Benchmarking Spherical Fourier Neural Operators for Regional Weather Forecasting

Abstract

Modern deep learning weather prediction models excel in prediction speed as well as accuracy. Among these Spherical Fourier Neural Operators have emerged as a promising architecture that implicitly model underlying atmospheric physics using Spherical Harmonics. However, implementations of such architecture are primarily evaluated on global benchmarks, while region specific analysis is left unexplored.

The thesis focuses on investigating the reproducibility and performance of SFNO based ForeCastNet model on both Global as well as Indian Subcontinent data. Experiments are conducted on two spatial resolutions, and the resulting model dynamics, stability and accuracy are analyzed. A complete training and evaluation pipeline is developed, which includes data preprocessing, dataset construction, model training, rollout, forecasting and evaluations across multiple variables.

Model performance is evaluated using standard meteorological verification metrics such as Root Mean Square Error (RMSE) and Anomaly Correlation Coefficient (ACC) across multiple forecast lead times. The results provide insights into the strengths and limitations of neural operator based forecasting models and contribute toward improving reproducibility and regional evaluation in scientific machine learning for weather prediction.