Original title: Missing the Butterfly and Predicting the Past: Features or Bugs of Accurate AI Weather Models?
Pedram Hassanzadeh, Weidong Li, Y. Qiang Sun, Jiangdi Wang +3 · August 2026
Computer-based artificial intelligence weather models match traditional physics-based models, but scientists do not fully understand why they work so well or how well they follow physical laws. To study this, researchers tested AI models across three systems, from simple chaotic setups to complex global weather data. They found that these AI tools can predict backwards in time to guess the past, though past predictions are systematically less accurate than future forecasts. Strangely, the models miss the butterfly effect, where tiny shifts spark huge changes. The researchers discovered that training AI on smoothed-out, coarse-grained data removes tiny, fast changes. This lets AI models learn large-scale patterns without catching the rapid errors that ruin traditional forecasts, though reducing this smoothing makes models act more physical while lowering accuracy.
You are watching a leaf fall from a tree, but you only look at it once every second, missing all the tiny wobbles in between. When you try to guess where it was a second ago, you can do it surprisingly well—even though you lost information. In this paper, the weather data is like that leaf video: it has been simplified by removing small-scale details. That simplification lets the model predict the past accurately, but it also means the model cannot see the tiny disturbances that would normally make long-term prediction impossible.
Weather forecasts from AI models are surprisingly accurate, but this paper shows they achieve that by ignoring fundamental physics like the butterfly effect and the arrow of time. That means we cannot trust them for long-term climate predictions without understanding this trade-off. The finding forces a rethink of how we evaluate AI weather models and what they actually learn.
The abstract does not specify whether the findings apply to all AI weather models beyond Pangu-Weather and the simplified systems tested. The explanation via coarse-graining is supported by experiments but may not capture all factors behind AI models' accuracy. The paper does not propose a solution to the trade-off between physical fidelity and forecast skill, nor does it demonstrate that backcasting is practically useful beyond analysis. Long-term climate emulation implications are suggested but not validated.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Missing the Butterfly and Predicting the Past: Features or Bugs of Accurate AI Weather Models? — Pedram Hassanzadeh, Weidong Li, Y. Qiang Sun 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.