AI Already Gets the Weather Right. Can It Do the Same for the Climate?

AI models already outperform many conventional systems in weather forecasting, but predicting the climate one or two decades ahead is another matter entirely. Some researchers argue that the key lies in pairing AI with physics.

Reaching for your phone to check the weekend forecast has become almost as routine as scrolling through cat videos on social media. Much of that forecasting now relies on artificial intelligence, which in less than five years has begun to go toe to toe with traditional weather models. Yet if AI can predict a few days ahead with remarkable accuracy, could it also reveal what the planet will look like in 30 years’ time—and help us manage the climate change already under way?

 

That was the premise behind Frontiers in Climate Science, an event hosted by Stanford University’s Doerr School of Sustainability in October 2026. The debate drew on the theoretical and practical expertise of its speakers: Isla Simpson, senior scientist at the US National Center for Atmospheric Research (NSF NCAR), and Tapio Schneider, professor at Caltech, principal scientist at Google and director of the Climate Modeling Alliance (CliMA). The panel was moderated by Marcia McNutt, former president of the US National Academy of Sciences. In short, the answer is yes—though, inevitably, with a few caveats.

 

In this article:

Ask climate researchers and you will find broad consensus that the climate is changing. The difficulty lies in the detail: how much, where and when? As Arun Majumdar, dean of the school, acknowledged in his opening remarks, not all the details of the climate dynamics are fully understood; all the more so when it comes to building the world of tomorrow.

 

Tapio Schneider offers a practical example. When a civil engineer is commissioned to design a drainage system for a city such as Palo Alto, they think in decades, not years. As he points out, the initial design needs to be able to accommodate storms 50 years from now. Yet depending on which climate model we rely on, the Earth will pass 2°C of warming either in the near term or several decades from now. What kind of rainfall should it be designed for, then?

 

Much of that gap in knowledge stems from something we see every day: clouds. If warming leads to more low clouds, they will reflect more solar radiation back into space and slow the rise in temperature. If they decrease, warming will speed up. Schneider explains that this uncertainty has not narrowed substantially since climate assessments began in the 1970s, “mostly because we don’t know how clouds respond to global warming”.

When a civil engineer is commissioned to design a drainage system for a city such as Palo Alto, they think in decades, not years.

There are other significant gaps, too. Aerosols, for example, create a parasol effect that masks part of the warming: in India, temperatures have risen less than in the rest of the world. “It’s highly desirable to clean up the air, but it will also mean it will get warmer,” says Schneider. Another source of uncertainty is how the carbon currently stored in forests and soils will behave. 

 

That said, researchers such as Simpson have spent years validating these simulations on the ground, largely thanks to satellite records. On the whole, the results are positive, though considerable shortcomings remain: the tropical Pacific and the Southern Ocean behave differently from expected, while the North Atlantic jet stream has strengthened far more than any model predicts. Even so, she warns against overdramatising: “I don’t want everyone to think all models are terrible. There are many good things that the models do.”

In weather forecasting, artificial intelligence has marked a qualitative leap forward. “Six years ago, these AI models started far behind the physical models. And by some metrics, not by all, [they are now] better than physical models,” argues Schneider. Official bodies already use them daily, among them the European Centre for Medium-Range Weather Forecasts (ECMWF), whose AI model, by its own estimates, consumes a thousand times less energy per forecast. 

Then there is the factor of trust. Why do we trust the weather forecast on our phone screen? Essentially, because it is put to the test every day. If the weather app does not fail tomorrow and the day after, we end up trusting its predictions. With climate models it is no longer so straightforward, since we cannot wait 50 years to see whether a simulation was right or wrong.

Artificial intelligence learns from past data, while the future has a habit of breaking patterns.

A second hurdle arises because artificial intelligence learns from past data, while the future has a habit of breaking patterns. “Clouds, a few decades from now, will be taller than any clouds you currently have on Earth,” Schneider predicts. Faced with an unprecedented situation, AI tends to play it safe: in a forecasting competition three to four weeks ahead, Schneider notes, one system ended up choosing the average scenario across much of the planet. “What’s happening is the AI model is regressing towards the mean.” As Marcia McNutt says, “science is the best and perhaps the only approach that humankind has ever created for predicting the future. But these AI models had very little physics in them.”

