NOTE: This story is a Grok summarized article on “Atmospheric Predictability Beyond 30 Days with Machine Learning” by P. Trent Vonich and Gregory J. Hakim. It was released in the Artificial Intelligence for the Earth Systems journal by the American Meteorologist Society, which I am a part of.
Weather Forecasts That Last Way Longer Than We Thought Possible
Scientists have long believed that weather forecasts can only be really accurate for about two weeks. After that, tiny mistakes in the starting weather data grow so fast because of chaos (the “butterfly effect”) that the forecast becomes useless. A new study challenges that old rule using an AI weather model.
What They Did
Researchers took a popular AI weather model called GraphCast. This model is special because it can “learn” from data and also figure out how small changes in the starting weather affect the forecast days later.
They didn’t just run normal forecasts. Instead, they tweaked the starting weather picture (called the initial conditions) for every day in 2020 — over 700 forecasts in total. They used math to find the best possible starting point that would make the forecast match what actually happened as closely as possible. They did this step by step, starting with short forecasts and slowly making them longer.
They tested this on a big computer using two levels of precision (like using more decimal places for more accuracy).
What They Found
The results were surprising:At 10 days ahead, the tweaked forecasts had 86% less error than normal forecasts that started from regular weather data. Every single forecast got better — none failed.
The forecasts stayed useful much longer. The weather patterns matched reality well out to about 27–28 days, and some skill lasted past 30 days. The errors grew at a steady rate, but started from a much better place, so they stayed smaller for longer.
They also checked if the AI was just making smoother, less detailed forecasts (which would look better but be less useful). It wasn’t — the improved forecasts kept the right amount of detail and activity.
What the “Tweaks” Looked Like
The changes they made to the starting weather weren’t random or tiny. On average, they showed a clear pattern:They made the big tropical air loop (warm air rising near the equator and sinking farther north and south) stronger than in the normal weather data.
This included:
Making some ocean areas near the tropics a bit cooler or warmer in the right places Adding or removing moisture in patterns that matched real weather systems Changing how air rises and sinks over large areas
These changes were about the same size as normal small errors in weather data. They weren’t crazy or unrealistic.
Testing in a Different AI Model
To make sure this wasn’t just a trick that worked only in one AI model, they took the improved starting points and used them in another AI model called Pangu-Weather.
The other model also got better forecasts — about 21% less error at its best point (around day 4). It wasn’t as big an improvement as in the first model, but it was still clearly better than normal. This shows the better starting points helped in general, but some of the improvement was specific to the first model.
Why This Matters
This study shows that better starting weather data can push accurate forecasts well past the old two-week limit — at least for big weather patterns across the whole planet. It suggests the atmosphere might be more predictable at large scales than we used to think, and that current weather data has some fixable mistakes in big patterns like the tropical air loop.
Important limits
The scientists knew what the weather actually did later (they were looking backward). This isn’t yet possible in real-time forecasts.
It works best for big, continent-sized weather features, not tiny storms or local details.
It took a lot of computer power for each forecast.
Bottom Line
The old idea that weather becomes unpredictable after about two weeks may not be as strict as we thought — at least when we use smart AI tools to find better starting points. This doesn’t mean we’ll have perfect 30-day forecasts tomorrow, but it opens the door to much longer useful forecasts in the future if scientists can figure out how to find these better starting points in real time.
The study is careful and well-tested. It doesn’t claim the atmosphere has no limits — it just shows those limits might be farther out than we believed when we start with smarter data.
