Met Office and Turing build AI weather model FastNet

TL;DR:

  • The Met Office and the Alan Turing Institute have developed FastNet, a machine learning weather model.
  • It matches — and on some metrics exceeds — the accuracy of the Met Office’s Global Model.
  • Physics principles embedded in training help it preserve sharp storm structures that other AI models blur.

Two of the UK’s flagship research institutions have produced a machine learning weather model that rivals conventional forecasting. FastNet, developed by the Met Office and the Alan Turing Institute and named after one of the Shipping Forecast’s sea areas, has achieved accuracy comparable to the Met Office’s physics-based Global Model, exceeding it on some measures.

Physics inside the machine learning

FastNet tackles a persistent weakness of AI weather prediction: the tendency to blur sharp fronts, gradients and storm structures. By embedding physical principles to guide the model during training, it keeps those features intact. Tested against Hurricane Ian (2022) and Storm Ciarán (2023), it produced more realistic storm cores, better pressure-wind relationships and higher, more accurate peak wind speeds.

The framing from the team is notably cautious. Met Office chief AI officer Professor Kirstine Dale called it “a significant step forward” but stressed an approach “centred on science and trust”, with the future likely a blend of AI and physics-based methods rather than a wholesale replacement. Dr Tom Dunstan described it as “a major step toward operational grade AI forecasting”, while the Turing’s Dr Scott Hosking highlighted a system that is “scientifically rigorous, computationally efficient, and capable of capturing sharp, detailed weather fronts”.

That trust-first tone echoes a wider UK theme Resultsense has tracked, where trust, not technology, is repeatedly cited as the barrier to scaling AI. It matters more in public forecasting than in most domains: a blurred storm core is not an abstract error but a missed warning. The Met Office also acknowledges the tension in the technology itself — balancing AI’s scientific promise against its rising energy demands and carbon footprint.

Looking forward

FastNet is presented as a research advance rather than an operational replacement, but it positions UK institutions at the front of physics-informed AI forecasting — a field where accuracy alone is not enough and public trust in the output is the real prize.

Share this article