Can AI Fight Climate Change Without Making It Worse? | Auroop Ganguly, Northeastern University
Auroop Ganguly, Director of AI for Climate and Sustainability at the Institute for Experiential AI and Professor of Civil and Environmental Engineering at Northeastern University, talks about how AI can help predict extreme weather and make infrastructure more resilient. He also discusses the energy demands of AI, the shortage of climate data in some regions, and why he thinks AI needs stronger governance.
Transcript
Welcome to Agents of Tech, the podcast where we explore cutting edge innovations shaping our world. I'm Autria Godfrey, coming to you from Washington, D.C.
Unfortunately, AI developments are often being powered primarily or even solely by considerations such as profitability and computation, neglecting to consider the sum of gains or losses to all of humanity. There are indications that green computing is a possibility. In fact a distinct possibility, if only we were to put all our concerted efforts in that.
So AI has certainly demonstrated significant promise to become a game changer for weather prediction across scales.
From catastrophic floods to raging wildfires, extreme weather events are pushing our infrastructure, economies and healthcare systems to their limits. There's also been growing concern about the massive energy demands of AI, how training and running these large scale models contribute to a rising carbon footprint. As AI adoption surges, so does its power consumption, raising critical questions about sustainability. But AI isn't just an energy intensive technology, it can also be a powerful tool for climate action.
From optimizing energy grids and refining climate models to detecting emissions and advancing conservation efforts, AI is already playing a key role in both mitigating and adapting to climate change. So the big question we're asking today, how do we balance AI's impact with its potential for good? We'll be speaking with Auroop Ganguly, the Director for AI for Climate and Sustainability at the Institute for Experiential AI, and Professor in Civil and Environmental Engineering at Northeastern University.
AI is neither a panacea, it's not a medicine for everything, nor is it a magic bullet. It is useful. It is powerful, but a lot of hard work still needs to go into AI systems to make them seriously actionable.
If you're enjoying today's podcast, make sure that you like, subscribe and share as this helps us better curate content for you. Now I am joined by Laila Rizvi, Hi Laila. Great to have you today.
Hi, Autria. Great to see you again.
All right. Well, Laila, unfortunately we are having this discussion at a time when it seems like natural disasters are just absolutely dominating the headlines from, you know, floods on our East Coast to raging wildfires on our West Coast. It seems like everywhere we're dealing with some sort of a natural disaster right now. So the fact that we're discussing how AI can maybe help mitigate some of these issues and help with climate change is a really topical one.
You have that hope in light of all of these kind of natural disasters, that maybe AI is that silver bullet which solves it all. And I think what would be really beneficial from today's episode is to get a realistic handle on what can it actually do, where can it actually help us, because there's a lot of hype about AI, and it will be able to do things that we've never seen before that we couldn't even imagine right now. So speaking to someone like Dr. Ganguly and getting a handle from his perspective on what is actually achievable and where is this, this technology being used because this stuff is complicated. If it was easy to solve, it would have been solved by now.
That's a great point. We'd already have a handle on it. Something that I think that's interesting is, you know, Dr. Ganguly's research really delves into understanding weather and hydrologic extremes in the context of climate change. So, you know, he employs AI and machine learning techniques to analyze these large data sets, enhancing the prediction and understanding of extreme weather events and their impacts on critical infrastructure. So, you know, by kind of using this predictive AI, we may be able to not completely, you know, avoid these large extreme weather events, but at least be able to have citizens be prepared and, you know, maybe avoid, you know, large scale disasters because of them.
Yeah. And I think that's kind of where we we want to look at it in the short term at least, because it's a chewable chunk. Completely reversing climate change and solving all of the Earth's climate change related issues going back to, you know, 100 years ago from the industrial revolution is a really difficult feat, whereas trying to get ahead of what's happening and ensure that our communities, our infrastructure, are resilient for what's inevitable, I think is the best step or small step to take, as of now. So I think it would be really interesting to see what Dr. Ganguly has to say, especially given his experience across kind of various different organizations and sectors from commercial government and academia.
Right, and you know, something we have discussed a lot here on Agents of Tech is the, you know, data storage concerns when it comes to AI and machine learning and the need for, you know, more and more space for all the storage of all this data. And in addition to that, the power that's needed to continue to, you know, sustain all of these machines. It's interesting to juxtapose that against the idea that, well, while that is an issue and we are still searching for solutions on that, we can actually utilize this type of technology to help with some of these climate induced events. So it's kind of using AI to help with some of the problems that perhaps AI might be causing. So it's a very interesting, you know, juxtaposition.
