AERIS: Argonne Earth Systems Model for Reliable and Skillful Predictions

Modeling complex nonlinear systems, such as the Earth system, to make skillful predictions beyond a few days is challenging due to computational costs and decreasing predictability over time caused by chaotic behavior inherent in the systems. To address this, researchers are using ALCF’s Aurora exascale supercomputer to develop AERIS, a multi-billion-parameter probabilistic foundation model for Earth system forecasting. Recognized as a finalist for the 2025 ACM Gordon Bell Special Prize for Climate Modeling, AERIS is trained to generate efficient, accurate forecasts from storm-scale events to subseasonal-to-seasonal forecasting.

Challenge

Earth system processes span vast spatial and temporal ranges, from localized convection to global circulation patterns. Current approaches often rely on specialized models at each scale, where errors accumulate across them, and computational complexity restricts the number of ensemble members that can be generated. Data-driven models show promise but face difficulties scaling to realistic resolutions, embedding physical constraints, and managing the extreme computational demands of AI training. Advancing research in this area requires a next-generation model capable of unifying local and global scales, maintaining physical fidelity, producing probabilistic forecasts, and training efficiently on extreme-scale supercomputers.

Approach

The team built AERIS using a generative, pixel-level Swin Transformer adapted for data-driven weather modeling. Central to its scalability is the team’s novel method called SWiPe (Sequence-Window Pipeline parallelism), which integrates window, sequence, and pipeline parallelism to efficiently shard the model without added communication overhead. Using ALCF’s Aurora system, the researchers trained AERIS on 16 terabytes of ECMWF Reanalysis v5 (ERA5) data encompassing several years of global atmospheric conditions. SWiPe reduced activation memory requirements and minimized communication costs, enabling stable training of models ranging from 1.3 to 80 billion parameters, marking AERIS as the first pixel-level diffusion model in the billion-parameter regime guided by physical priors for AI-driven Earth system forecasting.

Results

On Aurora, AERIS achieved a sustained, mixed-precision performance of 10.21 exaflops and a peak of 11.21 exaflops on 10,080 nodes (120,960 GPU tiles)—over 95 percent of the full machine. The training strategy demonstrates nearly linear scaling across thousands of GPUs while maintaining numerical stability with a generative, physics-informed loss function. In forecasting tests, AERIS delivered skillful medium-range predictions competitive with state-of-the-art numerical and machine learning models and remained stable in 90-day rollouts, capturing long-term ocean-atmosphere patterns and extreme weather indicators at high resolution. This work was recognized as a finalist for the 2025 ACM Gordon Bell Special Prize for Climate Modeling.

Impact

By demonstrating how generative models can perform across spatial scales, AERIS provides a blueprint for next-generation probabilistic forecasting tools for subseasonal-to-seasonal time scales. The model, trained on Aurora using a novel parallelization strategy, advances the scientific community’s ability to generate stable, high-resolution forecasts, improve subseasonal outlooks, and inform energy production and decision-making sensitive to extreme weather and environmental risk.