GDAS-Powered Machine Learning Weather Prediction: A Comparative Study on GraphCast initialized with GDAS States for Global Weather Forecasting

Abstract

The medium-range forecasting of global weather plays a pivotal role in decision-making processes across various societal and economic sectors. Recent years have witnessed a rapid evolution in machine learning (ML) models applications in weather prediction, demonstrating notably superior performance compared to traditional numerical weather prediction (NWP) models. These cutting-edge models leverage diverse ML architectures, such as Graph Neural Networks (GNNs), Convolutional Neural Networks (CNNs), Fourier Neural Operators (FNOs), and Transformers. Notably, Google DeepMind has pioneered a novel ML-based approach known as GraphCast, enabling direct training from reanalysis data and facilitating global predictions for numerous weather variables in less than a minute. Impressively, GraphCast forecasts show improved accuracy in predicting severe weather events, including phenomena like tropical cyclones, atmospheric rivers, and extreme heat. However, the efficiency of GraphCast relies on high-quality historical weather data for training, typically sourced from ECMWF's ERA5 reanalysis.  Concurrently, the National Centers for Environmental Prediction (NCEP) has initiated efforts in collaboration with the research community, focusing on developing Machine Learning Weather Prediction (MLWP). This study assesses the pre-trained GraphCast model, leveraging Global Data Assimilation System (GDAS) data from the operational GFSv16 model as initial states. Additionally, we explore the potential use of GDAS data as an alternative training source for GraphCast. Notably, GDAS data is available at a 0.25-degree latitude-longitude resolution and with a temporal resolution of 6 hours. Our investigation involves a comparative analysis of GraphCast's performance when initiated on GDAS data versus ERA5 and HRES data. Alongside this comparative analysis, we investigate the advantages and limitations of utilizing GDAS data for GraphCast while proposing potential approaches for enhancing future iterations of this study.