Spherical multigrid neural operator for improving autoregressive global weather forecasting

Area of Science:

• * Meteorology and Atmospheric Science
• * Artificial Intelligence and Machine Learning
• * Computational Science

Background:

• * Data-driven global weather forecasting models show promise but face challenges with spherical distortions, impacting forecast stability.
• * Conventional architectures and existing methods like spherical Fourier neural operator (SFNO) struggle with computational efficiency due to spherical harmonic convolution.
• * High computational costs limit the practical application of advanced neural operators in weather prediction.

Purpose of the Study:

• * To introduce a novel spherical multigrid neural operator (SMgNO) designed to overcome the limitations of existing weather forecasting models.
• * To enhance the accuracy and stability of autoregressive forecasts on a spherical domain.
• * To significantly reduce the computational resources required for global weather forecasting.

Main Methods:

• * Integration of spherical harmonic convolution with a low-resolution spherical Fourier neural operator (SFNO) within a multigrid framework.
• * Development of the spherical multigrid neural operator (SMgNO) to effectively mitigate spherical data distortions.
• * Experimental validation using spherical shallow water equations and medium-range global weather forecasting.

Main Results:

• * SMgNO demonstrated superior performance in medium-range global weather forecasting, showing a 9.31% improvement in anomaly correlation coefficient over IFS T42 for 500 hPa geopotential height (7-day lead time).
• * SMgNO achieved a 6.83% improvement in anomaly correlation coefficient over SFNO for the same forecasting task.
• * The proposed SMgNO requires only 10% of the floating-point operations of SFNO for forward propagation and 30.90% less GPU memory during training, indicating significant computational savings.

Conclusions:

• * SMgNO effectively alleviates data distortions inherent in spherical forecasting architectures.
• * The model offers a computationally efficient and robust solution for data-driven global weather forecasting.
• * SMgNO represents a significant advancement in applying neural operators to complex atmospheric prediction problems.