AI used to ‘democratize’ how we predict the weather 

Weather prediction systems provide critical information about dangerous storms, deadly heatwaves and potential droughts, among other climate emergencies.  

But they’re not always accurate. And, ironically, the supercomputers that generate forecasts are also energy-intensive, contributing to greenhouse gas emissions while predicting increasingly erratic weather caused by climate change.  

“The process right now is very computationally expensive,” says James Requeima, a post-doctoral researcher at the University of Toronto and the Vector Institute.  

Enter Aardvark Weather, a weather prediction model developed by Requeima and other researchers using artificial intelligence (AI). Described in a recent Nature article, the system produces results comparable to traditional methods, but is 10 times faster, uses a tiny fraction of the data and consumes 1,000 times less computing power.  

In fact, the model can be run on a regular computer or laptop. It’s also open-source and easily customizable, allowing small organizations, developing countries or people in remote regions to input the data they have and generate local forecasts on a minimal budget. 

The development could be a timely one. As Texas continues to deal with the fallout from catastrophic floods, Manitoba grapples with its most destructive wildfire season in 30 years and Europe reels from deadly heatwaves, there’s a clear need for accessible and accurate weather forecasting around the world.

“You hear a lot about the promise of AI to help people and hopefully make humanity better,” Requeima says. “We’re hoping to enact some of that promise with these weather prediction models.” 

Aardvark Weather is being developed at Cambridge University – where Requeima completed his PhD in engineering and machine learning – and the Alan Turing Institute. Requeima joined the project in 2023. He received post-doctoral funding for the project last year from U of T’s Data Science Institute, an institutional strategic initiative.  

U of T News recently spoke to Requeima about the project and his role. 

How is weather currently predicted? 

The big weather forecasters, such as the U.S. National Weather Service and the European Centre for Medium-Range Weather Forecasts, take initial conditions representing the current state of the atmosphere and put that information into a supercomputer. They then run a numerical simulation and propagate that forward into the future to get forecasts of the future states of the atmosphere.  

Then they take observations from real-world sensing instruments and incorporate them into their current belief about the atmosphere and re-run the forecast. There’s a constant iterative loop. From these atmospheric predictions, you can build a tornado forecaster or a precipitation forecaster. 

How can AI do better and with less computing power? 

End-to-end deep learning fundamentally changes how we approach weather prediction. Rather than the traditional, iterative process that relies on expensive numerical simulations, we train our model to map directly from sensor inputs to the weather variables we care about. We feed in raw observational data – from satellites, ships and weather stations – and the model learns to predict precipitation, atmospheric pressure, and other conditions directly. While training the initial model requires computational resources, once trained, it’s remarkably efficient. The resulting system is lightweight enough to run on a laptop, making predictions orders of magnitude faster and more accessible than traditional supercomputer-based methods.

This means communities can deploy these models locally to generate their own forecasts for the specific weather patterns that matter to them.

Have others used AI for weather prediction? 

Machine learning has been applied to climate modelling before, but previous approaches still depended on numerical simulations as their input. Our key breakthrough is demonstrating that you can move out of this paradigm and map directly from observation to targets. This proof of concept opens up a fundamentally new approach to forecasting – we've demonstrated that accurate weather prediction doesn’t require supercomputer simulations as an intermediate step.

How can this technology be used in practice? 

We are open sourcing this model – making it available to the community so others will improve upon our model to make changes and train it to do local modelling. We’re hoping this will help democratize weather prediction.  

Forecasting quality is correlated with wealth, so developing nations don't have access to as good forecasting as wealthier nations do. If we can help bring high-quality forecasting to areas that don't have it before, that’s a really big positive of this work.  

David [Duvenaud, an associate professor of computer science in U of T’s Faculty of Arts & Science] – my adviser – and I want to use AI in positive ways. Climate prediction is an important tool for assessing and developing ways of dealing with climate change – and the better climate models we have, the better our science can be around tackling that problem. That’s a driving motivation for me. 

What was your contribution to this work? 

During my PhD, I worked on neural processes – a type of neural network model that is effective for numerical forecasting. We discovered it was well-suited for scientific applications, especially climate modelling. For Aardvark, I helped design the model architecture and the multi-stage training scheme. 

Where did the name Aardvark Weather come from?  

The first author on this research, Anna Allen from Cambridge, did a lot of the heavy lifting on this – which is going out and finding the data sources, including a lot of Canadian data from weather stations, weather balloons and ship observations. She’s from Australia and is a lover of interesting animals like sloths – and aardvarks.