Breaking the Gridlock: Efficient Atmospheric Data Reconstruction and Prediction via Generative 3D Gaussian Splatting

Tao Han, Zhibin Wen, Zhenghao Chen, Hao Chen, Fenghua Ling, LEI BAI, Song Guo

Keywords: Generative 3D Gaussian Splatting, Numerical Weather Forecasting

Abstract: AI-based numerical weather prediction (NWP) models often rely on regular latitude–longitude grids that induce strong data redundancy, limiting scalability to higher resolutions and wasting computation. We present $\textit{GaussianCast}$, a generative 3D Gaussian Splatting (3DGS) framework for compact, continuous representation and efficient forecasting of high-dimensional atmospheric fields. To reduce redundancy while preserving global consistency, we place Gaussian centers on a Reduced Gaussian Grid (RGG), achieving equal-area sampling and enabling up to 14× compression. Conditioned on the current atmospheric state, multi-scale Graph Attention Transformers generate 3DGS covariances, occupancy, and attributes for both reconstruction and forecasting. On ERA5 dataset, GaussianCast achieves accurate weatehr reconstruction and skillful medium-range weather forecasting at substantially lower computational cost, and remains competitive on tropical cyclone tracks. To our knowledge, it is the first generative 3DGS NWP framework to place Gaussians on RGG and predict their parameters for reconstruction and forecasting. Code is available at: https://anonymous.4open.science/r/GaussianCast-9F7B

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 7603