Original title: Improving precipitation forecasts in an AI weather model using observational data
Julian F. Schmitt, Bertrand Delorme, Robert C. King, Yashica Patodia +3 · September 2026
Computer models powered by artificial intelligence can predict medium-range weather better than traditional physics systems. However, most global AI weather tools rely on a single historical weather dataset called ERA5. This dataset has known flaws, especially for rainfall. To fix this, researchers took an AI system built with a graph-transformer design and fine-tuned it using actual satellite precipitation data called IMERG. Training the system on real observation data helped it learn more realistic rainfall patterns. As a result, the model boosted medium-range probabilistic accuracy scores by up to 19%. It predicted drizzle and tropical storms with greater skill, and beat operational systems on extreme rain by 57%. Still, traditional physics models remained more reliable for the heaviest downpours.
Imagine a student preparing for a geography exam by studying from a textbook that contains outdated maps. The errors in the textbook lead to mistakes on the exam, especially on questions about rainfall patterns. Here, weather models are like the student, and they use a flawed dataset for learning. By incorporating direct satellite measurements of rain, the models correct those flaws and become more accurate, just as the student would with updated materials.
This helps meteorologists give more accurate rain predictions, which matters for flood warnings, farming, and daily planning. While it's a big step for AI models, the heaviest downpours still rely on traditional methods.
The model still does not outperform traditional physics-based models for the heaviest precipitation events. Also, the improvements are based on specific scores and may vary by region or condition.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Improving precipitation forecasts in an AI weather model using observational data — Julian F. Schmitt, Bertrand Delorme, Robert C. King 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.