From hurricanes to heatwaves, our ability to predict the weather helps guide critical decisions in energy, agriculture, public safety, and beyond. However, producing reliable forecasts, both quickly and far into the future, has long pushed the limits of science and computing.

Now, researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory introduce AERIS — the Argonne Earth Systems Model for Reliable and Skillful Predictions. AERIS is a breakthrough AI system that learns from decades of Earth systems data to deliver fast, high-resolution forecasts from hours to months into the future. In early tests, AERIS produced medium-range forecasts in about 30 seconds and was able to predict extreme events with high accuracy. 

AERIS is one of the largest AI models for science created to date. It was built and trained on Aurora, one of the world’s fastest supercomputers, located at the Argonne Leadership Computing Facility (ALCF), a DOE Office of Science user facility. Aurora is an exascale system capable of performing more than a quintillion — that is, a billion billion — calculations per second and ranks among the top-performing machines globally for AI workloads.

The combined scale of Aurora and the model itself, trained on high-resolution data and billions of parameters, enabled AERIS to push beyond the 10-day limit typical of many AI Earth system models.

“With AERIS, we’ve developed a model that can extend forecasts to subseasonal-to-seasonal scale, meaning it can produce useful predictions beyond two weeks,” said Argonne senior scientist Rao Kotamarthi.

Traditional models begin to lose accuracy around 10 days because the atmosphere and ocean are chaotic and sensitive to small changes. Even so, important signals appear weeks in advance, Kotamarthi noted. 

“By capturing those indicators, we can better forecast the effects in different parts of the world,” he said. “If you are in the energy sector, for example, you want to know how much gas to buy for winter, or how to plan electricity demand weeks in advance. These are the kinds of questions that motivate AERIS.”

The team’s work has been recognized as a finalist for the Association for Computing Machinery’s prestigious 2025 Gordon Bell Special Prize for Climate Modeling. Running on nearly the full Aurora system — 10,624 nodes and 63,744 graphics processing units (GPUs) — AERIS sustained 10.21 exaflops and peaked at 11.21 exaflops. That means AERIS achieved over 10 quintillion operations per second, far beyond Aurora’s 1 exaflop capacity, because AI workloads use smaller data types that enable the machine to handle many more tasks simultaneously

“Reaching over 10 exaflops on Aurora, AERIS is the highest performing model in AI for science that we've seen yet,” said Argonne computational scientist Jason Stock. “This is really the first billion-parameter model of its kind.” A billion-parameter model is a large AI system with a billion internal settings that it adjusts to recognize patterns and make predictions. 

A different kind of Earth systems model

AERIS takes a different path from traditional weather models. Numerical prediction systems solve complex equations to simulate atmospheric physics, which limits how fast they can run and how well they can use large archives of observations. AERIS is purely data-driven, learning patterns directly from decades of observations. The team trained AERIS on a massive dataset of high-resolution images of global weather conditions from 1979 to 2018, totaling 16 terabytes ­— equivalent to roughly 4 million photos stored on a smartphone. 

AERIS uses a modern AI technique called a diffusion model, best known for generating images, to create many plausible forecasts and estimate their uncertainty. Unlike traditional models that provide a single forecast, AERIS generates many possible scenarios, called ensembles, for uncertainty estimates.

It also reads the data pixel by pixel, keeping fine details that other methods often blur. This produces sharper, more realistic forecasts, but requires a lot of computing power.

Meeting those demands required both Aurora’s scale and a new way to divide the work. 

"We wanted to really push the boundaries of model size and resolution, but we needed to come up with a new approach to scaling because the existing solutions weren’t sufficient,” said Väinö Hatanpää, an assistant computer scientist at ALCF.

To run AERIS effectively on Aurora, the team developed a method called Sequence Window Parallelism (SWiPe). This approach efficiently distributes the model’s compute tasks and data across Aurora’s more than 60,000 GPUs while reducing communication between them. 

“SWiPe allows us to use more GPUs for the same problem without incurring additional communication between the devices, resulting in faster, more efficient training runs,” Hatanpää added.

The team also tested AERIS on the LUMI supercomputer at the CSC – IT Center for Science in Kajaani, Finland, showing the approach works well on different kinds of computer systems.

Early results, practical impact

In tests, AERIS outperformed a leading European forecasting system for predictions up to 10 days. It also stayed stable out to 90 days, capturing long‑term ocean-atmosphere patterns and realistic tropical wave behavior at high resolution.

“These factors drive extreme weather events that matter for infrastructure, safety, and energy demand,” Stock said. “The fact that our model can capture them at these time scales is really impressive.”

At the same time, the researchers note that AERIS is still early in its development and faces practical limits. Experiments with new configurations, higher resolutions, or added data can require enormous computing time, with some runs taking a week or more even on an exascale system. AERIS was designed so it could later be “distilled” into a smaller version that could run on less powerful machines, including laptops. They also see their approach as a blueprint that could help researchers build large-scale AI models for other areas of science.

“This effort validates the steps we’ve been taking over the past several years to demonstrate that an AI foundation model approach for Earth systems science can work,” Kotamarthi said. “Every time I look at the results, I’m amazed that this is possible. We’re at the beginning of a new path for doing big science at scale, and I’m excited to see where it goes.”

The study, “AERIS: Argonne Earth Systems Model for Reliable and Skillful Predictions,” was authored by Väinö Hatanpää, Eugene Ku, Jason Stock, Murali Emani, Sam Foreman, Chunyong Jung, Sandeep Madireddy, Varuni Sastry, Sam Wheeler, Huihuo Zheng, Troy Arcomano, Venkatram Vishwanath, and Rao Kotamarthi from Argonne National Laboratory; Ray A. O. Sinurat from the University of Chicago; and Tung Nguyen from the University of California, Los Angeles.

The team’s research was supported by DOE’s Office of Cybersecurity, Energy Security, and Emergency Response (CESER). Computing time on Aurora was supported by DOE’s Advanced Scientific Computing Research (ASCR) program