New AI Model Promises Faster, Smarter Weather Predictions

Aardvark Weather, a new AI model developed by researchers in the UK and Canada, could mark a turning point in global weather forecasting by replacing traditional weather simulations with artificial intelligence to maximize cost efficiency and accuracy.

Researchers from the University of Cambridge, the Vector Institute at the University of Toronto, and the Alan Turing Institute unveiled the new findings in a recent report published in Nature.

Unlike conventional forecasting tools that simulate atmospheric physics through complex equations, Aardvark Weather is a “deep learning” model that generates global forecasts for wind, humidity, geopotential, and temperature at multiple pressure levels.

It also delivers local station forecasts for 2-meter temperature and 10-meter wind speed. Deep learning is a subset of machine learning that teaches computers to recognize patterns in large amounts of data.

“At the moment, there are some computationally expensive components in the forecasting pipeline,” postdoctoral fellow at the University of Toronto’s Vector Institute James Requeima told Decrypt. “We’ve been able to replace many of these time-consuming parts with much lighter-weight models trained to perform the same tasks.”

By making those components more efficient, Aardvark could run forecasts more often and at higher resolutions, improving speed and accuracy.

As Requeima explained, the team designed components to replace each step in the forecasting pipeline, which involves turning raw observational data into a weather forecast.

“We found that once these machine learning components are chained together, the overall performance improves significantly,” he said. “By fine-tuning the entire pipeline for the final task we’re targeting, we can optimize each component not just for its isolated role, but for how it contributes to the outcome we care most about.”

The project also included researchers from Microsoft Research Cambridge, the European Centre for Medium-Range Weather Forecasts (ECMWF), and the British Antarctic Survey.

Aardvark Weather uses raw atmospheric data—like pressure, temperature, and relative humidity measurements—to produce high-resolution global and local forecasts.

The system is built around three neural components: an encoder, a processor, and a decoder.

• Encoder: Converts raw, unstructured observational data into a gridded representation of the atmosphere.
• Processor: Generates weather forecasts from the gridded data.
• Decoder: Translates the forecasts into specific local predictions.

To improve Aardvark’s performance and accuracy, components are first pre-trained on ERA5 reanalysis data—a high-quality historical dataset from ECMWF—and then fine-tuned using real-world weather observations.

“Data assimilation, in general, works like an autoregressive procedure. You start with the current atmospheric forecast, generated by large dynamical systems that estimate its present state. At time zero, you have this initial state,” Requeima said. “But data assimilation also needs to incorporate real-time measurements from remote sensors. So, you gather actual observations alongside the model’s forecast and adjust your atmosphere estimate accordingly.”

A Fraction of the Cost—and Time

According to the report, Aardvark can generate a full global forecast using four NVIDIA A100 GPUs in just one second compared to the hours needed by older models like the European Centre for Medium-Range Weather Forecasts’ high-resolution forecast.

This drastic reduction in computing requirements makes high-quality, customizable forecasting accessible to regions and agencies without the resources to operate full-scale NWP systems. It also enables much faster fine-tuning of the model.

Aardvark joins a growing suite of tools aimed at helping meteorologists predict and respond to extreme weather events. During recent storms, such as Hurricanes Helene and Milton, which battered the U.S. East Coast in October 2024, forecasters emphasized the importance of AI in improving storm intensity prediction.

Looking ahead, Requeima noted that the team plans to open source Aardvark to make the technology more widely accessible.

“I think it’s an important step toward democratizing weather modeling—making it more lightweight and accessible to the public,” he said. “That’s our hope. It also represents a major advancement in end-to-end weather modeling, particularly through a data-driven, machine learning approach.”

Edited by Sebastian Sinclair and Josh Quittner