Heatwaves, flash floods, wildfires, and hurricanes — extreme weather events are among the most devastating consequences of climate change today. Part of their destructiveness is due to their unpredictability. Without the ability to predict and prepare for extreme weather events, communities and governments across the world will be left vulnerable to the worst of their impacts. The urgency and scale of this problem might seem daunting, but for Aditya Grover, this only emphasizes the need for solutions.
“There will always be sustainability challenges, anticipated or otherwise — we saw that with the pandemic, and we are seeing that with climate change,” says Grover. “What AI and technology offer is the optimism to mitigate some of the effects of these challenges.”
Aditya Grover is a 2023 AI2050 Early Career Fellow and an assistant professor of computer science at UCLA. He was awarded a 2023 Kavli Fellowship by the US National Academy of Sciences, a 2022 AI Researcher of the Year Award by Samsung, and is one of 2024’s Forbes 30 Under 30 in Science. His work has been recognized with a best paper award at the Conference for Neural Information Processing Systems and a doctoral dissertation award from the Association for Computing Machinery’s Special Interest Group on Knowledge Discovery in Data in 2021.
Grover develops AI systems that can interact and reason with limited supervision in the real world and applies them to data-driven scientific discovery, with a particular focus on climate science and technology. The goal of his AI2050 project is to develop Atmos, a new AI model for the atmospheric sciences that can accurately forecast extreme weather events in advance. This research addresses Hard Problem #4, which aims to leverage AI to solve humanity’s greatest challenges.
The AI2050 initiative gratefully acknowledges Fayth Tan for assistance in producing this community perspective.
Your work uses AI to predict extreme weather events. What drew you to that specific application?
As a citizen of a global society, we see that extreme weather events are everywhere. Last year was the hottest year ever on record, [and] this year is breaking that record. No part of the world has been spared by an extreme weather event, so there is an urgency for all of us to think about how we can help address this challenge.
As a computer scientist and in particular, as an AI researcher, I am interested in finding patterns within the data and translating this technology to make predictions. With that perspective, weather is a very interesting domain, because it’s a domain that’s data rich. It’s become data rich especially due to remote sensing. There’s a ton of data that’s being collected every day and being made available to the global scientific community to understand, and in my case, use as substrate to train AI models to forecast future weather.
Those were the two motivations for me — the urgency of the challenge, and that this is a perfect fit for AI because of the immense amount of data that’s available for us to make innovations in this space.
Why do you think that AI is so well-poised to overcome the limitations of more traditional methods?
Traditional methods are based on physics. Essentially, these methods simulate systems of differential equations, which govern how matter and energy interact in the atmosphere. For a planetary scale system, these interactions are very, very complex and require fine-grained understanding of physics and massive computation.
AI, on the other hand, relies less on domain knowledge. It relies more on the availability of data and compute. With data, you can often learn things by observation, rather than having to write physical equations. I think that’s primarily why AI is poised to outperform traditional methods. Moreover, AI systems are much faster to execute. You can democratize AI technology much more easily than traditional methods which are confined to national labs and supercomputer systems. That has a multiplicative effect, because now you have a community of researchers, in various parts of the globe, who can play with these AI models, and easily iterate over them to improve the forecasts.
Are there specific challenges in working with earth science data?
One of the challenges of working with weather data is that it’s very heterogeneous. There are a lot of different variables that you care about and need to be modeled jointly. These variables can also exhibit different patterns over space and time. Different parts of the world have different geographies, different capabilities to record sensory information, and will have different responses to the changing climate.
With all this heterogeneity in the available data, it’s very hard to assimilate these data sources. That’s part of the research that goes into building AI systems that can work with such heterogeneous data and can extract patterns that work at a global scale.
As a computer scientist working in environmental sciences, are there challenges to doing this interdisciplinary work? Or are there perhaps more opportunities?
It starts with challenges that have potential to be converted into opportunities. From a personal standpoint, a big barrier going to a new domain outside traditional computer science and AI disciplines is learning to ask the right questions. That means both a data-driven framing of a problem statement, and also assessing its importance. Data is in every domain you can think of, whether it’s the social sciences, natural sciences, medicine, law. But what questions are the most pressing? How should we formulate these questions? How should we measure success? All of these are challenges which a computer scientist cannot do alone, but by working together with domain experts it can be immensely impactful— that’s been my experience.
How have you developed collaborations with domain experts?
We have two active collaborations. The first one is with researchers at the Argonne National Lab, who have a solid background in modeling regional and global weather. They are also compute-savvy with experience in high performance computing infrastructure, and working with them has significantly enriched [my] understanding of the challenges of working with atmospheric datasets at massive computational scales.
Another collaboration, which is closer to home at UCLA, is with a climate scientist, Dr. Karen McKinnon. She’s helping us think beyond weather — about how it transcends into climate as we go into longer timescales. At such timescales, the atmosphere gets very chaotic. It’s important to include that, because knowing whether you’re going to have an extreme weather event, a week from now or a month from now can make a world of a difference to people who are most at risk. Everyday citizens, and governments, can plan better, they can secure their surroundings better. Eventually, if you keep extrapolating to even longer time horizons, you start asking a deep and fundamental question whether these extreme weather events are an indicator (or not) of a changing climate in the long term.
Extreme weather events take place at different timescales — the conditions that create a drought might be different from the ones that create a forest fire or a flash flood. Could AI models work with all these different timescales?
There are some physical limits set by nature, and even an AI system is going to have to operate within those limits. For example, roughly around after 2 weeks, we cannot on average give point predictions better than historical averaging because there’s so much chaos. At such timescales, we ideally want to give a range of predictions. That range reflects our uncertainty around events in the atmosphere which are not in our control.
That being said, 3-4 years ago, the major focus of the field was to get AI systems to make accurate predictions up to a week from now. Slowly, the timescale of interest increased to 2 weeks. Now, it’s acknowledged by a lot of established weather agencies around the world, including the United States, that AI systems tend to outperform conventional prediction-based systems for up to 1-2 weeks in advance for major weather variables.
The next frontier is improving the generalizability of these models. We want weeks going into seasonal timescales, which affect a lot of industries in more fundamental ways. For example, in agriculture, when planting crops, predictions on that [timescale] are quite valuable. That’s where AI can offer benefits, even though the regime is more challenging.
A lot of environmental scientists studying climate change talk about the emotional impact of their work. Do you, as a computer scientist, feel the same way or share some of those experiences?
Due to the nature of my work, I live in two very different worlds. One is when you see these headlines about extreme weather events, and you think about the future of humanity. On the very next page, you will see headlines about how AI is going to transform our quality of life, and we will enter into this very productive and efficient era where almost everything from cooking and driving to coding and art will be automated. Sitting at the intersection of these two different worlds, my attitude has always been one of cautious optimism. As long as we can futureproof our communities to mitigate these existential challenges, I think that will be a big step ahead for AI and humanity.
A lot of young researchers have expressed wanting to pivot their research to climate science. In your opinion, what are some of the biggest problems or most exciting questions in the field right now?
If we look at climate science, at the most fundamental level, there are questions about predictability at longer timescales which are quite open for further exploration. There’s also a lot of questions around to what extent these events are being caused by anthropogenic activities. What kind of interventions should society enact such that we see less greenhouse emissions? Will this translate to fewer instances of extreme weather events? On a related note, I think evidence-based policymaking is also an important and impactful domain for upcoming young climate scientists.
Finally, the basic but profound realization that weather and climate affect almost every part of our lives is immensely valuable. From the availability of food and energy to planning our travel and migration — all these aspects of our life are critically influenced by our ability to understand weather and environment.