New research shows AI can produce trustworthy, physically realistic weather forecasts
Guiding AI weather models to better preserve physical consistency during training could be the key to addressing limitations of future AI forecasts, according to new research from the Met Office and the Alan Turing Institute.
AI weather models, which learn directly from data to determine forecasts, have accelerated in skill in recent years, proving their utility for the next generation of weather forecasts.
The Met Office and the Alan Turing Institute have co-developed a machine learning (ML) weather prediction model. FastNet is named after one of the 31 sea areas covered by the Shipping Forecast, and is a nod to the Met Office’s founder, Vice-Admiral Robert FitzRoy (the first professional weather forecaster, who laid the foundations for the Shipping Forecast).
The FastNet model is already demonstrating accuracy that is comparable to the Met Office’s Global Model, even exceeding it on performance for some metrics.
While development of FastNet continues, new research suggests that a different approach to training the AI model has proven beneficial and addresses some common issues with AI forecasting techniques.
While many AI systems achieve strong headline accuracy, they often struggle to reproduce the sharp fronts, gradients and storm structures essential for reliable medium range forecasts. AI models trained only to minimise average error commonly blur features such as cold fronts or tightly wound storm centres. This is important because forecast quality is not only about accuracy; forecasts must also remain physically realistic and maintain consistent relationships between weather variables.
The smoothing happens because the model is effectively rewarded for producing a safe, averaged prediction rather than committing to a precise, physically consistent structure that is reflective of meteorology - a phenomenon referred to as ‘blurring’.
Over time, these blurred edges reduce forecast usefulness, hiding important signals and masking errors that can grow with lead time. Existing methods to resolve this issue involve running the model multiple times and generating a set of possible forecasts, although this still does not guarantee a physically realistic structure for each individual forecast.
FastNet tackles this existing issue in a new way by incorporating physical principles to guide the model during training. The approach uses a modified spherical harmonic loss function, which means the model is guided to preserve the correct distribution of energy across different atmospheric scales, ensuring small scale features remain crisp rather than blurred. This makes the model reflective of the reality of the science of meteorology.
The research forms part of a broader effort to incorporate expert meteorological knowledge into machine-learning weather prediction models, helping improve the physical consistency of forecasts while retaining competitive forecast performance.
Why matching the Met Office Global Model is a major milestone
With this innovative approach to training FastNet now achieves performance comparable to the Met Office’s Global Model for metrics such as Root Mean Squared Error (RMSE). This is a major milestone since the Global Model is one of the world’s most advanced operational numerical weather prediction systems representing decades of scientific refinement, runs on one of the world’s most capable operational supercomputers, and is extensively validated for safety critical uses.
Although not yet operational, FastNet achieving similar skill on a common verification measures demonstrates the potential of AI forecasting approaches that emphasise physical consistency and supports further research towards future operational integration.
Better performance in extreme weather
FastNet also improves the modelling of extreme weather, where sharp gradients matter most. Tests on Hurricane Ian (2022) and Storm Ciarán (2023) showed:
- more realistic storm core structure
- improved pressure-wind relationships
- higher, more accurate peak wind speeds
- clearer depiction of intense gradients at longer lead times
These are areas where many AI models struggle because extremes are rare in training data. FastNet’s approach helps stabilise behaviour and retain structure even in high impact‑ events. The study found that these improvements helped produce forecasts that were more physically consistent while retaining strong predictive performance.
The image shows existing forecasting capability (ERA5), and compares it to the original version of FastNet (O96 baesline) and an updated version of FastNet, guided by physical consistency principles.
Dr Tom Dunstan, Met Office Manager, Data Science for Simulation, Dynamics Research, said: “FastNet demonstrates how AI systems can be guided during training to improve the representation of high-impact weather without compromising accuracy, a major step toward operational grade AI forecasting that shows the value of building physical understanding directly into machine learning, while helping build trust in future AI forecasting systems."
Dr Scott Hosking, Mission Director for Environmental Forecasting at the Alan Turing Institute, said: “FastNet represents an exciting advancement in AI weather modelling, embedding physics within machine learning to create a system that is scientifically rigorous, computationally efficient, and capable of capturing sharp, detailed weather fronts. By co-developing FastNet alongside expert meteorologists at the Met Office, we are prioritising the delivery of physically realistic and trustworthy outputs that are critical for supporting UK business decisions, protecting the public, and strengthening our national resilience.”
The future of forecasting
AI will play a significant part in the next generation of weather forecasts, with the evolving technology likely to accelerate improvements for accuracy.
Met Office Chief AI Officer Professor Kirstine Dale said: “This paper demonstrates a significant step forward for AI weather prediction. Where many previous efforts have been held back by ‘blurred’ features and unrealistic forecasts for extreme events, this shows a new approach which ensures the value and use of the model outputs when applying to real world situations.
“While the future of forecasting likely involves a blend of AI and physics-based forecasting methods, we’re taking an approach centred on science and trust, only making operational changes when the weight of scientific evidence is clear and performance is clearly demonstrated across a range of use cases and scenarios.”
FastNet will continue to develop, with a view to creating higher resolution forecasts for the UK and globally with future research continuing to explore approaches that improve physical consistency and help build trust in AI-generated weather forecasts.