AI in weather & climate
How does AI forecast tomorrow’s weather?
For 70 years, forecasts came from giant physics simulations running on supercomputers. Now AI models trained on decades of past weather can make a 10-day forecast in about a minute, and forecasters are learning when to trust them.
What AI is doing right now
Four ways AI is changing forecasts
Weather is one of the fields where AI has moved fastest from research into real, everyday use.
10-day forecasts in about a minute
Instead of solving physics equations step by step, AI learns how weather usually changes by studying about 40 years of past global weather.
Real exampleGoogle DeepMind’s GraphCast beat the leading physics-based forecast on 90% of 1,380 test measures and runs a 10-day forecast in under a minute on one machine (Science, 2023). Google DeepMind
AI forecasts go official
Major weather centers now run AI models alongside their traditional ones every day.
Real exampleEurope’s weather center (ECMWF) made its AI forecasting system, AIFS, operational in February 2025, running side by side with its physics model. ECMWF
Tracking hurricanes
Predicting where a hurricane will go helps officials decide who evacuates. AI models are now one of the tools forecasters compare.
Real exampleDuring the 2025 season, the National Hurricane Center tried out a Google DeepMind hurricane model and AI forecasts from NOAA and ECMWF as extra guidance, while keeping “trusted experts in the loop.” National Weather Service, 2026
Spotting tornado warning signs
AI can read many radar measurements at once to find the spinning winds that come before a tornado.
Real exampleIn a 2026 study using weather-radar data from China, an AI flagged tornado signs up to 54 minutes ahead in its case studies, and raised fewer false alarms than the methods it was compared with. The study had data from only 16 tornadoes, so 54 minutes is a best case, not a typical warning time. Read our explainer
Try it yourself
Why is AI so fast?
Step 1: click Forecast +6 hours (or Play 10 days) to watch an AI model roll a storm forward, one 6-hour step at a time. Step 2: start the race and see how long each method takes to make a full 10-day forecast.
Each square is one grid cell, with west on the left and east on the right. The storm is a simplified animation, not real GraphCast output.
GraphCast predicts the weather 6 hours ahead, then feeds its own prediction back in to predict the next 6 hours. 40 steps = 10 days.
Race to a 10-day forecast
The AI isn’t doing the physics. It learned, from about 40 years of past weather, what usually happens next. That shortcut is fast, but unusual, record-breaking weather can trip it up.
The pattern behind every AI tool
Data → pattern → prediction
Every “AI in…” page follows the same five steps. Learn them once and you can understand AI in any field.
Decades of global weather
Records of temperature, wind, humidity and pressure for every grid square, every few hours, going back about 40 years.
How weather moves
Fronts travel, storms spin up, warm air rises. The same patterns repeat in slightly different forms.
Predict the next 6 hours
The model predicts the next step, then uses its own answer to predict the step after that, out to 10 days.
Compare with reality
Forecasts are scored every day against what actually happened, and against physics models.
Warnings & decisions
Forecasters use it for storm warnings; farmers, airlines and power companies use it to plan.
Think like a scientist
What AI can’t do here
Knowing the limits is part of understanding the tool.
Record-breaking weather is hard
AI learns from the past. An event bigger than anything in its training data can be underestimated.
It still needs physics to start
AI forecasts begin from today’s weather, which comes from measurements combined using traditional methods.
Small, fast storms remain tough
A single thunderstorm can be smaller than a grid square and form within minutes.
It can’t explain itself
A physics model shows why it predicts rain. An AI model mostly can’t, which makes it harder to trust in surprises.
Where people stay in charge: meteorologists compare many models, weigh their known weaknesses, and decide what warnings to issue. The official forecast is still made by people.
Careers
Jobs where people work with this AI
Weather and climate jobs increasingly mix science with data and code.
Meteorologist
Forecasts the weather and issues warnings for storms, floods and heat.
Working with AICompares AI and physics models to judge which to trust for a given storm.
Climate data scientist
Studies long-term changes in temperature, rainfall and extremes.
Working with AIUses AI to analyze huge climate datasets and fill gaps in old records.
Emergency manager
Plans evacuations, shelters and response for disasters.
Working with AIUses AI-assisted forecasts to decide earlier when to act.
Renewable energy forecaster
Predicts how much wind and solar power will be available.
Working with AIUses AI weather forecasts to keep the power grid balanced.
Classroom activity
Beat the forecast
Students record forecasts from two weather apps for a week, compare them with what really happened, and calculate which was more accurate.
- Monday (10 min): As a class, choose one place to forecast (your school’s town) and find its forecast page on weather.gov. Each group picks two weather apps or websites and makes a table with a row for each day from Tuesday to Saturday and four columns: day, App A high, App B high, actual high. Write down each app’s forecast high for tomorrow.
- Tuesday to Friday (5 min a day): At about the same time each day, write down both apps’ forecast highs for tomorrow. From Wednesday on, also fill in yesterday’s actual high: open “3 Day History” on your town’s weather.gov page and find the highest temperature listed for that date. (Today’s high hasn’t happened yet.)
- Next Monday (20 min): Fill in Friday’s and Saturday’s actual highs. (“3 Day History” covers the last 72 hours; for anything older, use the “Past Weather” link.) You now have 5 complete days. For each day and each app, work out actual − forecast: forecast 72°, actual 75° → +3°; forecast 70°, actual 68° → −2°. Ignore the minus signs and average the sizes: for those two days, (3 + 2) ÷ 2 = 2.5°. That is the app’s average error, and the smaller one wins. Bonus: average the errors with their signs. Above zero means the app’s forecasts ran too cold; below zero, too warm.
- Share (10 min): Which app had the smaller average error? With only 5 days, is a 1-degree difference convincing? Did either app miss badly on a day with unusual weather? Check each app’s “About” or data-source page: does it say it uses AI?
- Optional rain challenge (+10 min): Each day, also write down each app’s chance of rain for tomorrow and, when you fill in an actual high, whether at least 0.01 inch of rain fell that day (see the precipitation columns in “3 Day History”). On the next Monday, score each app for each day: (chance of rain as a decimal − what happened)², where what happened is 1 if it rained and 0 if it didn’t. 70% and it rained: (0.7 − 1)² = 0.09. 70% and it stayed dry: 0.7² = 0.49. Average each app’s scores (this is called the Brier score): 0 is perfect, and an app that always said 50% would score 0.25. Apps measure chance of rain in different ways, and one dry week can’t tell you much, so treat this as a rough test.
Discussion questions
- Why might a forecast be better for temperature than for rain?
- If an AI forecast beats the physics models on most measures, when would you still want a human meteorologist?
- Who is responsible if an AI forecast misses a dangerous storm?
Fresh from the lab
Latest research in weather & climate
New studies are added every week and explained in plain language. Not every study uses AI, so ask yourself where AI could help. See all research →
New stories in this area are on the way. Browse all research →
Put curiosity to work
Read it. Try it. Question it.
Explore new AI research, or take a paper-first investigation into your classroom.