The Power of AI for Climate Modeling

Climate models — computer programs that simulate how the planet will change over time — aim to give us a glimpse of Earth decades from now. Simulating such a complex system is no easy feat, and the code for such programs is typically enough to fill 18,000 pages of printed text. 

At a Sept. 27 event held on Columbia University’s Morningside campus, Pierre Gentine discussed how artificial intelligence (AI) can improve the long-term projection and accuracy of climate models. The talk by Gentine, Maurice Ewing and J. Lamar Worzel Professor of Geophysics in the Department of Earth and Environmental Engineering and professor of earth and environmental sciences, kicked off a new initiative by Columbia Engineering, the Lecture Series in AI. 

The Lecture Series in AI explores the most cutting-edge topics in AI, bringing to campus thinkers and leaders who are shaping tomorrow’s technology landscape in a wide variety of fields. The next lecture — featuring Yann LeCun, Vice President & Chief AI Scientist at Meta — will be held on Oct.18. 

The inaugural lecture in the series, also part of Climate Week at Columbia Engineering, discussed the current roadblocks in climate modeling and opportunities to import some of the pillars of the AI revolution to advance the field. 

“The unfortunate truth is that, when you look at climate models, they tend to be actually very uncertain. What I mean by that is, that when you look at them and the different projections that they give you, whether for regional temperature or precipitation or whatever you are interested in, they give you a wide range of answers.”

Typically, scientists average across many climate models to make projections about the future, which in turn influence policy making decisions, ranging from international treaties like the Paris Agreement to local infrastructure planning and climate adaptation. Some models may be better than others, but they all are given the same weight nonetheless, leading to a large spread in projection data. 

Also, a typical global climate model divides up the Earth into a series of grid cells around 100 kilometers in longitude and latitude. Such a coarse model misses the details and creates a pixelated view of the world, creating uncertainty in climate calculations. 

“We may miss things that we are really interested in, like correctly characterizing extreme events, such extreme rainfall, flooding events, or droughts. We just won’t have that in those models, or we won’t have that to the magnitude that we need,” he said. “That’s problematic because that’s what you need to drive a lot of the risk assessment and physical risk to society. So, that’s something we need to improve.”

Developing a next-generation climate model

Gentine serves as director of Learning the Earth with Artificial Intelligence and Physics (LEAP), a National Science Foundation Science and Technology Center launched at Columbia in 2021. LEAP’s primary research strategy is to improve near-term climate projections by merging physical modeling with machine learning. Traditional climate models use physical modeling that includes thousands of variables in space and time, such as air pressure, temperature, and wind. 

LEAP wants to go a step further by incorporating physics into machine learning algorithms to develop a next-generation climate model. For example, LEAP researchers have trained neural networks to reproduce a high-resolution cloud model at lower resolution and found that plugging it into coarse climate models improves their accuracy, especially with extremes. 

The remainder of Gentine’s talk outlined the pillars of the AI revolution — modern algorithms, massive data, modern computing, and benchmark datasets — and how they can inform further advancements in climate modeling. Progress has certainly been made in the last several years, but from Gentine’s perspective, progress needs to accelerate in order to address the increasing number of extreme weather events and how humanity can successfully adapt to them.