A new AI model is outperforming the world’s top forecasting systems for weather, pollution and cyclones, according to a new study, boosting hopes of weather forecasting becoming cheaper and more accurate.
The model, called Aurora, accurately predicted cyclone paths and produced weather forecasts in a matter of seconds instead of hours.
It was trained on a vast collection of atmospheric data, like weather observations, climate simulations and satellite measurements, by researchers at Microsoft and the University of Pennsylvania.
When evaluated against global forecasting benchmarks, the AI system consistently produced faster forecasts than traditional models and, in many cases, offered greater accuracy, according to the new research published in Nature.
Aurora was able to predict the path of Doksuri, the costliest Pacific typhoon of 2023, four days before landfall. While official weather agencies forecast landfall in Taiwan, Aurora correctly placed it in the northern Philippines.
It also tracked the path and wind speeds of the storm Ciarán, which struck northwestern Europe last autumn, outperforming traditional models as well as newer systems based on AI like GraphCast and FourCastNet.
According to the study, Aurora was the only model to correctly estimate peak winds from the storm.
The results mark a major advance in modelling complex Earth systems with speed and accuracy. “Earth’s climate is perhaps the most complex system we study, with interactions spanning from quantum scales to planetary dynamics,” noted Dr Paris Perdikaris, associate professor at the University of Pennsylvania.
“With Aurora, we addressed a fundamental challenge in Earth system prediction: how to create forecasting tools that are both more accurate and dramatically more computationally efficient.”
The system is not limited to weather. Aurora has also been tested for forecasting air quality and ocean waves. In one case study, it accurately predicted a large sandstorm in Iraq, which closed airports and led to over 5,000 hospitalisations, a day before it occurred.
The model managed to do this despite being trained without explicit knowledge of atmospheric chemistry. Aurora “did not have any prior knowledge about atmospheric chemistry or how nitrogen dioxide, for instance, interacts with sunlight,” said study co-first author Dr Megan Stanley of Microsoft Research, “that wasn’t part of the original training.”
“And yet,” she said, “in fine-tuning, Aurora was able to adapt to that because it had already learned enough about all of the other processes”.
The model was also able to simulate complex ocean wave patterns generated by typhoons such as Nanmadol, which struck Japan in 2022. Aurora captured wave heights and direction with more detail and higher accuracy than the standard ocean forecasting systems in use today.
“When we compared Aurora to official forecasts from agencies like the National Hurricane Centre, China Meteorological Administration and others, Aurora outperformed all of them across different basins worldwide,” said Dr Perdikaris.
The model works by identifying patterns in large environmental datasets instead of solving physical equations. This allows it to generate 10-day weather forecasts and 5-day air quality predictions in under a minute, compared to the hours needed by traditional models running on supercomputers.
Unlike traditional systems that need supercomputers, a key advantage of Aurora is that it can run on simpler machines. This could make accurate local forecasts possible even in countries with limited resources.
“The most transformative aspect is democratising access to high-quality forecasts,” Dr Perdikaris said. “Traditional systems require supercomputers and specialised teams, putting them out of reach for many communities worldwide. Aurora can run on modest hardware while matching or exceeding traditional model performance.”
The new AI model’s foundation architecture allows it to be fine-tuned for various forecasting tasks, from local rain patterns to seasonal trends. “Knowledge gained from one area, such as atmospheric dynamics used in weather forecasting, enhances its predictive performance in other domains, including air quality modelling or predicting tropical cyclone formation,” noted Dr Perdikaris.
“This cross-domain learning is central to the foundation model philosophy that guides my broader research programme.”
Each new application requires only a small amount of additional data.
According to Microsoft, some fine-tuning experiments took only a few weeks compared to the years typically needed to build numerical models.
Although Aurora still needs existing data sources to generate forecasts, researchers say its speed and flexibility could make it useful for real-time applications in the future.
Microsoft says the source code and model weights are publicly available and Aurora is already being used to improve weather services on its MSN platform.
The researchers are interested in extending the model to generate predictions on a wider range of Earth system behaviours, including local and seasonal weather, extreme rainfall and urban flooding.
“What excites me most about this technology is its broader applicability,” Dr Perdikaris. “At Penn, we are exploring how similar foundation model approaches can address other prediction challenges beyond weather – from urban flooding to renewable energy forecasting to air quality management – making powerful predictive tools accessible to communities that need them most.”
Its developers believe that similar systems could eventually be adapted for other forecasting challenges, including floods, heatwaves and agriculture.
But while early results are promising, further evaluations are needed to understand how AI models like Aurora perform in diverse real-world settings.
