The Met Office has long been at the forefront of meteorological innovation, delivering trusted weather services to the public and supporting maritime safety.
In a rapidly evolving technological environment, the Met Office has partnered with Amazon Web Services (AWS) to explore how artificial intelligence (AI) can be used to extract information from complex weather model data and make it available in natural language.
A recently completed, non-operational Discovery Pathfinder project investigated the potential of generative AI – particularly Large Language Models (LLMs) and Vision Language Models (VLMs) – to extract meaningful information from raw gridded weather predictions and present them in the iconic Shipping Forecast format.
Why explore AI for the Shipping Forecast?
The Met Office currently delivers nearly 300 text-based products and services. Drafting these products is a highly skilled task, requiring expert interpretation of complex atmospheric and oceanic data. The non-operational Discovery Pathfinder project set out to examine whether AI tools could assist with the initial drafting of some of these products, freeing up meteorologists to focus their expertise where it matters most. The Shipping Forecast was chosen as a test case due to its strict format, concise length, and the challenge of combining multiple data sources.
A collaborative approach to innovation
Collaboration was central to the ethos and success of the project. The Met Office on its own cannot realize the value of breakthrough technology advances to the weather and climate endeavour, nor can it keep abreast of all the developments and opportunities associated with this fast-evolving environment: this can only be done by working with partners.
READ MORE: From data to decisions: AI factories for a climate ready world
Our prototyping engagement with Amazon Web Services (AWS) therefore not only helped accelerate innovation but also built valuable experience in developing generative AI architectures and capabilities for meteorological applications – leveraging the specific domain expertise of the two organisations. An impressive part of this work was going from concept to prototype and subsequent publication in just a few months; showing how working with partners can allow us to innovate at pace and alignment with our strategy.
The project and approach also align closely with the ambitions of the UK Government’s AI Opportunities Action Plan, which recognises the potential for adopting trustworthy AI in improving productivity and transforming citizen services at scale.
Technological breakthroughs: vision language models and meteorological data-to-text conversion
One of the most exciting outcomes from the six-week prototyping engagement was the development of a pioneering workflow for meteorological data-to-text conversion. The team explored the fine-tuning of Vision Language Models (VLMs) – which combine computer vision and natural language processing – to extract meaningful information and generate textual forecast bulletins from raw gridded weather data.
To overcome the challenge of conveying a day’s worth of hourly forecast information to a LLM, the approach that was developed involved encoding the data as video, enabling direct vision processing using the Amazon Nova Foundation Model.
“Essentially, using a combination of existing atmosphere and ocean model outputs – as well as an archive of manually generated and issued textual sea area bulletins – we taught the Foundation Model to watch videos of these data and write the forecast from that” said project lead and Met Office IT Fellow for Data Science, Dr Edward Steele.
This marked the first example of custom fine-tuning of the Amazon Nova Foundation Model for vision capabilities.
For comparison, an alternative LLM-based workflow (also using the Amazon Nova Foundation Model) that processed the data through an intermediate text representation before generating forecasts, was also developed.
The project evaluated both LLM and VLM approaches across four core weather attributes included in the sea area bulletins. The LLM approach scored 62% for matching the exact words used by human meteorologists, while the VLM approach scored 52%. Although the LLM did better in this test, improving the VLM approach could make it the better choice in the future, depending on application, as it could help solve problems that otherwise limit more basic automation methods, in a highly scalable way.
READ MORE: Cloud collaboration powering real-time weather intelligence
Ongoing research and next steps
This Discovery Pathfinder project is just the beginning. Initial results are hugely exciting and were presented at AWS Re:Invent, Amazon’s annual flagship technology conference, and in a scientific paper authored by the team from the Met Office and AWS. Further exploration is underway, with operational meteorologists working alongside AI specialists to refine our approach.
Professor Kirstine Dale, the Met Office’s Chief AI Officer said: “This project marks an important step forward in how we harness the power of AI to maximise the value of our data and intelligence for the public. Critically, this Discovery Pathfinder project isn’t about transforming the Shipping Forecast, it’s something much bigger than that. We are exploring ways to deploy AI solutions in scalable solutions so that we can draw on massive volumes of raw data to deliver efficient, effective and scalable products and services for our customers.”
Will AI replace marine forecasters?
A question often raised is whether AI will replace human forecasters, but this is not the case. As a technology discovery pathfinder, there are currently no plans to use AI for the operational generation of the Shipping Forecast. AI may help streamline parts of the workflow, but its implementation is designed to complement, not replace, the work of operational meteorologists. Human expertise will continue to be at the heart of our forecasting operations
The efficiencies gained from any operational adoption of AI should enable human meteorologists to spend more time to focus on areas where their expertise could have the most impact. The expertise and judgement of human forecasters remain central to the Met Office’s commitment to accuracy and public safety.
Although the role of operational meteorologists will likely change in the coming years. Computers didn’t replace forecasters in the 1960s, although it did shift their focus – and a similar change is likely as AI and LLMs are embedded into new processes. Operational meteorologists are constantly evolving their practices to deliver value where it counts, and adapting to new data sources, models or methods.
Keep up to date with weather warnings, and you can find the latest forecast on our website, on YouTube, by following us on X and Facebook, as well as on our mobile app which is available for iPhone from the App store and for Android from the Google Play store.
