November 2025
Design and Generation of Huge Ensembles of Weather Forecasts Using Spherical Fourier Neural Operators
A massive ensemble forecasting system to simulate and sample extreme weather events.
This figure shows air temperature and dewpoint forecasts for a heatwave over Kansas City, Missouri. The presented huge ensemble (blue dots) can sample the observed event (black triangle) better than existing models (red triangles).
The Science
Extreme weather events are rare and difficult to predict. Traditional weather models are accurate but expensive to run, which limits how many forecast scenarios (ensemble members) we can generate. To understand the drivers of extreme events, we need huge ensembles that cover many possible outcomes. We present an ensemble system that is fast, reliable, and detailed enough to capture future atmospheric states while also managing the enormous data volumes produced.
The Impact
Huge ensembles (HENS) deliver the sample sizes needed to probe extreme events with calibrated probabilistic forecasts. Existing models are too computationally expensive to sample extreme events at a large scale. We use machine learning methods, which are orders of magnitude faster than conventional methods. We develop a probabilistic model that represents different sources of uncertainty. The model is validated with diagnostics specifically designed to evaluate performance on extreme weather events. With a larger sample size, HENS can simulate extreme events that conventional systems often miss. It thus strengthens preparedness and advances our scientific understanding of when and why extreme weather occurs.
Summary
Simulating rare but severe weather events is difficult for today’s weather forecast systems, which typically use only about 100 ensemble members. Larger ensembles could better capture the full range of possible outcomes, especially extreme events, but running thousands of physics-based forecasts is too computationally expensive. This study replaces traditional simulations with machine learning (ML) to create massive ensemble forecasts. In Part I, we build a system using Spherical Fourier Neural Operators (SFNOs), which model uncertainty by slightly changing both model parameters and starting conditions. The ML ensemble performs competitively with existing systems but requires far fewer computing resources. The forecasts remain physically realistic over time and accurately represent extreme weather patterns.
In Part II, we generate a Huge Ensemble (HENS) of 7,424 members, initialized daily during summer 2023. This system provides detailed coverage of atmospheric variability and captures very rare (four standard deviation) events. HENS reduces cases where actual weather falls outside the predicted range. Overall, ML-based huge ensembles enable realistic, reliable, and computationally efficient simulations of extreme weather.
Contact
Travis O’Brien
Lawrence Berkeley National Laboratory / Indiana University
TAObrien@lbl.gov
Ankur Mahesh
Lawrence Berkeley National Laboratory
amahesh@lbl.gov
Funding
This research was supported by the Director, Office of Science, Office of Biological and Environmental Research of the US Department of Energy under contract no. DE-AC02-05CH11231 and by the Regional and Global Model Analysis Program area within the Earth and Environmental Systems Modeling Program. The research used resources of the National Energy Research Scientific Computing Center (NERSC), also supported by the Office of Science of the US Department of Energy, under contract no. DE-AC02-05CH11231. The computation for this paper was supported in part by the DOE Advanced Scientific Computing Research (ASCR) Leadership Computing Challenge (ALCC) 2023–2024 award “Huge Ensembles of Weather Extremes using the Fourier Forecasting Neural Network” to William Collins (LBNL).
This research was also supported in part by the Environmental Resilience Institute, funded by Indiana University’s Prepared for Environmental Change Grand Challenge initiative.
Publications
Mahesh, Ankur, Collins, W. D., et al. “Huge ensembles—Part 2: Properties of a huge ensemble of hindcasts generated with spherical Fourier neural operators.” Geoscientific Model Development 18, no. 17 (2025): 5605-5633. [https://doi.org/10.5194/gmd-18-5575-2025]
Mahesh, Ankur, Collins, W. D., et al. “Huge ensembles—Part 1: Design of ensemble weather forecasts using spherical Fourier neural operators.“ Geoscientific Model Development 18, no. 17 (2025): 5575-5603. [https://doi.org/10.5194/gmd-18-5605-2025]
Predict Extreme Weather Events in Minutes Without a Supercomputer, NVIDIA