Foundation models for addressing climate change

Hendrik Hamann spent his early career as a scientist making computers more energy efficient. He now uses computers to understand how the planet is changing and what we can do to cut carbon emissions and adapt to a warmer, and often more unpredictable, world.

Over more than two decades at IBM, Hamann developed technologies to reduce the power consumption of everything from largescale data centers to the microprocessors in your laptop. He has also invented new methods for integrating solar and wind power into the traditional grid. He is a fellow of the American Physical Society, and in 2016 awarded its prize for industrial applications of physics.

Today, he is chief science officer for climate and sustainability at IBM Research. In other physical challenges, he likes to compete in 50-mile and 100-mile ultramarathons. We caught up with him to talk about AI’s potential to improve our capacity to measure, mitigate, and adapt to climate change.

What drew you to the climate problem?

Climate change, in my view, is not only one of the most important challenges we’re facing, but it’s also a fascinating physics problem. This is especially true now, with the accelerating digitization of the physical world. We have data on just about everything — trees, oceans, the atmosphere — and the amount of data is only growing. We have a unique opportunity to use AI to extract insights from this data to discover how our climate works and to find solutions to climate change.

Foundation models let you mine this wealth of sensor data.

Exactly. Without AI, we wouldn't have a chance.

How can foundation models help us address climate change?

Climate change is an existential challenge and there’s no time to waste. We need to act quickly, and globally. That’s where foundation models come in. Unlike bespoke models, tailored for specialized tasks, foundation models can lead us to insights, discoveries, and different solutions much faster. They can create knowledge representations from petabytes and exabytes of climate data that we’ve collected. From that data, they can help us understand the carbon cycle, why certain extreme weather events happen, and how we can store more solar energy on the grid.

How are foundation models for climate different than for natural language?

Large language models (LLMs) extract meaning from sequences of words by understanding their relationships. Climate data is different. It’s highly dimensional and multi-modal; it contains information about time, space, temperature, clouds, atmospheric pressure, and so on. To unpack these relationships, we need new AI approaches and architectures. In other words, how do we train a model to understand the link between Tokyo winds in April and New York snowstorms in January.

How can AI facilitate Environmental Social Governance (ESG) reporting?

Foundation models can make ESG reporting much faster and easier, especially for “scope 3” CO₂ emissions in their supply chains. This is the amount of carbon embodied in the billions of products that we consume, that companies can choose to report publicly. A business can now use an LLM fine-tuned on ESG data to quickly estimate its carbon footprint. Faster reporting allows enterprises to integrate this information into their decision making, which can lead to reduced emissions.

Why is measuring carbon emissions so difficult?

Because it’s a gas, it's everywhere and diffused; it’s difficult to attribute carbon and other greenhouse gas emissions (GHGs) like methane to the places that produced them. We currently have two ways of tracking GHGs: indirectly, by estimating activities on-the-ground, and directly, using space-orbiting satellites. Foundation models also give us a real shot at direct measurements. By scaling this approach, we could enable carbon markets, which is a strategy for addressing climate change by putting a price on the industrial carbon we emit.

How is IBM innovating on technologies to mitigate climate change?

We are working on methods to increase renewable-energy capacity: improved methods for forecasting, which involves weather and climate prediction for many downstream tasks; and grid optimization, which involves bringing more solar and wind power on the grid.

We’re also working on fine-tuning our foundation models developed with NASA and others to understand how much CO₂ plants and trees are taking up. In essence, we’re using satellite-based observations from hyperspectral cameras and LiDAR sensors to measure changes in biomass, which is the amount of carbon that vegetation on Earth has absorbed from the atmosphere and stored.

How can biomass measurements help?

Many businesses and governments are interested in offsetting their carbon footprint by planting trees. If we can accurately track changes in biomass, we can verify that carbon that’s been sequestered, stays sequestered. Whether you’re storing carbon or just offsetting your own emissions, you can’t have a carbon market, or any commodity, if you can't measure it. That's why technologies for measuring carbon are so important. You cannot trade commodities like grain or oil if you can't measure them. IBM recently validated an AI method for tracking biomass by showing that it could calculate the amount of carbon released during California’s Caldor wildfire.

What is IBM doing on the adaptation side?

Foundation models can give us a window on the future, and that gives us time to plan. With NASA, we built a foundation model for flood detection. Better mapping of past floods can help us predict which areas are at high risk and where we should build in the future.

We are mapping urban heat islands to identify where city residents are most at risk and how should we plant trees or take other steps to cool the air. We’ve shown that you can fuse multiple data sources to estimate heat-island formation. Our test case was New York City; we are now extending this work to Abu Dhabi and places in Africa.

We are also building a weather and climate foundation model, which can help to predict extreme weather events like hurricanes and heat waves. Foundation models could potentially complement our current numerical weather prediction models. Foundation models can reduce high-performance computing loads because once they are trained and tuned, they will be much less computationally intensive than other models. They can help us make predictions faster and at a lower cost. IBM is working with NASA and Environment Canada on AI and foundation models for weather forecasting.

What is IBM doing to cut computing’s carbon footprint?

We are working hard to make computing more energy efficient. Efficiency has been a key innovation driver throughout the history of computing, from shrinking the size of transistors to inventing better ways of storing digital information, more sustainable packaging, and energy-efficient communication technologies. Today, cloud computing offers enormous opportunities for innovation. In the future, for example, we can shift workloads to data centers powered by renewable energy instead of fossil fuels.