For sure, and I think it just speaks to the fundamental importance of just interdisciplinary collaboration, because if you have people working at each of those disciplines, they are fixing each of the problems at each stage. So the energy problem might be solved by material scientists, and then the climate change problem might be solved by climate and Earth scientists. So if we have this interdisciplinary kind of hub, then you've not got as much of a juxtaposition because you're all working towards kind of progressing society and solving problems for the sake of humanity.
Right, right. Okay. Well, let's get to it. Here now is Professor Auroop Ganguly. Thank you so much for your time today, Professor.
Pleasure to be here.
All right. So I want to kind of start big because your work spans AI, climate resilience and sustainability. So how did these fields all intersect and what unique challenges or maybe unique opportunities arise at their crossroads?
So the successive hurricanes in the Carolinas and the recent wildfires in the Los Angeles regions are stark examples of why we need to bake in resilience into power grids, urban communities, and forest management. When the patterns of weather extremes intensify, as has been happening over the last few years, last several years sustaining economies, health and security becomes that much more difficult. This is especially true if the impacted infrastructures and social systems are already vulnerable. AI can help us develop predictive insights and suggest pathways to resilience, especially when the power of AI is combined with process understanding and the ability to analyze extreme events and interconnected systems.
So, Dr. Ganguly, I want to go on to more how AI has impacted your specific research? I mean, it's being hailed as a game changer when it comes to, climate change research. Can you share some specific examples where AI has made a tangible difference in one of your specific projects?
When you talk about climate change, you know, we talk about weather extremes intensifying and therefore our ability to deal with such weather extremes. One aspect of that is early warning systems. So AI has certainly demonstrated significant promise to become a game changer for weather prediction across scales.
So in one of our NASA funded work, where our stakeholder is the Tennessee Valley Authority or the TVA, we implemented what we called hybrid physics AI methods, physics and AI combined to improve precipitation nowcasting, short term rainfall forecasting, including for extremes. So this has implications for river management and dam operations and ultimately for informing flood emergency managers and power plant operators. The AI methods and our efforts to make them trustworthy by working with stakeholders and subject matter experts, were viewed favorably by the TVA and will be tested on their operational system.
This work was published in a Nature Partner journal and highlighted in reports and book chapters generated within a workshop of the United Nations Office of Disaster Reduction and the Conference of Parties. Given the loss of human lives and destruction of property over Appalachia owing to floods in recent years, the impacts, I think are supposed to be significant.
You mentioned the floods and, you know, North Carolina, obviously, the LA wildfires, you know, just a couple of examples of extreme weather events recently. Is AI applicable in any other regard other than the short term rainfall when it comes to helping predict and respond to these types of weather events?
Absolutely. So AI can help advance what we could call a longer term predictive understanding of statistics of extreme weather events, or how extreme events are changing things like 100 year events and understanding of these with AI and, different kind of data science methods, they have implications for the design of water resources systems. And then, of course, AI can help in short term prediction of such systems, which have implications for operational management.
The other aspect is AI can also help in understanding the consequences of such extreme events, such as for fuel networks, transportation networks, all of which can then lead to better ways to manage cascading failures caused by weather extremes in interconnected systems and also to recover from these damages when they do occur.
So, Dr. Ganguly, I want to move on from climate weather prediction onto net zero emissions, what would you say are the most promising innovations in helping us achieve globally net zero emissions in AI?
So I must confess, I am more optimistic about the ability of AI, together with other tools and with process understanding, to help us adapt to changing patterns of weather extremes and water stresses, than I am about the ability of AI to reduce emissions. While AI has shown value in sectors such as renewable energy, electric vehicles and energy storage technologies all of which can help reduce emissions, the race to build bigger data centers and larger compute powers can really be a dampener.
So, somewhat counterintuitively, the recent buzz around DeepSeek has led to suggestions that this may spawn greener AI. However, the devil appears to be in the details and owing to various factors, there is no guarantee yet the DeepSeek type approaches will lead to energy saving AI, but overall definitely AI can do something in these areas, but other factors seem to be dominating, which may make AI sort of not emissions or energy saving technology as of now. Hopefully it will change.
Yeah, I guess people would rather create problems and use AI to solve it rather than, you know, strip it back to start off with. That's kind of what's happening.
