Using AI to predict the weather

[Music] Welcome to Science Sessions, the podcast of the proceedings of the National Academy of Sciences, where we connect you with Academy members, researchers and policymakers. Join us as we explore the stories behind the science. I'm Matthew Hartcastle. Since their advent in the 1950s, physics-based weather forecasting models have enabled significant improvements in the accuracy of daily and long-term weather predictions. While powerful, these models can also require considerable processing power. In recent years, AI and machine learning methods have garnered interest as a less computationally intensive approach for weather forecasting. In this special episode, we will hear from atmospheric researchers, oceanographers and environmental scientists about recent advances in AI-enabled weather forecasting. With their ability to ingest large volumes of data, AI methods have the potential to capture patterns in daily or seasonal weather, helping to predict temperatures, precipitation, or extreme weather events. Jeffrey Schrader, an environmental economist at Columbia University, School of International and Public Affairs, in New York, is the author of a recent PNAS article. Schrader and his colleagues calculated the impact of day-ahead forecast accuracy on reducing heat-related mortality, finding that such forecasts improved by 34% between 2005 and 2023. Schrader explains what 48 expert forecasters had to say about how predictions might further improve through 2100, including the potential role of AI. We surveyed operational meteorologists, so people who are out there every day creating the weather forecasts that you and I and lots and lots of folks know and love and look at on an almost daily basis. What we found is basically they expect forecasts to keep getting better, a little bit slower than they've been getting better in the past, because they have been getting so much better, but they do expect forecasts to get really a lot better over the coming decades. In the last couple of decades, there's been a particular relationship between mortality and forecast errors, and that relationship really shows that forecast errors are especially deadly on hot days. In our central scenario, something like 18% of heat-related mortality is eliminated by expected improvements in forecasts by the end of this century. Those climate change makes hot days more common, forecasts are going to become more important. Under different climate change scenarios, you see that forecasts will help prevent even more deaths from heat. That number increases by up to 75% under even moderate climate change warming scenarios. Most of the experts that we talked to thought that AI and machine learning did hold promise for making forecasts better, but they also raised some caveats around that. One is the worry that some folks had that AI would cut humans out of the loop. One thing that human meteorologists are really good at doing is keeping these algorithms grounded into reality, all weather forecasts, whether they are AI generated, whether they're traditional physics-based forecasts, are fundamentally constrained by the quality of those forecasts is based on the quality of the input data that goes into them. Almost all of our experts talked about this as a really important potential downside risk. If we were to step back from investing in monitoring the weather, generating really high-quality weather measurements, then that would not only hurt our current forecasting infrastructure, it would hurt any possible future AI forecasts. Beyond daily weather forecasts, AI could improve the accuracy of long-term climate modeling. Ignacio Lopez Gomez, a research scientist at Google who studies weather forecasting, is the author of a recent PNAS article on using generative AI to downscale large-scale Earth system models into finer scale regional climate projections. Lopez Gomez explains the results of his and his colleagues approach. The main way that we currently have to obtain information or projections about the future climate is to use the so-called global climate models. These models enable us to peak hundreds of years into the future, but they do so at course resolution, so resolution of hundreds of kilometers. Regional climate models enable us to take that invaluable information at course scales and translate that into fine-scale information that then stakeholders can use to make actionable predictions. You can take these generative AI models, train them on real-world data, and they can produce, for instance, realistic videos or images if you train them from films. So the same way that we do that in media and in video, we can use generative AI models trained on historical weather patterns to essentially simulate the dynamics of weather systems at fine scales. Importantly, these models excel at capturing deep correlations in the data. We're interested, for instance, in the risk of wildfires. We need to predict what's the risk of high temperatures and how it correlates with winds and specific humidity. Using these generative AI tools in particular generative diffusion models, we can obtain very similar accuracy or skill compared to the physics-based models. Our model is about 90% less computationally expensive. We can then run this model many more times in order to understand risk better. Tropical cyclones are a dramatic example of weather. Accurately predicting how such systems will develop is essential for disaster preparedness. Xiao Fengli, an Oshogra firm at the Chinese Academy of Sciences Institute of Oceanology, is the author of another PNES article. Li and his colleagues use the machine learning model to forecast whether tropical cyclones will rapidly intensify, a difficult predict phenomenon that can increase the storm's destructive impact. Li describes the performance of the model. Traditional model direct predicts the tropical second intensity change over time from the input data. Our model asks a different question. What makes one storm