v1.0.0

Applying geometry-aware Clifford Fourier Neural Operators to weather forecasting

Version presented at the Machine Intelligence for GeoAnalytics and Remote Sensing (MIGARS) 2024 conference in Wellington on 9 April 2024

Abstract

To enable accurate medium-range (up to 15 days) forecasts of weather variables such as wind speed and pressure, we perform experiments using neural networks with Clifford Fourier transforms that are aware of the spherical geometry of the Earth. The neural network model is trained with the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) climate/weather dataset at an hourly temporal resolution. Taking sea-level pressure as the scalar field, and zonal wind and meridional wind as vector fields, we attempt to model the dependencies between different weather variables using Clifford Fourier Neural Operators that act as Partial Differential Equation surrogates. Forecast skill is evaluated using Root Mean Squared Error (RMSE) and Anomaly Correlation Coefficient (ACC) metrics, the latter used to better account for extreme weather events such as tropical cyclones. Once trained, we conduct further experiments using parameter-efficient fine-tuning techniques to adapt the MERRA-2 conditioned Clifford-based model weights to alternative climate/weather reanalysis datasets such as ECMWF Reanalysis v5 (ERA5), to test the viability of adapting pre-trained Weather Foundation Models to targets at different spatial and temporal scales.