Very interesting. You know, sustainability often relies on large scale modeling and data analysis. What role do you think AI plays in making these models more accurate, efficient or actionable?
So AI can analyze data from large model simulations, as well as observations to generate novel predictive understanding. AI can also go inside these models, and if there are some approximations which we call parameterizations, for example, then AI can in many cases help in making them better. The ability to exploit hidden patterns in data that may not be otherwise visible can make AI systems more accurate.
There's still a lot of work to be done, though, to make AI systems efficient, and perhaps even more work to make the systems trustworthy and hence actionable by stakeholders and decision makers. So the example that I gave with our work with TVA is one step by us in this, direction. There are others, of course, who have been working in this direction. So as AI researchers, there are AI researchers who talk about demystifying AI. They have indicated that AI is neither a panacea, it's not a medicine for everything, nor is it a magic bullet. It is useful. It is powerful, but a lot of hard work still needs to go into AI systems to make them seriously actionable.
And now I want to come on to kind of data of which we've mentioned, quite frequently. And AI requires a lot of data and a lot of the models, particularly the large language models, have relied on large scale, data and training on lots and lots of data. What are some of the challenges when it comes to collecting data, particularly for climate models and, especially in regions where there's perhaps not as robust infrastructure for collecting data, does that impact models? Does it result in us not being able to produce models for those areas which lack that infrastructure? What's your take on that?
It's a good point. So there are no easy answers to this question again. But there are approaches such as transfer learning and methods in what are often called foundational models that can occasionally be leveraged to learn insights or even methods from data rich areas in ways that generalizes to data poor areas. So in one of our works, this was joint with a group at NASA and others, we showed how insights about air pollution can be learned from cities in developed nations and transferred to cities where data and resources are scarce.
However, as they say in statistics, and this applies to AI and machine learning as well, there is no free lunch theorem. So these generalizations may be subject to caveats and assumptions and may not always even be feasible. So even if we have some data in the less resource regions, they can help in fine tuning. You talked about large language models. Sometimes you also talk about small language models which learn from these large language models and then are fine tuned for one purpose, the same kind of thing. They can learn from all the data that's available everywhere, including, places where we have more data. But then they can be fine tuned for one region.
There is an element of looking under the lamppost here, right? I mean, I look under the lamppost because there is light there, as opposed to where I actually may have lost my keys. But that being said, if there is enough data, sometimes AI methods have shown the ability to learn and then generalize to other areas. And the other thing is, simulation and AI studies can jointly point to data sampling strategies such as, for example, how limited resources can be spent to generate or leverage data in ways that add most value for the AI method or for the end user.
How adaptable are these AI models when it comes to, taking into consideration the different, I guess, landscapes or topography of different regions when it comes to making these predictions?
If the AI has seen lots of data, if the AI has seen situations which are similar to these other situations where there is less data. So you have seen in places where there is more data, conditions, situations that are similar, they may be able to generalize, but there is also where fine tuning comes in. So this kind of thing about, you know, this has this kind of topography, so can you primarily pick from those cases where the topography is like that. This also, is where physics and AI can work together. If these things are important, if this kind of rainfall is so important, this kind of, you know, flooding is so important, can we embed that into the system in addition to the AI based learning?
Yeah, I'm actually glad you mentioned the physics because I was just thinking, of course, there's the option to, you know, train a large generalized model, then fine tune. But if we're aiming to create a model which effectively simulates the physics of the world and particular, atmospheric interactions, then there's actually no need to get really specific, because if it's understanding just the physics of kind of weather and climate systems, then it can hopefully predict anything because it's just trying to understand the world.
Right. Well physics, of course, is not perfect either. And then again, some of these physics models were built originally when we did not have so much data or so much compute power. So the idea here is that now that we have data and compute power, this this ability of AI to do things, we don't want to miss that opportunity. This can, you know, lead to many different opportunities. At the same time. The physics that we have built over the years, some of them may still be useful.
So yes, exactly how to combine them in the most effective way is something that some of us have been thinking, and they may be more useful in some areas than in others. Definitely in earth sciences, climate, weather, and consequences to infrastructures and societies, they may be more useful than others. And when I say physics, by the way, just to qualify, this is process understanding. So this can be understanding of natural systems and human engineered systems, social systems, you know all of that I'm sort of lumping in this word physics in this case.