intensify? Well, another similar does not. By comparing these cases, the model learns to isolate the key physical signals behind the rapid intensification. This makes it more robust, especially for rare and extreme events where traditional model often struggled. We use three main categories of data. The first one is atmospheric and oceanic environment data. We also use seven observations. The third data set we use is historical storm information. By combining environment structure and history information, the model can capture both the physical conditions and the internal dynamics of rapid intensification. Li's model achieved around 92% detection accuracy for rapid intensification events, while keeping the false alarm rate below 10%. Previous model often had much higher false alarm rate around 25 to 30%. This is important because rapid intensification is relatively rare. If we issue too many false alarms, it reduces the trust in the forecast, but missing real events can be also dangerous. So, improving both accuracy and reducing the false alarm rate, our model achieves a better balance between the two. On a smaller scale, thunderstorms are another example of rapidly developing, potentially dangerous weather that has historically been difficult to accurately predict. In another recent PNAS article, "Queesu, an environmental scientist at Hong Kong University of Science and Technology," inner colleagues use a deep diffusion model to develop four-hour ahead now-casting of thunderstorms. Suu explains what now-casting is and how our model works. Now-casting is a weather forecasting method that predicts conditions for the immediate future, typically in the next 0-6 hours. Now-casting relies on real-time data from radar or satellite to track the current movement of storms. Unlike traditional methods that would struggle with storm growth and dissipation, and also some AI models tend to blur predictions of extreme events. Our deep diffusion model for satellite data uses a two-step process that explicitly separates the predictable large-scale motion from chaotic, stochastic evolution of storms. The first step is to use your network to generate a deterministic baseline prediction of future satellite images. This captures the general movement of clouds. The second step is the diffusion model that predicts the difference between baseline prediction and the actual observation. Instead of predicting the cloud directly, it learns the complex patterns of growth decay and intensity change of storms. We found our model performs other traditional model by 2% to 20%, especially at a larger spatial scale. In general, it successfully predicted the growth and dissipation of convective cells, we can predict the occurrence of convective storm for our subhead. AI methods are fundamentally limited, but the data used to train them. If that data is flawed or inaccurate, the performance of the AI will suffer. Yet another Pianes article looked at the ability of AI to predict rare weather events called "grace swans" that are missing from the training data completely. Headdram Hassan Zada, a geophysicist at the University of Chicago, and one of the article's authors explains what "grace swans are" and why they may be challenging for AI to predict. AI weather models are really transforming weather forecasting. They learn from the past patterns. So they're doing a very good job predicting things that are not that different from what they have seen before. The question is if these models haven't seen something in the past, can they still predict them in short term forecasting? Grace swans are things that are physically possible. We just haven't seen them. We might see them. If you just observe the system more, climate change can increase the likelihood of getting these strong early events. Chong Sun, an atmospheric researcher at the University of Chicago, is another author of the article. Sun explains how Pianes' colleagues tested the ability of AI to predict "grace swans". Instead of like a waiting for some really extreme event to happen and test them, we actually designed the training this set for the AI weather model by reviewing very strong topo-cycle from the 20 sets. We can test whether this model trained without this event can predict these strong topo-cycle. What we found is AI really have issues with predicting this strong topo-cycle if you remove them from the 20 sets. They always give what we call a force negative. They predict a topo-cycle but it's very weak instead of a strong topo-cycle. We also have some encouraging findings. What we found is the AI model can learn across different regions. If we only add data back in the West Pacific region, the model can learn from that and predict strong hurricanes in the West Atlantic Ocean. Hassan Zahra expands on this unexpected and finding. We call this translocation. If there is an event that nobody around the world had seen it, the models cannot forecast it. But there are lots of extreme events that happen in different parts of the world. We have been looking at this rainfall in Dubai in April of 2024, which was twice more rain than any rain in that region. Yet we have found AI models to actually work well because that kind of rain happened in other parts of the world. These models are very powerful and they always do better than expected, like the translocation was something we didn't expect, but they're not magical. They cannot do extrapolation. As a supplement to physics-based models, human expertise and high-quality data collection, AI-enabled methods have the potential to continue improving the accuracy of weather and climate forecasts. With climate change increasing the likelihood of high temperatures and extreme weather events, the value of accurate predictions will only increase. Thanks for tuning in to Science Sessions. You can subscribe to Science Sessions on Spotify, Apple Podcasts, or wherever you get your podcasts. If you like this episode, please consider leaving a review and helping us spread the word. [BLANK_AUDIO]