Democracy of AI Numerical Weather Models: An Example of Global Forecasting with FourCastNetv2 Made by a University Research Lab Using GPU

Iman Khadir, Shane Stevenson, Henry Li, Kyle Krick, Abram Burrows, David Hall, Stan Posey, Samuel S. P. ShenarXiv 2025 · DOI · arXiv

Introduction

Traditional Numerical Weather Prediction (NWP) has long been the exclusive domain of national supercomputing centers. The computational intensity of solving complex fluid dynamics equations creates a high barrier to entry, limiting operational forecasting to major governmental agencies. However, the advent of AI-driven models like NVIDIA’s FourCastNet offers a faster, more cost-effective alternative. This paper demonstrates that high-resolution global weather forecasting is now accessible to university research groups by leveraging modern GPU clusters and open-source tools, effectively "democratizing" the field.

Infrastructure and Methodology

To bridge the "last-mile" technical hurdles, our team utilized the TIDE project, part of the National Research Platform (NRP) Nautilus cluster. This infrastructure uses Kubernetes to manage isolated "pods," allowing researchers to scale from simple inference to fine-tuning complex models.

The TIDE Cluster architecture within the National Research Platform Nautilus cluster, illustrating the network flow from public datasets to GPU compute nodes.

We compared two generations of NVIDIA’s weather models:

  • FourCastNet (v1): Based on an Adaptive Fourier Neural Operator (AFNO) and trained on 20 atmospheric variables.
  • FourCastNetv2: An advanced iteration utilizing a Spherical Fourier Neural Operator (SFNO), trained on 73 variables for improved stability and accuracy.

While the original training of these models required massive resources (e.g., 64 GPUs), we demonstrated that a university lab can successfully perform inference and fine-tuning using a modest setup of 4x NVIDIA A100 GPUs.

Data Management and Workflow

The primary dataset used is ERA5 (ECMWF Reanalysis v5), providing global hourly estimates at a 0.25° resolution (~31 km). Managing the approximately 6 terabytes of data required for FourCastNetv2 is a significant challenge. We utilized a shared CephFS volume to ensure accessibility across the research team. The workflow involves a rigorous data normalization pipeline and an autoregressive inference process, where the model iteratively uses its own output to generate long-term forecasts.

Key Results: Hurricane Florence Case Study

A critical test for these models is their ability to handle complex land-sea interactions. In a case study of 2018’s Hurricane Florence, we found that FourCastNetv2 significantly outperformed its predecessor. While v1 struggled with accuracy after the hurricane made landfall, v2 remained stable and accurately predicted the hurricane's eye location over a 102-hour period.

Comparison of Hurricane Florence track predictions between FourCastNet (v1) and FourCastNetv2, showing the superior stability of the v2 model after landfall.

This stability is likely due to the more robust SFNO architecture in v2. Beyond specific storm tracks, the models maintained physical coherence in global forecasts for zonal winds and surface pressure even at long lead times (up to 150 hours).

Discussion and Future Outlook

Our findings suggest that for resource-constrained labs, focusing on inference and fine-tuning is more viable than training from scratch. We observed a training loss plateau when using limited compute, suggesting the model quickly converges on existing weights without a more diverse dataset.

However, the ECMWF AI-Models API has emerged as a game-changer, simplifying data acquisition and allowing students to generate 10-day forecasts on consumer-grade hardware. Despite current limitations—such as the five-day latency in ERA5 data and the technical complexity of translating supercomputer scripts to Kubernetes—this roadmap enables immediate deployment by students and researchers. Future work will focus on applying these low-cost models to provide localized forecasting for regions currently lacking high-resolution weather data infrastructure.

Summary updated

BibTeX
@misc{khadir2025democracy,
  author = {Iman Khadir and Shane Stevenson and Henry Li and Kyle Krick and Abram Burrows and David Hall and Stan Posey and Samuel S. P. Shen},
  title = {{Democracy of AI Numerical Weather Models: An Example of Global Forecasting with FourCastNetv2 Made by a University Research Lab Using GPU}},
  year = {2025},
  eprint = {2504.17028},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.17028},
}
Figure