Professor, I want to shift just a tiny bit to talk a little bit more about the energy demands of AI technologies, because this is, you know, really a hot topic when you're talking about the, the taxing that, you know, is put on the power grid when you're talking about, you know, powering AI. So my question to you is, how do you balance those energy demands of all these AI technologies with your overall goal of sustainability?
I wish there was a more concerted desire and effort to balance these. And if there were, there are ways to start doing that. And this is an area where more research and development is definitely needed. And really where there may be a need to improve AI governance and industry best practices. Unfortunately, AI developments are often being powered primarily or even solely by considerations such as profitability and computation, neglecting to consider the sum of gains or losses to all of humanity.
And I'm not saying that as any ism, right? It's not like I'm ranting against capitalism or anything. It's just, you know, we have to also think about overall, you know, what it means for us as humans. That being said, I'm an eternal optimist. So certainly AI has shown promise in adapting, for example, to weather and water extremes and change adapting meaning, coming up with ways for us to adapt. And in terms of mitigating multi-sector risks.
There are indications that green computing is a possibility. In fact a distinct possibility, if only we were to put all our concerted efforts in that. But that would require a governance framework that all parties, whether by parties, we mean companies or nations, they can trust and they'll follow.
In order to comply, I guess, with sustainability standards or, you know, become one of these green companies, do you think that it's going to have to be the result of mandates and not companies, you know, self prioritizing those goals?
It's just the nature of the way these things work, companies are fairly competitive entities. So they'll primarily think of their bottom line. They'll primarily think of how I can outperform my competitors. So yes, just like, you know, monopolies have to be sometimes broken down. Similarly, there has to be at some level, you know, governance structures, whether it's at national levels or whether it's at international level, you know, like various treaties that exist made for similar kinds of things. Is it doable? Hopefully.
Dr. Ganguly, I actually want to flip the script on governance a little bit now. Now, AI obviously has the potential to actually influence policy on sustainability itself. So what are your thoughts on AI driven insights when it comes to, sustainable policies and governance? Do you think AI might be the thing that actually drives us to stop putting so much energy into AI?
On the one side, right, if you think of AI driven insights, they are already helping and they can certainly help, show a lot of promise to help, policies related to resource conservation such as food, energy, water, policies related to managing hazards from design of infrastructures, preparedness before the hazard strikes, recovery and relief during the hazard, and assuring infrastructure, services and even insights have started to help in understanding urbanization, land use change, migration policies related to those as well as their broader implications.
So in these areas, there is already some ways we already see some promise and some real implementations of yes, this is happening. Now, where AI can also add values through our understanding of mitigation and emissions, which we already mentioned, while also ensuring sustainable computing. There is a need to develop these areas further. It's not. It's both the R&D and like we were discussing earlier, the ability to whether it's a mandate, whether it's a soft mandate, hard mandate, but there has to be something that says, well, yes, we understand your corporate responsibilities or national responsibilities to outperform the competition, but there is also something broader than that. We do not want to do it at the expense of humanity at large.
The focus right now is on winning the AI race rather than sustaining.
Well, one useful analogy there to remember is that there have been various cases in general, right? If you think of, for example, you know, the consensus that was built around, let's say, not smoking or in environmental areas, consensus that was built around reducing certain kinds of chemicals which lead to the ozone hole. So there are cases where public consciousness has led to changes in policy on certain kinds of mandates. So there's hope.
Yeah, but I guess in those examples, historically, we often cause the damage first before we go back on it.
Right. Kind of on that on that vein, when it comes to ethical concerns or biases that may arise when deploying AI in some of these critical disaster management areas, how do you address that? How do you, you know, plan for that and try to minimize that?
So this is a vast topic. But I'll keep my response relatively short here. The implementation of AI systems in critical areas that you mentioned can easily go off track. So just to give an example, if AI algorithms primarily maximize net monetary gain, then it may often lull the allocation system to consistently award resources to those who may not need it much at the cost of others who do. Similarly, an AI algorithm that learns from unrepresentative data samples, too many samples of one type, may attempt to channel preparedness efforts, relief and rescue efforts, or recovery efforts before, during, or in the aftermath of disasters, to areas or to communities where these efforts may not be as useful as they would be elsewhere.
These discussions are issues which almost I have seen never fails to raise the proverbial temperatures in societal discourses and debates, especially when extreme polarization dominates and trust deficits are high. So you have to be careful here how we communicate the core issues. Perhaps the key is common sense rules going out and explaining why these are common sense and governance, including how AI results are tuned, understood and deployed.
