EllipWeather: Gaussian Ellipsoid Representation for Weather Modeling

ICLR 2026 Conference Submission21271 Authors

Keywords: representation learning, weather modeling

Abstract: Accurate weather forecasting plays a critical role in a variety of sectors, including disaster management, agriculture, transportation, and energy consumption. Previous deep learning approaches have primarily relied on pixel-based representations of weather data, which can lead to significant data redundancy and inefficiencies in capturing the continuous nature of weather phenomena. To address these challenges, we propose a novel approach, EllipWeather, which leverages Gaussian ellipsoids to represent weather patterns, as weather phenomena can be effectively modeled using a mixture of Gaussian distributions. With this representation, we first develop an equivariant graph neural network to capture the intrinsic equivariance of weather variances, specifically tailored to process Gaussian ellipsoids for weather prediction tasks. Then we also demonstrate the potential of EllipWeather in downstream tasks such as data compression and downscaling (super-resolution). Extensive experiments on commonly used datasets show that EllipWeather achieves superior performance over previous works.

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

Submission Number: 21271