HomeAISTORM Wins the 2026 ACM Gordon Bell Prize for Climate Modelling

STORM Wins the 2026 ACM Gordon Bell Prize for Climate Modelling

ACM has awarded the 2026 Gordon Bell Prize for Climate Modelling to a 13-member team, most of it at Oak Ridge National Laboratory, for STORM, a generative AI method that recasts data assimilation as a single-stage sampling problem. The prize was presented on 19 November 2026 at SC26 in Chicago, and is described in ACM’s release of the same day and on the conference paper page. The scaling figures are the paper’s, as tested. STORM is a research method, not an operational forecast service.

The paper, “Global Attention with Linear Complexity for Exascale Generative Data Assimilation in Earth System Prediction,” calls STORM “a one-stage generative AI framework that reformulates DA as diffusion-based posterior sampling, replacing PDE forecasts with scalable AI inference.” Conventional two-stage data assimilation, the authors say, is computationally prohibitive at high resolution because repeated physics-based ensemble forecasts, observation updates and data movement limit how large an ensemble can be.

As tested, the paper reports: “STORM scales to 74,400 GPUs on Frontier with 96%–99% strong-scaling efficiency and up to 6 ExaFLOPs sustained BF16 throughput”, and “32,768-member ensembles for uncertainty quantification in 34 seconds on 4,096 GPUs.” It also says the method scales to 20 billion spatiotemporal tokens and 177,000 temporal frames. Those numbers describe a run on the Frontier supercomputer. They are not a schedule of official weather forecasts.

The authors then present their own results on hurricane tracking and long-term climate reanalysis, including, in the paper’s wording, “benefits from longer temporal context and recovery of temperature extremes missed by forecast-only predictions.” Those accuracy gains are the team’s. ACM’s release adds that the methodology may be applicable in other fields, including fusion energy, astrophysics and materials science; that is a claim about breadth, not a report of those applications having been delivered.

The corresponding author is Guannan Zhang of ORNL. The 13-member team also includes researchers from Auburn University, Florida State University, Colorado State University, Tulane University and AMD. A cash prize of $10,000 accompanies the award, which ACM says was conceived and funded by Gordon Bell and will be given annually for the next ten years, with the winning research published in The International Journal of High Performance Computing Applications.

STORM belongs with other AI weather and climate methods that have been published as research, not as a national forecast service. ECMWF has a prototype that makes AI-DOP forecasts straight from observations, still framed as a prototype. DeepMind has opened the PARDIG and SatWeave weather datasets, which are research data releases, not official guidance. Closer still as a method paper is MIT’s work on training climate emulators on far less data. None of those posts, and not this prize, converts a tested algorithm into an operational forecast.

A prize-winning method is not an operational forecast

The figures that will travel with this award are the ones that have to be kept in their setting. STORM, the authors say, scaled to 74,400 GPUs on Frontier with 96%–99% strong-scaling efficiency and up to 6 exaFLOPs of sustained BF16 throughput, and produced 32,768-member ensembles in 34 seconds on 4,096 GPUs. Those are as-tested performance numbers from a prize paper. They tell you what the team measured on a particular supercomputer, under the conditions of that study. They do not tell you that a weather service has adopted the method, that a hurricane forecast you will see tomorrow was made this way, or that the accuracy gains reported for tracking and reanalysis have been confirmed by an operational centre.

A Gordon Bell Prize for Climate Modelling recognises, in ACM’s words, innovative parallel computing contributions toward solving the global climate crisis. Winning it is a serious mark in high-performance computing. It is not a procurement, a forecast licence, or a finding that physics-based models have been retired. The paper is open about the problem it is trying to fix: two-stage data assimilation that cannot grow ensembles cheaply enough at high resolution. Whether STORM is later taken up as a component of an operational system is a later fact. The next things to watch are an independent reproduction of the scaling and ensemble timings, or a named weather or climate centre saying it will run the method in operations. Neither is in the ACM release or the paper page.

How we sourced this

The description of STORM and the scaling and ensemble figures are taken from the SC26 paper abstract and from ACM’s 19 November release, as set out in Research’s brief. Hurricane and reanalysis accuracy claims are attributed to the authors. No operational-forecast claim has been added.

Sources

Related stories

ECMWF Prototype Makes AI-DOP Forecasts Straight From Observations is the closest recent TSN account of an AI weather method that is still a prototype, and is the next read.

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.