Accurate weather prediction plays a vital role in our lives, more so than ever as we adapt to a changing climate.
Wherever we are in the world, our regular forecasts influence everything from crucial decisions in agriculture, transport and energy through to warnings for communities facing floods or extreme heat.
With this scope for societal benefit, it’s unsurprising to see the excitement generated by AI weather initiatives, from Huawei’s Pangu-Weather to Google DeepMind’s GenCast.
Here at the Turing, using AI to enhance weather prediction is a key aspect of our Environment and Sustainability grand challenge. Taking forward work developed at the University of Cambridge, we have set up a team to deliver the first of a new generation of AI weather prediction systems, called Aardvark.
Aardvark is exciting researchers as it could lead to a paradigm shift in weather prediction that is faster, more accurate, more flexible and less computationally expensive than all other traditional and AI-enabled approaches to date.
How do we predict the weather?
To understand Aardvark’s value, it’s important to explain how weather forecasts are currently created.
Meteorological services around the world use a long-established approach called numerical weather prediction, which requires three steps:
- Step one: meteorologists gather information from satellites, weather stations, weather balloons, boats, buoys, ships and aircraft, and combine this with a recent forecast to estimate the current state of the atmosphere.
- Step two: a complex computational model of the atmosphere evolves the current state forward in time to produce weather forecasts.
- Step three: the forecasts are processed to ensure that they are ready for use at particular locations, correcting biases and increasing spatial granularity, and input from human forecasters is incorporated.
This pipeline requires huge supercomputers, complex software and large support teams. Whilst this is possible in countries like the UK, the requirement for substantial computational power means numerical weather prediction systems are often less practical in developing regions of the world such as sub-Saharan Africa.
Over the last two years, both tech companies and meteorology services have started to use AI models to replace the second step in the pipeline, making it far faster and more accurate. However, the first step of the pipeline has remained untouched by AI and since it is every bit as expensive as the second, this first generation of AI forecasting models still has to be supported by huge supercomputers and large teams.
What’s new about Aardvark?
Aardvark is the first system that replaces all of the steps in the weather forecasting pipeline with a single AI model that can be trained and run on a desktop computer. Aardvark’s forecasts are thousands of times faster than all existing traditional and AI-based forecasting systems.
Aardvark takes in multimodal data from satellites, weather stations and weather balloons, and produces a ten-day global forecast.
This data is complicated, often originating from specific locations around the world at different times and having complex patterns of missing data. So a novel deep learning architecture is necessary to process it. This architecture is trained to forecast weather up to ten days into the future using a large historical dataset.
Why does this matter? First and foremost, this could offer real value to the developing world where access to supercomputers and complex infrastructure and expertise is more limited. But the technology can have utility anywhere, improving efficiency and accuracy, and even slashing the substantial carbon footprint of weather prediction.
Comparing an Aardvark wind speed forecast (left) with the actual weather (ground truth, right) shows that Aardvark can accurately predict wind speeds around the world, including the formation of 2018 tropical cyclone Berguitta
Assessing opportunities and challenges
Aardvark reimagines current weather prediction methods, offering a range of opportunities.
Alongside requiring less computing power, Aardvark is fast. Traditional forecasts can take hours to produce on a supercomputer whereas, once trained, Aardvark can create forecasts within minutes and can be run on a desktop computer.
There are highly promising signs of Aardvark’s accuracy too. Globally, Aardvark is already as accurate as America’s Global Forecast System (GFS), but it is only using about 10% of the available data to make its forecasts, meaning that further improvements in accuracy should be possible. We are excited to see what happens as we increase the amount of data and optimise Aardvark end-to-end to provide more accurate forecasts. This new paradigm could replace the traditional numerical approach in developing countries.
A streamlined system like this could also play a significant role in democratising access to advanced forecasting tools, empowering developing or data-sparse countries to build capacity and create bespoke weather forecasting systems that previously would have required large teams to operate, deploy and maintain.
There are of course challenges and it’s important to acknowledge that, whilst machine learning weather tools are moving at great pace, this is still an experimental technology that will require rigorous evaluation over a period of time.
Weather prediction tools must accurately predict all types of weather, and extremes like hurricanes and floods are especially important. Unfortunately, rare events like these are less represented in the training data, meaning that AI systems may struggle more on these phenomena.
We also need to ensure we account for our changing climate, which could render models trained on past data less accurate.
Early signs suggest that we can rise to these challenges.
What’s next for Aardvark?
Bringing the ideas behind Aardvark into the Turing will advance the institute’s wider work to target a step-change in high-precision environmental forecasting for weather, oceans and sea ice.
A specific objective will be seeking to ensure Aardvark can support livelihoods in developing countries, such as communities in the Global South and the Arctic.
And researchers will keep improving Aardvark’s functionality, incorporating additional data and improving accuracy, particularly when it comes to specialised tasks such as predicting events like hurricanes and floods, as well as providing seasonal and long-range forecasts that could facilitate longer-term planning for governments and businesses.
Looking forward, AI- and machine learning-powered weather prediction is an incredibly exciting field, with the potential to achieve significant benefits to society.
The range of organisations striving for new developments will give the world the best chance of achieving breakthroughs that will save countless lives and protect our economies.
At the Turing, we want to be at the forefront of this, working with our ecosystem partners to play a role in ensuring that the UK remains at the cutting edge of environmental prediction in the years ahead.
Find out more about Aardvark in the project’s new Nature paper.
Top image: Martina / Adobe Stock
