No one can run a controlled experiment on a spare Earth. So scientists build one in a computer. A climate model is a set of the laws of physics, chemistry and biology, written as equations and solved for the whole planet. Models are not crystal balls, and they are not just curves fitted to past data. They are tools for asking "what if?", and their track record can be checked.
Inside a model
A global climate model divides the atmosphere and ocean into a three-dimensional grid of boxes, often around 100 kilometers across and dozens of layers deep. In each box, the computer calculates temperature, wind, moisture and pressure, using basic principles such as conservation of energy, mass and momentum, and the way gases absorb and emit radiation. It then steps forward in time, minutes at a stride, for centuries.
Some processes, like individual clouds or thunderstorms, are far smaller than a grid box. These are represented with simplified rules called parameterizations, built from observations and detailed studies. This is where much of the remaining uncertainty lives, and different modeling centers make different choices, which is one reason scientists compare many models at once.
Weather forecasts versus climate projections
A common objection runs: "Forecasters can't predict the weather two weeks out, so how can anyone predict climate in 50 years?" The answer lies in what is being predicted. In the 1960s the meteorologist Edward Lorenz showed that weather is chaotic: tiny errors in today's conditions grow until a forecast loses its value after about two weeks. But climate is about averages and odds, not a particular day. You cannot say whether it will rain in Chicago on July 9 next year, but you can say with confidence that July will be warmer than January. Climate projections work like that, tracking how the statistics shift when the energy balance changes.
A short history
In 1967 Syukuro Manabe and Richard Wetherald built one of the first models to include both radiation and the rising of warm air. It estimated about 2°C of warming for doubled CO2 and predicted that the stratosphere would cool, a fingerprint later confirmed by observation. In 2021 Manabe shared the Nobel Prize in Physics, together with Klaus Hasselmann, who developed methods to detect the human signal in noisy climate data. In 1988 NASA scientist James Hansen told the US Senate that warming had begun and presented model projections for the decades ahead.
How models are tested
- Hindcasts: starting a model in the past and seeing if it reproduces the climate that actually followed, including the seasons, ocean currents and the twentieth century's warming.
- Natural experiments: after Mount Pinatubo erupted in 1991, Hansen's team predicted global cooling of around half a degree over the next couple of years, followed by recovery. That is roughly what happened.
- Past climates: models are asked to simulate the last ice age and warm periods millions of years ago, then compared with the proxy records from Lesson 3.
- Predictions made long ago: a 2020 study led by Zeke Hausfather checked 17 model projections published between 1970 and 2007. Once differences between assumed and actual emissions were accounted for, 14 of them matched the warming that followed.
Scenarios: models answer "if," not "will"
A model cannot know how much coal, oil and gas people will burn. So scientists run models under several scenarios of future emissions, from rapid cuts to continued growth. The latest IPCC set uses scenarios called Shared Socioeconomic Pathways. The spread across scenarios is mostly a spread in human choices, not in physics. Through mid-century, the scenarios look broadly similar. By 2100 they diverge sharply, from about 1.4°C of warming with very deep cuts to about 4.4°C with very high emissions.
Honest limitations
Models are better at global averages than at local details such as the rainfall in a particular valley. Some of the newest models ran hotter than other evidence supports, and the IPCC's 2021 report responded by weighting its projections using observed warming rather than taking every model at face value. Scientists publish these limits openly, which is part of why the overall picture can be trusted.
Small models, big lessons
Not every climate model is a giant. In 1969 the Soviet climatologist Mikhail Budyko and the American meteorologist William Sellers independently published simple energy balance models that treated each band of latitude as a single box, balancing incoming sunlight against escaping heat. Both found that the ice-albedo feedback was strong enough that, if sunlight dropped by a modest amount, ice could spread toward the equator and lock the whole planet into a frozen state. Many scientists doubted that such a snowball Earth had ever happened, but geologists later found evidence that ice reached the tropics more than 600 million years ago. Simple models like these cannot forecast the rainfall in a valley, but they make the core logic of the climate system easy to see and check. Scientists still use them alongside the full models to make sure that complex results make physical sense.
Worked example: checking an old projection
Suppose a 1988 model projected about 0.9°C of warming by 2020, and the world warmed about 0.7°C over that period. Was the model wrong?
- Separate the physics from the inputs: the projection assumed certain future emissions. Ask what actually happened.
- Check the inputs: in this kind of case, emissions of methane and ozone-destroying chemicals grew less than assumed, partly thanks to the 1987 Montreal Protocol, so the real forcing was smaller.
- Compare like with like: scale the projection to the forcing that actually occurred, and the warming per unit of forcing comes out close to what was observed.
- Draw the lesson: the physics held up, while the guess about human behavior was off.
Judging a model fairly means asking whether it got the physics right given what people actually did. By that test, climate models have a solid record.