SIParCS 2024 - Arnold Kazadi
Graph Machine Learning for Global Weather Prediction: Graph Residual Transformer + Gated Recurrent Unit (GRU)
In recent years, there has been significant interest in leveraging machine learning (ML) methods for global weather prediction. These methods are demonstrating competitive skills compared to top physics-based medium-range weather prediction systems, such as ECMWF’s IFS-HRES, with an improvement of up to 10% in RMSE on the ERA5 dataset. However, most existing methods, initially developed for image and video processing tasks, require a regular-grid representation, which is not suitable for representing spherical objects like the Earth. A grid representation of the globe overrepresents the poles and may require padding from opposite boundaries to ensure continuity of the spherical domain representing the Earth. Therefore, this work proposes a Graph Neural Network (GNN) combined with a Gated Recurrent Unit (GRU) model. This model, relatively smaller than other ML models for global weather prediction (~3 M parameters), adapts to diverse discretization schemes of the globe, and the GRU component helps mitigate the oversmoothing phenomenon typically associated with GNNs.
Mentors: David John Gagne, John Schreck, Charlie Becker, Gabrielle Gantos, Will Chapman
Slides and poster