In the relentless pursuit of improving weather forecasting, recent breakthroughs have illuminated a new path for high-resolution prediction models that stand to revolutionize how we understand and anticipate severe weather events. Traditionally, weather prediction has been constrained by the resolution limits of computational models and the availability of high-quality observational data. However, the advent of machine learning and powerful foundation models tailored to the Earth system signals a paradigm shift poised to enhance accuracy and extend the horizons of operational weather forecasts.
At the forefront of this advancement is Aurora, an innovative AI-driven forecasting system meticulously developed to operate at an unprecedented spatial resolution of approximately 0.1° in mid-latitudes. This resolution is comparable to the high-resolution configuration of the Integrated Forecasting System (IFS) known as HRES, which currently represents the gold standard in numerical weather prediction with its Gaussian grid (TCo1279) framework. By matching this granularity, Aurora confronts the limitations that have historically relegated AI weather models to coarser scales, predominantly 0.25°, dictated by data availability and computational feasibility.
The significance of operating at a finer resolution cannot be overstated. Weather phenomena such as convective storms, boundary layer processes, and other mesoscale dynamics unfold at spatial scales that are poorly resolved at coarser grids. Capturing these features with fidelity is crucial for accurately forecasting high-impact weather events, including severe storms and rapid atmospheric changes. Aurora’s capacity to work directly with 0.1° resolution data, which has only become accessible since 2016, illustrates a leap forward beyond previous AI systems tethered to more abundant but less detailed datasets.
Achieving this feat required employing a novel pretraining–fine-tuning protocol. Initially, Aurora was pretrained on extensive datasets at coarser resolutions to learn generalized atmospheric representations robustly. It was then fine-tuned on the relatively newer high-resolution IFS HRES analysis data spanning from 2016 to 2022. This approach harnesses the strengths of both data regimes, enabling Aurora to efficiently adapt to finer scales without necessitating prohibitively large volumes of high-resolution data for training from scratch. The result is a model that surpasses the forecasting skill of the operational IFS HRES when evaluated under established protocols.
Evaluation of Aurora’s performance reveals compelling advantages. When measured against the root mean squared error (RMSE) across a comprehensive set of meteorological target variables, pressure levels, and forecast lead times, Aurora outperforms IFS HRES in over 92% of comparisons. This superiority becomes especially pronounced beyond the 12-hour lead time mark, with RMSE reductions reaching up to 24%. Such improvements signal enhanced reliability in medium- to long-range forecasts, a critical window for disaster preparedness and mitigation efforts.
Interestingly, at the shortest lead times, classical numerical methods represented by IFS HRES maintain an edge, a pattern consistent with other contemporary AI forecasting models. This disparity underscores the complementary nature of traditional and AI-based predictive approaches, with AI models delivering significant enhancements as forecast horizons extend. The interplay between model types may pave the way for hybrid systems that exploit the strengths of each methodology.
To further validate Aurora’s real-world applicability, researchers conducted extensive evaluations using the WeatherReal-ISD dataset, a rich compilation of in situ measurements from over 13,000 weather observation stations worldwide. These assessments focused on key surface variables—10-meter wind speed and 2-meter air temperature—across forecast lead times extending up to 10 days. Aurora consistently outperformed IFS HRES in reducing forecast error across this entire range, indicating robust skill in capturing surface atmospheric phenomena crucial for everyday weather impacts.
The case for pretraining is further reinforced by quantitative analyses demonstrating that models trained with a pretraining stage hold a 25% performance advantage over those trained from scratch using only high-resolution data. This finding emphasizes the value of leveraging historical, abundant datasets to bootstrap learning before specializing in new, high-resolution regimes with sparser data availability.
The practical strengths of Aurora are vividly illustrated in a detailed case study of Storm Ciarán, a powerful mid-latitude storm that swept across Northwest Europe in late 2023. The storm generated record-breaking low pressure readings in England during November, raising significant forecasting challenges. When initialized at 31 October 00 UTC, comparisons among several AI models highlighted Aurora as uniquely capable of capturing the abrupt surge in maximum 10-meter wind speeds, closely aligning with the ground truth provided by the IFS analysis. Other AI models, including FourCastNet, GraphCast, and Pangu-Weather, notably failed to reproduce this rapid intensification.
This breakthrough not only demonstrates Aurora’s superior predictive skill but also underscores the crucial role of spatial resolution and methodological innovations in capturing extreme weather dynamics. The accurate representation of such rapid and localized phenomena is vital for issuing warnings and protecting communities from the devastating impacts of severe storms and other high-impact weather events.
Methodologically, it is worth noting that for the Storm Ciarán predictions, Aurora was run without Low-Rank Adaptation (LoRA), a model compression technique often employed to reduce computational costs. This decision was aimed at maximizing the model’s sensitivity to extreme event dynamics, highlighting the flexibility embedded in Aurora’s design to balance efficiency and precision based on situational requirements.
The success of Aurora, a foundation AI model for the Earth system, signals a new era in atmospheric science where machine learning models are not merely complementary but can supersede traditional forecasting systems in accuracy and temporal reach. By bridging the data and resolution gap that has long hindered AI weather prediction, this approach unlocks latent capacity for more reliable, higher fidelity forecasts.
Looking ahead, the integration of such foundation models promises transformative impacts across sectors reliant on weather information—from disaster risk management and agriculture to renewable energy and transportation. As data acquisition systems continue to improve and resolutions increase, the potential for AI systems like Aurora to leverage ever more granular observations will only grow, leading to smarter, faster, and more actionable weather forecasts worldwide.
In conclusion, Aurora’s demonstrated capabilities exemplify a pivotal step towards realizing the vision of foundation models that encapsulate the complex, multiscale intricacies of the Earth system. Through innovative training protocols and leveraging advances in data availability, Aurora establishes a new benchmark for operational weather prediction, fostering optimism for future developments at the nexus of atmospheric science and artificial intelligence.
Subject of Research:
Development and evaluation of a high-resolution AI-based weather forecasting model (Aurora) designed to surpass state-of-the-art numerical weather prediction systems.
Article Title:
A foundation model for the Earth system
Article References:
Bodnar, C., Bruinsma, W.P., Lucic, A. et al. A foundation model for the Earth system. Nature (2025). https://doi.org/10.1038/s41586-025-09005-y
Image Credits:
AI Generated
