As climate extremes become more common — from typhoons and heatwaves to dust storms and rising seas — the tools we use to understand and prepare for these events are evolving fast. The latest innovation from Microsoft Research is a new AI model called Aurora, setting a new bar for how we forecast not just the weather, but a wide range of environmental phenomena.
Aurora is a large-scale AI foundation model trained on a massive and varied set of data and built to handle far more than temperature or rainfall predictions. While it can do standard weather forecasting well, one of its main features features is its ability to tackle multiple forecasting tasks at once: tracking tropical cyclones, predicting air pollution, modeling wave patterns, and more.
Another is its learning. It was trained on over a million hours of data from satellites, radar, ground sensors, and simulations (likely the largest dataset ever used for an AI weather model) allowing quick fine-tuning for specific jobs using just a small amount of extra data.
Fine-tuning Aurora for new tasks has taken engineers as little as several weeks compared to traditional weather models that can take years. That speed opens the door for faster innovation, especially in places that don’t have access to high-end forecasting tools.
In tests, Aurora outperformed traditional forecasting models in 91% of cases when predicting medium-range weather — that’s forecasts out to 14 days, like what you’d see in a weather app. And it does this incredibly fast! Forecasts that used to take hours on a supercomputer now take just seconds on powerful GPUs.
It’s also proving better at spotting extreme events, which are typically harder to predict. For example, when Typhoon Doksuri hit in 2023, Aurora correctly predicted its landfall in the Philippines four days before it happened while official forecasts missed the mark. These kinds of early warnings can make a real difference on the ground.
Beyond Weather: Air Quality, Ocean Waves, and More
Aurora has also shown strong results in less conventional forecasting areas. In Iraq, where a series of intense sandstorms hit in 2022, the model accurately predicted a major event a full day ahead — and did it using only limited air quality data. That’s significant, since modeling things like smog and dust usually involves complex chemistry simulations and detailed emissions tracking.
In another test, Aurora was used to predict ocean wave activity, including the impact of Typhoon Nanmadol in Japan. Despite being trained on just a few years of wave data, it matched or beat current models in 86% of cases — a sign that the model is picking up meaningful patterns with minimal input.
Open-Source and Ready to Build On
Microsoft has released the model’s code and weights, so researchers and developers can explore it, run it, and build on it. Whether working on renewable energy forecasts, agricultural planning, or even flood risk modeling, Aurora could be a valuable new tool.
Microsoft’s own MSN Weather has already integrated a version of Aurora to offer more detailed and timely forecasts to users. And on Azure’s AI Foundry Labs, developers can experiment with Aurora to solve problems in their own industries.
Researchers say Aurora won’t be replacing existing systems but it’s a powerful addition to the forecasting toolkit. As the model continues to improve and evolve, it could help bring better, faster forecasts to more people, especially in parts of the world that need them most. And this might just be the beginning. Aurora could help in laying groundwork for a whole new generation of environmental models that understand weather and the planet as a whole.
Learn more about Aurora’s innovation and research on Microsoft’s news post.
Leave a Reply