Can AI Predict Hurricanes and Extreme Weather?

Yes, in some specific cases. An AI hurricane model is already giving forecasters about a day more warning on where a storm will go and how strong it will get, and a machine learning method has improved forecasts three to four weeks ahead. Neither replaces human forecasters or physics-based models, and the far future is still a matter of better odds, not certainty.

Below are the cases we have reported and checked on this beat. They are not the whole field, but each names who it could help: coastal families, farmers, aid groups, conservationists and the fishers who depend on living reefs.

Hurricanes and the weeks ahead

For people on a low stretch of coast, a forecast is what sets the clock on boarding up or leaving. A Google DeepMind system called WeatherNext Cyclones, described in Nature with co-authors from the US National Hurricane Center, Colorado State University and the UK Met Office, reaches the accuracy of leading models for track, strength and size about a day or more earlier. Its average five-day track error was 230 km, against 370 km for the ECMWF ensemble. Forecasters at the Hurricane Center began folding its experimental guidance into their work in 2025. It does not yet forecast rainfall or storm surge, and most of the authors are Google employees. Read how an AI forecast could give coastal families about a day more hurricane warning.

A farmer deciding when to plant needs a sense of the weather three or four weeks out, a stretch forecasters have long called a “predictability desert”. A lightweight machine learning method from Microsoft Research, Harvard, MIT, ECMWF and others learns how an existing weather model has gone wrong before and corrects its odds. It was the best entry in ECMWF’s 2025 AI Weather Quest, and boosted flood forecast skill by 76 to 92 percent across catalogued flood events. Skill is still modest, several authors are Microsoft employees or shareholders, and no agency is reported to run it yet. See how farmers could get better odds three to four weeks ahead.

Drought, hunger and early help

In Kenya’s dry northern counties, a bad season can turn into hungry children faster than aid arrives. A team from UC Berkeley, Cornell, the International Food Policy Research Institute, the US Census Bureau and Kenya’s National Drought Management Authority built machine learning forecasts of child malnutrition from years of monthly arm measurements, plus satellite rainfall and vegetation data, conflict data and food prices. In 2026 a dashboard built on it was integrated into the agency’s drought early warning processes. According to Cornell, roughly half of one-month alerts matched real crises, and the cash transfers it now helps trigger are not yet evaluated. Two authors work for the agency, a declared competing interest. Read how forecasts could help reach hungry Kenyan children sooner.

Protecting wildlife and reefs

Off South Sulawesi, homemade bombs stun fish and shatter the reefs that feed coastal families. Researchers from universities in the UK and Indonesia, working with the Zoological Society of London and Mars Sustainable Solutions (a Mars program that has run coral restoration there), used underwater microphones and a machine learning detector called BombDetect to find 3,567 verified blasts over a 16-month pilot. Blasts dropped on Fridays, the local prayer day. The yearly totals are extrapolated, and the study does not yet show that patrols guided by the data reduce bombing. Read how underwater microphones could help guard reefs.

Elephant counts in Botswana often come from costly survey flights every four years or so. A team including the University of Edinburgh and Elephants Without Borders ran what the authors describe as the country’s first synchronous satellite and aerial wildlife count, using an AI elephant detector on 30 cm satellite imagery. Satellites were fast and needed no permits, but elephants resting under trees made the estimates uncertain. It is a single case study, and the authors say satellites should work alongside planes for now. See why counting elephants from space still needs planes.

Storing clean power

When the sun sets and the wind drops, stored power has to carry the load. MIT researchers led by Ju Li report in Joule a new electrolyte solvent, DMFSA, that helps sodium-metal batteries with both stability and speed. According to MIT News, an AI-guided search designed 100,000 candidate molecules, narrowed them to 200, and 27 were tested in the lab. This is lab-scale cell testing, not a commercial battery, and long-term safety, cost and manufacturing at scale are untested. Read about the step toward sodium batteries for renewable power.

How close is this to everyday life?

• Already in use: the hurricane model, as experimental guidance for human forecasters, and Kenya’s malnutrition forecasts inside its drought early warning system.
• Getting close: the three to four week forecasts, which have open code but no reported agency running them, and the reef listening system, a pilot whose next test is whether patrols guided by it reduce bombing.
• Years away: counting elephants from space without planes, and sodium batteries built around the new solvent, which still need larger, practical cells.

The next real step in several of these stories is people deciding to use the tools: a forecasting center running the open code, a patrol boat sent where the blasts cluster, a herding family in Marsabit getting help before the worst of a dry spell. We will follow those steps as they happen on the climate page.