Microsoft’s Aurora AI Model Sets New Bar for Forecasting Weather and Beyond

Microsoft Research has introduced a new AI foundation model called Aurora, designed to predict not only the weather but a wide range of environmental phenomena—from air pollution to ocean wave activity. The model, described in a recent Nature paper, leverages vast, diverse datasets to deliver faster, more accurate forecasts than traditional methods—at a fraction of the computational cost.

Aurora’s architecture and training approach mark a shift in how AI models can be used to understand the Earth system. Rather than focusing solely on atmospheric data, the model was built to handle multiple environmental domains, offering new capabilities for governments, industries, and researchers preparing for an era of increasingly extreme weather.

A Foundation Model for the Earth System

Aurora is what researchers call a foundation model—a large AI system trained on massive, varied datasets and fine-tuned for specific applications. What sets Aurora apart is that it was built from the start to be flexible, trained on over one million hours of atmospheric data from satellites, radar, weather stations, and simulations.

While many AI models are fine-tuned to forecast a single variable, Aurora has been adapted to predict a range of environmental conditions, including:

• Air pollution and smog

• Ocean wave patterns

• Cyclones and hurricanes

• Medium-range weather (up to 14 days out)

Aurora outperformed existing numerical and AI-based weather models on 91% of forecasting targets when fine-tuned for mid-range forecasting at a .25-degree resolution.

Incorporating many diverse sources of data results in "not only greater accuracy in general, but it also means we are better at forecasting extreme events,” said Megan Stanley, a senior Microsoft researcher on the project, citing Aurora’s performance across varied test cases.

Accuracy in High-Stakes Scenarios

Tropical Cyclone Forecasting 

In a striking example of Aurora’s capabilities, the model accurately predicted the landfall of Typhoon Doksuri in the Philippines four days ahead of the event—while official forecasts incorrectly placed the storm off the coast of Taiwan.

In global testing, Aurora also beat the National Hurricane Center and seven major forecasting agencies in predicting cyclone tracks for the 2022–2023 season—making it the first machine learning model to do so.

Air Quality and Sandstorm Prediction

Aurora also proved effective in lower-data environments. When a major sandstorm hit Iraq in June 2022, the model predicted the event a day in advance at a much lower cost than traditional air quality forecasts.

Forecasting air quality is especially difficult, as it requires modeling complex chemical interactions and human-caused emissions. Despite not being trained on atmospheric chemistry, Aurora adapted during fine-tuning and delivered reliable predictions—highlighting the power of its foundational training.

“It didn’t learn anything about atmospheric chemistry, or how nitrogen dioxide, for instance, interacts with sunlight — that wasn’t part of the original training,” Stanley noted, “And yet, in fine tuning, Aurora was able to adapt to that, because it had already learned enough about all of the other processes.”

Ocean Wave Forecasting

Aurora has also demonstrated strong performance in forecasting ocean wave activity, a task that typically requires large, complex models and extensive data. The model was fine-tuned to predict wave details such as height, direction, and pattern dynamics, helping it detect subtle shifts in ocean behavior with a level of detail that exceeds traditional systems.

Despite the fact that training data for wave activity was only available from 2016 onward—a relatively short window for such a complex task—Aurora still matched or outperformed the current gold standard in 86% of test comparisons conducted over a full year.

That includes high-impact scenarios such as Typhoon Nanmadol, the most intense typhoon of 2022, which made landfall in Japan in September. The storm brought record-breaking rainfall that triggered landslides, widespread flooding, and power outages that extended as far as South Korea. Aurora was able to model the resulting wave conditions with an accuracy that surpassed the best existing forecasting models, a significant achievement given the storm's scale and complexity.

Researchers attributed Aurora’s success in wave prediction to its ability to perceive intricate wave interactions and coastal dynamics, even with limited fine-tuning data. This level of sensitivity makes the model particularly valuable in preparing for severe maritime weather, helping protect infrastructure, shipping routes, and coastal communities.

Fast, Efficient, and Open to the Public

One of Aurora’s most significant advantages lies in its speed and scalability. Once trained, the model can generate forecasts in just seconds—a sharp contrast to traditional numerical weather prediction systems, which often require hours of computation on large supercomputers.

This leap in efficiency is powered by Aurora’s encoder architecture, which standardizes diverse types of input data—like satellite imagery, radar, and sensor readings—into a unified format the model can process rapidly. Aurora then runs on high-bandwidth GPUs, enabling it to deliver high-resolution predictions with much lower operational costs than legacy systems.

According to Microsoft researchers, Aurora is approximately 5,000 times faster than conventional systems during inference, making it suitable for real-time forecasting at global scale.

To encourage broader use and experimentation, Microsoft has:

• Released Aurora’s source code and model weights publicly, giving researchers and developers full access to the foundation model.

• Integrated Aurora into MSN Weather, where a customized version now provides more frequent, accurate hourly forecasts, including detailed parameters like precipitation and cloud cover.

• Made Aurora available through Azure AI Foundry Labs, Microsoft’s platform for applied AI research, allowing teams across industries to test, adapt, and build on the model.

• Aurora is also available through the European Centre for Medium-Range Weather Forecasts (ECMWF), a widely used platform in the forecasting community to further integrate Aurora into the global forecasting ecosystem.

By making the model both open and fast, Microsoft is helping lower the barrier to advanced environmental forecasting—particularly in regions or sectors without access to supercomputing infrastructure. Aurora’s flexibility also means organizations can tailor it to their own needs, from localized flood modeling to energy grid protection, without having to build forecasting systems from scratch.

What This Means

As weather grows more volatile and interconnected with other global systems, the demand for faster, more adaptable forecasting is only increasing. Aurora demonstrates how foundation models—trained broadly, then refined for specific tasks—can deliver precision and versatility across multiple forecasting domains, even with limited specialized data.

Its early success offers several key advantages:

• Speed and affordability over traditional forecasting systems

• Adaptability across low-data scenarios and diverse regions

• Public access for researchers and institutions worldwide

By significantly lowering the barrier to high-quality environmental forecasting, Aurora has the potential to transform how we prepare for floods, storms, pollution events, and more—especially in regions that currently lack access to advanced forecasting tools.

“It’s the first of its kind,” said Stanley. “But it doesn’t mean it will be the last.”

Looking Ahead

Aurora’s open release positions it as both a research tool and a practical solution. Microsoft sees potential applications in rainfall prediction, crop management, energy grid protection, and other areas where accurate environmental modeling can save lives or resources.

As climate risks rise, AI systems like Aurora may play a growing role in helping societies plan for—and adapt to—a more uncertain world.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.