New research from the Met Office and the Alan Turing Institute suggests that guiding AI weather models to better preserve physical consistency during training could address key limitations in AI-generated forecasts.
The research centers on FastNet, a machine learning weather prediction model co-developed by the two organizations and named after one of the 31 sea areas covered by the Shipping Forecast, a nod to Met Office founder Vice-Admiral Robert FitzRoy. FastNet is already demonstrating accuracy comparable to the Met Office’s Global Model, even exceeding it on some performance metrics.
Many AI weather systems achieve strong headline accuracy but struggle to reproduce the sharp fronts, gradients and storm structures needed for reliable medium-range forecasts. Models trained only to minimize average error commonly blur features such as cold fronts or tightly wound storm centers, a phenomenon known as ‘blurring’, which can hide important signals and mask errors that grow with lead time.
FastNet addresses this using a modified spherical harmonic loss function, guiding the model to preserve the correct distribution of energy across atmospheric scales so small-scale that features remain crisp rather than blurred. This approach means FastNet now achieves performance comparable to the Global Model on metrics such as root mean squared error (RMSE) – a significant milestone given the Global Model’s decades of scientific refinement and its extensive validation for safety-critical use.
FastNet is not yet operational but tests on Hurricane Ian (2022) and Storm Ciarán (2023) showed more realistic storm core structure, improved pressure-wind relationships, higher and more accurate peak wind speeds and clearer depiction of intense gradients at longer lead times.
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 the model embeds physics within machine learning to create a system that is scientifically rigorous and computationally efficient while capturing sharp, detailed weather fronts.
Met Office chief AI officer Prof. Kirstine Dale said the research marks a significant step forward for AI weather prediction, with the Met Office is taking an approach “centered on science and trust, only making operational changes when the weight of scientific evidence is clear.”
FastNet’s development will continue, with future research aimed at higher-resolution forecasts for the UK and globally.
In related news, NSF NCAR researchers develop advanced model for neighborhood-scale low-altitude wind prediction