Can I circle back and ask you one final question? You briefly mentioned DeepSeek and of course, you know, like DeepSeek has kind of stormed on to the AI stage. I get the sense that you perhaps don't feel like it's the next like AI golden child.
In fields like AI, they will change very rapidly. Golden childs have come and gone and become brass or copper very soon. So it's early. It's very early. But what I like about DeepSeek in general, just that concept is it shows that, you know, there are opportunities for the smaller people, the entrepreneurial minded, you know, not the big monopolies to do something, especially when there are constraints. If you think this was an innovation motivated by necessity and that always gives hope that, you know, some of the inventions, some of the innovations can be a result of frugality.
So these kind of things here in this case, it came from one part of the world, but this could happen in anywhere. So that's the hope. Now is DeepSeek itself something that will either completely change the way we do things, in general for AI or in general for how we manage energy and AI? Well, there are a lot of details there that have to be addressed. But is this something that's worth exploring that maybe we don't always have to go for these large data centers? Maybe there are ways to do it in more frugal, innovative ways while achieving the same result. Maybe even better.
The brain, DNA. All of these amazing things came from selectionary pressures. Evolution didn't have many resources to work with, and I agree that I remember I saw Sam Altman say maybe a year ago that don't even bother trying competing with OpenAI on the foundation model level. Right. And I think that's what DeepSeek represents, actually. Yeah. If you're innovative enough, you can compete because that's how you make major leaps.
Well, yes. And, there has been this hubris that we have this biggest everything, and there is no way you guys can do anything that has been the, you know, the major source of innovation in all of humanity. You mentioned evolution. That made me think of something. So personally, I always think that as a force of change, evolution is better than revolution unless there is no other way out. That being said, if we are not careful with the way these AI systems are going and what can they do for humanity? I fear that at some, at some point, it may come to a stage when the evolution of these technologies and the way we use them may no longer be sufficient. I just hope that it will not be to the detriment of humanity.
And to your point, what you're trying to use AI for is to try to mitigate these, you know, extreme climate events, so. Thank you so much, Dr. Ganguly, for giving us your time. It was a pleasure to interview.
My pleasure. And thank you so much for having me.
Absolutely. Autria, what did you think of that? Super interesting right?
Very interesting. I think that, you know, his work could not be more timely with the whole extreme climate events. When you look at the things that have been happening here in the United States, it's like there's either a flood event or there's a hurricane or you know, or like snow events, like I think that, you know, utilizing AI to try to mitigate those events and, you know, preserve human life. I feel like he's really, you know, on the pulse of trying to correct what is happening right now.
Yeah. And what I really liked about, that interview was how Dr. Ganguly kind of brought it out to the macro level in terms of, yeah, the specific work he was doing and how he's leveraging AI in his work in collaboration with NASA, but also just the implications for humanity. And balancing that and also, you know, is trying to achieve the best possible AI model at everyone's expense, the ultimate goal? I think we all acknowledge, yeah, it makes people feel anxious that someone else might be able to get ahead in the AI race. And it's a powerful technology and it can be used in diverse ways. But should the rest of humanity be sacrificed for that, for that one goal?
I think that's a little bit concerning that I gather that his opinion was, yes, once we get to a point where we can see the damage caused, then maybe we might implement some measures to try to reduce that. Despite that, I do think that human beings throughout history generate knowledge and develop new technology to deal with those problems that may not have previously existed. So, you know, not being able to feed everyone. We just came up with new technology to grow more food and improve agriculture. That's not the natural state of the world. And I think we're going to do that with climate change, where I agree with him that we're going to become more resilient to it. And I think we'll probably do that with whatever problems come about with AI, we'll develop a new technology which is beyond AI. But it's hard to conceive that idea at this current point in time.
Yeah, yeah. Really fascinating. I think, you know, his work is incredible. And I think that, you know, it will be interesting to see how the use of AI to try to treat and combat and prevent climate change, to kind of see how that adapts and becomes implemented. I think it's really, really fascinating. All right. A really fascinating conversation today that does it for today's episode. If you enjoyed our discussion with Auroop Ganguly, don't forget to like, subscribe and share as this really helps us to improve the show for you. That does it for today's episode. Thanks so much for listening today. Until next time, goodbye, goodbye.
Agents of Tech is brought to you by WebsEdge. Special thanks to our studio engineers Adam Deen and Sam Saris, editor Abbie Harries and producer Cath Sheehan.