AI for scienceMiddleLesson 2545 min read

Weather and climate models

Machine-learning weather models learn from decades of past weather and can forecast very fast. Physics-based models still matter, and every forecast carries uncertainty.

✦ Ask the tutor 🧪 Lab · 20 min 📝 Lesson test

Speed

Lesson in motion

In 60 seconds

Weather and climate models

Machine-learning weather models learn from decades of past weather and can forecast very fast. Physics-based models still matter, and every forecast carries uncertainty.

1/3

In simple words

An old farmer may look at the sky and say, "It will rain tomorrow." She has seen thousands of days and learned the patterns. A machine-learning weather model does something similar, but it has "seen" many years of weather data from the whole planet.

How forecasts are usually made

For decades, weather forecasts have come from physics-based models, also called numerical weather prediction. They start from the current state of the air, oceans and land, measured by satellites, weather stations, balloons and ships. Then they use the laws of physics to calculate, step by step, how the atmosphere will change. This needs very large supercomputers.

What machine-learning weather models do

A machine-learning weather model learns from many years of past weather data. It learns how the weather at one moment tends to turn into the weather a few hours later. To forecast, it repeats that step many times. GraphCast, from Google DeepMind, is one well-known example. Other groups and weather agencies have built similar models.

Once trained, such models can produce a forecast very fast on much smaller computers than a full physics model. Researchers have shown that, in some settings, they can match or beat traditional forecasts on common measures. This is a real and exciting change.

Why physics models still matter

Starting point. ML models usually start from an analysis of the current weather that is built with physics-based systems and observations.
Training data. The past weather data they learn from is itself often reconstructed with physics models.
Rare and new events. Extreme events are rare in the data. A physics model follows the laws of nature even in situations it has not seen before.
Understanding. Physics models let scientists ask "why" and test ideas about how the atmosphere works.

Many experts see the future as a mix: physics and machine learning working together, each checking the other. Weather agencies test new models carefully before they rely on them.

Uncertainty: every forecast is a range

The atmosphere is chaotic: tiny differences at the start can grow into big differences later. So forecasters often run a model many times with slightly different starting points. This set is called an ensemble. If most runs agree, confidence is higher. If they spread out, the forecast is less certain.

Note

"60% chance of rain" does not mean it will rain for 60% of the day. It means that, on days with forecasts like this, rain is expected in about six out of ten cases. A forecast is a probability, not a promise.

Weather is not climate

	Weather forecast	Climate projection
Question	Will it rain in Pune on Friday?	How might rainfall patterns change over the coming decades?
Time scale	Hours to about two weeks	Decades
Role of ML today	Fast forecasts that learn from past weather	Helpful for speeding up parts of models and analysing data
Big challenge	Chaos and rare extremes	The future climate may not look like the past data a model learned from

Why this matters for India

The monsoon shapes farming, water supply, power and daily life across India. When it starts, how much it rains and where, and breaks in the rains all matter to farmers deciding when to sow. Heavy rain, cyclones and heatwaves put lives at risk. Better and faster forecasts, with honest uncertainty, can help people prepare.

The India Meteorological Department (IMD) issues India's official forecasts and warnings. Researchers in India and elsewhere are exploring how machine learning can improve monsoon and extreme-weather prediction. For decisions that affect safety, always follow official warnings.

Real example

A farmer in Vidarbha hears two apps give different forecasts for next week. That is not a scandal: it is uncertainty. The useful question is how confident each forecast is, and what the official agency advises.

Watch and read more

Lab

A one-week forecast diary comparing forecasts with what really happened where you live.

~20 min

The problem

This lab takes a few minutes a day for a week. You need a phone or a newspaper.

Choose two forecast sources, for example the official IMD forecast and one weather app.
Each morning, write down each source's forecast for tomorrow: rain yes or no (or the chance of rain) and the maximum temperature.
The next day, write down what actually happened (did it rain where you live? roughly how hot was it?).
After seven days, make a table: date, source A, source B, what happened.
Answer: which source was closer more often? Did the sources ever disagree? When they disagreed, what happened?

You are done when

Your diary has at least five days of forecasts and outcomes
You recorded forecasts before the day, not after
You compared both sources fairly with the same rules
You avoided a big conclusion from a small sample of days

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Source A was right on six of seven days. Can you say it is the better forecast?Reveal

Not with confidence. Seven days is a very small sample, and weather in one week may be easy to forecast (for example, a dry spell). Forecast agencies judge models over many months, many places and many kinds of weather, and they pay special attention to difficult days like heavy rain.

Q2Why is it important to write the forecast down before the day happens?Reveal

If you look back after the fact, it is easy to remember forecasts as closer than they were, or to pick the source that turned out right. Recording predictions in advance is a basic rule of honest science. It is the same idea as registering a study plan before collecting data.

Please sign in to continue.

Questions people ask

If ML models are so fast, why not use only them?

They depend on physics-based systems for their starting conditions and much of their training data. They may also struggle with extreme events that are rare in the past data. Agencies test them side by side with physics models, and many see the two as partners.

Can AI predict the monsoon months ahead?

Long-range forecasts, like seasonal monsoon outlooks, are harder than next-week forecasts and have more uncertainty. Researchers are studying how machine learning can help. Treat any seasonal forecast as a probability, not a fixed answer.

Why do different weather apps disagree?

They may use different models, different starting data or different ways of turning model output into a simple icon. Because the atmosphere is chaotic, small differences can grow. Look for the official forecast and pay attention to uncertainty.

Can ML help with climate change research?

Yes, in parts: speeding up slow pieces of climate models, analysing satellite data, or downscaling to local detail. But climate projections ask about a future that may differ from the past, so physics-based understanding stays central.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.