The Met Office Makes Leaps in Last Mile Forecast Generation with Amazon Nova

The Met Office has a reputation for preserving its rich heritage while pushing the boundaries of scientific, technological, and operational expertise. It has a goal of radically enhancing weather and climate decision making by transforming its data infrastructure and establishing a foundation for innovative, personalized weather experiences at scale. To meet future demand, the organization saw an opportunity to experiment with AI to augment and automate textual weather forecast generation.

Issued by the Met Office on behalf of the Maritime & Coastguard Agency, The Shipping Forecast is an iconic 150-year-old service that’s both operationally critical for maritime safety and culturally significant in the UK. Like many textual products and services, its current forecast script is manually generated. For some of the key marine warning’s services alone, expert meteorologists spend approximately 2,920 hours a year evaluating vast datasets spanning different weather variables, from the winds to waves, visibility, and other weather types across 31 different sea areas. They must then accurately condense data into a prescribed text format for broadcast. While changes to the way that the Shipping Forecast is presently delivered were not in scope of the project, as an example of one of its many services, the Met Office recognized that by tackling common text generation challenges, it could apply learnings to other products in the future.  

With hands-on support from AWS, the organization aims to open the door to what’s possible with AI. As Dr Edward Steele, IT Fellow for Data Science at the Met Office, explains: “This discovery pathfinder fundamentally cuts to the core of how we can deliver even more efficient, effective, and scalable services.” He adds, “The time we can save meteorologists in writing the forecast can be invested in enabling our highly qualified staff to focus their input where it can have the most impact, which is hugely exciting.”

In future, the Met Office sees further opportunity to boost relevance and personalize weather services at scale. “Rather just providing information about what the weather will do (‘Is it going to rain?’), we anticipate unlocking additional benefit by providing information about what the weather will do (‘Do I need to take an umbrella?’)” explains Steele.

As a government agency, the Met Office is trusted to maintain a high bar of accuracy, meaning text generation must meet strict quality and consistency standards, and the experiments conducted therefore represent an early (albeit significant) milestone in exploring how AI could be used across its numerous products and services. 

To innovate rapidly while maintaining integrity, the organization collaborated closely with the AWS Prototyping team. “It took just four weeks from spinning up the environment to the final set of initial experiments. It is hugely impressive what was achieved in that time by leveraging our combined expertise in meteorological science and scalable technology.” says Steele. After testing foundation models, Amazon Nova emerged as the frontrunner for performance while allowing forecast information to be provided in multimodal formats. 

The technical experiments conducted for the Shipping Forecast prototype considered a range of candidate approaches—from leveraging a vision language model (VLM) for meteorological data-to-text conversion, encoding gridded weather data in 24-frame video for direct vision processing, to using a large language model (LLM) benchmark, that processed data through an intermediate text representation stage prior to generating forecasts. As an early adopter, Steele notes that, “The timing of this could not be more perfect with the release of Amazon Nova, particularly with the vision language opportunity.” 

For its LLM, the Met Office used Nova Pro in Amazon Bedrock with maritime-specific prompt engineering to turn gridded data into sophisticated text descriptions and forecasts. For its VLM, it fine-tuned Nova Lite using Amazon SageMaker distributed training. By doing so, it could process hourly weather evolution through visual pattern recognition. Rather than manually producing statements, the resulting prototype was a single simplified workflow that can easily be customized to other use cases.

As the first example of custom fine-tuning Amazon Nova for vision capabilities, the development process uncovered powerful learnings. It was able to use different numerical weather prediction model types and explore different evaluation approaches. “The engagement with both the AWS Specialist Prototyping the Amazon Nova teams was hugely positive because of the shared organizational values of scientific integrity and operational excellence. We were mutually transparent in discussing emergent product feature requests to help us and future users,” says Steele.

Against the context of growing demand for multimodal products and services, the project proves that AI can help address scalability challenges, improve consistency, and free highly skilled meteorologists to focus on complex analysis and quality assurance. Having benchmarked Amazon Nova, Steele shares that the models are both cost-efficient and “demonstrate competitive performance, particularly given the exactness of the word-based evaluation used.” 

Strict word-level comparisons—explicitly counting the exact number of matched words (true positives), missed words (false negatives), extra words (false positives)—against equivalent human-written and issued forecasts over a three month period demonstrated 62 percent LLM accuracy for complete forecast generation and 52 percent VLM accuracy in data to text conversion across a number of different weather components. Meanwhile, it saw 83 percent gale warning accuracy compared with forecasters’ observations. While already impressive, the team anticipate that this performance may be further enhanced by leveraging some of the recent advances announced at AWS re:Invent 2025, such as upgrading to Amazon Nova 2 models, and exploiting some of the additional training opportunities offered by Amazon Nova Forge. 

 

Following the success of the prototype, the Met Office is equipped with the knowledge to enhance results, including expanding its training data set and refining prompts for machine interpretation. Because almost 300 of its current products and services involve transforming raw grid data to text, the Shipping Forecast discovery pathfinder provides a valuable foundation for possible efficiencies and automation across its portfolio. As Steele notes, “The infrastructure and scalable pattern for developing standard formatted products and services is hugely valuable, and Amazon Nova has certainly demonstrated its value in that.” 

Beyond exploring how AI can be scaled, the organization is seeking to expand its capabilities to meet the growing appetite for innovation. As Steele says, “The AWS environment demonstrates the ability to easily change components for new use cases.” It sees the collaboration as a pathway to personalized, context-aware weather services with LLMs like Amazon Nova enabling people to interact with meteorological data using natural language. Steele adds, “Our organizational strategy focuses on customer-driven agility, innovation, and purposeful data and intelligence. This project absolutely sits at the core of that.” 

Further refinement and testing is required before any potential AI-based improvements to the Shipping Forecast can be made operational, but the project demonstrates remarkable potential that could mean improvements for the Shipping Forecast and beyond.