On March 4, 2025, experts from diverse fields gathered at Columbia University to explore major questions in the rapidly evolving field of artificial intelligence at the inaugural Columbia AI Summit. Covering topics from healthcare, business and policy, to the sciences, engineering and the humanities, the summit offered a 360-degree view of AI’s transformative impact on society.
Featuring Climate School researchers, the afternoon session, From Chaos to Code: How AI Can Tame the Climate Crisis, addressed how AI is emerging as a powerful tool in climate science, disaster preparedness and building resilience across interconnected systems. Read on for highlights from the session or watch the video below.
Speakers
Introductory Remarks: David Sandalow, School of International and Public Affairs
» View all Columbia AI Summit videos
The Power of AI
“There are a number of different transformational [AI] gains that are possible. One I’m particularly excited about is in materials innovation … Famously, 150 years ago, when Thomas Edison invented the modern light bulb, he spent months and months and months physically testing different types of elements or materials, running electricity through them to see how much light and heat would be produced. And he finally came up with what he thought was the optimal solution. Today, we can simulate a million of those interactions in a second using modern AI tools. And that ability gives us two benefits. One of them is that we can test materials that don’t actually exist. And then, if it looks like they would be beneficial in a certain application, we can fabricate them and see if they work. And then we can downselect much more quickly. The potential for AI to help us accelerate the pace of energy innovation, I think, is outsized, and we need to figure out how to mobilize all kinds of tools to do that.” —David Sandalow
“What’s happened now is that the AI [weather prediction] models—which essentially didn’t exist something like five years ago—have become as good by many metrics or even arguably better by some than the [traditional] physics-based models. The AI models don’t know more or less about the physics of the atmosphere explicitly. They don’t know about conservation of energy or momentum or any of that. Just like all other AI, they’re trained on data. You give them lots of historical weather data, and by knowing what happened in the past and then seeing what happened later, they find the patterns, in the mysterious way that machine learning and AI do. And so as with other AI, the training is very expensive, but the forecasting, as David said, the actual running of the models in real time is very cheap, and they’ve gone from nowhere to being really, really powerful in a short time.” —Adam Sobel
“I think equality is critical. Because in the past, when a power system has a problem—and this is exactly what happened in 2021 during winter storm Uri—power operators cannot supply enough energy and they need to cut demand. I think during that time, Texas cut off around 30% of the demand—because they lost a lot of generation—to keep the rest of the system performing. But if you think about an energy system—which is basically a branching network structure—then the easiest way to cut demand is to cut off the branches. But it is usually, or historically, more disadvantaged communities that live in the branched areas, which usually have less up-to-date infrastructure … On the other hand, with new deployments of batteries and solar PV, we are seeing affluent neighborhoods that are deploying microgrids—like a Tesla Powerwall that can supply a home. We should begin to rethink how we cut off demand in an emergency and take into consideration that some families or buildings can supply themselves through a microgrid, but we also have disadvantaged communities. In these cases, AI can help a lot with making recommendations to power system operators, making quick decisions and sometimes even simulating different scenarios. Those are areas where I think AI is already showing a lot of promise.” —Bolun Xu
“As a behavioral ecologist, I’m very interested in how individual organisms respond to environmental change. And I think to start making real predictions, we need to understand variation at the individual level to be able to predict what populations and what ecosystems are going to do in response to climate change or land use change. And I think that’s where [species identification] technology can be very useful. We’re moving from individual species recognition to individual identification using images from camera traps or from recording pods … We can use images to capture individual zebras—I think that was the first species because they have almost a barcode, like a fingerprint—and you can make really good population projections and follow those animals and see where they go, so to see what happens during droughts or periods of land use change … I think as we move to projections, we can scale up from actual individual level variation and differences up to populations, up to ecosystems, and make much better projections of how organisms are going to respond to change.” —Dustin Rubenstein
“This is a truly unique opportunity with the help of AI to establish methodologies in order to determine optimal adaptation and resilient strategies for the infrastructure against weather and climate related hazards. The problem is extremely challenging, it is highly multidisciplinary, it will involve contributions from engineering, from the physical science, from social sciences and many other fields. And the only way to approach it is through a novel Monte Carlo approach, including stochastic optimization. AI is going to help us with integrating all these different disciplines contributing to the eventual solution of the problem.” —George Deodatis
The Challenges of AI
“We all know that predicting a hurricane, especially as far as climate is concerned—when is a hurricane going to happen over the next 5, 10 or 20 years, and what is the frequency and the anticipated intensity—is very tough. And so far we have some models, probabilistic models, describing this uncertainty. But these models are assuming a stationary climate, a climate that is not changing. And we know now that the climate is clearly changing, so these probabilistic models about extreme events are going to be changing in time. This makes the problem an order of magnitude more challenging. So we are relying on artificial intelligence approaches in order to establish, in the near future, some models being able to quantify in a probabilistic way this evolution of occurrence and intensity of these extreme events.” —George Deodatis
“If you look at food systems, you see a lot of inequities. Smallholder farmers, indigenous populations, and peoples who very much shepherd and are the custodians of ecosystems and our food systems are often left behind in this race towards better data and decision making. I think we have to be very careful about how we employ these technologies, particularly because we have so many smallholder farmers in the world that are often being left behind. So this data democratization around AI is going to be so important.” —Jessica Fanzo
“With AI moving so fast there’s potential to use things people don’t understand and I think there’s also a tension now between public and private … There is a move now from the White House to privatize the weather service, which has been talked about for years, but there’s now a real threat of it happening … New [AI models] come from the private sector, but it’s still totally dependent on the public sector infrastructure for the underlying data and the physics models that do a lot of the background work. So there’s a great tension here and I think a real danger of breaking things in the infrastructure that we depend on to keep people safe.” —Adam Sobel
“I think about 30% of daily generation capacity in California now comes from giant batteries. And it’s fair to say that many of those batteries are now operated by AI—of course with human monitoring … And in many cases, we cannot explain [what the AI is doing], making the power system operators worried. It’s really about understanding the transparency and how to regulate this AI, and I think that’s a real concern now with many power system operators.” —Bolun Xu
“We know the models rely on the data we train them with, and I think most people here are talking about one species: humans. What happens when we train our models for one species or one family of species, and then do we cut corners and try to use that to project with other species? I think that’s a real concern. How much can we refine our models to be able to work with another ecosystem or other species when the training data are from something different?” —Dustin Rubenstein
The Columbia AI Summit was organized by Columbia AI, a new initiative that aims to promote Columbia’s work on artificial intelligence with courses, curricula, events, digital tools and more. Columbia AI is a university-wide effort led through a partnership between the Data Science Institute, Columbia Engineering, and the Executive Vice President for Research.
*Highlights have been edited for clarity