The outlook is sunny for weather apps... thanks to AI

AI-based weather forecasting systems aren't only cheaper than supercomputers, they could also bring more accurate predictions to the world. Nicole Kobie reports on the winds of change

- 

Look at your weather app and it says rain soon - and then you end up carrying an umbrella in the sunshine. Or, more likely in this country, the app predicts sun and you end up drenched. What if weather forecasts were more accurate?

Now imagine you live in a country that doesn't have access to its own weather models and relies on systems designed for other parts of the world. Local storms can't be predicted and seasonal forecasts required for agriculture aren't accurate enough.

There's a solution and, predictably, it's based on AI. Less predictably, it's called Aardvark - an AI-based system that ditches supercomputers and allows local models to be run on a desktop by a single researcher.

The prediction problem

At the moment, weather forecasts are generated in a multipart process using supercomputers. Each section can take several hours to run, and requires large teams of experts to use, develop, maintain and deploy. That means weather prediction is expensive and time-consuming, effectively cutting it off for countries without the means to own and operate supercomputers.

Ghanaian meteorologist Michael Baidu is currently a research fellow at the University of Leeds. "We do not run our own models, mainly because of computational costs and how much power is required to run these models," he said, speaking at the Turing Institute's AI UK conference.

The challenge is exacerbated by unstable power supplies; if it cuts out part way through a model running, the whole process must begin again. "So we've mainly relied on the global north for the output of the models that we use for forecasts," said Baidu.