Artificial Intelligence and Machine Learning

AI Is In the Weather Forecast

Artificial intelligence can improve climate modeling, boosting both the speed and accuracy of predictions.

stylized tornado, illustration

The path of a tornado or the intensity of a flood may seem somewhat mysterious and unpredictable. Yet amid countless variables—air temperature, humidity, wind, barometric pressure, ocean currents, topography, and solar intensity—there are critical factors that lead to a specific conditions.

Understanding these dynamics, which can change minute by minute, has long been the goal of climate scientists. That is where digital twins and other forms of artificial intelligence (AI) blow into the picture. These tools deliver sophisticated visual simulations that help weather forecasters and climate scientists identify atmospheric patterns more precisely.

Improved climate modeling can boost both the speed and accuracy of predictions. It can identify risks for floods or wildfires in specific places, help government agencies and citizens plan and prepare for weather-related events and reduce long-term climate risks such as crop failures, deforestation, water shortages, wear on buildings, and energy availability.

“The biggest advantage of digital twins and advanced AI tools is that they enable new research and understanding that cannot be achieved with traditional simulation and analysis methodologies,” said Lucas Harris, deputy division leader at the Geophysical Fluid Dynamics Laboratory of the National Oceanic and Atmospheric Administration (NOAA).

Modeling Reigns Supreme

Weather forecasting and climate modeling are undergoing a quiet revolution. Decades ago, meteorologists could forecast only two or three days out. Today, five- to seven-day predictions are the norm—with a much higher degree of accuracy. “We’re combining complex numerical models of the earth with large volumes of physical weather data,” said Irina Sandu, director of the Destination Earth Project (DestinE) at the European Centre for Medium-Range Weather Forecasting (ECMWF).

Emerging modeling methods, including digital twins, deliver a far more detailed picture of climate and weather behavior. Instead of modeling the Earth system at a 10- or 100-kilometer scale, the digital twin can generate simulations at a scale of 5 kilometers or less, globally. At the same time, it allows scientists to test different hypothesis and examine how they might affect the Earth system, depending on different scenarios. “We can do bespoke simulations and examine critical factors in a highly interactive way,” Sandu said.

For example, the European Union’s DestinE project, launched in 2021 and entering a second phase in 2024, strives for highly accurate, interactive, dynamic digital twin simulations that can be used for forecasting extreme weather events and multi-decadal climate simulations. These digital replicas of Earth systems are valuable for urban planning, energy management, and emergency relief plans, Sandu explained. For example, researchers might examine how to optimize the use of wind turbines and solar panels in specific areas.

In the U.S., NOAA also is deploying digital twins and other AI models to advance climate science. One project, Earth Observations, is a digital twin that models climate, weather, and specific ecosystems. Another project, eXperimental System for High-resolution prediction on Earth-to-Local Domains (X-SHiELD), performs kilometer-scale simulations for specific events. “We can see how small-scale features like thunderstorms or terrain can interact with the larger-scale climate circulations,” NOAA’s Harris noted.

Meanwhile, Nvidia has introduced Earth-2, a full-stack, open platform digital twin that generates highly interactive, AI-augmented, high-resolution climate and weather simulations. The system—which uses the company’s CorrDiff AI diffusion model to render detailed images and scenes using data interpolation—simulates floods, heat waves, probable hurricane tracks, and numerous other events, said Dion Harris, director of the Accelerated Data Center at Nvidia.

At the University of Pittsburgh, researchers are using a digital twin to better understand how climate change and extreme weather conditions can affect buildings. Scientists will run simulations to understand specific scenarios as well as long-term trends. “This shows us how we can improve designs and operations to adapt to varying conditions,” said Alessandro Fascetti, assistant professor in the Department of Civil and Environmental Engineering at the University of Pittsburgh.

Winds of Change

Powering digital twins are high-performance computing (HPC), GPUs, algorithms, improved software, and large volumes of data. For example, Nvidia’s Earth-2 runs on DGX Cloud supercomputers, including DGX GH200, HGX H100, and OVX systems. It incorporates hundreds of GPUs.

ECMWF plugs in data from satellites, weather stations, and other sources. High-performance computers in Barcelona, Spain, Bologna, Italy, and Kajaani, Finland are equipped to handle 352 billion points of information and 1 petabyte of data per day.  This recently allows researchers to produce the first multi-decade prototype project at a 5-kilometer scale. Eventually, scientists would like to achieve a resolution of 1.4 kilometers.

In the future, ECMWF hopes to plug in additional data sources, including backyard-connected weather stations, while improving the current numerical models that underpin the digital twins. Thanks to rapid advances in AI, data-driven models can complement existing physical models. With this more-complete picture of weather and climate dynamics, forecasters and government officials can get closer to the goal of real-time forecasting and earlier warnings for extreme events.

For now, the biggest challenges revolve around the cost and availability of compute power that is used to train data models and run digital twins. Not surprisingly, scientists are unable to push every climate and weather scenario through a digital twin because the cost of GPU hours in a high-performance computing center is exorbitantly high. This makes the technology feasible for large government weather services, private forecasting services, and major corporations studying specialized use cases.

“Global kilometer-scale models are still too computationally expensive for routine weather forecasting,” said NOAA’s Harris. “We can get part of the way there with new and more advanced modeling technologies that ‘zoom in’ on a region.” Yet, in order to arrive at precise weather and climate models, there is an ongoing need to plug in more data, refine algorithms, and further validate results. This includes comparing the performance of a digital twin simulation with actual events and the outcomes they model.

Yet, over the next few years, climate scientists are certain digital twins and other AI-powered visual simulations will play a significant role in identifying extreme weather risks and long-term climate solutions. They might pinpoint an outbreak of tornadoes, identify the specific path of a typhoon, or model the impact of a drought on land, mining, and agriculture.

Concluded Nvidia’s Harris, “There is a virtuous cycle with this technology. As we gain more information and feed it back into these digital twin models, they get better and better.”

Samuel Greengard is an author and journalist based in West Linn, OR, USA.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More