Original title: Online Reinforcement Learning in the Met Office Unified Model through Distributed Model-Agent Coupling
Pritthijit Nath, Sebastian Schemm, Peter Haynes, Emily Shuckburgh +1 · September 2026
Computer weather models make small errors that add up over time. Researchers wanted to fix these errors using reinforcement learningA type of AI where an agent learns by interacting with an environment to maximize rewards., where an artificial intelligence agent learns by trying actions and getting feedback. They connected reinforcement learning agents directly into the Met Office Unified ModelA complex computer program used for weather forecasting that simulates atmospheric conditions., a global weather simulator. Each agent watched an atmospheric column across 70 vertical height levels. As the model ran, the agent applied small, bounded temperature nudges to correct mistakes, learning from real operational weather data over ten training runs. Then, they tested the trained agent in a normal forecast without extra help. The combined system stayed numerically stable. At six hours in, it cut pressure errors and reduced height errors in four of six latitude zones, performing best across the tropics.
Imagine you're baking a cake and you taste the batter as you go. If it's too sweet, you add a bit of flour to balance it, but you have to be careful not to ruin the texture. Here, the weather model is like the cake recipe, and the AI agents are like the baker tasting and adjusting the ingredients in real time to fix errors without breaking the chemistry of the batter.
Better weather forecasts help farmers, pilots, and communities prepare for storms. This method could make predictions more reliable without breaking the complex simulations used by forecasters.
This is a single-case experiment, so results need to be confirmed in more scenarios. Only evaluated at +6 hours forecast, so longer-term performance is unknown. The method might not generalize to all weather conditions or regions.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Online Reinforcement Learning in the Met Office Unified Model through Distributed Model-Agent Coupling — Pritthijit Nath, Sebastian Schemm, Peter Haynes et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.