 

That does not mean we should do without AI, but rather define the boundaries within which it is applied. You do not need to understand the thermodynamics of storms to grab an umbrella when leaving the house: a reliable forecast is enough. When it comes to the climate, Schneider qualifies, it is harder because “I don’t think you can separate the understanding from the predicting part”. That is where AI’s black boxes collide with that goal. In other words, we can obtain a prediction without properly understanding how the system arrived at it, making it difficult to assess whether it will be reliable under conditions different from those it learned from

That is precisely the starting point for CliMA, the consortium led by Schneider from Caltech. Its approach combines the physical laws explaining how the climate works with AI’s ability to learn from data. Its original aim was limited to improving physical cloud models, but bit by bit the team ended up building a radically new model. In his view, it is “the first genuinely new climate model in at least 40 years. It’s starting from zero lines of code. Now it’s 600,000.”

 

The physical laws guaranteeing the conservation of mass, momentum and energy serve as guides that allow projections to advance into uncharted states. For example, if the model calculates that snow is melting, it must account for where the heat comes from and what happens to the resulting water. Within those constraints, AI learns from data which microscale processes behave similarly under any climate condition, such as turbulent flows or the way air masses inside a cloud interact with the air surrounding it. AI helps represent the effect of those small movements without having to calculate them one by one in every cloud on the planet.

Its approach combines the physical laws explaining how the climate works with AI’s ability to learn from data.

The balance between physics and data depends largely on the problem. Clouds require robust physical representation; snow, by contrast, is an ideal fit for AI: observing it at different altitudes and latitudes provides a good representative sample of how it will respond to warming. Schneider, a keen skier, explains that a neural network with just a handful of lines of code has achieved “something like a 30% or so error reduction relative to state-of-the-art snow models”.

 

The model offers other practical improvements. It runs entirely on GPUs, meaning it can be executed on an average university’s computing cluster without needing a large supercomputing centre. It is also open source and, according to its creator, reduces costs noticeably. Even the research process can be faster, using AI agents to write equations, test them and introduce improvements; all without sacrificing an understanding of what is happening: “In the end, you want to understand what you have learned here,” he concludes.

Although this type of research seeks a fairly experimental balance between AI and physical models, managing the causes and consequences of climate change is already seeing a direct impact from artificial intelligence. Here are some examples:

  • Renewable energy and power grids. AI forecasts solar and wind generation cycles, optimises energy storage and matches supply and demand in smart grids such as Copenhagen’s.  These and other advances are evident in initiatives aimed at a lower-carbon economy.
  • Water management. Technology makes it possible to optimise water resources during prolonged droughts. For example, the PRISTINE project, coordinated by ACCIONA, determines the concentration of emerging contaminants at wastewater treatment plants in real time.
  • Weather forecasting and early warnings. The ECMWF’s AI model, fully operational since February 2025, increases accuracy when tracking cyclone paths by up to 20%.
  • Wildfire detection and suppression. Data analytics and drones can act as a genuine technological firebreak: orbital sensors spot an outbreak within minutes, and platforms such as FireMap use algorithms to map a fire’s spread in real time, backed up by simulators that calculate which way the flames will travel.

So, back to where we started: can AI help us understand and tackle climate change? At Stanford, the conclusion is yes, provided it relies on physical models and verifiable empirical series. When it comes to adapting to its consequences, the results are already tangible. Either way, AI promises to be a formidable ally in facing one of humanity’s great challenges.

 

Sources:

David is a journalist specializing in innovation. From his early days as a mobile technology analyst to his latest role as Country Manager at Terraview, an AI-driven startup focused on viticulture, he has always been closely linked to innovation and emerging technologies.

He contributes to El Confidencial and cultural outlets such as Frontera D and El Estado Mental, driven by the belief that the human and the technological can—and should—go hand in hand.