Synthetic Generation of Extratropical Cyclones’ Fields with Generative Adversarial Networks

Abstract Extratropical cyclones (ETCs) are large atmospheric systems that can bring heavy precipitation, strong winds, and storm surges. Historical records of intense ETCs are scarce, hindering the development of robust machine learning algorithms to study them. Here, we mitigate this data scarcity by generating synthetic fields of mean sea level pressure, wind speed at 10 m, and rainfall rate. We employ a deep learning framework, the progressive growing generative adversarial network (PG-GAN), and use ERA5 reanalysis data and historical cyclone tracks to train our network. The network performance is evaluated in terms of diversity (the ability to generate distinct event scenarios) and fidelity (the ability to generate realistic maps). Results show that PG-GAN can capture detailed ETC structures and the large-scale circulation characteristics associated with ETCs, successfully balancing diversity and fidelity in its output. Our work offers potential for enhanced risk assessment and deeper insights into the atmospheric dynamics associated with ETCs. Significance Statement Extratropical cyclones significantly impact weather and climate but are challenging to study due to scarce data. Our research employs artificial intelligence to generate synthetic atmospheric data, enhancing our understanding of these storms. This method not only improves predictions and risk assessments but also suggests new pathways for climate research. Our findings demonstrate the potential to better prepare for weather-related disasters and offer insights into broader climatic impacts.