Case study: AI within the Met Office

Weather prediction is fundamental to the smooth running of daily life, from ensuring the safe transport of cargo and allowing farmers to optimise planting and pesticide application to helping people plan days out. The way forecasts are presented is equally as important in making this information accessible.

The Met Office has long been at the forefront of delivering these services in the UK, with the ways it has been developing and presenting weather and climate forecasts constantly evolving. The current AI era is no exception.

Much of the technology the agency has been using for many years could now be classed as AI, although it wasn’t originally acknowledged under that label. However, recent developments are permitting it to build on these foundations, promising a future where, one expert says, weather and climate prediction, and even more so weather and climate products, will become more precise and useful than ever.

“Weather is such a large part of our national conversation for a reason”, Dr Edward Steele, Met Office IT Fellow for Data Science, said.

He added: “Forecast advances will help further support government, industry and citizens make better decisions to stay safe and thrive.”

One pioneering project the agency recently undertook was its exploration of the use of generative AI to create textual forecasts from the raw data output of weather models. It chose to use the Shipping Forecast to test this due to its iconic status as the world's longest running continuous weather forecast, as well as its complexity.

In collaboration with employees from AWS specialist prototyping team, the Met Office experimented with two ways of creating these forecasts.

The first was through the usage of vision language models (VLMs), which merge computer vision and natural language processing. To pilot this, the team encoded the weather model data as a video, having the computer ‘watch’ a day’s worth of hourly forecast information from a multitude of data sources and write a report.

The second approach they tested was a large language model (LLM), which processed data by converting it into basic text descriptions based on hard-coded rules established by software engineers, before then turning those descriptions into a more coherent and accessible report.

After four weeks of prototyping, the system was able to complete this task in under five minutes with between 52% and 62% accuracy. In contrast, this would generally take an expert meteorologist from a few hours to a day to produce.

Both models were developed using the Amazon Nova Foundation Model. Given that the Met Office leverages, among others, the cloud-computing infrastructure of Amazon Web Services, the decision to use Amazon technology was largely part of a strategy to “reuse, before buy, before build”, Dr Steele, who was the lead on the project, said.

He said: “[In doing this,] we are minimising time building something that already exists and maximising time building something new that usefully extends it, exploiting our particular expertise and deep weather and climate domain knowledge.”

This project demonstrates but one of the new potential applications of AI in the Met Office’s work. The team continues to develop technology to streamline tasks ranging from weather and climate prediction/projection, to downscaling, to post-processing.

Dr Steele said that the biggest opportunity that AI offers meteorology is the difference it makes to the user experience of weather products, describing this as a change from thinking about “what the weather is going to be” to “what the weather is going to do”.

He said: “ Ultimately, rather than just “doing what we do already, just in some cases better and faster”, AI also gives us the opportunity for “doing entirely new and different things” - for example, providing bespoke insight in a fully integrated, contextual, decision-by-decision capacity.

“Excitingly, this lays the foundations for being able to move from what I affectionately refer to as ‘giving out homework’ (i.e. 'it is going to rain at 10am, figure out what this means for you') to being able to provide a proactive situational answer (i.e. 'your calendar indicates you will walk to the shops at 10am, you may want to take an umbrella')."

Improvements in the capabilities of machine learning weather prediction models are, however, not negligible. They have increased the accuracy, efficiency and timeliness of weather predictions at a much lower computational cost, with some data-driven models outperforming physics-based models for a subset of parameters for the first time at Christmas 2022.

Nevertheless, the use of AI technology for weather prediction does not come without its challenges. With machine learning models lacking an inherent understanding of meteorology, care must be taken to ensure that the predictions they generate are coherent. Furthermore, they often struggle with predicting the development of extreme weather events, which cannot be sufficiently represented in training datasets.

It is for this reason that it is crucial that physics-based models remain a very important part of the forecasting approach, Dr Steele insisted.

He added: “Underpinning all of this, there is an absolute need to ensure the reliability and trustworthiness of outputs, which is critical to both data and textual forecast products - an area we invest considerable attention.”

Going forward, the Met Office will be using the findings from its textual forecast generation pilot to support the operational production of a number of their regional forecast scripts, narratives and bulletins, which are set to be evaluated in real time by expert operational meteorologists in a live pilot next year.