Work

Application of Artificial Intelligence in Weather Prediction: A Mini-Review of Data-Driven Global Forecasting Models

Year: 2026
Type: review
Abstract: Over the past four years, artificial intelligence (AI)-based data-driven models have moved from proof-of-concept demonstrations to systems that consistently outperform operational numerical weather prediction (NWP) on standard verification benchmarks. This mini-review synthesises 62 peer-reviewed publications from 2020 to 2026 to assess the current state of AI weather prediction across four thematic areas: neural network architectures, quantitative benchmark performance, downstream applications, and structural limitations. Four principal architecture families, namely Fourier Neural Operators, Graph Neural Networks, Transformer-based models, and diffusion-based generative systems, have been demonstrated at global scale, producing ten-day forecasts at 0.25-degree resolution in under one minute on commodity graphics processing unit (GPU) hardware. Evaluated against the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System High Resolution (IFS-HRES) operational forecast using the WeatherBench 2 framework, leading models including GraphCast, Pangu-Weather, FuXi, Aurora, and FengWu outperform IFS-HRES at five-day lead times for 500 hPa geopotential height, with root mean square error improvements of 8 to 14 percent relative to the operational benchmark. The probabilistic model GenCast further outperforms the ECMWF ensemble system on 97.4 percent of probabilistic verification targets. Beyond standard benchmarking, AI models demonstrate consistent advantages for tropical cyclone track prediction and precipitation nowcasting, but exhibit systematic underperformance for record-breaking surface extremes, tropical cyclone intensity, and convective-scale precipitation, limitations that reflect training exclusively on the ECMWF Reanalysis v5 (ERA5) dataset, the absence of physical conservation laws, and incomplete extreme-event evaluation frameworks. Emerging hybrid physics-AI architectures, end-to-end data assimilation systems, and foundation models trained on diverse multi-source datasets represent the most viable pathways to addressing these limitations and realising the operational potential of AI weather prediction.
Language: English
Open Access status: bronze
OpenAlex ID: W7222942552
FWCI:
Cites: 59
Cited by:
Related to: 9