ON THE GENERALIZATION CAPABILITIES OF GRAPH TRANSFORMER MODELS IN WEATHER FORECASTING WITH SYNTHETIC DATA

Abstract

This study investigates the generalization capabilities of Graph-Transformer models for weather forecasting using synthetic data generated by the MPAS Atmosphere Core. The research addresses gaps in real-world data representation, particularly for extreme weather events, by proposing the use of synthetic datasets for model training. The methodology includes the generation of graph-based data, the development of Graph Transformer architecture, and performance evaluation using real-world datasets such as ERA5. The results demonstrate that while the models perform well for short-term predictions, challenges related to error propagation in autoregressive sequences persist. The study suggests future development of hybrid models, integrating machine learning with physical knowledge, to enhance robustness and interpretability. This research contributes to advancements in climate modeling and machine learning, offering new perspectives for extreme weather forecasting and spatiotemporal analysis.