Wednesday, November 18, 2026·Focal News

The Boston Focal

Boston's independent voice

Local

MIT's AI Approach Cracks the 'Predictability Desert' in Long-Range Weather Forecasting

Researchers at MIT have developed an AI system that corrects errors in existing weather models, extending useful forecasts up to six weeks ahead — a window meteorologists call the 'predictability desert.' The breakthrough could help utilities, farmers, and emergency managers prepare for extreme weather earlier.

MIT's AI Approach Cracks the 'Predictability Desert' in Long-Range Weather Forecasting

CAMBRIDGE — For decades, meteorologists have hit a wall at roughly two weeks. Beyond that point, forecast accuracy collapses into what scientists call the "predictability desert" — the sub-seasonal window between two and six weeks out where traditional models offer little useful guidance.

Now, a research team at MIT led by climate scientist Judah Cohen says artificial intelligence can help existing weather models push through that barrier. Rather than replacing conventional forecasting, the new approach uses machine learning to identify and correct recurring patterns of error in model predictions, learning from past forecasts and comparing them to what actually happened.

"Here really for the first time, I feel like in this paper we showed by using AI, we can make an improvement in this sub-seasonal two to six week forecast," Cohen said.

The team's method was tested in an international forecasting competition, where it demonstrated twice the improvement over historical weather averages compared with a leading forecasting model, according to Cohen. He cautioned, however, that the gain is measured against a baseline of historical conditions and that overall forecast skill remains modest.

"It sounds amazing, but again, the accuracy is modest," he said.

The system is not designed to predict whether it will snow on a specific day six weeks from now. Instead, it focuses on broader patterns — whether a given week is likely to be warmer, colder, wetter, or drier than normal. That kind of probabilistic guidance can still be valuable: a signal that colder-than-normal temperatures are more likely several weeks out gives utilities, farmers, and emergency managers time to prepare, even without knowing the exact timing or location of a storm.

Consider an approaching ice storm. Power companies often need days to bring in additional crews and equipment from other parts of the country. More reliable advance warning could let them stage resources earlier — and avoid committing them unnecessarily when conditions shift.

The same information could help agricultural operations brace for unusual temperatures or precipitation, and give energy companies a better sense of future demand.

Cohen, who has spent much of his career studying long-range weather patterns, recalled how far short-range forecasting has come. "When I was in grad school, we would say two days, that was it. Three days it was gone, there was no more accuracy or skill," he said. Today, useful forecasts routinely extend seven to 10 days, sometimes longer.

Weather remains inherently chaotic, and no computer model can eliminate that uncertainty. But Cohen believes combining traditional forecasting with AI could help scientists make progress in one of meteorology's most stubborn challenges.

The predictability desert hasn't disappeared. But researchers may finally be finding a way through it — and while meteorologists still won't promise snow on a specific Tuesday six weeks out, they may be getting closer to identifying the patterns that make it more likely.

More from The Boston Focal