“Indispensable for our transition to a green economy”: Inside the New Era of Climate Modeling
Our window to a sustainable future is narrowing.
If we’re going to combat the catastrophic effects of climate change, we can’t wait for the collective action of nations to solve our most pressing climate challenges. One of the many ways we can take action now is by improving our capacity to predict how exactly climate will change by making critical advancements in the science of climate modeling.
Over the past 50 years, large numerical and computational models, known as climate models, have become an indispensable tool for climate science. Climate modeling is similar to weather forecasting, but focuses on changes over much longer periods of time. Using hundreds of factors, ranging from air temperature and solar radiation to CO2 and deforestation levels, climate models help us predict future climate, and related weather patterns at a regional level. However, we still need to improve their accuracy.
With accurate climate modeling, we can pinpoint how a changing climate influences things affecting human survival, such as agriculture, ecosystems, and the threat of new diseases.
Schmidt Futures’ Virtual Earth System Research Institute (VESRI), will accelerate the pace of climate research by assembling hundreds of scientists across eleven countries, and fifty research institutions to tackle some of the most difficult scientific and computational problems in climate modeling. With more reliable models that are widely accessible, more informed decisions on how to prepare and prevent climate change-related disasters can be made. The Institute complements the work of The Schmidt Family Foundation to restore a balanced relationship between people and planet, and the 11th Hour Project’s support for innovative climate change mitigation solutions.
We spoke to Distinguished Fellow V. Balaji to understand the field of climate modeling today, and what the future holds.
Tell us about the field of climate modeling. Why is it important?
I’d like to stress the importance of one word: counterfactuals. A counterfactual is something that cannot be measured or observed because it never happened. For example, in order to track cause and effect, let’s say the carbon emissions attributable to industry, we need to know what the planet might have looked like, had the Industrial Revolution never taken place. Of course, we don’t have knowledge of that hypothetical climate. But we can simulate it: if we have a reasonably realistic model, we can set it to pre-industrial conditions, say from 1850, and run it forward to the present, producing a counterfactual non-industrial Earth. We similarly look at the future of planet Earth under different future possible scenarios: a world of cooperation to stabilize the climate, a world of conflict, a world of “business as usual.”
These are the experiments we do with climate models, since we cannot do them with the one planet we have. The importance of climate models to generate these counterfactuals cannot be overstated. Many people think of climate models as tools to predict the future, but they are way more than that. It’s the comparison against a counterfactual that makes policy: you need a prediction of what would happen under a given policy, but also what would happen without it. Climate models are indispensable for our transition to a green economy. And that’s why our work at the Virtual Earth System Research Institute (VESRI) is focused on improving the accuracy and credibility of major climate models.
How has advanced computation and AI impacted the field of climate modeling?
In 2017, VESRI was conceived at a moment when computing technology had hit an inflection point that led both hardware and software toward machine learning (ML), and large language models (LLMs).
The field is now reflecting broader shifts defined by the transition from the period referred to as the “AI winter” (when progress stagnated) to the current AI renaissance, in which a confluence of techniques and computing power has advanced progress at a breakneck pace. This era favors the applications of AI over conventional computations, so we are supporting projects that discover fundamentally new approaches inspired by machine learning.
These new approaches are necessary because we’re taking on complex interdisciplinary problems- across physics, math, and more – to address shortcomings in current climate models.
Has AI also had a negative impact on this work?
In general this impact has been broadly felt across all sciences, and the impact has been generally anywhere between positive and revolutionary. Negative outcomes are possible of course: while the LLMs burst on the public scene with the release of ChatGPT, I already see people increasingly relying on such tools to summarize large amounts of material. I understand the impetus, since one cannot possibly keep pace with the scientific literature, but there’s a danger of losing details and nuances, if no one reads the primary literature anymore. Details matter, it often turns out.
The UN’s 2022 climate report was widely misinterpreted, because a summary for policymakers stated that “Global greenhouse gasses are projected to peak between 2020 and at the latest by 2025, in global modeled pathways that limit warming to 1.5C”. Now, that wording is ambiguous. You could infer that emissions can rise until 2025 and the world could still stay under 1.5C warming compared to pre-industrial levels. However, emissions actually need to fall immediately to avoid the most dangerous levels of warming. This is another reason why we need more accurate climate models. As Dr. Joeri Rogelj, one of the report’s authors noted: “Because models work on 5-year increments, we can’t derive statements with higher precision.”
What do you see as the next big step or hurdle for climate modeling?
The key challenge for climate modeling has been the ability to discern small changes accumulating slowly over time, slower even than the entire instrumental record. It has not yet been 70 years since Sputnik, the first satellite! By comparison, the climate, particularly the oceans and forests, change over far longer durations. We are making exciting progress in calibrating our models using the new techniques coming from machine learning to all the available data, while still being able to generalize outside the observations. This is an essential feature of climate models, to be able to project the kinds of climate events that haven’t been seen during the satellite period, or even a couple of centuries of weather data.
What predictions do you have for the advancement or application of climate modeling over the next 5-10 years?
“Predictions are hard, especially about the future” as a wise person once said. This is particularly true in an era of technological upheaval. But since you asked, I’ll go out on a limb and say, it’s likely that AI will be doing the weather forecasts for you on your device in 10 years. Climate, as I stated above, is a different problem: it’s hard to reason from data about the recent past because the climate is changing, faster than we have seen for 10,000 years. But we are making tremendous leaps, and something tells me that AI & machine learning applications in climate in the next few years will not only let us learn new things about the climate, but also give us fundamental new insights about AI itself.
For example, models have a limit on how far they can “zoom in” – how much detail of the Earth they can include. AI and machine learning helps us look at very high resolution data, so that it can fill in details in the models without actually simulating them. As another example, we have some parameters in our models whose value is unknown or hard to measure in nature: Machine learning can help us pick the optimal value that gives the most accurate climate projections.
