In February 2021, three brutal winter storms slammed Texas, killing at least 246 people and knocking out power, water, and heat across the state. Energy officials had little time to prepare. If they’d known what was coming weeks earlier, they might have brought in extra fuel, protected equipment, and saved lives. But until now, forecasts beyond two weeks were little better than guesses.
That’s changing. A new AI-powered model developed by researchers at MIT, Microsoft, and other institutions has doubled the accuracy of weather predictions between two and six weeks out. For the first time, forecasts in this range are consistently better than just using long-term climate averages. This window—called the "predictability desert"—has long been a blind spot for meteorologists. It’s too far out for short-term models to work, and too soon for long-term climate patterns to matter. But now, with this breakthrough, farmers, energy providers, and emergency planners can finally see into it.
The team created a machine-learning "wrapper" that corrects errors in existing forecasting models. Instead of building a new model from scratch, they trained an AI to learn from past mistakes in a gold-standard system used by the European Center for Medium-Range Weather Forecasts. The AI spots patterns in the errors and adjusts the forecasts accordingly. When tested, the improved model outperformed all others in an international competition and became the first to beat climatological averages in the two-to-six-week range.
This isn’t just about knowing if it’ll rain next month. The model also improves predictions of extreme events—like cold snaps, floods, and heatwaves—that hit hardest. Judah Cohen, a research scientist at MIT and co-author of the study, says, "Extreme events might be the most impactful part of this for society." Farmers can plan planting around droughts. Grid operators can stockpile energy before a deep freeze. Cities can pre-deploy emergency supplies.
The model was developed by a global team including Hannah Guan, a Harvard undergraduate; Soukayna Mouatadid from the University of Toronto; researchers from Microsoft; and experts from Brazil, Canada, and Europe. Their work, published in the journal Nature, marks a turning point in weather science. For decades, the two-week forecasting limit—set by chaos theory—was seen as a hard wall. But this AI tool shows that with the right approach, we can push beyond it.
Now, the focus shifts to real-world use. The team hopes to integrate their system into national forecasting agencies and make it available to communities worldwide. As climate change brings more unpredictable weather, the need for longer, more accurate forecasts has never been greater. With this tool, we’re not just seeing further into the future—we’re learning how to prepare for it.