI in Weather and Climate Prediction

Narrative

The wave of deep-learning weather models that emerged from 2022 onwards represents the most consequential shift in meteorology since ensemble numerical weather prediction. The first significant contact between big-tech AI methods and operational meteorology came in 2022, when NVIDIA released FourCastNet (Pathak et al.), Huawei published Pangu-Weather (Bi et al.), and Google DeepMind circulated an early GraphCast preprint. FourCastNet used an Adaptive Fourier Neural Operator architecture on a Vision Transformer backbone, training on ERA5 at 0.25-degree resolution and achieving inference speeds five orders of magnitude faster than conventional NWP. Pangu-Weather followed with a 3D Earth-Specific Transformer, becoming the first model to outperform ECMWF's deterministic IFS on all tested variables at that resolution. Both were trained primarily on ERA5 reanalysis, which provided the decades of consistent global atmospheric state data that made supervised learning feasible at scale.

The landmark publication came in November 2023, when Remi Lam and the Google DeepMind team published GraphCast in Science. The model used a message-passing graph neural network on a multi-scale spherical mesh, training directly from ERA5 and fine-tuning on ECMWF HRES analyses, and demonstrated superior accuracy on 90 percent of 1,380 verification targets against HRES, completing a 10-day global forecast in under one minute. GenCast, described in a 2024 Nature paper, extended the architecture to probabilistic ensemble forecasting using a conditional diffusion transformer, outperforming ECMWF's 51-member ENS on 97.2 percent of verification targets and producing physically coherent ensemble members with appropriate spherical harmonic power spectra. Microsoft Research then published Aurora in Nature in May 2025, a 1.3-billion-parameter foundation model pre-trained on over one million hours of geophysical data that beat existing numerical and AI models across 91 percent of forecasting targets and extended to air quality, ocean wave, and hurricane prediction through fine-tuning. Google's NeuralGCM, published in Nature in August 2024, took a different path, embedding learned ML physics parameterisations inside a classical differentiable dynamical core, achieving accurate 1-to-15-day ensemble weather forecasts while also running stable multi-decade climate simulations, the only hybrid model to bridge weather and climate timescales credibly.

Operationalisation moved from research demonstration to production across 2024 and 2025. ECMWF launched AIFS 1.0 on 25 February 2025, running the GNN-based model alongside the physics-based IFS and reporting tropical cyclone track improvements of up to 20 percent and an approximately 1,000-fold reduction in energy use per forecast. A deterministic AIFS Single and a probabilistically trained AIFS ENS were both upgraded to version 2 by May 2026. NOAA declared three AI-driven models operational in December 2025: AIGFS (deterministic), AIGEFS (31-member ensemble), and HGEFS (a 62-member hybrid of AI and physics traces), with AIGFS completing a 16-day global forecast using only 0.3 percent of the computing resources of the traditional GFS. Early AIGEFS skill scores showed forecast lead-time extensions of 18 to 24 hours against the traditional Global Ensemble Forecast System, though tropical cyclone intensity forecasts remained a known weakness.

The fundamental question of whether ML models alter or merely approach the intrinsic predictability limit imposed by atmospheric chaos remains scientifically contested. A 2023 Geophysical Research Letters paper (Selz and Craig) showed that current AI models cannot simulate the butterfly effect and incorrectly suggest unlimited atmospheric predictability, because their error growth rate and structure differ fundamentally from those of high-resolution physics-based models. A 2025 arxiv study explored whether ML models could extend predictability beyond 30 days by using backpropagation to optimise initial conditions, exploiting the differentiability that physics models lack. The consensus is that ML models can approach the theoretical two-week deterministic limit more cheaply than NWP, but do not extend it; the Lorenz chaos horizon remains intact. Ensemble-based ML models such as GenCast address uncertainty quantification more credibly than deterministic counterparts, but challenges around physical conservation, out-of-distribution extreme events, and climate-timescale generalisation persist across all current architectures.